linux/fs/pnode.c

611 lines
15 KiB
C
Raw Normal View History

/*
* linux/fs/pnode.c
*
* (C) Copyright IBM Corporation 2005.
* Released under GPL v2.
* Author : Ram Pai (linuxram@us.ibm.com)
*
*/
#include <linux/mnt_namespace.h>
#include <linux/mount.h>
#include <linux/fs.h>
#include <linux/nsproxy.h>
#include <uapi/linux/mount.h>
#include "internal.h"
#include "pnode.h"
/* return the next shared peer mount of @p */
static inline struct mount *next_peer(struct mount *p)
{
return list_entry(p->mnt_share.next, struct mount, mnt_share);
}
static inline struct mount *first_slave(struct mount *p)
{
return list_entry(p->mnt_slave_list.next, struct mount, mnt_slave);
}
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
static inline struct mount *last_slave(struct mount *p)
{
return list_entry(p->mnt_slave_list.prev, struct mount, mnt_slave);
}
static inline struct mount *next_slave(struct mount *p)
{
return list_entry(p->mnt_slave.next, struct mount, mnt_slave);
}
static struct mount *get_peer_under_root(struct mount *mnt,
struct mnt_namespace *ns,
const struct path *root)
{
struct mount *m = mnt;
do {
/* Check the namespace first for optimization */
if (m->mnt_ns == ns && is_path_reachable(m, m->mnt.mnt_root, root))
return m;
m = next_peer(m);
} while (m != mnt);
return NULL;
}
/*
* Get ID of closest dominating peer group having a representative
* under the given root.
*
* Caller must hold namespace_sem
*/
int get_dominating_id(struct mount *mnt, const struct path *root)
{
struct mount *m;
for (m = mnt->mnt_master; m != NULL; m = m->mnt_master) {
struct mount *d = get_peer_under_root(m, mnt->mnt_ns, root);
if (d)
return d->mnt_group_id;
}
return 0;
}
static int do_make_slave(struct mount *mnt)
{
struct mount *master, *slave_mnt;
if (list_empty(&mnt->mnt_share)) {
if (IS_MNT_SHARED(mnt)) {
mnt_release_group_id(mnt);
CLEAR_MNT_SHARED(mnt);
}
master = mnt->mnt_master;
if (!master) {
struct list_head *p = &mnt->mnt_slave_list;
while (!list_empty(p)) {
slave_mnt = list_first_entry(p,
struct mount, mnt_slave);
list_del_init(&slave_mnt->mnt_slave);
slave_mnt->mnt_master = NULL;
}
return 0;
}
} else {
struct mount *m;
/*
* slave 'mnt' to a peer mount that has the
* same root dentry. If none is available then
* slave it to anything that is available.
*/
for (m = master = next_peer(mnt); m != mnt; m = next_peer(m)) {
if (m->mnt.mnt_root == mnt->mnt.mnt_root) {
master = m;
break;
}
}
list_del_init(&mnt->mnt_share);
mnt->mnt_group_id = 0;
CLEAR_MNT_SHARED(mnt);
}
list_for_each_entry(slave_mnt, &mnt->mnt_slave_list, mnt_slave)
slave_mnt->mnt_master = master;
list_move(&mnt->mnt_slave, &master->mnt_slave_list);
list_splice(&mnt->mnt_slave_list, master->mnt_slave_list.prev);
INIT_LIST_HEAD(&mnt->mnt_slave_list);
mnt->mnt_master = master;
return 0;
}
/*
* vfsmount lock must be held for write
*/
void change_mnt_propagation(struct mount *mnt, int type)
{
if (type == MS_SHARED) {
set_mnt_shared(mnt);
return;
}
do_make_slave(mnt);
if (type != MS_SLAVE) {
list_del_init(&mnt->mnt_slave);
mnt->mnt_master = NULL;
if (type == MS_UNBINDABLE)
mnt->mnt.mnt_flags |= MNT_UNBINDABLE;
else
mnt->mnt.mnt_flags &= ~MNT_UNBINDABLE;
}
}
/*
* get the next mount in the propagation tree.
* @m: the mount seen last
* @origin: the original mount from where the tree walk initiated
*
* Note that peer groups form contiguous segments of slave lists.
* We rely on that in get_source() to be able to find out if
* vfsmount found while iterating with propagation_next() is
* a peer of one we'd found earlier.
*/
static struct mount *propagation_next(struct mount *m,
struct mount *origin)
{
/* are there any slaves of this mount? */
if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
return first_slave(m);
while (1) {
struct mount *master = m->mnt_master;
if (master == origin->mnt_master) {
struct mount *next = next_peer(m);
return (next == origin) ? NULL : next;
} else if (m->mnt_slave.next != &master->mnt_slave_list)
return next_slave(m);
/* back at master */
m = master;
}
}
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
static struct mount *skip_propagation_subtree(struct mount *m,
struct mount *origin)
{
/*
* Advance m such that propagation_next will not return
* the slaves of m.
*/
if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
m = last_slave(m);
return m;
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
static struct mount *next_group(struct mount *m, struct mount *origin)
{
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
while (1) {
while (1) {
struct mount *next;
if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
return first_slave(m);
next = next_peer(m);
if (m->mnt_group_id == origin->mnt_group_id) {
if (next == origin)
return NULL;
} else if (m->mnt_slave.next != &next->mnt_slave)
break;
m = next;
}
/* m is the last peer */
while (1) {
struct mount *master = m->mnt_master;
if (m->mnt_slave.next != &master->mnt_slave_list)
return next_slave(m);
m = next_peer(master);
if (master->mnt_group_id == origin->mnt_group_id)
break;
if (master->mnt_slave.next == &m->mnt_slave)
break;
m = master;
}
if (m == origin)
return NULL;
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
/* all accesses are serialized by namespace_sem */
static struct user_namespace *user_ns;
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
static struct mount *last_dest, *first_source, *last_source, *dest_master;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
static struct mountpoint *mp;
static struct hlist_head *list;
static inline bool peers(struct mount *m1, struct mount *m2)
{
return m1->mnt_group_id == m2->mnt_group_id && m1->mnt_group_id;
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
static int propagate_one(struct mount *m)
{
struct mount *child;
int type;
/* skip ones added by this propagate_mnt() */
if (IS_MNT_NEW(m))
return 0;
/* skip if mountpoint isn't covered by it */
if (!is_subdir(mp->m_dentry, m->mnt.mnt_root))
return 0;
if (peers(m, last_dest)) {
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
type = CL_MAKE_SHARED;
} else {
struct mount *n, *p;
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
bool done;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
for (n = m; ; n = p) {
p = n->mnt_master;
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
if (p == dest_master || IS_MNT_MARKED(p))
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
break;
}
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
do {
struct mount *parent = last_source->mnt_parent;
if (last_source == first_source)
break;
done = parent->mnt_master == p;
if (done && peers(n, parent))
break;
last_source = last_source->mnt_master;
} while (!done);
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
type = CL_SLAVE;
/* beginning of peer group among the slaves? */
if (IS_MNT_SHARED(m))
type |= CL_MAKE_SHARED;
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
/* Notice when we are propagating across user namespaces */
if (m->mnt_ns->user_ns != user_ns)
type |= CL_UNPRIVILEGED;
child = copy_tree(last_source, last_source->mnt.mnt_root, type);
if (IS_ERR(child))
return PTR_ERR(child);
child->mnt.mnt_flags &= ~MNT_LOCKED;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
mnt_set_mountpoint(m, mp, child);
last_dest = m;
last_source = child;
if (m->mnt_master != dest_master) {
read_seqlock_excl(&mount_lock);
SET_MNT_MARK(m->mnt_master);
read_sequnlock_excl(&mount_lock);
}
hlist_add_head(&child->mnt_hash, list);
mnt: Add a per mount namespace limit on the number of mounts CAI Qian <caiqian@redhat.com> pointed out that the semantics of shared subtrees make it possible to create an exponentially increasing number of mounts in a mount namespace. mkdir /tmp/1 /tmp/2 mount --make-rshared / for i in $(seq 1 20) ; do mount --bind /tmp/1 /tmp/2 ; done Will create create 2^20 or 1048576 mounts, which is a practical problem as some people have managed to hit this by accident. As such CVE-2016-6213 was assigned. Ian Kent <raven@themaw.net> described the situation for autofs users as follows: > The number of mounts for direct mount maps is usually not very large because of > the way they are implemented, large direct mount maps can have performance > problems. There can be anywhere from a few (likely case a few hundred) to less > than 10000, plus mounts that have been triggered and not yet expired. > > Indirect mounts have one autofs mount at the root plus the number of mounts that > have been triggered and not yet expired. > > The number of autofs indirect map entries can range from a few to the common > case of several thousand and in rare cases up to between 30000 and 50000. I've > not heard of people with maps larger than 50000 entries. > > The larger the number of map entries the greater the possibility for a large > number of active mounts so it's not hard to expect cases of a 1000 or somewhat > more active mounts. So I am setting the default number of mounts allowed per mount namespace at 100,000. This is more than enough for any use case I know of, but small enough to quickly stop an exponential increase in mounts. Which should be perfect to catch misconfigurations and malfunctioning programs. For anyone who needs a higher limit this can be changed by writing to the new /proc/sys/fs/mount-max sysctl. Tested-by: CAI Qian <caiqian@redhat.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-09-28 05:27:17 +00:00
return count_mounts(m->mnt_ns, child);
}
/*
* mount 'source_mnt' under the destination 'dest_mnt' at
* dentry 'dest_dentry'. And propagate that mount to
* all the peer and slave mounts of 'dest_mnt'.
* Link all the new mounts into a propagation tree headed at
* source_mnt. Also link all the new mounts using ->mnt_list
* headed at source_mnt's ->mnt_list
*
* @dest_mnt: destination mount.
* @dest_dentry: destination dentry.
* @source_mnt: source mount.
* @tree_list : list of heads of trees to be attached.
*/
int propagate_mnt(struct mount *dest_mnt, struct mountpoint *dest_mp,
struct mount *source_mnt, struct hlist_head *tree_list)
{
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
struct mount *m, *n;
int ret = 0;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
/*
* we don't want to bother passing tons of arguments to
* propagate_one(); everything is serialized by namespace_sem,
* so globals will do just fine.
*/
user_ns = current->nsproxy->mnt_ns->user_ns;
last_dest = dest_mnt;
propogate_mnt: Handle the first propogated copy being a slave When the first propgated copy was a slave the following oops would result: > BUG: unable to handle kernel NULL pointer dereference at 0000000000000010 > IP: [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > PGD bacd4067 PUD bac66067 PMD 0 > Oops: 0000 [#1] SMP > Modules linked in: > CPU: 1 PID: 824 Comm: mount Not tainted 4.6.0-rc5userns+ #1523 > Hardware name: Bochs Bochs, BIOS Bochs 01/01/2007 > task: ffff8800bb0a8000 ti: ffff8800bac3c000 task.ti: ffff8800bac3c000 > RIP: 0010:[<ffffffff811fba4e>] [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP: 0018:ffff8800bac3fd38 EFLAGS: 00010283 > RAX: 0000000000000000 RBX: ffff8800bb77ec00 RCX: 0000000000000010 > RDX: 0000000000000000 RSI: ffff8800bb58c000 RDI: ffff8800bb58c480 > RBP: ffff8800bac3fd48 R08: 0000000000000001 R09: 0000000000000000 > R10: 0000000000001ca1 R11: 0000000000001c9d R12: 0000000000000000 > R13: ffff8800ba713800 R14: ffff8800bac3fda0 R15: ffff8800bb77ec00 > FS: 00007f3c0cd9b7e0(0000) GS:ffff8800bfb00000(0000) knlGS:0000000000000000 > CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 > CR2: 0000000000000010 CR3: 00000000bb79d000 CR4: 00000000000006e0 > Stack: > ffff8800bb77ec00 0000000000000000 ffff8800bac3fd88 ffffffff811fbf85 > ffff8800bac3fd98 ffff8800bb77f080 ffff8800ba713800 ffff8800bb262b40 > 0000000000000000 0000000000000000 ffff8800bac3fdd8 ffffffff811f1da0 > Call Trace: > [<ffffffff811fbf85>] propagate_mnt+0x105/0x140 > [<ffffffff811f1da0>] attach_recursive_mnt+0x120/0x1e0 > [<ffffffff811f1ec3>] graft_tree+0x63/0x70 > [<ffffffff811f1f6b>] do_add_mount+0x9b/0x100 > [<ffffffff811f2c1a>] do_mount+0x2aa/0xdf0 > [<ffffffff8117efbe>] ? strndup_user+0x4e/0x70 > [<ffffffff811f3a45>] SyS_mount+0x75/0xc0 > [<ffffffff8100242b>] do_syscall_64+0x4b/0xa0 > [<ffffffff81988f3c>] entry_SYSCALL64_slow_path+0x25/0x25 > Code: 00 00 75 ec 48 89 0d 02 22 22 01 8b 89 10 01 00 00 48 89 05 fd 21 22 01 39 8e 10 01 00 00 0f 84 e0 00 00 00 48 8b 80 d8 00 00 00 <48> 8b 50 10 48 89 05 df 21 22 01 48 89 15 d0 21 22 01 8b 53 30 > RIP [<ffffffff811fba4e>] propagate_one+0xbe/0x1c0 > RSP <ffff8800bac3fd38> > CR2: 0000000000000010 > ---[ end trace 2725ecd95164f217 ]--- This oops happens with the namespace_sem held and can be triggered by non-root users. An all around not pleasant experience. To avoid this scenario when finding the appropriate source mount to copy stop the walk up the mnt_master chain when the first source mount is encountered. Further rewrite the walk up the last_source mnt_master chain so that it is clear what is going on. The reason why the first source mount is special is that it it's mnt_parent is not a mount in the dest_mnt propagation tree, and as such termination conditions based up on the dest_mnt mount propgation tree do not make sense. To avoid other kinds of confusion last_dest is not changed when computing last_source. last_dest is only used once in propagate_one and that is above the point of the code being modified, so changing the global variable is meaningless and confusing. Cc: stable@vger.kernel.org fixes: f2ebb3a921c1ca1e2ddd9242e95a1989a50c4c68 ("smarter propagate_mnt()") Reported-by: Tycho Andersen <tycho.andersen@canonical.com> Reviewed-by: Seth Forshee <seth.forshee@canonical.com> Tested-by: Seth Forshee <seth.forshee@canonical.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-05-05 14:29:29 +00:00
first_source = source_mnt;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
last_source = source_mnt;
mp = dest_mp;
list = tree_list;
dest_master = dest_mnt->mnt_master;
/* all peers of dest_mnt, except dest_mnt itself */
for (n = next_peer(dest_mnt); n != dest_mnt; n = next_peer(n)) {
ret = propagate_one(n);
if (ret)
goto out;
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
/* all slave groups */
for (m = next_group(dest_mnt, dest_mnt); m;
m = next_group(m, dest_mnt)) {
/* everything in that slave group */
n = m;
do {
ret = propagate_one(n);
if (ret)
goto out;
n = next_peer(n);
} while (n != m);
}
out:
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
read_seqlock_excl(&mount_lock);
hlist_for_each_entry(n, tree_list, mnt_hash) {
m = n->mnt_parent;
if (m->mnt_master != dest_mnt->mnt_master)
CLEAR_MNT_MARK(m->mnt_master);
}
smarter propagate_mnt() The current mainline has copies propagated to *all* nodes, then tears down the copies we made for nodes that do not contain counterparts of the desired mountpoint. That sets the right propagation graph for the copies (at teardown time we move the slaves of removed node to a surviving peer or directly to master), but we end up paying a fairly steep price in useless allocations. It's fairly easy to create a situation where N calls of mount(2) create exactly N bindings, with O(N^2) vfsmounts allocated and freed in process. Fortunately, it is possible to avoid those allocations/freeings. The trick is to create copies in the right order and find which one would've eventually become a master with the current algorithm. It turns out to be possible in O(nodes getting propagation) time and with no extra allocations at all. One part is that we need to make sure that eventual master will be created before its slaves, so we need to walk the propagation tree in a different order - by peer groups. And iterate through the peers before dealing with the next group. Another thing is finding the (earlier) copy that will be a master of one we are about to create; to do that we are (temporary) marking the masters of mountpoints we are attaching the copies to. Either we are in a peer of the last mountpoint we'd dealt with, or we have the following situation: we are attaching to mountpoint M, the last copy S_0 had been attached to M_0 and there are sequences S_0...S_n, M_0...M_n such that S_{i+1} is a master of S_{i}, S_{i} mounted on M{i} and we need to create a slave of the first S_{k} such that M is getting propagation from M_{k}. It means that the master of M_{k} will be among the sequence of masters of M. On the other hand, the nearest marked node in that sequence will either be the master of M_{k} or the master of M_{k-1} (the latter - in the case if M_{k-1} is a slave of something M gets propagation from, but in a wrong peer group). So we go through the sequence of masters of M until we find a marked one (P). Let N be the one before it. Then we go through the sequence of masters of S_0 until we find one (say, S) mounted on a node D that has P as master and check if D is a peer of N. If it is, S will be the master of new copy, if not - the master of S will be. That's it for the hard part; the rest is fairly simple. Iterator is in next_group(), handling of one prospective mountpoint is propagate_one(). It seems to survive all tests and gives a noticably better performance than the current mainline for setups that are seriously using shared subtrees. Cc: stable@vger.kernel.org Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2014-02-27 14:35:45 +00:00
read_sequnlock_excl(&mount_lock);
return ret;
}
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
static struct mount *find_topper(struct mount *mnt)
{
/* If there is exactly one mount covering mnt completely return it. */
struct mount *child;
if (!list_is_singular(&mnt->mnt_mounts))
return NULL;
child = list_first_entry(&mnt->mnt_mounts, struct mount, mnt_child);
if (child->mnt_mountpoint != mnt->mnt.mnt_root)
return NULL;
return child;
}
/*
* return true if the refcount is greater than count
*/
static inline int do_refcount_check(struct mount *mnt, int count)
{
return mnt_get_count(mnt) > count;
}
/*
* check if the mount 'mnt' can be unmounted successfully.
* @mnt: the mount to be checked for unmount
* NOTE: unmounting 'mnt' would naturally propagate to all
* other mounts its parent propagates to.
* Check if any of these mounts that **do not have submounts**
* have more references than 'refcnt'. If so return busy.
*
fs: scale mntget/mntput The problem that this patch aims to fix is vfsmount refcounting scalability. We need to take a reference on the vfsmount for every successful path lookup, which often go to the same mount point. The fundamental difficulty is that a "simple" reference count can never be made scalable, because any time a reference is dropped, we must check whether that was the last reference. To do that requires communication with all other CPUs that may have taken a reference count. We can make refcounts more scalable in a couple of ways, involving keeping distributed counters, and checking for the global-zero condition less frequently. - check the global sum once every interval (this will delay zero detection for some interval, so it's probably a showstopper for vfsmounts). - keep a local count and only taking the global sum when local reaches 0 (this is difficult for vfsmounts, because we can't hold preempt off for the life of a reference, so a counter would need to be per-thread or tied strongly to a particular CPU which requires more locking). - keep a local difference of increments and decrements, which allows us to sum the total difference and hence find the refcount when summing all CPUs. Then, keep a single integer "long" refcount for slow and long lasting references, and only take the global sum of local counters when the long refcount is 0. This last scheme is what I implemented here. Attached mounts and process root and working directory references are "long" references, and everything else is a short reference. This allows scalable vfsmount references during path walking over mounted subtrees and unattached (lazy umounted) mounts with processes still running in them. This results in one fewer atomic op in the fastpath: mntget is now just a per-CPU inc, rather than an atomic inc; and mntput just requires a spinlock and non-atomic decrement in the common case. However code is otherwise bigger and heavier, so single threaded performance is basically a wash. Signed-off-by: Nick Piggin <npiggin@kernel.dk>
2011-01-07 06:50:11 +00:00
* vfsmount lock must be held for write
*/
int propagate_mount_busy(struct mount *mnt, int refcnt)
{
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
struct mount *m, *child, *topper;
struct mount *parent = mnt->mnt_parent;
if (mnt == parent)
return do_refcount_check(mnt, refcnt);
/*
* quickly check if the current mount can be unmounted.
* If not, we don't have to go checking for all other
* mounts
*/
if (!list_empty(&mnt->mnt_mounts) || do_refcount_check(mnt, refcnt))
return 1;
for (m = propagation_next(parent, parent); m;
m = propagation_next(m, parent)) {
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
int count = 1;
child = __lookup_mnt(&m->mnt, mnt->mnt_mountpoint);
if (!child)
continue;
/* Is there exactly one mount on the child that covers
* it completely whose reference should be ignored?
*/
topper = find_topper(child);
if (topper)
count += 1;
else if (!list_empty(&child->mnt_mounts))
continue;
if (do_refcount_check(child, count))
return 1;
}
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
return 0;
}
/*
* Clear MNT_LOCKED when it can be shown to be safe.
*
* mount_lock lock must be held for write
*/
void propagate_mount_unlock(struct mount *mnt)
{
struct mount *parent = mnt->mnt_parent;
struct mount *m, *child;
BUG_ON(parent == mnt);
for (m = propagation_next(parent, parent); m;
m = propagation_next(m, parent)) {
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
child = __lookup_mnt(&m->mnt, mnt->mnt_mountpoint);
if (child)
child->mnt.mnt_flags &= ~MNT_LOCKED;
}
}
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
static void umount_one(struct mount *mnt, struct list_head *to_umount)
{
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
CLEAR_MNT_MARK(mnt);
mnt->mnt.mnt_flags |= MNT_UMOUNT;
list_del_init(&mnt->mnt_child);
list_del_init(&mnt->mnt_umounting);
list_move_tail(&mnt->mnt_list, to_umount);
}
/*
* NOTE: unmounting 'mnt' naturally propagates to all other mounts its
* parent propagates to.
*/
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
static bool __propagate_umount(struct mount *mnt,
struct list_head *to_umount,
struct list_head *to_restore)
{
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
bool progress = false;
struct mount *child;
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/*
* The state of the parent won't change if this mount is
* already unmounted or marked as without children.
*/
if (mnt->mnt.mnt_flags & (MNT_UMOUNT | MNT_MARKED))
goto out;
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* Verify topper is the only grandchild that has not been
* speculatively unmounted.
*/
list_for_each_entry(child, &mnt->mnt_mounts, mnt_child) {
if (child->mnt_mountpoint == mnt->mnt.mnt_root)
continue;
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
if (!list_empty(&child->mnt_umounting) && IS_MNT_MARKED(child))
continue;
/* Found a mounted child */
goto children;
}
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* Mark mounts that can be unmounted if not locked */
SET_MNT_MARK(mnt);
progress = true;
mnt: Tuck mounts under others instead of creating shadow/side mounts. Ever since mount propagation was introduced in cases where a mount in propagated to parent mount mountpoint pair that is already in use the code has placed the new mount behind the old mount in the mount hash table. This implementation detail is problematic as it allows creating arbitrary length mount hash chains. Furthermore it invalidates the constraint maintained elsewhere in the mount code that a parent mount and a mountpoint pair will have exactly one mount upon them. Making it hard to deal with and to talk about this special case in the mount code. Modify mount propagation to notice when there is already a mount at the parent mount and mountpoint where a new mount is propagating to and place that preexisting mount on top of the new mount. Modify unmount propagation to notice when a mount that is being unmounted has another mount on top of it (and no other children), and to replace the unmounted mount with the mount on top of it. Move the MNT_UMUONT test from __lookup_mnt_last into __propagate_umount as that is the only call of __lookup_mnt_last where MNT_UMOUNT may be set on any mount visible in the mount hash table. These modifications allow: - __lookup_mnt_last to be removed. - attach_shadows to be renamed __attach_mnt and its shadow handling to be removed. - commit_tree to be simplified - copy_tree to be simplified The result is an easier to understand tree of mounts that does not allow creation of arbitrary length hash chains in the mount hash table. The result is also a very slight userspace visible difference in semantics. The following two cases now behave identically, where before order mattered: case 1: (explicit user action) B is a slave of A mount something on A/a , it will propagate to B/a and than mount something on B/a case 2: (tucked mount) B is a slave of A mount something on B/a and than mount something on A/a Histroically umount A/a would fail in case 1 and succeed in case 2. Now umount A/a succeeds in both configurations. This very small change in semantics appears if anything to be a bug fix to me and my survey of userspace leads me to believe that no programs will notice or care of this subtle semantic change. v2: Updated to mnt_change_mountpoint to not call dput or mntput and instead to decrement the counts directly. It is guaranteed that there will be other references when mnt_change_mountpoint is called so this is safe. v3: Moved put_mountpoint under mount_lock in attach_recursive_mnt As the locking in fs/namespace.c changed between v2 and v3. v4: Reworked the logic in propagate_mount_busy and __propagate_umount that detects when a mount completely covers another mount. v5: Removed unnecessary tests whose result is alwasy true in find_topper and attach_recursive_mnt. v6: Document the user space visible semantic difference. Cc: stable@vger.kernel.org Fixes: b90fa9ae8f51 ("[PATCH] shared mount handling: bind and rbind") Tested-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-01-20 05:28:35 +00:00
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* If a mount is without children and not locked umount it. */
if (!IS_MNT_LOCKED(mnt)) {
umount_one(mnt, to_umount);
} else {
children:
list_move_tail(&mnt->mnt_umounting, to_restore);
}
out:
return progress;
}
static void umount_list(struct list_head *to_umount,
struct list_head *to_restore)
{
struct mount *mnt, *child, *tmp;
list_for_each_entry(mnt, to_umount, mnt_list) {
list_for_each_entry_safe(child, tmp, &mnt->mnt_mounts, mnt_child) {
/* topper? */
if (child->mnt_mountpoint == mnt->mnt.mnt_root)
list_move_tail(&child->mnt_umounting, to_restore);
else
umount_one(child, to_umount);
}
}
}
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
static void restore_mounts(struct list_head *to_restore)
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
{
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* Restore mounts to a clean working state */
while (!list_empty(to_restore)) {
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
struct mount *mnt, *parent;
struct mountpoint *mp;
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
mnt = list_first_entry(to_restore, struct mount, mnt_umounting);
CLEAR_MNT_MARK(mnt);
list_del_init(&mnt->mnt_umounting);
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* Should this mount be reparented? */
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
mp = mnt->mnt_mp;
parent = mnt->mnt_parent;
while (parent->mnt.mnt_flags & MNT_UMOUNT) {
mp = parent->mnt_mp;
parent = parent->mnt_parent;
}
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
if (parent != mnt->mnt_parent)
mnt_change_mountpoint(parent, mp, mnt);
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
}
}
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
static void cleanup_umount_visitations(struct list_head *visited)
{
while (!list_empty(visited)) {
struct mount *mnt =
list_first_entry(visited, struct mount, mnt_umounting);
list_del_init(&mnt->mnt_umounting);
}
}
/*
* collect all mounts that receive propagation from the mount in @list,
* and return these additional mounts in the same list.
* @list: the list of mounts to be unmounted.
*
* vfsmount lock must be held for write
*/
int propagate_umount(struct list_head *list)
{
struct mount *mnt;
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
LIST_HEAD(to_restore);
LIST_HEAD(to_umount);
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
LIST_HEAD(visited);
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
/* Find candidates for unmounting */
list_for_each_entry_reverse(mnt, list, mnt_list) {
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
struct mount *parent = mnt->mnt_parent;
struct mount *m;
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
/*
* If this mount has already been visited it is known that it's
* entire peer group and all of their slaves in the propagation
* tree for the mountpoint has already been visited and there is
* no need to visit them again.
*/
if (!list_empty(&mnt->mnt_umounting))
continue;
list_add_tail(&mnt->mnt_umounting, &visited);
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
for (m = propagation_next(parent, parent); m;
m = propagation_next(m, parent)) {
struct mount *child = __lookup_mnt(&m->mnt,
mnt->mnt_mountpoint);
if (!child)
continue;
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
if (!list_empty(&child->mnt_umounting)) {
/*
* If the child has already been visited it is
* know that it's entire peer group and all of
* their slaves in the propgation tree for the
* mountpoint has already been visited and there
* is no need to visit this subtree again.
*/
m = skip_propagation_subtree(m, parent);
continue;
} else if (child->mnt.mnt_flags & MNT_UMOUNT) {
/*
* We have come accross an partially unmounted
* mount in list that has not been visited yet.
* Remember it has been visited and continue
* about our merry way.
*/
list_add_tail(&child->mnt_umounting, &visited);
continue;
}
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
/* Check the child and parents while progress is made */
while (__propagate_umount(child,
&to_umount, &to_restore)) {
/* Is the parent a umount candidate? */
child = child->mnt_parent;
if (list_empty(&child->mnt_umounting))
break;
}
}
}
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
umount_list(&to_umount, &to_restore);
restore_mounts(&to_restore);
mnt: Make propagate_umount less slow for overlapping mount propagation trees Andrei Vagin pointed out that time to executue propagate_umount can go non-linear (and take a ludicrious amount of time) when the mount propogation trees of the mounts to be unmunted by a lazy unmount overlap. Make the walk of the mount propagation trees nearly linear by remembering which mounts have already been visited, allowing subsequent walks to detect when walking a mount propgation tree or a subtree of a mount propgation tree would be duplicate work and to skip them entirely. Walk the list of mounts whose propgatation trees need to be traversed from the mount highest in the mount tree to mounts lower in the mount tree so that odds are higher that the code will walk the largest trees first, allowing later tree walks to be skipped entirely. Add cleanup_umount_visitation to remover the code's memory of which mounts have been visited. Add the functions last_slave and skip_propagation_subtree to allow skipping appropriate parts of the mount propagation tree without needing to change the logic of the rest of the code. A script to generate overlapping mount propagation trees: $ cat runs.h set -e mount -t tmpfs zdtm /mnt mkdir -p /mnt/1 /mnt/2 mount -t tmpfs zdtm /mnt/1 mount --make-shared /mnt/1 mkdir /mnt/1/1 iteration=10 if [ -n "$1" ] ; then iteration=$1 fi for i in $(seq $iteration); do mount --bind /mnt/1/1 /mnt/1/1 done mount --rbind /mnt/1 /mnt/2 TIMEFORMAT='%Rs' nr=$(( ( 2 ** ( $iteration + 1 ) ) + 1 )) echo -n "umount -l /mnt/1 -> $nr " time umount -l /mnt/1 nr=$(cat /proc/self/mountinfo | grep zdtm | wc -l ) time umount -l /mnt/2 $ for i in $(seq 9 19); do echo $i; unshare -Urm bash ./run.sh $i; done Here are the performance numbers with and without the patch: mhash | 8192 | 8192 | 1048576 | 1048576 mounts | before | after | before | after ------------------------------------------------ 1025 | 0.040s | 0.016s | 0.038s | 0.019s 2049 | 0.094s | 0.017s | 0.080s | 0.018s 4097 | 0.243s | 0.019s | 0.206s | 0.023s 8193 | 1.202s | 0.028s | 1.562s | 0.032s 16385 | 9.635s | 0.036s | 9.952s | 0.041s 32769 | 60.928s | 0.063s | 44.321s | 0.064s 65537 | | 0.097s | | 0.097s 131073 | | 0.233s | | 0.176s 262145 | | 0.653s | | 0.344s 524289 | | 2.305s | | 0.735s 1048577 | | 7.107s | | 2.603s Andrei Vagin reports fixing the performance problem is part of the work to fix CVE-2016-6213. Cc: stable@vger.kernel.org Fixes: a05964f3917c ("[PATCH] shared mounts handling: umount") Reported-by: Andrei Vagin <avagin@openvz.org> Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 22:25:19 +00:00
cleanup_umount_visitations(&visited);
mnt: In propgate_umount handle visiting mounts in any order While investigating some poor umount performance I realized that in the case of overlapping mount trees where some of the mounts are locked the code has been failing to unmount all of the mounts it should have been unmounting. This failure to unmount all of the necessary mounts can be reproduced with: $ cat locked_mounts_test.sh mount -t tmpfs test-base /mnt mount --make-shared /mnt mkdir -p /mnt/b mount -t tmpfs test1 /mnt/b mount --make-shared /mnt/b mkdir -p /mnt/b/10 mount -t tmpfs test2 /mnt/b/10 mount --make-shared /mnt/b/10 mkdir -p /mnt/b/10/20 mount --rbind /mnt/b /mnt/b/10/20 unshare -Urm --propagation unchaged /bin/sh -c 'sleep 5; if [ $(grep test /proc/self/mountinfo | wc -l) -eq 1 ] ; then echo SUCCESS ; else echo FAILURE ; fi' sleep 1 umount -l /mnt/b wait %% $ unshare -Urm ./locked_mounts_test.sh This failure is corrected by removing the prepass that marks mounts that may be umounted. A first pass is added that umounts mounts if possible and if not sets mount mark if they could be unmounted if they weren't locked and adds them to a list to umount possibilities. This first pass reconsiders the mounts parent if it is on the list of umount possibilities, ensuring that information of umoutability will pass from child to mount parent. A second pass then walks through all mounts that are umounted and processes their children unmounting them or marking them for reparenting. A last pass cleans up the state on the mounts that could not be umounted and if applicable reparents them to their first parent that remained mounted. While a bit longer than the old code this code is much more robust as it allows information to flow up from the leaves and down from the trunk making the order in which mounts are encountered in the umount propgation tree irrelevant. Cc: stable@vger.kernel.org Fixes: 0c56fe31420c ("mnt: Don't propagate unmounts to locked mounts") Reviewed-by: Andrei Vagin <avagin@virtuozzo.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2016-10-24 21:16:13 +00:00
list_splice_tail(&to_umount, list);
mnt: In umount propagation reparent in a separate pass It was observed that in some pathlogical cases that the current code does not unmount everything it should. After investigation it was determined that the issue is that mnt_change_mntpoint can can change which mounts are available to be unmounted during mount propagation which is wrong. The trivial reproducer is: $ cat ./pathological.sh mount -t tmpfs test-base /mnt cd /mnt mkdir 1 2 1/1 mount --bind 1 1 mount --make-shared 1 mount --bind 1 2 mount --bind 1/1 1/1 mount --bind 1/1 1/1 echo grep test-base /proc/self/mountinfo umount 1/1 echo grep test-base /proc/self/mountinfo $ unshare -Urm ./pathological.sh The expected output looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 The output without the fix looks like: 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 49 54 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 50 53 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 51 49 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 54 47 0:25 /1/1 /mnt/1/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 53 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 50 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 46 31 0:25 / /mnt rw,relatime - tmpfs test-base rw,uid=1000,gid=1000 47 46 0:25 /1 /mnt/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 48 46 0:25 /1 /mnt/2 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 52 48 0:25 /1/1 /mnt/2/1 rw,relatime shared:1 - tmpfs test-base rw,uid=1000,gid=1000 That last mount in the output was in the propgation tree to be unmounted but was missed because the mnt_change_mountpoint changed it's parent before the walk through the mount propagation tree observed it. Cc: stable@vger.kernel.org Fixes: 1064f874abc0 ("mnt: Tuck mounts under others instead of creating shadow/side mounts.") Acked-by: Andrei Vagin <avagin@virtuozzo.com> Reviewed-by: Ram Pai <linuxram@us.ibm.com> Signed-off-by: "Eric W. Biederman" <ebiederm@xmission.com>
2017-05-15 19:42:07 +00:00
return 0;
}