linux/net/rds/ib_rdma.c

852 lines
22 KiB
C
Raw Normal View History

/*
* Copyright (c) 2006 Oracle. All rights reserved.
*
* This software is available to you under a choice of one of two
* licenses. You may choose to be licensed under the terms of the GNU
* General Public License (GPL) Version 2, available from the file
* COPYING in the main directory of this source tree, or the
* OpenIB.org BSD license below:
*
* Redistribution and use in source and binary forms, with or
* without modification, are permitted provided that the following
* conditions are met:
*
* - Redistributions of source code must retain the above
* copyright notice, this list of conditions and the following
* disclaimer.
*
* - Redistributions in binary form must reproduce the above
* copyright notice, this list of conditions and the following
* disclaimer in the documentation and/or other materials
* provided with the distribution.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
* BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
* ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
* CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*
*/
#include <linux/kernel.h>
include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h percpu.h is included by sched.h and module.h and thus ends up being included when building most .c files. percpu.h includes slab.h which in turn includes gfp.h making everything defined by the two files universally available and complicating inclusion dependencies. percpu.h -> slab.h dependency is about to be removed. Prepare for this change by updating users of gfp and slab facilities include those headers directly instead of assuming availability. As this conversion needs to touch large number of source files, the following script is used as the basis of conversion. http://userweb.kernel.org/~tj/misc/slabh-sweep.py The script does the followings. * Scan files for gfp and slab usages and update includes such that only the necessary includes are there. ie. if only gfp is used, gfp.h, if slab is used, slab.h. * When the script inserts a new include, it looks at the include blocks and try to put the new include such that its order conforms to its surrounding. It's put in the include block which contains core kernel includes, in the same order that the rest are ordered - alphabetical, Christmas tree, rev-Xmas-tree or at the end if there doesn't seem to be any matching order. * If the script can't find a place to put a new include (mostly because the file doesn't have fitting include block), it prints out an error message indicating which .h file needs to be added to the file. The conversion was done in the following steps. 1. The initial automatic conversion of all .c files updated slightly over 4000 files, deleting around 700 includes and adding ~480 gfp.h and ~3000 slab.h inclusions. The script emitted errors for ~400 files. 2. Each error was manually checked. Some didn't need the inclusion, some needed manual addition while adding it to implementation .h or embedding .c file was more appropriate for others. This step added inclusions to around 150 files. 3. The script was run again and the output was compared to the edits from #2 to make sure no file was left behind. 4. Several build tests were done and a couple of problems were fixed. e.g. lib/decompress_*.c used malloc/free() wrappers around slab APIs requiring slab.h to be added manually. 5. The script was run on all .h files but without automatically editing them as sprinkling gfp.h and slab.h inclusions around .h files could easily lead to inclusion dependency hell. Most gfp.h inclusion directives were ignored as stuff from gfp.h was usually wildly available and often used in preprocessor macros. Each slab.h inclusion directive was examined and added manually as necessary. 6. percpu.h was updated not to include slab.h. 7. Build test were done on the following configurations and failures were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my distributed build env didn't work with gcov compiles) and a few more options had to be turned off depending on archs to make things build (like ipr on powerpc/64 which failed due to missing writeq). * x86 and x86_64 UP and SMP allmodconfig and a custom test config. * powerpc and powerpc64 SMP allmodconfig * sparc and sparc64 SMP allmodconfig * ia64 SMP allmodconfig * s390 SMP allmodconfig * alpha SMP allmodconfig * um on x86_64 SMP allmodconfig 8. percpu.h modifications were reverted so that it could be applied as a separate patch and serve as bisection point. Given the fact that I had only a couple of failures from tests on step 6, I'm fairly confident about the coverage of this conversion patch. If there is a breakage, it's likely to be something in one of the arch headers which should be easily discoverable easily on most builds of the specific arch. Signed-off-by: Tejun Heo <tj@kernel.org> Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org> Cc: Ingo Molnar <mingo@redhat.com> Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
2010-03-24 08:04:11 +00:00
#include <linux/slab.h>
#include <linux/rculist.h>
#include <linux/llist.h>
#include "rds.h"
#include "ib.h"
static DEFINE_PER_CPU(unsigned long, clean_list_grace);
#define CLEAN_LIST_BUSY_BIT 0
/*
* This is stored as mr->r_trans_private.
*/
struct rds_ib_mr {
struct rds_ib_device *device;
struct rds_ib_mr_pool *pool;
struct ib_fmr *fmr;
struct llist_node llnode;
/* unmap_list is for freeing */
struct list_head unmap_list;
unsigned int remap_count;
struct scatterlist *sg;
unsigned int sg_len;
u64 *dma;
int sg_dma_len;
};
/*
* Our own little FMR pool
*/
struct rds_ib_mr_pool {
unsigned int pool_type;
struct mutex flush_lock; /* serialize fmr invalidate */
struct delayed_work flush_worker; /* flush worker */
atomic_t item_count; /* total # of MRs */
atomic_t dirty_count; /* # dirty of MRs */
struct llist_head drop_list; /* MRs that have reached their max_maps limit */
struct llist_head free_list; /* unused MRs */
struct llist_head clean_list; /* global unused & unamapped MRs */
wait_queue_head_t flush_wait;
atomic_t free_pinned; /* memory pinned by free MRs */
unsigned long max_items;
unsigned long max_items_soft;
unsigned long max_free_pinned;
struct ib_fmr_attr fmr_attr;
};
static struct workqueue_struct *rds_ib_fmr_wq;
int rds_ib_fmr_init(void)
{
rds_ib_fmr_wq = create_workqueue("rds_fmr_flushd");
if (!rds_ib_fmr_wq)
return -ENOMEM;
return 0;
}
/* By the time this is called all the IB devices should have been torn down and
* had their pools freed. As each pool is freed its work struct is waited on,
* so the pool flushing work queue should be idle by the time we get here.
*/
void rds_ib_fmr_exit(void)
{
destroy_workqueue(rds_ib_fmr_wq);
}
static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool, int free_all, struct rds_ib_mr **);
static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr);
static void rds_ib_mr_pool_flush_worker(struct work_struct *work);
static struct rds_ib_device *rds_ib_get_device(__be32 ipaddr)
{
struct rds_ib_device *rds_ibdev;
struct rds_ib_ipaddr *i_ipaddr;
rcu_read_lock();
list_for_each_entry_rcu(rds_ibdev, &rds_ib_devices, list) {
list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
if (i_ipaddr->ipaddr == ipaddr) {
RDS/IB: add refcount tracking to struct rds_ib_device The RDS IB client .remove callback used to free the rds_ibdev for the given device unconditionally. This could race other users of the struct. This patch adds refcounting so that we only free the rds_ibdev once all of its users are done. Many rds_ibdev users are tied to connections. We give the connection a reference and change these users to reference the device in the connection instead of looking it up in the IB client data. The only user of the IB client data remaining is the first lookup of the device as connections are built up. Incrementing the reference count of a device found in the IB client data could race with final freeing so we use an RCU grace period to make sure that freeing won't happen until those lookups are done. MRs need the rds_ibdev to get at the pool that they're freed in to. They exist outside a connection and many MRs can reference different devices from one socket, so it was natural to have each MR hold a reference. MR refs can be dropped from interrupt handlers and final device teardown can block so we push it off to a work struct. Pool teardown had to be fixed to cancel its pending work instead of deadlocking waiting for all queued work, including itself, to finish. MRs get their reference from the global device list, which gets a reference. It is left unprotected by locks and remains racy. A simple global lock would be a significant bottleneck. More scalable (complicated) locking should be done carefully in a later patch. Signed-off-by: Zach Brown <zach.brown@oracle.com>
2010-05-18 22:48:51 +00:00
atomic_inc(&rds_ibdev->refcount);
rcu_read_unlock();
return rds_ibdev;
}
}
}
rcu_read_unlock();
return NULL;
}
static int rds_ib_add_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
{
struct rds_ib_ipaddr *i_ipaddr;
i_ipaddr = kmalloc(sizeof *i_ipaddr, GFP_KERNEL);
if (!i_ipaddr)
return -ENOMEM;
i_ipaddr->ipaddr = ipaddr;
spin_lock_irq(&rds_ibdev->spinlock);
list_add_tail_rcu(&i_ipaddr->list, &rds_ibdev->ipaddr_list);
spin_unlock_irq(&rds_ibdev->spinlock);
return 0;
}
static void rds_ib_remove_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
{
struct rds_ib_ipaddr *i_ipaddr;
struct rds_ib_ipaddr *to_free = NULL;
spin_lock_irq(&rds_ibdev->spinlock);
list_for_each_entry_rcu(i_ipaddr, &rds_ibdev->ipaddr_list, list) {
if (i_ipaddr->ipaddr == ipaddr) {
list_del_rcu(&i_ipaddr->list);
to_free = i_ipaddr;
break;
}
}
spin_unlock_irq(&rds_ibdev->spinlock);
if (to_free)
kfree_rcu(to_free, rcu);
}
int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
{
struct rds_ib_device *rds_ibdev_old;
rds_ibdev_old = rds_ib_get_device(ipaddr);
if (!rds_ibdev_old)
return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
if (rds_ibdev_old != rds_ibdev) {
rds_ib_remove_ipaddr(rds_ibdev_old, ipaddr);
RDS/IB: add refcount tracking to struct rds_ib_device The RDS IB client .remove callback used to free the rds_ibdev for the given device unconditionally. This could race other users of the struct. This patch adds refcounting so that we only free the rds_ibdev once all of its users are done. Many rds_ibdev users are tied to connections. We give the connection a reference and change these users to reference the device in the connection instead of looking it up in the IB client data. The only user of the IB client data remaining is the first lookup of the device as connections are built up. Incrementing the reference count of a device found in the IB client data could race with final freeing so we use an RCU grace period to make sure that freeing won't happen until those lookups are done. MRs need the rds_ibdev to get at the pool that they're freed in to. They exist outside a connection and many MRs can reference different devices from one socket, so it was natural to have each MR hold a reference. MR refs can be dropped from interrupt handlers and final device teardown can block so we push it off to a work struct. Pool teardown had to be fixed to cancel its pending work instead of deadlocking waiting for all queued work, including itself, to finish. MRs get their reference from the global device list, which gets a reference. It is left unprotected by locks and remains racy. A simple global lock would be a significant bottleneck. More scalable (complicated) locking should be done carefully in a later patch. Signed-off-by: Zach Brown <zach.brown@oracle.com>
2010-05-18 22:48:51 +00:00
rds_ib_dev_put(rds_ibdev_old);
return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
RDS/IB: add refcount tracking to struct rds_ib_device The RDS IB client .remove callback used to free the rds_ibdev for the given device unconditionally. This could race other users of the struct. This patch adds refcounting so that we only free the rds_ibdev once all of its users are done. Many rds_ibdev users are tied to connections. We give the connection a reference and change these users to reference the device in the connection instead of looking it up in the IB client data. The only user of the IB client data remaining is the first lookup of the device as connections are built up. Incrementing the reference count of a device found in the IB client data could race with final freeing so we use an RCU grace period to make sure that freeing won't happen until those lookups are done. MRs need the rds_ibdev to get at the pool that they're freed in to. They exist outside a connection and many MRs can reference different devices from one socket, so it was natural to have each MR hold a reference. MR refs can be dropped from interrupt handlers and final device teardown can block so we push it off to a work struct. Pool teardown had to be fixed to cancel its pending work instead of deadlocking waiting for all queued work, including itself, to finish. MRs get their reference from the global device list, which gets a reference. It is left unprotected by locks and remains racy. A simple global lock would be a significant bottleneck. More scalable (complicated) locking should be done carefully in a later patch. Signed-off-by: Zach Brown <zach.brown@oracle.com>
2010-05-18 22:48:51 +00:00
}
rds_ib_dev_put(rds_ibdev_old);
return 0;
}
void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
{
struct rds_ib_connection *ic = conn->c_transport_data;
/* conn was previously on the nodev_conns_list */
spin_lock_irq(&ib_nodev_conns_lock);
BUG_ON(list_empty(&ib_nodev_conns));
BUG_ON(list_empty(&ic->ib_node));
list_del(&ic->ib_node);
spin_lock(&rds_ibdev->spinlock);
list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
spin_unlock(&rds_ibdev->spinlock);
spin_unlock_irq(&ib_nodev_conns_lock);
ic->rds_ibdev = rds_ibdev;
RDS/IB: add refcount tracking to struct rds_ib_device The RDS IB client .remove callback used to free the rds_ibdev for the given device unconditionally. This could race other users of the struct. This patch adds refcounting so that we only free the rds_ibdev once all of its users are done. Many rds_ibdev users are tied to connections. We give the connection a reference and change these users to reference the device in the connection instead of looking it up in the IB client data. The only user of the IB client data remaining is the first lookup of the device as connections are built up. Incrementing the reference count of a device found in the IB client data could race with final freeing so we use an RCU grace period to make sure that freeing won't happen until those lookups are done. MRs need the rds_ibdev to get at the pool that they're freed in to. They exist outside a connection and many MRs can reference different devices from one socket, so it was natural to have each MR hold a reference. MR refs can be dropped from interrupt handlers and final device teardown can block so we push it off to a work struct. Pool teardown had to be fixed to cancel its pending work instead of deadlocking waiting for all queued work, including itself, to finish. MRs get their reference from the global device list, which gets a reference. It is left unprotected by locks and remains racy. A simple global lock would be a significant bottleneck. More scalable (complicated) locking should be done carefully in a later patch. Signed-off-by: Zach Brown <zach.brown@oracle.com>
2010-05-18 22:48:51 +00:00
atomic_inc(&rds_ibdev->refcount);
}
void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
{
struct rds_ib_connection *ic = conn->c_transport_data;
/* place conn on nodev_conns_list */
spin_lock(&ib_nodev_conns_lock);
spin_lock_irq(&rds_ibdev->spinlock);
BUG_ON(list_empty(&ic->ib_node));
list_del(&ic->ib_node);
spin_unlock_irq(&rds_ibdev->spinlock);
list_add_tail(&ic->ib_node, &ib_nodev_conns);
spin_unlock(&ib_nodev_conns_lock);
ic->rds_ibdev = NULL;
RDS/IB: add refcount tracking to struct rds_ib_device The RDS IB client .remove callback used to free the rds_ibdev for the given device unconditionally. This could race other users of the struct. This patch adds refcounting so that we only free the rds_ibdev once all of its users are done. Many rds_ibdev users are tied to connections. We give the connection a reference and change these users to reference the device in the connection instead of looking it up in the IB client data. The only user of the IB client data remaining is the first lookup of the device as connections are built up. Incrementing the reference count of a device found in the IB client data could race with final freeing so we use an RCU grace period to make sure that freeing won't happen until those lookups are done. MRs need the rds_ibdev to get at the pool that they're freed in to. They exist outside a connection and many MRs can reference different devices from one socket, so it was natural to have each MR hold a reference. MR refs can be dropped from interrupt handlers and final device teardown can block so we push it off to a work struct. Pool teardown had to be fixed to cancel its pending work instead of deadlocking waiting for all queued work, including itself, to finish. MRs get their reference from the global device list, which gets a reference. It is left unprotected by locks and remains racy. A simple global lock would be a significant bottleneck. More scalable (complicated) locking should be done carefully in a later patch. Signed-off-by: Zach Brown <zach.brown@oracle.com>
2010-05-18 22:48:51 +00:00
rds_ib_dev_put(rds_ibdev);
}
void rds_ib_destroy_nodev_conns(void)
{
struct rds_ib_connection *ic, *_ic;
LIST_HEAD(tmp_list);
/* avoid calling conn_destroy with irqs off */
spin_lock_irq(&ib_nodev_conns_lock);
list_splice(&ib_nodev_conns, &tmp_list);
spin_unlock_irq(&ib_nodev_conns_lock);
list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node)
rds_conn_destroy(ic->conn);
}
struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *rds_ibdev,
int pool_type)
{
struct rds_ib_mr_pool *pool;
pool = kzalloc(sizeof(*pool), GFP_KERNEL);
if (!pool)
return ERR_PTR(-ENOMEM);
pool->pool_type = pool_type;
init_llist_head(&pool->free_list);
init_llist_head(&pool->drop_list);
init_llist_head(&pool->clean_list);
mutex_init(&pool->flush_lock);
init_waitqueue_head(&pool->flush_wait);
INIT_DELAYED_WORK(&pool->flush_worker, rds_ib_mr_pool_flush_worker);
if (pool_type == RDS_IB_MR_1M_POOL) {
/* +1 allows for unaligned MRs */
pool->fmr_attr.max_pages = RDS_FMR_1M_MSG_SIZE + 1;
pool->max_items = RDS_FMR_1M_POOL_SIZE;
} else {
/* pool_type == RDS_IB_MR_8K_POOL */
pool->fmr_attr.max_pages = RDS_FMR_8K_MSG_SIZE + 1;
pool->max_items = RDS_FMR_8K_POOL_SIZE;
}
pool->max_free_pinned = pool->max_items * pool->fmr_attr.max_pages / 4;
pool->fmr_attr.max_maps = rds_ibdev->fmr_max_remaps;
pool->fmr_attr.page_shift = PAGE_SHIFT;
pool->max_items_soft = rds_ibdev->max_fmrs * 3 / 4;
return pool;
}
void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo)
{
struct rds_ib_mr_pool *pool_1m = rds_ibdev->mr_1m_pool;
iinfo->rdma_mr_max = pool_1m->max_items;
iinfo->rdma_mr_size = pool_1m->fmr_attr.max_pages;
}
void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *pool)
{
cancel_delayed_work_sync(&pool->flush_worker);
rds_ib_flush_mr_pool(pool, 1, NULL);
WARN_ON(atomic_read(&pool->item_count));
WARN_ON(atomic_read(&pool->free_pinned));
kfree(pool);
}
static inline struct rds_ib_mr *rds_ib_reuse_fmr(struct rds_ib_mr_pool *pool)
{
struct rds_ib_mr *ibmr = NULL;
struct llist_node *ret;
unsigned long *flag;
preempt_disable();
flag = this_cpu_ptr(&clean_list_grace);
set_bit(CLEAN_LIST_BUSY_BIT, flag);
ret = llist_del_first(&pool->clean_list);
if (ret)
ibmr = llist_entry(ret, struct rds_ib_mr, llnode);
clear_bit(CLEAN_LIST_BUSY_BIT, flag);
preempt_enable();
return ibmr;
}
static inline void wait_clean_list_grace(void)
{
int cpu;
unsigned long *flag;
for_each_online_cpu(cpu) {
flag = &per_cpu(clean_list_grace, cpu);
while (test_bit(CLEAN_LIST_BUSY_BIT, flag))
cpu_relax();
}
}
static struct rds_ib_mr *rds_ib_alloc_fmr(struct rds_ib_device *rds_ibdev,
int npages)
{
struct rds_ib_mr_pool *pool;
struct rds_ib_mr *ibmr = NULL;
int err = 0, iter = 0;
if (npages <= RDS_FMR_8K_MSG_SIZE)
pool = rds_ibdev->mr_8k_pool;
else
pool = rds_ibdev->mr_1m_pool;
if (atomic_read(&pool->dirty_count) >= pool->max_items / 10)
queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
/* Switch pools if one of the pool is reaching upper limit */
if (atomic_read(&pool->dirty_count) >= pool->max_items * 9 / 10) {
if (pool->pool_type == RDS_IB_MR_8K_POOL)
pool = rds_ibdev->mr_1m_pool;
else
pool = rds_ibdev->mr_8k_pool;
}
while (1) {
ibmr = rds_ib_reuse_fmr(pool);
if (ibmr)
return ibmr;
/* No clean MRs - now we have the choice of either
* allocating a fresh MR up to the limit imposed by the
* driver, or flush any dirty unused MRs.
* We try to avoid stalling in the send path if possible,
* so we allocate as long as we're allowed to.
*
* We're fussy with enforcing the FMR limit, though. If the driver
* tells us we can't use more than N fmrs, we shouldn't start
* arguing with it */
if (atomic_inc_return(&pool->item_count) <= pool->max_items)
break;
atomic_dec(&pool->item_count);
if (++iter > 2) {
if (pool->pool_type == RDS_IB_MR_8K_POOL)
rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_depleted);
else
rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_depleted);
return ERR_PTR(-EAGAIN);
}
/* We do have some empty MRs. Flush them out. */
if (pool->pool_type == RDS_IB_MR_8K_POOL)
rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_wait);
else
rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_wait);
rds_ib_flush_mr_pool(pool, 0, &ibmr);
if (ibmr)
return ibmr;
}
ibmr = kzalloc_node(sizeof(*ibmr), GFP_KERNEL, rdsibdev_to_node(rds_ibdev));
if (!ibmr) {
err = -ENOMEM;
goto out_no_cigar;
}
ibmr->fmr = ib_alloc_fmr(rds_ibdev->pd,
(IB_ACCESS_LOCAL_WRITE |
IB_ACCESS_REMOTE_READ |
IB_ACCESS_REMOTE_WRITE|
IB_ACCESS_REMOTE_ATOMIC),
&pool->fmr_attr);
if (IS_ERR(ibmr->fmr)) {
err = PTR_ERR(ibmr->fmr);
ibmr->fmr = NULL;
printk(KERN_WARNING "RDS/IB: ib_alloc_fmr failed (err=%d)\n", err);
goto out_no_cigar;
}
ibmr->pool = pool;
if (pool->pool_type == RDS_IB_MR_8K_POOL)
rds_ib_stats_inc(s_ib_rdma_mr_8k_alloc);
else
rds_ib_stats_inc(s_ib_rdma_mr_1m_alloc);
return ibmr;
out_no_cigar:
if (ibmr) {
if (ibmr->fmr)
ib_dealloc_fmr(ibmr->fmr);
kfree(ibmr);
}
atomic_dec(&pool->item_count);
return ERR_PTR(err);
}
static int rds_ib_map_fmr(struct rds_ib_device *rds_ibdev, struct rds_ib_mr *ibmr,
struct scatterlist *sg, unsigned int nents)
{
struct ib_device *dev = rds_ibdev->dev;
struct scatterlist *scat = sg;
u64 io_addr = 0;
u64 *dma_pages;
u32 len;
int page_cnt, sg_dma_len;
int i, j;
int ret;
sg_dma_len = ib_dma_map_sg(dev, sg, nents,
DMA_BIDIRECTIONAL);
if (unlikely(!sg_dma_len)) {
printk(KERN_WARNING "RDS/IB: dma_map_sg failed!\n");
return -EBUSY;
}
len = 0;
page_cnt = 0;
for (i = 0; i < sg_dma_len; ++i) {
unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
if (dma_addr & ~PAGE_MASK) {
if (i > 0)
return -EINVAL;
else
++page_cnt;
}
if ((dma_addr + dma_len) & ~PAGE_MASK) {
if (i < sg_dma_len - 1)
return -EINVAL;
else
++page_cnt;
}
len += dma_len;
}
page_cnt += len >> PAGE_SHIFT;
if (page_cnt > ibmr->pool->fmr_attr.max_pages)
return -EINVAL;
dma_pages = kmalloc_node(sizeof(u64) * page_cnt, GFP_ATOMIC,
rdsibdev_to_node(rds_ibdev));
if (!dma_pages)
return -ENOMEM;
page_cnt = 0;
for (i = 0; i < sg_dma_len; ++i) {
unsigned int dma_len = ib_sg_dma_len(dev, &scat[i]);
u64 dma_addr = ib_sg_dma_address(dev, &scat[i]);
for (j = 0; j < dma_len; j += PAGE_SIZE)
dma_pages[page_cnt++] =
(dma_addr & PAGE_MASK) + j;
}
ret = ib_map_phys_fmr(ibmr->fmr,
dma_pages, page_cnt, io_addr);
if (ret)
goto out;
/* Success - we successfully remapped the MR, so we can
* safely tear down the old mapping. */
rds_ib_teardown_mr(ibmr);
ibmr->sg = scat;
ibmr->sg_len = nents;
ibmr->sg_dma_len = sg_dma_len;
ibmr->remap_count++;
if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
rds_ib_stats_inc(s_ib_rdma_mr_8k_used);
else
rds_ib_stats_inc(s_ib_rdma_mr_1m_used);
ret = 0;
out:
kfree(dma_pages);
return ret;
}
void rds_ib_sync_mr(void *trans_private, int direction)
{
struct rds_ib_mr *ibmr = trans_private;
struct rds_ib_device *rds_ibdev = ibmr->device;
switch (direction) {
case DMA_FROM_DEVICE:
ib_dma_sync_sg_for_cpu(rds_ibdev->dev, ibmr->sg,
ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
break;
case DMA_TO_DEVICE:
ib_dma_sync_sg_for_device(rds_ibdev->dev, ibmr->sg,
ibmr->sg_dma_len, DMA_BIDIRECTIONAL);
break;
}
}
static void __rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
{
struct rds_ib_device *rds_ibdev = ibmr->device;
if (ibmr->sg_dma_len) {
ib_dma_unmap_sg(rds_ibdev->dev,
ibmr->sg, ibmr->sg_len,
DMA_BIDIRECTIONAL);
ibmr->sg_dma_len = 0;
}
/* Release the s/g list */
if (ibmr->sg_len) {
unsigned int i;
for (i = 0; i < ibmr->sg_len; ++i) {
struct page *page = sg_page(&ibmr->sg[i]);
/* FIXME we need a way to tell a r/w MR
* from a r/o MR */
WARN_ON(!page->mapping && irqs_disabled());
set_page_dirty(page);
put_page(page);
}
kfree(ibmr->sg);
ibmr->sg = NULL;
ibmr->sg_len = 0;
}
}
static void rds_ib_teardown_mr(struct rds_ib_mr *ibmr)
{
unsigned int pinned = ibmr->sg_len;
__rds_ib_teardown_mr(ibmr);
if (pinned) {
struct rds_ib_mr_pool *pool = ibmr->pool;
atomic_sub(pinned, &pool->free_pinned);
}
}
static inline unsigned int rds_ib_flush_goal(struct rds_ib_mr_pool *pool, int free_all)
{
unsigned int item_count;
item_count = atomic_read(&pool->item_count);
if (free_all)
return item_count;
return 0;
}
/*
* given an llist of mrs, put them all into the list_head for more processing
*/
static unsigned int llist_append_to_list(struct llist_head *llist,
struct list_head *list)
{
struct rds_ib_mr *ibmr;
struct llist_node *node;
struct llist_node *next;
unsigned int count = 0;
node = llist_del_all(llist);
while (node) {
next = node->next;
ibmr = llist_entry(node, struct rds_ib_mr, llnode);
list_add_tail(&ibmr->unmap_list, list);
node = next;
count++;
}
return count;
}
/*
* this takes a list head of mrs and turns it into linked llist nodes
* of clusters. Each cluster has linked llist nodes of
* MR_CLUSTER_SIZE mrs that are ready for reuse.
*/
static void list_to_llist_nodes(struct rds_ib_mr_pool *pool,
struct list_head *list,
struct llist_node **nodes_head,
struct llist_node **nodes_tail)
{
struct rds_ib_mr *ibmr;
struct llist_node *cur = NULL;
struct llist_node **next = nodes_head;
list_for_each_entry(ibmr, list, unmap_list) {
cur = &ibmr->llnode;
*next = cur;
next = &cur->next;
}
*next = NULL;
*nodes_tail = cur;
}
/*
* Flush our pool of MRs.
* At a minimum, all currently unused MRs are unmapped.
* If the number of MRs allocated exceeds the limit, we also try
* to free as many MRs as needed to get back to this limit.
*/
static int rds_ib_flush_mr_pool(struct rds_ib_mr_pool *pool,
int free_all, struct rds_ib_mr **ibmr_ret)
{
struct rds_ib_mr *ibmr, *next;
struct llist_node *clean_nodes;
struct llist_node *clean_tail;
LIST_HEAD(unmap_list);
LIST_HEAD(fmr_list);
unsigned long unpinned = 0;
unsigned int nfreed = 0, dirty_to_clean = 0, free_goal;
int ret = 0;
if (pool->pool_type == RDS_IB_MR_8K_POOL)
rds_ib_stats_inc(s_ib_rdma_mr_8k_pool_flush);
else
rds_ib_stats_inc(s_ib_rdma_mr_1m_pool_flush);
if (ibmr_ret) {
DEFINE_WAIT(wait);
while (!mutex_trylock(&pool->flush_lock)) {
ibmr = rds_ib_reuse_fmr(pool);
if (ibmr) {
*ibmr_ret = ibmr;
finish_wait(&pool->flush_wait, &wait);
goto out_nolock;
}
prepare_to_wait(&pool->flush_wait, &wait,
TASK_UNINTERRUPTIBLE);
if (llist_empty(&pool->clean_list))
schedule();
ibmr = rds_ib_reuse_fmr(pool);
if (ibmr) {
*ibmr_ret = ibmr;
finish_wait(&pool->flush_wait, &wait);
goto out_nolock;
}
}
finish_wait(&pool->flush_wait, &wait);
} else
mutex_lock(&pool->flush_lock);
if (ibmr_ret) {
ibmr = rds_ib_reuse_fmr(pool);
if (ibmr) {
*ibmr_ret = ibmr;
goto out;
}
}
/* Get the list of all MRs to be dropped. Ordering matters -
* we want to put drop_list ahead of free_list.
*/
dirty_to_clean = llist_append_to_list(&pool->drop_list, &unmap_list);
dirty_to_clean += llist_append_to_list(&pool->free_list, &unmap_list);
if (free_all)
llist_append_to_list(&pool->clean_list, &unmap_list);
free_goal = rds_ib_flush_goal(pool, free_all);
if (list_empty(&unmap_list))
goto out;
/* String all ib_mr's onto one list and hand them to ib_unmap_fmr */
list_for_each_entry(ibmr, &unmap_list, unmap_list)
list_add(&ibmr->fmr->list, &fmr_list);
ret = ib_unmap_fmr(&fmr_list);
if (ret)
printk(KERN_WARNING "RDS/IB: ib_unmap_fmr failed (err=%d)\n", ret);
/* Now we can destroy the DMA mapping and unpin any pages */
list_for_each_entry_safe(ibmr, next, &unmap_list, unmap_list) {
unpinned += ibmr->sg_len;
__rds_ib_teardown_mr(ibmr);
if (nfreed < free_goal ||
ibmr->remap_count >= pool->fmr_attr.max_maps) {
if (ibmr->pool->pool_type == RDS_IB_MR_8K_POOL)
rds_ib_stats_inc(s_ib_rdma_mr_8k_free);
else
rds_ib_stats_inc(s_ib_rdma_mr_1m_free);
list_del(&ibmr->unmap_list);
ib_dealloc_fmr(ibmr->fmr);
kfree(ibmr);
nfreed++;
}
}
if (!list_empty(&unmap_list)) {
/* we have to make sure that none of the things we're about
* to put on the clean list would race with other cpus trying
* to pull items off. The llist would explode if we managed to
* remove something from the clean list and then add it back again
* while another CPU was spinning on that same item in llist_del_first.
*
* This is pretty unlikely, but just in case wait for an llist grace period
* here before adding anything back into the clean list.
*/
wait_clean_list_grace();
list_to_llist_nodes(pool, &unmap_list, &clean_nodes, &clean_tail);
if (ibmr_ret)
*ibmr_ret = llist_entry(clean_nodes, struct rds_ib_mr, llnode);
/* more than one entry in llist nodes */
if (clean_nodes->next)
llist_add_batch(clean_nodes->next, clean_tail, &pool->clean_list);
}
atomic_sub(unpinned, &pool->free_pinned);
atomic_sub(dirty_to_clean, &pool->dirty_count);
atomic_sub(nfreed, &pool->item_count);
out:
mutex_unlock(&pool->flush_lock);
if (waitqueue_active(&pool->flush_wait))
wake_up(&pool->flush_wait);
out_nolock:
return ret;
}
static void rds_ib_mr_pool_flush_worker(struct work_struct *work)
{
struct rds_ib_mr_pool *pool = container_of(work, struct rds_ib_mr_pool, flush_worker.work);
rds_ib_flush_mr_pool(pool, 0, NULL);
}
void rds_ib_free_mr(void *trans_private, int invalidate)
{
struct rds_ib_mr *ibmr = trans_private;
struct rds_ib_mr_pool *pool = ibmr->pool;
struct rds_ib_device *rds_ibdev = ibmr->device;
rdsdebug("RDS/IB: free_mr nents %u\n", ibmr->sg_len);
/* Return it to the pool's free list */
if (ibmr->remap_count >= pool->fmr_attr.max_maps)
llist_add(&ibmr->llnode, &pool->drop_list);
else
llist_add(&ibmr->llnode, &pool->free_list);
atomic_add(ibmr->sg_len, &pool->free_pinned);
atomic_inc(&pool->dirty_count);
/* If we've pinned too many pages, request a flush */
if (atomic_read(&pool->free_pinned) >= pool->max_free_pinned ||
atomic_read(&pool->dirty_count) >= pool->max_items / 5)
queue_delayed_work(rds_ib_fmr_wq, &pool->flush_worker, 10);
if (invalidate) {
if (likely(!in_interrupt())) {
rds_ib_flush_mr_pool(pool, 0, NULL);
} else {
/* We get here if the user created a MR marked
* as use_once and invalidate at the same time.
*/
queue_delayed_work(rds_ib_fmr_wq,
&pool->flush_worker, 10);
}
}
RDS/IB: add refcount tracking to struct rds_ib_device The RDS IB client .remove callback used to free the rds_ibdev for the given device unconditionally. This could race other users of the struct. This patch adds refcounting so that we only free the rds_ibdev once all of its users are done. Many rds_ibdev users are tied to connections. We give the connection a reference and change these users to reference the device in the connection instead of looking it up in the IB client data. The only user of the IB client data remaining is the first lookup of the device as connections are built up. Incrementing the reference count of a device found in the IB client data could race with final freeing so we use an RCU grace period to make sure that freeing won't happen until those lookups are done. MRs need the rds_ibdev to get at the pool that they're freed in to. They exist outside a connection and many MRs can reference different devices from one socket, so it was natural to have each MR hold a reference. MR refs can be dropped from interrupt handlers and final device teardown can block so we push it off to a work struct. Pool teardown had to be fixed to cancel its pending work instead of deadlocking waiting for all queued work, including itself, to finish. MRs get their reference from the global device list, which gets a reference. It is left unprotected by locks and remains racy. A simple global lock would be a significant bottleneck. More scalable (complicated) locking should be done carefully in a later patch. Signed-off-by: Zach Brown <zach.brown@oracle.com>
2010-05-18 22:48:51 +00:00
rds_ib_dev_put(rds_ibdev);
}
void rds_ib_flush_mrs(void)
{
struct rds_ib_device *rds_ibdev;
down_read(&rds_ib_devices_lock);
list_for_each_entry(rds_ibdev, &rds_ib_devices, list) {
if (rds_ibdev->mr_8k_pool)
rds_ib_flush_mr_pool(rds_ibdev->mr_8k_pool, 0, NULL);
if (rds_ibdev->mr_1m_pool)
rds_ib_flush_mr_pool(rds_ibdev->mr_1m_pool, 0, NULL);
}
up_read(&rds_ib_devices_lock);
}
void *rds_ib_get_mr(struct scatterlist *sg, unsigned long nents,
struct rds_sock *rs, u32 *key_ret)
{
struct rds_ib_device *rds_ibdev;
struct rds_ib_mr *ibmr = NULL;
int ret;
rds_ibdev = rds_ib_get_device(rs->rs_bound_addr);
if (!rds_ibdev) {
ret = -ENODEV;
goto out;
}
if (!rds_ibdev->mr_8k_pool || !rds_ibdev->mr_1m_pool) {
ret = -ENODEV;
goto out;
}
ibmr = rds_ib_alloc_fmr(rds_ibdev, nents);
if (IS_ERR(ibmr)) {
rds_ib_dev_put(rds_ibdev);
return ibmr;
}
ret = rds_ib_map_fmr(rds_ibdev, ibmr, sg, nents);
if (ret == 0)
*key_ret = ibmr->fmr->rkey;
else
printk(KERN_WARNING "RDS/IB: map_fmr failed (errno=%d)\n", ret);
ibmr->device = rds_ibdev;
RDS/IB: add refcount tracking to struct rds_ib_device The RDS IB client .remove callback used to free the rds_ibdev for the given device unconditionally. This could race other users of the struct. This patch adds refcounting so that we only free the rds_ibdev once all of its users are done. Many rds_ibdev users are tied to connections. We give the connection a reference and change these users to reference the device in the connection instead of looking it up in the IB client data. The only user of the IB client data remaining is the first lookup of the device as connections are built up. Incrementing the reference count of a device found in the IB client data could race with final freeing so we use an RCU grace period to make sure that freeing won't happen until those lookups are done. MRs need the rds_ibdev to get at the pool that they're freed in to. They exist outside a connection and many MRs can reference different devices from one socket, so it was natural to have each MR hold a reference. MR refs can be dropped from interrupt handlers and final device teardown can block so we push it off to a work struct. Pool teardown had to be fixed to cancel its pending work instead of deadlocking waiting for all queued work, including itself, to finish. MRs get their reference from the global device list, which gets a reference. It is left unprotected by locks and remains racy. A simple global lock would be a significant bottleneck. More scalable (complicated) locking should be done carefully in a later patch. Signed-off-by: Zach Brown <zach.brown@oracle.com>
2010-05-18 22:48:51 +00:00
rds_ibdev = NULL;
out:
if (ret) {
if (ibmr)
rds_ib_free_mr(ibmr, 0);
ibmr = ERR_PTR(ret);
}
RDS/IB: add refcount tracking to struct rds_ib_device The RDS IB client .remove callback used to free the rds_ibdev for the given device unconditionally. This could race other users of the struct. This patch adds refcounting so that we only free the rds_ibdev once all of its users are done. Many rds_ibdev users are tied to connections. We give the connection a reference and change these users to reference the device in the connection instead of looking it up in the IB client data. The only user of the IB client data remaining is the first lookup of the device as connections are built up. Incrementing the reference count of a device found in the IB client data could race with final freeing so we use an RCU grace period to make sure that freeing won't happen until those lookups are done. MRs need the rds_ibdev to get at the pool that they're freed in to. They exist outside a connection and many MRs can reference different devices from one socket, so it was natural to have each MR hold a reference. MR refs can be dropped from interrupt handlers and final device teardown can block so we push it off to a work struct. Pool teardown had to be fixed to cancel its pending work instead of deadlocking waiting for all queued work, including itself, to finish. MRs get their reference from the global device list, which gets a reference. It is left unprotected by locks and remains racy. A simple global lock would be a significant bottleneck. More scalable (complicated) locking should be done carefully in a later patch. Signed-off-by: Zach Brown <zach.brown@oracle.com>
2010-05-18 22:48:51 +00:00
if (rds_ibdev)
rds_ib_dev_put(rds_ibdev);
return ibmr;
}