qemu/block/stream.c

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

407 lines
13 KiB
C
Raw Normal View History

/*
* Image streaming
*
* Copyright IBM, Corp. 2011
*
* Authors:
* Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>
*
* This work is licensed under the terms of the GNU LGPL, version 2 or later.
* See the COPYING.LIB file in the top-level directory.
*
*/
#include "qemu/osdep.h"
#include "trace.h"
#include "block/block_int.h"
#include "block/blockjob_int.h"
2016-03-14 08:01:28 +00:00
#include "qapi/error.h"
#include "qapi/qmp/qdict.h"
#include "qemu/ratelimit.h"
#include "sysemu/block-backend.h"
#include "block/copy-on-read.h"
enum {
/*
* Maximum chunk size to feed to copy-on-read. This should be
* large enough to process multiple clusters in a single call, so
* that populating contiguous regions of the image is efficient.
*/
STREAM_CHUNK = 512 * 1024, /* in bytes */
};
typedef struct StreamBlockJob {
BlockJob common;
BlockBackend *blk;
BlockDriverState *base_overlay; /* COW overlay (stream from this) */
BlockDriverState *above_base; /* Node directly above the base */
BlockDriverState *cor_filter_bs;
BlockDriverState *target_bs;
BlockdevOnError on_error;
char *backing_file_str;
bool bs_read_only;
} StreamBlockJob;
static int coroutine_fn stream_populate(BlockBackend *blk,
int64_t offset, uint64_t bytes)
{
assert(bytes < SIZE_MAX);
return blk_co_preadv(blk, offset, bytes, NULL, BDRV_REQ_PREFETCH);
}
static int stream_prepare(Job *job)
{
StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
BlockDriverState *unfiltered_bs = bdrv_skip_filters(s->target_bs);
BlockDriverState *unfiltered_bs_cow = bdrv_cow_bs(unfiltered_bs);
BlockDriverState *base;
BlockDriverState *unfiltered_base;
Error *local_err = NULL;
int ret = 0;
/* We should drop filter at this point, as filter hold the backing chain */
bdrv_cor_filter_drop(s->cor_filter_bs);
s->cor_filter_bs = NULL;
/*
* bdrv_set_backing_hd() requires that the unfiltered_bs and the COW child
* of unfiltered_bs is drained. Drain already here and use
* bdrv_set_backing_hd_drained() instead because the polling during
* drained_begin() might change the graph, and if we do this only later, we
* may end up working with the wrong base node (or it might even have gone
* away by the time we want to use it).
*/
bdrv_drained_begin(unfiltered_bs);
if (unfiltered_bs_cow) {
bdrv_ref(unfiltered_bs_cow);
bdrv_drained_begin(unfiltered_bs_cow);
}
block/stream: Drain subtree around graph change When the stream block job cuts out the nodes between top and base in stream_prepare(), it does not drain the subtree manually; it fetches the base node, and tries to insert it as the top node's backing node with bdrv_set_backing_hd(). bdrv_set_backing_hd() however will drain, and so the actual base node might change (because the base node is actually not part of the stream job) before the old base node passed to bdrv_set_backing_hd() is installed. This has two implications: First, the stream job does not keep a strong reference to the base node. Therefore, if it is deleted in bdrv_set_backing_hd()'s drain (e.g. because some other block job is drained to finish), we will get a use-after-free. We should keep a strong reference to that node. Second, even with such a strong reference, the problem remains that the base node might change before bdrv_set_backing_hd() actually runs and as a result the wrong base node is installed. Both effects can be seen in 030's TestParallelOps.test_overlapping_5() case, which has five nodes, and simultaneously streams from the middle node to the top node, and commits the middle node down to the base node. As it is, this will sometimes crash, namely when we encounter the above-described use-after-free. Taking a strong reference to the base node, we no longer get a crash, but the resuling block graph is less than ideal: The expected result is obviously that all middle nodes are cut out and the base node is the immediate backing child of the top node. However, if stream_prepare() takes a strong reference to its base node (the middle node), and then the commit job finishes in bdrv_set_backing_hd(), supposedly dropping that middle node, the stream job will just reinstall it again. Therefore, we need to keep the whole subtree drained in stream_prepare(), so that the graph modification it performs is effectively atomic, i.e. that the base node it fetches is still the base node when bdrv_set_backing_hd() sets it as the top node's backing node. Verify this by asserting in said 030's test case that the base node is always the top node's immediate backing child when both jobs are done. Signed-off-by: Hanna Reitz <hreitz@redhat.com> Message-Id: <20220324140907.17192-1-hreitz@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Acked-by: Vladimir Sementsov-Ogievskiy <v.sementsov-og@mail.ru>
2022-03-24 14:09:07 +00:00
base = bdrv_filter_or_cow_bs(s->above_base);
unfiltered_base = bdrv_skip_filters(base);
if (bdrv_cow_child(unfiltered_bs)) {
const char *base_id = NULL, *base_fmt = NULL;
if (unfiltered_base) {
base_id = s->backing_file_str ?: unfiltered_base->filename;
if (unfiltered_base->drv) {
base_fmt = unfiltered_base->drv->format_name;
}
}
block/stream: Drain subtree around graph change When the stream block job cuts out the nodes between top and base in stream_prepare(), it does not drain the subtree manually; it fetches the base node, and tries to insert it as the top node's backing node with bdrv_set_backing_hd(). bdrv_set_backing_hd() however will drain, and so the actual base node might change (because the base node is actually not part of the stream job) before the old base node passed to bdrv_set_backing_hd() is installed. This has two implications: First, the stream job does not keep a strong reference to the base node. Therefore, if it is deleted in bdrv_set_backing_hd()'s drain (e.g. because some other block job is drained to finish), we will get a use-after-free. We should keep a strong reference to that node. Second, even with such a strong reference, the problem remains that the base node might change before bdrv_set_backing_hd() actually runs and as a result the wrong base node is installed. Both effects can be seen in 030's TestParallelOps.test_overlapping_5() case, which has five nodes, and simultaneously streams from the middle node to the top node, and commits the middle node down to the base node. As it is, this will sometimes crash, namely when we encounter the above-described use-after-free. Taking a strong reference to the base node, we no longer get a crash, but the resuling block graph is less than ideal: The expected result is obviously that all middle nodes are cut out and the base node is the immediate backing child of the top node. However, if stream_prepare() takes a strong reference to its base node (the middle node), and then the commit job finishes in bdrv_set_backing_hd(), supposedly dropping that middle node, the stream job will just reinstall it again. Therefore, we need to keep the whole subtree drained in stream_prepare(), so that the graph modification it performs is effectively atomic, i.e. that the base node it fetches is still the base node when bdrv_set_backing_hd() sets it as the top node's backing node. Verify this by asserting in said 030's test case that the base node is always the top node's immediate backing child when both jobs are done. Signed-off-by: Hanna Reitz <hreitz@redhat.com> Message-Id: <20220324140907.17192-1-hreitz@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Acked-by: Vladimir Sementsov-Ogievskiy <v.sementsov-og@mail.ru>
2022-03-24 14:09:07 +00:00
bdrv_set_backing_hd_drained(unfiltered_bs, base, &local_err);
/*
* This call will do I/O, so the graph can change again from here on.
* We have already completed the graph change, so we are not in danger
* of operating on the wrong node any more if this happens.
*/
ret = bdrv_change_backing_file(unfiltered_bs, base_id, base_fmt, false);
if (local_err) {
error_report_err(local_err);
block/stream: Drain subtree around graph change When the stream block job cuts out the nodes between top and base in stream_prepare(), it does not drain the subtree manually; it fetches the base node, and tries to insert it as the top node's backing node with bdrv_set_backing_hd(). bdrv_set_backing_hd() however will drain, and so the actual base node might change (because the base node is actually not part of the stream job) before the old base node passed to bdrv_set_backing_hd() is installed. This has two implications: First, the stream job does not keep a strong reference to the base node. Therefore, if it is deleted in bdrv_set_backing_hd()'s drain (e.g. because some other block job is drained to finish), we will get a use-after-free. We should keep a strong reference to that node. Second, even with such a strong reference, the problem remains that the base node might change before bdrv_set_backing_hd() actually runs and as a result the wrong base node is installed. Both effects can be seen in 030's TestParallelOps.test_overlapping_5() case, which has five nodes, and simultaneously streams from the middle node to the top node, and commits the middle node down to the base node. As it is, this will sometimes crash, namely when we encounter the above-described use-after-free. Taking a strong reference to the base node, we no longer get a crash, but the resuling block graph is less than ideal: The expected result is obviously that all middle nodes are cut out and the base node is the immediate backing child of the top node. However, if stream_prepare() takes a strong reference to its base node (the middle node), and then the commit job finishes in bdrv_set_backing_hd(), supposedly dropping that middle node, the stream job will just reinstall it again. Therefore, we need to keep the whole subtree drained in stream_prepare(), so that the graph modification it performs is effectively atomic, i.e. that the base node it fetches is still the base node when bdrv_set_backing_hd() sets it as the top node's backing node. Verify this by asserting in said 030's test case that the base node is always the top node's immediate backing child when both jobs are done. Signed-off-by: Hanna Reitz <hreitz@redhat.com> Message-Id: <20220324140907.17192-1-hreitz@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Acked-by: Vladimir Sementsov-Ogievskiy <v.sementsov-og@mail.ru>
2022-03-24 14:09:07 +00:00
ret = -EPERM;
goto out;
}
}
block/stream: Drain subtree around graph change When the stream block job cuts out the nodes between top and base in stream_prepare(), it does not drain the subtree manually; it fetches the base node, and tries to insert it as the top node's backing node with bdrv_set_backing_hd(). bdrv_set_backing_hd() however will drain, and so the actual base node might change (because the base node is actually not part of the stream job) before the old base node passed to bdrv_set_backing_hd() is installed. This has two implications: First, the stream job does not keep a strong reference to the base node. Therefore, if it is deleted in bdrv_set_backing_hd()'s drain (e.g. because some other block job is drained to finish), we will get a use-after-free. We should keep a strong reference to that node. Second, even with such a strong reference, the problem remains that the base node might change before bdrv_set_backing_hd() actually runs and as a result the wrong base node is installed. Both effects can be seen in 030's TestParallelOps.test_overlapping_5() case, which has five nodes, and simultaneously streams from the middle node to the top node, and commits the middle node down to the base node. As it is, this will sometimes crash, namely when we encounter the above-described use-after-free. Taking a strong reference to the base node, we no longer get a crash, but the resuling block graph is less than ideal: The expected result is obviously that all middle nodes are cut out and the base node is the immediate backing child of the top node. However, if stream_prepare() takes a strong reference to its base node (the middle node), and then the commit job finishes in bdrv_set_backing_hd(), supposedly dropping that middle node, the stream job will just reinstall it again. Therefore, we need to keep the whole subtree drained in stream_prepare(), so that the graph modification it performs is effectively atomic, i.e. that the base node it fetches is still the base node when bdrv_set_backing_hd() sets it as the top node's backing node. Verify this by asserting in said 030's test case that the base node is always the top node's immediate backing child when both jobs are done. Signed-off-by: Hanna Reitz <hreitz@redhat.com> Message-Id: <20220324140907.17192-1-hreitz@redhat.com> Reviewed-by: Eric Blake <eblake@redhat.com> Acked-by: Vladimir Sementsov-Ogievskiy <v.sementsov-og@mail.ru>
2022-03-24 14:09:07 +00:00
out:
if (unfiltered_bs_cow) {
bdrv_drained_end(unfiltered_bs_cow);
bdrv_unref(unfiltered_bs_cow);
}
bdrv_drained_end(unfiltered_bs);
return ret;
}
static void stream_clean(Job *job)
{
StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
if (s->cor_filter_bs) {
bdrv_cor_filter_drop(s->cor_filter_bs);
s->cor_filter_bs = NULL;
}
blk_unref(s->blk);
s->blk = NULL;
/* Reopen the image back in read-only mode if necessary */
if (s->bs_read_only) {
/* Give up write permissions before making it read-only */
bdrv_reopen_set_read_only(s->target_bs, true, NULL);
}
g_free(s->backing_file_str);
}
static int coroutine_fn stream_run(Job *job, Error **errp)
{
StreamBlockJob *s = container_of(job, StreamBlockJob, common.job);
BlockDriverState *unfiltered_bs = bdrv_skip_filters(s->target_bs);
int64_t len;
int64_t offset = 0;
int error = 0;
int64_t n = 0; /* bytes */
if (unfiltered_bs == s->base_overlay) {
/* Nothing to stream */
return 0;
}
WITH_GRAPH_RDLOCK_GUARD() {
len = bdrv_co_getlength(s->target_bs);
if (len < 0) {
return len;
}
}
job_progress_set_remaining(&s->common.job, len);
for ( ; offset < len; offset += n) {
bool copy;
int ret;
/* Note that even when no rate limit is applied we need to yield
* with no pending I/O here so that bdrv_drain_all() returns.
*/
block_job_ratelimit_sleep(&s->common);
if (job_is_cancelled(&s->common.job)) {
break;
}
copy = false;
WITH_GRAPH_RDLOCK_GUARD() {
ret = bdrv_co_is_allocated(unfiltered_bs, offset, STREAM_CHUNK, &n);
if (ret == 1) {
/* Allocated in the top, no need to copy. */
} else if (ret >= 0) {
/*
* Copy if allocated in the intermediate images. Limit to the
* known-unallocated area [offset, offset+n*BDRV_SECTOR_SIZE).
*/
ret = bdrv_co_is_allocated_above(bdrv_cow_bs(unfiltered_bs),
s->base_overlay, true,
offset, n, &n);
/* Finish early if end of backing file has been reached */
if (ret == 0 && n == 0) {
n = len - offset;
}
copy = (ret > 0);
}
}
trace_stream_one_iteration(s, offset, n, ret);
if (copy) {
ret = stream_populate(s->blk, offset, n);
}
if (ret < 0) {
BlockErrorAction action =
block_job_error_action(&s->common, s->on_error, true, -ret);
if (action == BLOCK_ERROR_ACTION_STOP) {
n = 0;
continue;
}
if (error == 0) {
error = ret;
}
if (action == BLOCK_ERROR_ACTION_REPORT) {
break;
}
}
/* Publish progress */
job_progress_update(&s->common.job, n);
if (copy) {
block_job_ratelimit_processed_bytes(&s->common, n);
Improve block job rate limiting for small bandwidth values ratelimit_calculate_delay() previously reset the accounting every time slice, no matter how much data had been processed before. This had (at least) two consequences: 1. The minimum speed is rather large, e.g. 5 MiB/s for commit and stream. Not sure if there are real-world use cases where this would be a problem. Mirroring and backup over a slow link (e.g. DSL) would come to mind, though. 2. Tests for block job operations (e.g. cancel) were rather racy All block jobs currently use a time slice of 100ms. That's a reasonable value to get smooth output during regular operation. However this also meant that the state of block jobs changed every 100ms, no matter how low the configured limit was. On busy hosts, qemu often transferred additional chunks until the test case had a chance to cancel the job. Fix the block job rate limit code to delay for more than one time slice to address the above issues. To make it easier to handle oversized chunks we switch the semantics from returning a delay _before_ the current request to a delay _after_ the current request. If necessary, this delay consists of multiple time slice units. Since the mirror job sends multiple chunks in one go even if the rate limit was exceeded in between, we need to keep track of the start of the current time slice so we can correctly re-compute the delay for the updated amount of data. The minimum bandwidth now is 1 data unit per time slice. The block jobs are currently passing the amount of data transferred in sectors and using 100ms time slices, so this translates to 5120 bytes/second. With chunk sizes usually being O(512KiB), tests have plenty of time (O(100s)) to operate on block jobs. The chance of a race condition now is fairly remote, except possibly on insanely loaded systems. Signed-off-by: Sascha Silbe <silbe@linux.vnet.ibm.com> Message-id: 1467127721-9564-2-git-send-email-silbe@linux.vnet.ibm.com Reviewed-by: Max Reitz <mreitz@redhat.com> Signed-off-by: Max Reitz <mreitz@redhat.com>
2016-06-28 15:28:41 +00:00
}
}
/* Do not remove the backing file if an error was there but ignored. */
return error;
}
static const BlockJobDriver stream_job_driver = {
.job_driver = {
.instance_size = sizeof(StreamBlockJob),
.job_type = JOB_TYPE_STREAM,
.free = block_job_free,
.run = stream_run,
.prepare = stream_prepare,
.clean = stream_clean,
.user_resume = block_job_user_resume,
},
};
void stream_start(const char *job_id, BlockDriverState *bs,
BlockDriverState *base, const char *backing_file_str,
BlockDriverState *bottom,
int creation_flags, int64_t speed,
BlockdevOnError on_error,
const char *filter_node_name,
Error **errp)
{
StreamBlockJob *s = NULL;
BlockDriverState *iter;
bool bs_read_only;
int basic_flags = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED;
BlockDriverState *base_overlay;
BlockDriverState *cor_filter_bs = NULL;
BlockDriverState *above_base;
QDict *opts;
int ret;
GLOBAL_STATE_CODE();
assert(!(base && bottom));
assert(!(backing_file_str && bottom));
if (bottom) {
/*
* New simple interface. The code is written in terms of old interface
* with @base parameter (still, it doesn't freeze link to base, so in
* this mean old code is correct for new interface). So, for now, just
* emulate base_overlay and above_base. Still, when old interface
* finally removed, we should refactor code to use only "bottom", but
* not "*base*" things.
*/
assert(!bottom->drv->is_filter);
base_overlay = above_base = bottom;
} else {
GRAPH_RDLOCK_GUARD_MAINLOOP();
base_overlay = bdrv_find_overlay(bs, base);
if (!base_overlay) {
error_setg(errp, "'%s' is not in the backing chain of '%s'",
base->node_name, bs->node_name);
return;
}
/*
* Find the node directly above @base. @base_overlay is a COW overlay,
* so it must have a bdrv_cow_child(), but it is the immediate overlay
* of @base, so between the two there can only be filters.
*/
above_base = base_overlay;
if (bdrv_cow_bs(above_base) != base) {
above_base = bdrv_cow_bs(above_base);
while (bdrv_filter_bs(above_base) != base) {
above_base = bdrv_filter_bs(above_base);
}
}
}
/* Make sure that the image is opened in read-write mode */
bs_read_only = bdrv_is_read_only(bs);
if (bs_read_only) {
/* Hold the chain during reopen */
if (bdrv_freeze_backing_chain(bs, above_base, errp) < 0) {
return;
}
ret = bdrv_reopen_set_read_only(bs, false, errp);
/* failure, or cor-filter will hold the chain */
bdrv_unfreeze_backing_chain(bs, above_base);
if (ret < 0) {
return;
}
}
opts = qdict_new();
qdict_put_str(opts, "driver", "copy-on-read");
qdict_put_str(opts, "file", bdrv_get_node_name(bs));
/* Pass the base_overlay node name as 'bottom' to COR driver */
qdict_put_str(opts, "bottom", base_overlay->node_name);
if (filter_node_name) {
qdict_put_str(opts, "node-name", filter_node_name);
}
cor_filter_bs = bdrv_insert_node(bs, opts, BDRV_O_RDWR, errp);
if (!cor_filter_bs) {
goto fail;
}
if (!filter_node_name) {
cor_filter_bs->implicit = true;
}
s = block_job_create(job_id, &stream_job_driver, NULL, cor_filter_bs,
0, BLK_PERM_ALL,
speed, creation_flags, NULL, NULL, errp);
if (!s) {
goto fail;
}
s->blk = blk_new_with_bs(cor_filter_bs, BLK_PERM_CONSISTENT_READ,
basic_flags | BLK_PERM_WRITE, errp);
if (!s->blk) {
goto fail;
}
/*
* Disable request queuing in the BlockBackend to avoid deadlocks on drain:
* The job reports that it's busy until it reaches a pause point.
*/
blk_set_disable_request_queuing(s->blk, true);
blk_set_allow_aio_context_change(s->blk, true);
/*
* Prevent concurrent jobs trying to modify the graph structure here, we
* already have our own plans. Also don't allow resize as the image size is
* queried only at the job start and then cached.
*/
bdrv_graph_wrlock(bs);
if (block_job_add_bdrv(&s->common, "active node", bs, 0,
basic_flags | BLK_PERM_WRITE, errp)) {
bdrv_graph_wrunlock();
goto fail;
}
/* Block all intermediate nodes between bs and base, because they will
* disappear from the chain after this operation. The streaming job reads
* every block only once, assuming that it doesn't change, so forbid writes
* and resizes. Reassign the base node pointer because the backing BS of the
* bottom node might change after the call to bdrv_reopen_set_read_only()
* due to parallel block jobs running.
* above_base node might change after the call to
* bdrv_reopen_set_read_only() due to parallel block jobs running.
*/
base = bdrv_filter_or_cow_bs(above_base);
for (iter = bdrv_filter_or_cow_bs(bs); iter != base;
iter = bdrv_filter_or_cow_bs(iter))
{
ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
basic_flags, errp);
if (ret < 0) {
bdrv_graph_wrunlock();
goto fail;
}
}
bdrv_graph_wrunlock();
s->base_overlay = base_overlay;
s->above_base = above_base;
s->backing_file_str = g_strdup(backing_file_str);
s->cor_filter_bs = cor_filter_bs;
s->target_bs = bs;
s->bs_read_only = bs_read_only;
s->on_error = on_error;
trace_stream_start(bs, base, s);
job_start(&s->common.job);
return;
fail:
if (s) {
job_early_fail(&s->common.job);
}
if (cor_filter_bs) {
bdrv_cor_filter_drop(cor_filter_bs);
}
if (bs_read_only) {
bdrv_reopen_set_read_only(bs, true, NULL);
}
}