From: "Benoît Canet" <benoit.canet@nodalink.com>
To: Kevin Wolf <kwolf@redhat.com>
Cc: famz@redhat.com, jcody@redhat.com, stefanha@redhat.com,
qemu-devel@nongnu.org, "Benoît Canet" <benoit.canet@nodalink.com>
Subject: Re: [Qemu-devel] [PATCH] block: Make op blockers recursive
Date: Tue, 9 Sep 2014 14:28:22 +0000 [thread overview]
Message-ID: <20140909142822.GA7953@nodalink.com> (raw)
In-Reply-To: <20140909115646.GI4847@noname.str.redhat.com>
On Tue, Sep 09, 2014 at 01:56:46PM +0200, Kevin Wolf wrote:
> Am 22.08.2014 um 18:11 hat Benoît Canet geschrieben:
> > Since the block layer code is starting to modify the BDS graph right in the
> > middle of BDS chains (block-mirror's replace parameter for example) QEMU needs
> > to properly block and unblock whole BDS subtrees; recursion is a neat way to
> > achieve this task.
> >
> > This patch also takes care of modifying the op blockers users.
> >
> > Signed-off-by: Benoit Canet <benoit.canet@nodalink.com>
> > ---
> > block.c | 69 ++++++++++++++++++++++++++++++++++++++++++++---
> > block/blkverify.c | 21 +++++++++++++++
> > block/commit.c | 3 +++
> > block/mirror.c | 17 ++++++++----
> > block/quorum.c | 25 +++++++++++++++++
> > block/stream.c | 3 +++
> > block/vmdk.c | 34 +++++++++++++++++++++++
> > include/block/block.h | 2 +-
> > include/block/block_int.h | 6 +++++
> > 9 files changed, 171 insertions(+), 9 deletions(-)
> >
> > diff --git a/block.c b/block.c
> > index 6fa0201..d964b6c 100644
> > --- a/block.c
> > +++ b/block.c
> > @@ -5446,7 +5446,9 @@ bool bdrv_op_is_blocked(BlockDriverState *bs, BlockOpType op, Error **errp)
> > return false;
> > }
> >
> > -void bdrv_op_block(BlockDriverState *bs, BlockOpType op, Error *reason)
> > +/* do the real work of blocking a BDS */
> > +static void bdrv_do_op_block(BlockDriverState *bs, BlockOpType op,
> > + Error *reason)
> > {
> > BdrvOpBlocker *blocker;
> > assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX);
> > @@ -5456,7 +5458,9 @@ void bdrv_op_block(BlockDriverState *bs, BlockOpType op, Error *reason)
> > QLIST_INSERT_HEAD(&bs->op_blockers[op], blocker, list);
> > }
> >
> > -void bdrv_op_unblock(BlockDriverState *bs, BlockOpType op, Error *reason)
> > +/* do the real work of unblocking a BDS */
> > +static void bdrv_do_op_unblock(BlockDriverState *bs, BlockOpType op,
> > + Error *reason)
> > {
> > BdrvOpBlocker *blocker, *next;
> > assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX);
> > @@ -5468,6 +5472,65 @@ void bdrv_op_unblock(BlockDriverState *bs, BlockOpType op, Error *reason)
> > }
> > }
> >
> > +static bool bdrv_op_is_blocked_by(BlockDriverState *bs, BlockOpType op,
> > + Error *reason)
> > +{
> > + BdrvOpBlocker *blocker, *next;
> > + assert((int) op >= 0 && op < BLOCK_OP_TYPE_MAX);
> > + QLIST_FOREACH_SAFE(blocker, &bs->op_blockers[op], list, next) {
>
> This doesn't actually need the SAFE version.
>
> > + if (blocker->reason == reason) {
> > + return true;
> > + }
> > + }
> > + return false;
> > +}
> > +
> > +/* block recursively a BDS */
> > +void bdrv_op_block(BlockDriverState *bs, BlockOpType op, Error *reason)
> > +{
> > + if (!bs) {
> > + return;
> > + }
> > +
> > + /* prevent recursion loop */
> > + if (bdrv_op_is_blocked_by(bs, op, reason)) {
> > + return;
> > + }
> > +
> > + /* block first for recursion loop protection to work */
> > + bdrv_do_op_block(bs, op, reason);
> > +
> > + bdrv_op_block(bs->file, op, reason);
> > + bdrv_op_block(bs->backing_hd, op, reason);
> > +
> > + if (bs->drv && bs->drv->bdrv_op_recursive_block) {
> > + bs->drv->bdrv_op_recursive_block(bs, op, reason);
> > + }
>
> Here you block bs->file/bs->backing_hd automatically, no matter whether
> the block driver implements the callback or not. I'm not sure if we can
> do it like this for all times, but for now that should be okay.
>
> > +}
> > +
> > +/* unblock recursively a BDS */
> > +void bdrv_op_unblock(BlockDriverState *bs, BlockOpType op, Error *reason)
> > +{
> > + if (!bs) {
> > + return;
> > + }
> > +
> > + /* prevent recursion loop */
> > + if (!bdrv_op_is_blocked_by(bs, op, reason)) {
> > + return;
> > + }
> > +
> > + /* unblock first for recursion loop protection to work */
> > + bdrv_do_op_unblock(bs, op, reason);
> > +
> > + bdrv_op_unblock(bs->file, op, reason);
> > + bdrv_op_unblock(bs->backing_hd, op, reason);
> > +
> > + if (bs->drv && bs->drv->bdrv_op_recursive_unblock) {
> > + bs->drv->bdrv_op_recursive_unblock(bs, op, reason);
> > + }
> > +}
> > +
> > void bdrv_op_block_all(BlockDriverState *bs, Error *reason)
> > {
> > int i;
> > @@ -5848,7 +5911,7 @@ BlockDriverState *check_to_replace_node(const char *node_name, Error **errp)
> > return NULL;
> > }
> >
> > - if (bdrv_op_is_blocked(to_replace_bs, BLOCK_OP_TYPE_REPLACE, errp)) {
> > + if (bdrv_op_is_blocked(to_replace_bs, BLOCK_OP_TYPE_MIRROR_REPLACE, errp)) {
>
> That rename could have been a separate patch.
>
> > return NULL;
> > }
> >
> > diff --git a/block/blkverify.c b/block/blkverify.c
> > index 621b785..75ec3df 100644
> > --- a/block/blkverify.c
> > +++ b/block/blkverify.c
> > @@ -320,6 +320,24 @@ static void blkverify_attach_aio_context(BlockDriverState *bs,
> > bdrv_attach_aio_context(s->test_file, new_context);
> > }
> >
> > +static void blkverify_op_recursive_block(BlockDriverState *bs, BlockOpType op,
> > + Error *reason)
> > +{
> > + BDRVBlkverifyState *s = bs->opaque;
> > +
> > + bdrv_op_block(bs->file, op, reason);
> > + bdrv_op_block(s->test_file, op, reason);
> > +}
> > +
> > +static void blkverify_op_recursive_unblock(BlockDriverState *bs, BlockOpType op,
> > + Error *reason)
> > +{
> > + BDRVBlkverifyState *s = bs->opaque;
> > +
> > + bdrv_op_unblock(bs->file, op, reason);
> > + bdrv_op_unblock(s->test_file, op, reason);
> > +}
>
> Here you block bs->file again, even though the block.c function has
> already done it. Nothing bad happens, because bdrv_op_block() simply
> stops when the blocker is already set, but the code would be nicer if we
> did block bs->file only in one place. (I'm leaning towards doing it in
> the drivers, but that's your decision.)
>
> > static BlockDriver bdrv_blkverify = {
> > .format_name = "blkverify",
> > .protocol_name = "blkverify",
> > @@ -337,6 +355,9 @@ static BlockDriver bdrv_blkverify = {
> > .bdrv_attach_aio_context = blkverify_attach_aio_context,
> > .bdrv_detach_aio_context = blkverify_detach_aio_context,
> >
> > + .bdrv_op_recursive_block = blkverify_op_recursive_block,
> > + .bdrv_op_recursive_unblock = blkverify_op_recursive_unblock,
> > +
> > .is_filter = true,
> > .bdrv_recurse_is_first_non_filter = blkverify_recurse_is_first_non_filter,
> > };
> > diff --git a/block/commit.c b/block/commit.c
> > index 91517d3..8a122b7 100644
> > --- a/block/commit.c
> > +++ b/block/commit.c
> > @@ -142,6 +142,9 @@ wait:
> >
> > if (!block_job_is_cancelled(&s->common) && sector_num == end) {
> > /* success */
> > + /* unblock only BDS to be dropped */
> > + bdrv_op_unblock_all(top, s->common.blocker);
> > + bdrv_op_block_all(base, s->common.blocker);
>
> This suggests that bdrv_op_(un)block_all() doesn't provide the right
> interface, but we rather want an optional base argument in it, so that
> you can (un)block a specific part of the backing file chain.
Good idea.
>
> > ret = bdrv_drop_intermediate(active, top, base, s->backing_file_str);
> > }
> >
> > diff --git a/block/mirror.c b/block/mirror.c
> > index 5e7a166..28ed47d 100644
> > --- a/block/mirror.c
> > +++ b/block/mirror.c
> > @@ -324,6 +324,7 @@ static void coroutine_fn mirror_run(void *opaque)
> > uint64_t last_pause_ns;
> > BlockDriverInfo bdi;
> > char backing_filename[1024];
> > + BlockDriverState *to_replace;
> > int ret = 0;
> > int n;
> >
> > @@ -512,14 +513,16 @@ immediate_exit:
> > g_free(s->in_flight_bitmap);
> > bdrv_release_dirty_bitmap(bs, s->dirty_bitmap);
> > bdrv_iostatus_disable(s->target);
> > + to_replace = s->common.bs;
> > + if (s->to_replace) {
> > + bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
> > + to_replace = s->to_replace;
> > + }
> > if (s->should_complete && ret == 0) {
> > - BlockDriverState *to_replace = s->common.bs;
> > - if (s->to_replace) {
> > - to_replace = s->to_replace;
> > - }
> > if (bdrv_get_flags(s->target) != bdrv_get_flags(to_replace)) {
> > bdrv_reopen(s->target, bdrv_get_flags(to_replace), NULL);
> > }
> > + bdrv_op_unblock_all(to_replace, bs->job->blocker);
>
> You need to unblock all of the BDSes that are going to be removed.
> Aren't you unblocking more than that here?
>
> It's not a real problem because block_job_completed() would unblock the
> rest anyway and its bdrv_op_unblock_all() call is simply ignored now,
> but it would be nicer to just unblock here what's actually needed.
>
> > bdrv_swap(s->target, to_replace);
> > if (s->common.driver->job_type == BLOCK_JOB_TYPE_COMMIT) {
> > /* drop the bs loop chain formed by the swap: break the loop then
> > @@ -530,7 +533,6 @@ immediate_exit:
> > }
> > }
> > if (s->to_replace) {
> > - bdrv_op_unblock_all(s->to_replace, s->replace_blocker);
> > error_free(s->replace_blocker);
> > bdrv_unref(s->to_replace);
> > }
> > @@ -648,6 +650,11 @@ static void mirror_start_job(BlockDriverState *bs, BlockDriverState *target,
> > return;
> > }
> >
> > + /* remove BLOCK_OP_TYPE_MIRROR_REPLACE from the list of
> > + * blocked operations so the replaces parameter can work
> > + */
> > + bdrv_op_unblock(bs, BLOCK_OP_TYPE_MIRROR_REPLACE, bs->job->blocker);
>
> What purpose does a blocker serve when it's disabled before it is
> checked? I would only ever expect a bdrv_op_unblock() after some
> operation on the BDS has finished, but not when starting an operation
> that needs it.
BLOCK_OP_TYPE_MIRROR_REPLACE is checked and blocked by block-job-complete
during the time the mirror finish when an arbitrary node of the graph must be
replaced.
The start of the mirroring block everything including
BLOCK_OP_TYPE_MIRROR_REPLACE so this hunk relax the blocking so block-job-complete
can have a chance of being able to block it.
The comment of this hunk seems to be deficient and the code not self explanatory.
On the other way maybe block-job-complete is done wrong from the start but
relaxing BLOCK_OP_TYPE_MIRROR_REPLACE when the mirror start is the only way
to make block-job-complete work as intended.
Thanks
Benoît
>
> > s->replaces = g_strdup(replaces);
> > s->on_source_error = on_source_error;
> > s->on_target_error = on_target_error;
> > diff --git a/block/stream.c b/block/stream.c
> > index cdea3e8..2c917b7 100644
> > --- a/block/stream.c
> > +++ b/block/stream.c
> > @@ -192,6 +192,9 @@ wait:
> > }
> > }
> > ret = bdrv_change_backing_file(bs, base_id, base_fmt);
> > + /* unblock only BDS to be dropped */
> > + bdrv_op_unblock_all(bs->backing_hd, s->common.blocker);
> > + bdrv_op_block_all(base, s->common.blocker);
>
> Same thing as for commit.
>
> > close_unused_images(bs, base, base_id);
> > }
> >
> > diff --git a/block/vmdk.c b/block/vmdk.c
> > index 01412a8..bc5b17f 100644
> > --- a/block/vmdk.c
> > +++ b/block/vmdk.c
> > @@ -2220,6 +2220,38 @@ static QemuOptsList vmdk_create_opts = {
> > }
> > };
> >
> > +static void vmdk_op_recursive_block(BlockDriverState *bs, BlockOpType op,
> > + Error *reason)
> > +{
> > + BDRVVmdkState *s = bs->opaque;
> > + int i;
> > +
> > + bdrv_op_block(bs->file, op, reason);
> > + bdrv_op_block(bs->backing_hd, op, reason);
>
> Already blocked in block.c, remove one of the two
>
> > + for (i = 0; i < s->num_extents; i++) {
> > + if (s->extents[i].file) {
> > + bdrv_op_block(s->extents[i].file, op, reason);
> > + }
> > + }
> > +}
>
> Kevin
next prev parent reply other threads:[~2014-09-09 14:28 UTC|newest]
Thread overview: 21+ messages / expand[flat|nested] mbox.gz Atom feed top
2014-08-22 16:11 [Qemu-devel] [PATCH] Implement recursive op blockers Benoît Canet
2014-08-22 16:11 ` [Qemu-devel] [PATCH] block: Make op blockers recursive Benoît Canet
2014-08-25 6:04 ` Fam Zheng
2014-08-25 9:06 ` Benoît Canet
2014-08-25 9:37 ` Fam Zheng
2014-08-25 12:12 ` Benoît Canet
2014-08-26 4:42 ` Fam Zheng
2014-08-26 6:45 ` Benoît Canet
2014-09-04 20:42 ` Stefan Hajnoczi
2014-09-10 8:54 ` Fam Zheng
2014-09-10 14:18 ` Benoît Canet
2014-09-10 15:14 ` Eric Blake
2014-09-10 15:49 ` Benoît Canet
2014-09-11 11:22 ` Kevin Wolf
2014-09-11 0:50 ` Fam Zheng
2014-09-09 11:56 ` Kevin Wolf
2014-09-09 14:28 ` Benoît Canet [this message]
2014-09-12 3:48 ` Fam Zheng
2014-09-15 15:17 ` Benoît Canet
2014-09-18 2:57 ` Fam Zheng
2014-09-22 12:33 ` Benoît Canet
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=20140909142822.GA7953@nodalink.com \
--to=benoit.canet@nodalink.com \
--cc=famz@redhat.com \
--cc=jcody@redhat.com \
--cc=kwolf@redhat.com \
--cc=qemu-devel@nongnu.org \
--cc=stefanha@redhat.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).