Re: [PATCH v3 0/6] migration/multifd: Fix channel creation vs. cleanup races
On Wed, Feb 07, 2024 at 10:31:51AM -0300, Fabiano Rosas wrote:
> Peter Xu writes:
>
> > On Tue, Feb 06, 2024 at 06:51:12PM -0300, Fabiano Rosas wrote:
> >> Based-on: 20240202102857.110210-1-pet...@redhat.com
> >> [PATCH v2 00/23] migration/multifd: Refactor ->send_prepare() and cleanups
> >> https://lore.kernel.org/r/20240202102857.110210-1-pet...@redhat.com
> >>
> >> Hi,
> >>
> >> For v3 I fixed the refcounting issue spotted by Avihai. The situation
> >> there is a bit clunky due to historical reasons. The gist is that we
> >> have an assumption that channel creation never fails after p->c has
> >> been set, so when 'p->c == NULL' we have to unref and when 'p->c !=
> >> NULL' the cleanup code will do the unref.
> >
> > Yes, this looks good to me. That's a good catch.
> >
> > I'll leave at least one more day for Avihai and/or Dan to have another
> > look. My r-b persist as of now on patch 5.
> >
> > Actually I think the conditional unref is slightly tricky, but it's not its
> > own fault, IMHO, OTOH it's more about a1af605bd5ad where p->c is slightly
> > abused. My understanding is we can avoid that conditional unref with below
> > patch 1 as a cleanup (on top of this series). Then patch 2 comes all
> > alongside.
>
> Yes, I even wrote some code to always set p->c and leave the unref to
> the cleanup routine. Doing reference counting in the middle of the code
> like that leaves us exposed to new code relying on p->c being valid
> during cleanup. There's already yank and the cleanup itself which expect
> p->c to be valid.
>
> However, I couldn't get past the uglyness of overwriting p->c, so I kept
> the changes minimal for this version.
Yep. The good part of "only set p->c if channel fully established" is that
then the migration thread fully owns the ioc as long as set, and no
overwritting possible.
>
> I'm also wondering whether we can remove the TLS handshake thread
> altogether now that we moved the multifd_send_setup call into the
> migration thread. My (poor) understanding is that a1af605bd5ad hit the
> issue that the QIOTask completion would just execute after
> multifd_save_setup returned. Otherwise I don't see how adding a thread
> to an already async task would have helped. But I need to think about
> that a bit more.
It could be even simpler than that, iiuc. Note that in the stack showed in
that commit message, it hasn't even reached the async handling:
Src: (multifd_send_0)
multifd_channel_connect
multifd_tls_channel_connect
multifd_tls_channel_connect
qio_channel_tls_handshake
qio_channel_tls_handshake_task < async handling provided
here..
qcrypto_tls_session_handshake
gnutls_handshake <-- but we're still at sync
phase..
...
recvmsg (Blocking I/O waiting for response)
IMHO it'll be much easier if we can remove those threads. Please keep the
adventure there, just a heads-up that the async paths seem to have a close
dependency so far on gmainloop contexts, while the migration thread may not
provide that anymore (and I hope we don't introduce anything either; I
think it's better we stick with a full threaded model in migration rather
than event based).
>
> >
> > We don't need to rush on these, though, we should fix the thread race
> >first because multiple of us hit it, and all cleanups can be done
> >later.
>
> I said we should not merge these two changes right now, but I take that
> back. I'll leave it up to you, there doesn't seem to be that much impact
> in adding them.
Thanks. I just sent the pull without them, as I don't yet have plan to
queue them so soon; I'll be accused to abuse the maintainership. :-)
If you think worthwhile, I can still repost them as formal patches later
and put there on the list. If your explore on a bigger hammer works then I
think we can ignore these two patches. But if you or anyone thinks we
could have these as good cleanups, we can also merge them first for 9.0 and
leave whatever else for later.
>
> >
> > =
> > From 0830819d86e08c5175d6669505aa712a0a09717f Mon Sep 17 00:00:00 2001
> > From: Peter Xu
> > Date: Wed, 7 Feb 2024 10:08:35 +0800
> > Subject: [PATCH 1/2] migration/multifd: Cleanup TLS iochannel referencing
> >
> > Commit a1af605bd5 ("migration/multifd: fix hangup with TLS-Multifd due to
> > blocking handshake") introduced a thread for TLS channels, which will
> > resolve the issue on blocking the main thread. However in the same commit
> > p->c is slightly abused just to be able to pass over the pointer "p" into
> > the thread.
> >
> > That's the major reason we'll need to conditionally free the io channel in
> > the fault paths.
> >
> > To clean it up, using a separate structure to pass over both "p" and "tioc"
> > in the tls handshake thread. Then we can make it a rule that p->c will
> > never be set until the channel is completely setup. With that, we can drop
> > the tricky
Re: [PATCH v3 0/6] migration/multifd: Fix channel creation vs. cleanup races
On Tue, Feb 06, 2024 at 06:51:12PM -0300, Fabiano Rosas wrote: > Based-on: 20240202102857.110210-1-pet...@redhat.com > [PATCH v2 00/23] migration/multifd: Refactor ->send_prepare() and cleanups > https://lore.kernel.org/r/20240202102857.110210-1-pet...@redhat.com > > Hi, > > For v3 I fixed the refcounting issue spotted by Avihai. The situation > there is a bit clunky due to historical reasons. The gist is that we > have an assumption that channel creation never fails after p->c has > been set, so when 'p->c == NULL' we have to unref and when 'p->c != > NULL' the cleanup code will do the unref. > > CI run: https://gitlab.com/farosas/qemu/-/pipelines/1166889341 Apologize if I queue this too fast, but i'll disappear tomorrow, so I want to have this thread race fixed soon. I hope that's already complete from angle of all race can happen, but if otherwise we work on top. queued, thanks. -- Peter Xu
Re: [PATCH v3 0/6] migration/multifd: Fix channel creation vs. cleanup races
Peter Xu writes:
> On Tue, Feb 06, 2024 at 06:51:12PM -0300, Fabiano Rosas wrote:
>> Based-on: 20240202102857.110210-1-pet...@redhat.com
>> [PATCH v2 00/23] migration/multifd: Refactor ->send_prepare() and cleanups
>> https://lore.kernel.org/r/20240202102857.110210-1-pet...@redhat.com
>>
>> Hi,
>>
>> For v3 I fixed the refcounting issue spotted by Avihai. The situation
>> there is a bit clunky due to historical reasons. The gist is that we
>> have an assumption that channel creation never fails after p->c has
>> been set, so when 'p->c == NULL' we have to unref and when 'p->c !=
>> NULL' the cleanup code will do the unref.
>
> Yes, this looks good to me. That's a good catch.
>
> I'll leave at least one more day for Avihai and/or Dan to have another
> look. My r-b persist as of now on patch 5.
>
> Actually I think the conditional unref is slightly tricky, but it's not its
> own fault, IMHO, OTOH it's more about a1af605bd5ad where p->c is slightly
> abused. My understanding is we can avoid that conditional unref with below
> patch 1 as a cleanup (on top of this series). Then patch 2 comes all
> alongside.
Yes, I even wrote some code to always set p->c and leave the unref to
the cleanup routine. Doing reference counting in the middle of the code
like that leaves us exposed to new code relying on p->c being valid
during cleanup. There's already yank and the cleanup itself which expect
p->c to be valid.
However, I couldn't get past the uglyness of overwriting p->c, so I kept
the changes minimal for this version.
I'm also wondering whether we can remove the TLS handshake thread
altogether now that we moved the multifd_send_setup call into the
migration thread. My (poor) understanding is that a1af605bd5ad hit the
issue that the QIOTask completion would just execute after
multifd_save_setup returned. Otherwise I don't see how adding a thread
to an already async task would have helped. But I need to think about
that a bit more.
>
> We don't need to rush on these, though, we should fix the thread race
>first because multiple of us hit it, and all cleanups can be done
>later.
I said we should not merge these two changes right now, but I take that
back. I'll leave it up to you, there doesn't seem to be that much impact
in adding them.
>
> =
> From 0830819d86e08c5175d6669505aa712a0a09717f Mon Sep 17 00:00:00 2001
> From: Peter Xu
> Date: Wed, 7 Feb 2024 10:08:35 +0800
> Subject: [PATCH 1/2] migration/multifd: Cleanup TLS iochannel referencing
>
> Commit a1af605bd5 ("migration/multifd: fix hangup with TLS-Multifd due to
> blocking handshake") introduced a thread for TLS channels, which will
> resolve the issue on blocking the main thread. However in the same commit
> p->c is slightly abused just to be able to pass over the pointer "p" into
> the thread.
>
> That's the major reason we'll need to conditionally free the io channel in
> the fault paths.
>
> To clean it up, using a separate structure to pass over both "p" and "tioc"
> in the tls handshake thread. Then we can make it a rule that p->c will
> never be set until the channel is completely setup. With that, we can drop
> the tricky conditional unref of the io channel in the error path.
>
> Signed-off-by: Peter Xu
> ---
> migration/multifd.c | 37 +++--
> 1 file changed, 23 insertions(+), 14 deletions(-)
>
> diff --git a/migration/multifd.c b/migration/multifd.c
> index adfe8c9a0a..4a85a6b7b3 100644
> --- a/migration/multifd.c
> +++ b/migration/multifd.c
> @@ -873,16 +873,22 @@ out:
>
> static void multifd_new_send_channel_async(QIOTask *task, gpointer opaque);
>
> +typedef struct {
> +MultiFDSendParams *p;
> +QIOChannelTLS *tioc;
> +} MultiFDTLSThreadArgs;
> +
> static void *multifd_tls_handshake_thread(void *opaque)
> {
> -MultiFDSendParams *p = opaque;
> -QIOChannelTLS *tioc = QIO_CHANNEL_TLS(p->c);
> +MultiFDTLSThreadArgs *args = opaque;
>
> -qio_channel_tls_handshake(tioc,
> +qio_channel_tls_handshake(args->tioc,
>multifd_new_send_channel_async,
> - p,
> + args->p,
>NULL,
>NULL);
> +g_free(args);
> +
> return NULL;
> }
>
> @@ -892,6 +898,7 @@ static bool multifd_tls_channel_connect(MultiFDSendParams
> *p,
> {
> MigrationState *s = migrate_get_current();
> const char *hostname = s->hostname;
> +MultiFDTLSThreadArgs *args;
> QIOChannelTLS *tioc;
>
> tioc = migration_tls_client_create(ioc, hostname, errp);
> @@ -906,11 +913,14 @@ static bool
> multifd_tls_channel_connect(MultiFDSendParams *p,
> object_unref(OBJECT(ioc));
> trace_multifd_tls_outgoing_handshake_start(ioc, tioc, hostname);
> qio_channel_set_name(QIO_CHANNEL(tioc), "multifd-tls-outgoing");
> -p->c = QIO_CHANNEL(tioc);
This assignment also meant multifd_send_channel_destroy() would call
Re: [PATCH v3 0/6] migration/multifd: Fix channel creation vs. cleanup races
On Tue, Feb 06, 2024 at 06:51:12PM -0300, Fabiano Rosas wrote:
> Based-on: 20240202102857.110210-1-pet...@redhat.com
> [PATCH v2 00/23] migration/multifd: Refactor ->send_prepare() and cleanups
> https://lore.kernel.org/r/20240202102857.110210-1-pet...@redhat.com
>
> Hi,
>
> For v3 I fixed the refcounting issue spotted by Avihai. The situation
> there is a bit clunky due to historical reasons. The gist is that we
> have an assumption that channel creation never fails after p->c has
> been set, so when 'p->c == NULL' we have to unref and when 'p->c !=
> NULL' the cleanup code will do the unref.
Yes, this looks good to me. That's a good catch.
I'll leave at least one more day for Avihai and/or Dan to have another
look. My r-b persist as of now on patch 5.
Actually I think the conditional unref is slightly tricky, but it's not its
own fault, IMHO, OTOH it's more about a1af605bd5ad where p->c is slightly
abused. My understanding is we can avoid that conditional unref with below
patch 1 as a cleanup (on top of this series). Then patch 2 comes all
alongside.
We don't need to rush on these, though, we should fix the thread race first
because multiple of us hit it, and all cleanups can be done later.
=
>From 0830819d86e08c5175d6669505aa712a0a09717f Mon Sep 17 00:00:00 2001
From: Peter Xu
Date: Wed, 7 Feb 2024 10:08:35 +0800
Subject: [PATCH 1/2] migration/multifd: Cleanup TLS iochannel referencing
Commit a1af605bd5 ("migration/multifd: fix hangup with TLS-Multifd due to
blocking handshake") introduced a thread for TLS channels, which will
resolve the issue on blocking the main thread. However in the same commit
p->c is slightly abused just to be able to pass over the pointer "p" into
the thread.
That's the major reason we'll need to conditionally free the io channel in
the fault paths.
To clean it up, using a separate structure to pass over both "p" and "tioc"
in the tls handshake thread. Then we can make it a rule that p->c will
never be set until the channel is completely setup. With that, we can drop
the tricky conditional unref of the io channel in the error path.
Signed-off-by: Peter Xu
---
migration/multifd.c | 37 +++--
1 file changed, 23 insertions(+), 14 deletions(-)
diff --git a/migration/multifd.c b/migration/multifd.c
index adfe8c9a0a..4a85a6b7b3 100644
--- a/migration/multifd.c
+++ b/migration/multifd.c
@@ -873,16 +873,22 @@ out:
static void multifd_new_send_channel_async(QIOTask *task, gpointer opaque);
+typedef struct {
+MultiFDSendParams *p;
+QIOChannelTLS *tioc;
+} MultiFDTLSThreadArgs;
+
static void *multifd_tls_handshake_thread(void *opaque)
{
-MultiFDSendParams *p = opaque;
-QIOChannelTLS *tioc = QIO_CHANNEL_TLS(p->c);
+MultiFDTLSThreadArgs *args = opaque;
-qio_channel_tls_handshake(tioc,
+qio_channel_tls_handshake(args->tioc,
multifd_new_send_channel_async,
- p,
+ args->p,
NULL,
NULL);
+g_free(args);
+
return NULL;
}
@@ -892,6 +898,7 @@ static bool multifd_tls_channel_connect(MultiFDSendParams
*p,
{
MigrationState *s = migrate_get_current();
const char *hostname = s->hostname;
+MultiFDTLSThreadArgs *args;
QIOChannelTLS *tioc;
tioc = migration_tls_client_create(ioc, hostname, errp);
@@ -906,11 +913,14 @@ static bool multifd_tls_channel_connect(MultiFDSendParams
*p,
object_unref(OBJECT(ioc));
trace_multifd_tls_outgoing_handshake_start(ioc, tioc, hostname);
qio_channel_set_name(QIO_CHANNEL(tioc), "multifd-tls-outgoing");
-p->c = QIO_CHANNEL(tioc);
+
+args = g_new0(MultiFDTLSThreadArgs, 1);
+args->tioc = tioc;
+args->p = p;
p->tls_thread_created = true;
qemu_thread_create(>tls_thread, "multifd-tls-handshake-worker",
- multifd_tls_handshake_thread, p,
+ multifd_tls_handshake_thread, args,
QEMU_THREAD_JOINABLE);
return true;
}
@@ -923,6 +933,7 @@ static bool multifd_channel_connect(MultiFDSendParams *p,
migration_ioc_register_yank(ioc);
p->registered_yank = true;
+/* Setup p->c only if the channel is completely setup */
p->c = ioc;
p->thread_created = true;
@@ -976,14 +987,12 @@ out:
trace_multifd_new_send_channel_async_error(p->id, local_err);
multifd_send_set_error(local_err);
-if (!p->c) {
-/*
- * If no channel has been created, drop the initial
- * reference. Otherwise cleanup happens at
- * multifd_send_channel_destroy()
- */
-object_unref(OBJECT(ioc));
-}
+/*
+ * For error cases (TLS or non-TLS), IO channel is always freed here
+ * rather than when cleanup multifd: since p->c is not set, multifd
+ * cleanup code doesn't even know its existance.
+ */
+object_unref(OBJECT(ioc));
[PATCH v3 0/6] migration/multifd: Fix channel creation vs. cleanup races
Based-on: 20240202102857.110210-1-pet...@redhat.com [PATCH v2 00/23] migration/multifd: Refactor ->send_prepare() and cleanups https://lore.kernel.org/r/20240202102857.110210-1-pet...@redhat.com Hi, For v3 I fixed the refcounting issue spotted by Avihai. The situation there is a bit clunky due to historical reasons. The gist is that we have an assumption that channel creation never fails after p->c has been set, so when 'p->c == NULL' we have to unref and when 'p->c != NULL' the cleanup code will do the unref. CI run: https://gitlab.com/farosas/qemu/-/pipelines/1166889341 v2: https://lore.kernel.org/r/20240205194929.28963-1-faro...@suse.de In this v2 I made sure NO channel is created after the semaphores are posted. Feel free to call me out if that's not the case. Not much changes, except that now both TLS and non-TLS go through the same code, so there's a centralized place to do error handling and releasing the semaphore. CI run: https://gitlab.com/farosas/qemu/-/pipelines/1165206107 based on Peter's code: https://gitlab.com/farosas/qemu/-/pipelines/1165303276 v1: https://lore.kernel.org/r/20240202191128.1901-1-faro...@suse.de This contains 2 patches from my previous series addressing the p->running misuse and the TLS thread leak and 3 new patches to fix the cleanup-while-creating-threads race. For the p->running I'm keeping the idea from the other series to remove p->running and use a more narrow p->thread_created flag. This flag is used only inform whether the thread has been created so we can join it. For the cleanup race I have moved some code around and added a semaphore to make multifd_save_setup() only return once all channel creation tasks have started. The idea is that after multifd_save_setup() returns, no new creations are in flight and the p->thread_created flags will never change again, so they're enough to cause the cleanup code to wait for the threads to join. CI run: https://gitlab.com/farosas/qemu/-/pipelines/1162798843 @Peter: I can rebase this on top of your series once we decide about it. Fabiano Rosas (6): migration/multifd: Join the TLS thread migration/multifd: Remove p->running migration/multifd: Move multifd_send_setup error handling in to the function migration/multifd: Move multifd_send_setup into migration thread migration/multifd: Unify multifd and TLS connection paths migration/multifd: Add a synchronization point for channel creation migration/migration.c | 14 ++-- migration/multifd.c | 168 +- migration/multifd.h | 11 ++- 3 files changed, 109 insertions(+), 84 deletions(-) -- 2.35.3
