On Wed, Feb 23, 2022 at 09:52:08AM +0000, Dr. David Alan Gilbert wrote:
> * Peter Xu (pet...@redhat.com) wrote:
> > On Tue, Feb 22, 2022 at 11:32:10AM +0000, Dr. David Alan Gilbert wrote:
> > > * Peter Xu (pet...@redhat.com) wrote:
> > > > To allow postcopy recovery, the ram fast load (preempt-only) dest QEMU
> > > > thread
> > > > needs similar handling on fault tolerance. When ram_load_postcopy()
> > > > fails,
> > > > instead of stopping the thread it halts with a semaphore, preparing to
> > > > be
> > > > kicked again when recovery is detected.
> > > >
> > > > A mutex is introduced to make sure there's no concurrent operation upon
> > > > the
> > > > socket. To make it simple, the fast ram load thread will take the
> > > > mutex during
> > > > its whole procedure, and only release it if it's paused. The fast-path
> > > > socket
> > > > will be properly released by the main loading thread safely when there's
> > > > network failures during postcopy with that mutex held.
> > >
> > > I *think* this is mostly OK; but I worry I don't understand all the
> > > cases; e.g.
> > > a) If the postcopy channel errors first
> > > b) If the main channel errors first
> >
> > Ah right, I don't think I handled all the cases. Sorry.
> >
> > We always check the main channel, but if the postcopy channel got faulted,
> > we may not fall into paused mode as expected.
> >
> > I'll fix that up.
>
> Thanks.
>
> > >
> > > Can you add some docs to walk through those and explain the locking ?
> >
> > Sure.
> >
> > The sem is mentioned in the last sentence of paragraph 1, where it's purely
> > used for a way to yield the fast ram load thread so that when something
> > wrong happens it can sleep on that semaphore. Then when we recover we'll
> > post to the semaphore to kick it up. We used it like that in many places,
> > e.g. postcopy_pause_sem_dst to yield the main load thread.
> >
> > The 2nd paragraph above was for explaining why we need the mutex; it's
> > basically the same as rp_mutex protecting to_src_file, so that we won't
> > accidentally close() the qemufile during some other thread using it. So
> > the fast ram load thread needs to take that new mutex for mostly the whole
> > lifecycle of itself (because it's loading from that qemufile), meanwhile
> > only drop the mutex when it prepares to sleep. Then the main load thread
> > can recycle the postcopy channel using qemu_fclose() safely.
>
> Yes, that feels like it needs to go in the code somewhere.
Sure, I'll further squash below comment update into the same patch. I
reworded some places but mostly it should be telling the same thing:
---8<---
diff --git a/migration/migration.h b/migration/migration.h
index 945088064a..91f845e9e4 100644
--- a/migration/migration.h
+++ b/migration/migration.h
@@ -118,7 +118,17 @@ struct MigrationIncomingState {
/* Postcopy priority thread is used to receive postcopy requested pages */
QemuThread postcopy_prio_thread;
bool postcopy_prio_thread_created;
- /* Used to sync with the prio thread */
+ /*
+ * Used to sync between the ram load main thread and the fast ram load
+ * thread. It protects postcopy_qemufile_dst, which is the postcopy
+ * fast channel.
+ *
+ * The ram fast load thread will take it mostly for the whole lifecycle
+ * because it needs to continuously read data from the channel, and
+ * it'll only release this mutex if postcopy is interrupted, so that
+ * the ram load main thread will take this mutex over and properly
+ * release the broken channel.
+ */
QemuMutex postcopy_prio_thread_mutex;
/*
* An array of temp host huge pages to be used, one for each postcopy
@@ -149,6 +159,12 @@ struct MigrationIncomingState {
/* notify PAUSED postcopy incoming migrations to try to continue */
QemuSemaphore postcopy_pause_sem_dst;
QemuSemaphore postcopy_pause_sem_fault;
+ /*
+ * This semaphore is used to allow the ram fast load thread (only when
+ * postcopy preempt is enabled) fall into sleep when there's network
+ * interruption detected. When the recovery is done, the main load
+ * thread will kick the fast ram load thread using this semaphore.
+ */
QemuSemaphore postcopy_pause_sem_fast_load;
/* List of listening socket addresses */
---8<---
>
> > [...]
> >
> > > > @@ -3466,6 +3468,17 @@ static MigThrError postcopy_pause(MigrationState
> > > > *s)
> > > > qemu_file_shutdown(file);
> > > > qemu_fclose(file);
> > > >
> > > > + /*
> > > > + * Do the same to postcopy fast path socket too if there is.
> > > > No
> > > > + * locking needed because no racer as long as we do this
> > > > before setting
> > > > + * status to paused.
> > > > + */
> > > > + if (s->postcopy_qemufile_src) {
> > > > +
> > > > migration_ioc_unregister_yank_from_file(s->postcopy_qemufile_src);
> > >
> > > Shouldn't this do a qemu_file_shutdown on here first?
> >
> > Yes I probably should.
> >
> > With all above, I plan to squash below changes into this patch:
> >
> > ---8<---
> > diff --git a/migration/migration.c b/migration/migration.c
> > index c68a281406..69778cab23 100644
> > --- a/migration/migration.c
> > +++ b/migration/migration.c
> > @@ -3475,6 +3475,7 @@ static MigThrError postcopy_pause(MigrationState *s)
> > */
> > if (s->postcopy_qemufile_src) {
> >
> > migration_ioc_unregister_yank_from_file(s->postcopy_qemufile_src);
> > + qemu_file_shutdown(s->postcopy_qemufile_src);
> > qemu_fclose(s->postcopy_qemufile_src);
> > s->postcopy_qemufile_src = NULL;
> > }
> > @@ -3534,8 +3535,13 @@ static MigThrError
> > migration_detect_error(MigrationState *s)
> > return MIG_THR_ERR_FATAL;
> > }
> >
> > - /* Try to detect any file errors */
> > - ret = qemu_file_get_error_obj(s->to_dst_file, &local_error);
> > + /*
> > + * Try to detect any file errors. Note that postcopy_qemufile_src will
> > + * be NULL when postcopy preempt is not enabled.
> > + */
> > + ret = qemu_file_get_error_obj_any(s->to_dst_file,
> > + s->postcopy_qemufile_src,
> > + &local_error);
> > if (!ret) {
> > /* Everything is fine */
> > assert(!local_error);
> > diff --git a/migration/qemu-file.c b/migration/qemu-file.c
> > index 1479cddad9..397652f0ba 100644
> > --- a/migration/qemu-file.c
> > +++ b/migration/qemu-file.c
> > @@ -139,6 +139,33 @@ int qemu_file_get_error_obj(QEMUFile *f, Error **errp)
> > return f->last_error;
> > }
> >
> > +/*
> > + * Get last error for either stream f1 or f2 with optional Error*.
> > + * The error returned (non-zero) can be either from f1 or f2.
> > + *
> > + * If any of the qemufile* is NULL, then skip the check on that file.
> > + *
> > + * When there is no error on both qemufile, zero is returned.
> > + */
> > +int qemu_file_get_error_obj_any(QEMUFile *f1, QEMUFile *f2, Error **errp)
> > +{
> > + int ret = 0;
> > +
> > + if (f1) {
> > + ret = qemu_file_get_error_obj(f1, errp);
> > + /* If there's already error detected, return */
> > + if (ret) {
> > + return ret;
> > + }
> > + }
> > +
> > + if (f2) {
> > + ret = qemu_file_get_error_obj(f2, errp);
> > + }
> > +
> > + return ret;
> > +}
> > +
> > /*
> > * Set the last error for stream f with optional Error*
> > */
> > diff --git a/migration/qemu-file.h b/migration/qemu-file.h
> > index 3f36d4dc8c..2564e5e1c7 100644
> > --- a/migration/qemu-file.h
> > +++ b/migration/qemu-file.h
> > @@ -156,6 +156,7 @@ void qemu_file_update_transfer(QEMUFile *f, int64_t
> > len);
> > void qemu_file_set_rate_limit(QEMUFile *f, int64_t new_rate);
> > int64_t qemu_file_get_rate_limit(QEMUFile *f);
> > int qemu_file_get_error_obj(QEMUFile *f, Error **errp);
> > +int qemu_file_get_error_obj_any(QEMUFile *f1, QEMUFile *f2, Error **errp);
> > void qemu_file_set_error_obj(QEMUFile *f, int ret, Error *err);
> > void qemu_file_set_error(QEMUFile *f, int ret);
> > int qemu_file_shutdown(QEMUFile *f);
> > diff --git a/migration/savevm.c b/migration/savevm.c
> > index 2d32340d28..24b69a1008 100644
> > --- a/migration/savevm.c
> > +++ b/migration/savevm.c
> > @@ -2651,8 +2651,8 @@ retry:
> > while (true) {
> > section_type = qemu_get_byte(f);
> >
> > - if (qemu_file_get_error(f)) {
> > - ret = qemu_file_get_error(f);
> > + ret = qemu_file_get_error_obj_any(f, mis->postcopy_qemufile_dst,
> > NULL);
> > + if (ret) {
> > break;
> > }
> > ---8<---
> >
> > Does it look sane? Let me know if there's still things missing.
>
> Reviewed-by: Dr. David Alan Gilbert <dgilb...@redhat.com>
Thanks!
--
Peter Xu
