diff mbox series

[v4,3/4] migration: zero-copy flush only at the end of bitmap scanning

Message ID 20220620053944.257547-4-leobras@redhat.com (mailing list archive)
State New, archived
Headers show
Series MSG_ZEROCOPY fixes & improvements | expand

Commit Message

Leonardo Bras June 20, 2022, 5:39 a.m. UTC
When sending memory pages with MSG_ZEROCOPY, it's necessary to flush
to make sure all dirty pages are sent before a future version of them
happens to be sent.

Currently, the flush happens every time at the end of ram_save_iterate(),
which usually happens around 20x per second, due to a timeout.

Change so it flushes only after a whole scanning of the dirty bimap,
so it never sends a newer version of a page before an older one, while
avoiding unnecessary overhead.

Signed-off-by: Leonardo Bras <leobras@redhat.com>
---
 migration/multifd.h |  1 +
 migration/multifd.c | 58 ++++++++++++++++++++++++++-------------------
 migration/ram.c     |  7 ++++++
 3 files changed, 41 insertions(+), 25 deletions(-)

Comments

Juan Quintela June 20, 2022, 9:23 a.m. UTC | #1
Leonardo Bras <leobras@redhat.com> wrote:
> When sending memory pages with MSG_ZEROCOPY, it's necessary to flush
> to make sure all dirty pages are sent before a future version of them
> happens to be sent.
>
> Currently, the flush happens every time at the end of ram_save_iterate(),
> which usually happens around 20x per second, due to a timeout.
>
> Change so it flushes only after a whole scanning of the dirty bimap,
> so it never sends a newer version of a page before an older one, while
> avoiding unnecessary overhead.
>
> Signed-off-by: Leonardo Bras <leobras@redhat.com>

I agree that previous one is too many flushes, but this one changes to too
much memory to be uncommitted, and that is important because otherwise we
have unlimited amount of dirty memory.

> +/*
> + * Set zero_copy_flush = true for every multifd channel
> + *
> + * When using zero-copy, it's necessary to flush the pages before any of
> + * the pages can be sent again, so we'll make sure the new version of the
> + * pages will always arrive _later_ than the old pages.
> + *
> + * Should be called only after we finished one whole scanning of
> + * all the dirty bitmaps.
> + */
> +int multifd_zero_copy_flush(void)
> +{
> +    int i;
> +    Error *local_err = NULL;
> +
> +    if (!migrate_use_multifd()) {
> +        return 0;
> +    }
> +
> +    for (i = 0; i < migrate_multifd_channels(); i++) {
> +        MultiFDSendParams *p = &multifd_send_state->params[i];
> +        int ret;
> +
> +        ret = qio_channel_flush(p->c, &local_err);
> +        if (ret < 0) {
> +            error_report_err(local_err);
> +            return ret;
> +        }
> +    }
> +
> +    return 0;
> +}


Here you flush every channel, Only at the end of a range you want to do this.


>  int multifd_send_sync_main(QEMUFile *f)
>  {
>      int i;
> -    bool flush_zero_copy;
>  
>      if (!migrate_use_multifd()) {
>          return 0;
> @@ -581,19 +613,6 @@ int multifd_send_sync_main(QEMUFile *f)
>          }
>      }
>  
> -    /*
> -     * When using zero-copy, it's necessary to flush the pages before any of
> -     * the pages can be sent again, so we'll make sure the new version of the
> -     * pages will always arrive _later_ than the old pages.
> -     *
> -     * Currently we achieve this by flushing the zero-page requested writes
> -     * per ram iteration, but in the future we could potentially optimize it
> -     * to be less frequent, e.g. only after we finished one whole scanning of
> -     * all the dirty bitmaps.
> -     */
> -
> -    flush_zero_copy = migrate_use_zero_copy_send();
> -
>      for (i = 0; i < migrate_multifd_channels(); i++) {
>          MultiFDSendParams *p = &multifd_send_state->params[i];
>  
> @@ -615,17 +634,6 @@ int multifd_send_sync_main(QEMUFile *f)
>          ram_counters.transferred += p->packet_len;
>          qemu_mutex_unlock(&p->mutex);
>          qemu_sem_post(&p->sem);
> -
> -        if (flush_zero_copy && p->c) {
> -            int ret;
> -            Error *err = NULL;
> -
> -            ret = qio_channel_flush(p->c, &err);
> -            if (ret < 0) {
> -                error_report_err(err);
> -                return -1;
> -            }
> -        }

This synchronization already happens once every iteration through all
ram.

</me checks how>

And low and behold, it doesn't.

The problem here is that we are calling multifd_send_sync_main() in the
wrong place, i.e. we are being too conservative.

We need to call multifd_send_sync_main() just before doing
migration_bitmap_sync().  The reason that we need to call that function
is exactly the same that we need to flush for zero_copy.

So, what we need to change is remove the call to
multifd_send_sync_main(), not how it handles zero_copy.

>      }
>      for (i = 0; i < migrate_multifd_channels(); i++) {
>          MultiFDSendParams *p = &multifd_send_state->params[i];
> diff --git a/migration/ram.c b/migration/ram.c
> index 5f5e37f64d..514584e44f 100644
> --- a/migration/ram.c
> +++ b/migration/ram.c
> @@ -2288,6 +2288,13 @@ static int ram_find_and_save_block(RAMState *rs)
>      rs->last_seen_block = pss.block;
>      rs->last_page = pss.page;
>  
> +    if (pss.complete_round && migrate_use_zero_copy_send()) {
> +        int ret = multifd_zero_copy_flush();
> +        if (ret < 0) {
> +            return ret;
> +        }
> +    }
> +

This place is not right either, because we can have a sync in the middle
for other reasons.

We call migration_bitmap_sync() in save_live_pending(), and that is not
when we finish the complete_round.

Once discussed this, what I asked in the past is that you are having too
much dirty memory on zero_copy.  When you have a Multiterabyte guest, in
a single round you have a "potentially" dirty memory on each channel of:

   total_amount_memory / number of channels.

In a Multiterabyte guest, this is going to be more that probably in the
dozens of gigabytes.  As far as I know there is no card/driver that will
benefit for so many pages in zero_copy, and kernel will move to
synchronous copy at some point.  (In older threads, daniel showed how to
test for this case).

What I proposed is that you check in the migration_send_thread() how
much memory you have written since last synchronization.  Once that it
is big enough (I don't know the limits for card, in the docs that you
showed suggested the size is a few megabytes), you just sync at that
point and continue.

You still need to synchronize all threads at bitmap sync, but as said,
that is handled by multifd_send_sync_main(), and we should fix that
instead of changing where zero_copy flushes.

/* Removed not relevant bits of the function */

static void *multifd_send_thread(void *opaque)
{
    size_t zero_copy_bytes_sent = 0;

    ...

    while (true) {

            ....

            trace_multifd_send(p->id, packet_num, p->normal_num, p->zero_num,
                               p->flags, p->next_packet_size);

            if (use_zero_copy_send) {
                /* Send header first, without zerocopy */
                ret = qio_channel_write_all(p->c, (void *)p->packet,
                                            p->packet_len, &local_err);
                if (ret != 0) {
                    break;
                }

****** Note aside

Did you answered my question here of what happens when you do:

write_with_zero_copy(1MB); write_without_zero_copy(4KB); write_with_zero_copy(1MB); write_without_zero_copy(4KB);

My guess is that write_without_zero_copy(4KB) will just flush the
socket.  I can't think how it can work without doing that.
If so, we are not getting what we want.

            } else {
                /* Send header using the same writev call */
                p->iov[0].iov_len = p->packet_len;
                p->iov[0].iov_base = p->packet;
            }

            ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
                                              0, p->write_flags, &local_err);
            if (ret != 0) {
                break;
            }

            qemu_mutex_lock(&p->mutex);
            p->num_packets++;
            p->total_normal_pages += p->normal_num;
            p->total_zero_pages += p->zero_num;
            p->pages->num = 0;
            p->pages->block = NULL;
            p->sent_bytes += p->packet_len;;
            p->sent_bytes += p->next_packet_size;

**** Addition
            zero_copy_bytes_sent += p->packet_len + p->next_packet_size;

            p->pending_job--;
            qemu_mutex_unlock(&p->mutex);
***** Addition
            if (zero_copy_bytes_sent > Threshold) // 2MB/4MB?  I don't know
                    ret = qio_channel_flush(p->c, &local_err);
                    // Handle error somehow
                    //  If you want to be a pro, just change the
                    // Threshold depending on what the kernel answers.
                    // If it has to revert to synchronous sent, just
                    // decrease the threshold, otherwise increase it.

            if (p->flags & MULTIFD_FLAG_SYNC) {
                qemu_sem_post(&p->sem_sync);
            }
            qemu_sem_post(&multifd_send_state->channels_ready);
        } else if (p->quit) {
            qemu_mutex_unlock(&p->mutex);
            break;
        } else {
            qemu_mutex_unlock(&p->mutex);
            /* sometimes there are spurious wakeups */
        }
    }
    .............
}

What do you think?

Later, Juan.
Peter Xu June 20, 2022, 3:44 p.m. UTC | #2
On Mon, Jun 20, 2022 at 11:23:53AM +0200, Juan Quintela wrote:
> Once discussed this, what I asked in the past is that you are having too
> much dirty memory on zero_copy.  When you have a Multiterabyte guest, in
> a single round you have a "potentially" dirty memory on each channel of:
> 
>    total_amount_memory / number of channels.
> 
> In a Multiterabyte guest, this is going to be more that probably in the
> dozens of gigabytes.  As far as I know there is no card/driver that will
> benefit for so many pages in zero_copy, and kernel will move to
> synchronous copy at some point.  (In older threads, daniel showed how to
> test for this case).

I was wondering whether the kernel needs to cache a lot of messages for
zero copy if we don't flush it for a long time, as recvmsg(MSG_ERRQUEUE)
seems to be fetching one message from the kernel one at a time.  And,
whether that queue has a limit in length or something.

Does it mean that when the kernel could have cached enough of these
messages then it'll fallback to the no-zero-copy mode?  And probably that's
the way how kernel protects itself from using too much buffer for the error
msgs?

This reminded me - Leo, have you considered adding the patch altogether to
detect the "fallback to non-zero-copy" condition?  Because when with it and
when the fallback happens at some point (e.g. when the guest memory is
larger than some value) we'll know.

Thanks,
Leonardo Bras June 21, 2022, 3:26 a.m. UTC | #3
On Mon, 2022-06-20 at 11:23 +0200, Juan Quintela wrote:
> Leonardo Bras <leobras@redhat.com> wrote:
> > When sending memory pages with MSG_ZEROCOPY, it's necessary to flush
> > to make sure all dirty pages are sent before a future version of them
> > happens to be sent.
> > 
> > Currently, the flush happens every time at the end of ram_save_iterate(),
> > which usually happens around 20x per second, due to a timeout.
> > 
> > Change so it flushes only after a whole scanning of the dirty bimap,
> > so it never sends a newer version of a page before an older one, while
> > avoiding unnecessary overhead.
> > 
> > Signed-off-by: Leonardo Bras <leobras@redhat.com>
> 
> I agree that previous one is too many flushes, but this one changes to too
> much memory to be uncommitted, and that is important because otherwise we
> have unlimited amount of dirty memory.

I don't quite understand what you meant here.
What does it mean to be uncommitted at this context?
I don't see how we get unlimited amounts of dirty memory here. 

Zero-copy is not supposed to copy memory pages, so let's say all pages are dirty
and enqueued to send, but our network interface is stalling the send:
All memory we should have allocated is VM ram pagecount x sizeof(iov), because
then we flush.

Unless... are you referring to locked memory (as dirty enqueued memory -> locked
memory)? 
That would be a point: If we enqueue more memory than the locked memory amount
our user support, the migration will fail. 

But that would mean a very fast CPU (lots of sendmsgs) and a very slow network
interface. 

> 
> > +/*
> > + * Set zero_copy_flush = true for every multifd channel
> > + *
> > + * When using zero-copy, it's necessary to flush the pages before any of
> > + * the pages can be sent again, so we'll make sure the new version of the
> > + * pages will always arrive _later_ than the old pages.
> > + *
> > + * Should be called only after we finished one whole scanning of
> > + * all the dirty bitmaps.
> > + */
> > +int multifd_zero_copy_flush(void)
> > +{
> > +    int i;
> > +    Error *local_err = NULL;
> > +
> > +    if (!migrate_use_multifd()) {
> > +        return 0;
> > +    }
> > +
> > +    for (i = 0; i < migrate_multifd_channels(); i++) {
> > +        MultiFDSendParams *p = &multifd_send_state->params[i];
> > +        int ret;
> > +
> > +        ret = qio_channel_flush(p->c, &local_err);
> > +        if (ret < 0) {
> > +            error_report_err(local_err);
> > +            return ret;
> > +        }
> > +    }
> > +
> > +    return 0;
> > +}
> 
> 
> Here you flush every channel, Only at the end of a range you want to do this.

Yes, the idea is to flush after a full scan of the dirty-bitmap.

> 
> 
> >  int multifd_send_sync_main(QEMUFile *f)
> >  {
> >      int i;
> > -    bool flush_zero_copy;
> >  
> >      if (!migrate_use_multifd()) {
> >          return 0;
> > @@ -581,19 +613,6 @@ int multifd_send_sync_main(QEMUFile *f)
> >          }
> >      }
> >  
> > -    /*
> > -     * When using zero-copy, it's necessary to flush the pages before any
> > of
> > -     * the pages can be sent again, so we'll make sure the new version of
> > the
> > -     * pages will always arrive _later_ than the old pages.
> > -     *
> > -     * Currently we achieve this by flushing the zero-page requested writes
> > -     * per ram iteration, but in the future we could potentially optimize
> > it
> > -     * to be less frequent, e.g. only after we finished one whole scanning
> > of
> > -     * all the dirty bitmaps.
> > -     */
> > -
> > -    flush_zero_copy = migrate_use_zero_copy_send();
> > -
> >      for (i = 0; i < migrate_multifd_channels(); i++) {
> >          MultiFDSendParams *p = &multifd_send_state->params[i];
> >  
> > @@ -615,17 +634,6 @@ int multifd_send_sync_main(QEMUFile *f)
> >          ram_counters.transferred += p->packet_len;
> >          qemu_mutex_unlock(&p->mutex);
> >          qemu_sem_post(&p->sem);
> > -
> > -        if (flush_zero_copy && p->c) {
> > -            int ret;
> > -            Error *err = NULL;
> > -
> > -            ret = qio_channel_flush(p->c, &err);
> > -            if (ret < 0) {
> > -                error_report_err(err);
> > -                return -1;
> > -            }
> > -        }
> 
> This synchronization already happens once every iteration through all
> ram.
> 
> </me checks how>
> 
> And low and behold, it doesn't.
> 
> The problem here is that we are calling multifd_send_sync_main() in the
> wrong place, i.e. we are being too conservative.
> 
> We need to call multifd_send_sync_main() just before doing
> migration_bitmap_sync().  The reason that we need to call that function
> is exactly the same that we need to flush for zero_copy.
> 
> So, what we need to change is remove the call to
> multifd_send_sync_main(), not how it handles zero_copy.

So, IIUC, multifd have been syncing in a conservative way (i.e. more than
actually needed), and we need to do the same improvement (sync less) for multifd
too, instead of just changing it for zero-copy. Is that correct?

> 
> >      }
> >      for (i = 0; i < migrate_multifd_channels(); i++) {
> >          MultiFDSendParams *p = &multifd_send_state->params[i];
> > diff --git a/migration/ram.c b/migration/ram.c
> > index 5f5e37f64d..514584e44f 100644
> > --- a/migration/ram.c
> > +++ b/migration/ram.c
> > @@ -2288,6 +2288,13 @@ static int ram_find_and_save_block(RAMState *rs)
> >      rs->last_seen_block = pss.block;
> >      rs->last_page = pss.page;
> >  
> > +    if (pss.complete_round && migrate_use_zero_copy_send()) {
> > +        int ret = multifd_zero_copy_flush();
> > +        if (ret < 0) {
> > +            return ret;
> > +        }
> > +    }
> > +
> 
> This place is not right either, because we can have a sync in the middle
> for other reasons.

What do you mean by "in the middle" here? 
Does it mean we should not have multifd code in this function?

> 
> We call migration_bitmap_sync() in save_live_pending(), and that is not
> when we finish the complete_round.

Agree. 
We could add/set a flag in multifd, say in the above area, and then sync/flush
in the correct place (in the future).

But at my experience debugging my code, I found that the loop at
ram_save_iterate() will not stop/break at "bitmap scan end", so we may end up
sending the same page twice before a flush, which is what we want to avoid.

> 
> Once discussed this, what I asked in the past is that you are having too
> much dirty memory on zero_copy.  When you have a Multiterabyte guest, in
> a single round you have a "potentially" dirty memory on each channel of:
> 
>    total_amount_memory / number of channels.

If dirty memory -> locked memory, yes, max_locked_memory == guest_ram.

> 
> In a Multiterabyte guest, this is going to be more that probably in the
> dozens of gigabytes.  As far as I know there is no card/driver that will
> benefit for so many pages in zero_copy, and kernel will move to
> synchronous copy at some point.  (In older threads, daniel showed how to
> test for this case).

The idea of MSG_ZEROCOPY is to avoid copying the pages content over to the
kernel space. I don't recall any reason a network card would drop to copying
instead of keep using zero-copy, unless it does not support it (scatter-gather
not supported, for example), but in this case it would never use zero-copy from
the start. 

IIRC, as long as we have enough locked memory, there should be no problem for
the driver to keep using zero-copy send.

Well, unless you mean the gain of using zero-copy should be irrelevant compared
to the cost of locked memory in some scenarios, say very powerful CPU with slow
network interface, where using zero-copy should not interfere with migration
speed. But that is not a scenario for zero-copy anyway

> 
> What I proposed is that you check in the migration_send_thread() how
> much memory you have written since last synchronization.  Once that it
> is big enough (I don't know the limits for card, in the docs that you
> showed suggested the size is a few megabytes), you just sync at that
> point and continue.

This would help us keep the locked memory under control? yes.
But it would come with the cost of flushing *much* more often than needed.

I mean, it would only be useful if the network card is completely stalling the
send, or if the cpu is scheduling more zero-copy sends than the network card can
process.

The point here is: we don't have to keep track of how much has been sent, but
instead of how much is enqueued (i.e. locked memory). Maybe there is cheap way
to track this value, and we could flush when we detect it's too high.

> 
> You still need to synchronize all threads at bitmap sync, but as said,
> that is handled by multifd_send_sync_main(), and we should fix that
> instead of changing where zero_copy flushes.

Ok, so we keep the zero-copy-flush inside the multifd_send_sync_main() and
change where it's called in order to match the end of the dirty-bitmap scan.
Is that correct?

> 
> /* Removed not relevant bits of the function */
> 
> static void *multifd_send_thread(void *opaque)
> {
>     size_t zero_copy_bytes_sent = 0;
> 
>     ...
> 
>     while (true) {
> 
>             ....
> 
>             trace_multifd_send(p->id, packet_num, p->normal_num, p->zero_num,
>                                p->flags, p->next_packet_size);
> 
>             if (use_zero_copy_send) {
>                 /* Send header first, without zerocopy */
>                 ret = qio_channel_write_all(p->c, (void *)p->packet,
>                                             p->packet_len, &local_err);
>                 if (ret != 0) {
>                     break;
>                 }
> 
> ****** Note aside
> 
> Did you answered my question here of what happens when you do:
> 
> write_with_zero_copy(1MB); write_without_zero_copy(4KB);
> write_with_zero_copy(1MB); write_without_zero_copy(4KB);
> 
> My guess is that write_without_zero_copy(4KB) will just flush the
> socket.  I can't think how it can work without doing that.

To be honest, I did not follow the MSG_ZEROCOPY code to this extent. 

But it depends. If write_without_zero_copy() has MSG_DONTWAIT we could have
everything being sent in a non-blocking way.

But, of course, taking it into account that in multifd we are using a blocking
sendmsg(), I also agree that it is probably flushing the packets previously sent
with MSG_ZEROCOPY (or we could have issues in the receiving part).

But that's an assumption. I still need to dig the code in that aspect to be
sure. I will work on that asap.

> If so, we are not getting what we want.

In fact yes, we are.
We are avoiding copying the buffer, which is what we intend to do, and avoiding
the cost of the 'flush' syscall happening too often. 

If we think multifd is supposed to have blocking sendmsg(), then we are doing
exactly that here, with zero-copy for the heavy part, which is ideal IMHO.

We could try adding MSG_DONTWAIT to the sendmsg() in multifd (which would take
the blocking / synchronous send away), but I have no idea on the impact it would
have at multifd inner workings.

> 
>             } else {
>                 /* Send header using the same writev call */
>                 p->iov[0].iov_len = p->packet_len;
>                 p->iov[0].iov_base = p->packet;
>             }
> 
>             ret = qio_channel_writev_full_all(p->c, p->iov, p->iovs_num, NULL,
>                                               0, p->write_flags, &local_err);
>             if (ret != 0) {
>                 break;
>             }
> 
>             qemu_mutex_lock(&p->mutex);
>             p->num_packets++;
>             p->total_normal_pages += p->normal_num;
>             p->total_zero_pages += p->zero_num;
>             p->pages->num = 0;
>             p->pages->block = NULL;
>             p->sent_bytes += p->packet_len;;
>             p->sent_bytes += p->next_packet_size;
> 
> **** Addition
>             zero_copy_bytes_sent += p->packet_len + p->next_packet_size;
> 
>             p->pending_job--;
>             qemu_mutex_unlock(&p->mutex);
> ***** Addition
>             if (zero_copy_bytes_sent > Threshold) // 2MB/4MB?  I don't know
>                     ret = qio_channel_flush(p->c, &local_err);
>                     // Handle error somehow
>                     //  If you want to be a pro, just change the
>                     // Threshold depending on what the kernel answers.
>                     // If it has to revert to synchronous sent, just
>                     // decrease the threshold, otherwise increase it.
> 
>             if (p->flags & MULTIFD_FLAG_SYNC) {
>                 qemu_sem_post(&p->sem_sync);
>             }
>             qemu_sem_post(&multifd_send_state->channels_ready);
>         } else if (p->quit) {
>             qemu_mutex_unlock(&p->mutex);
>             break;
>         } else {
>             qemu_mutex_unlock(&p->mutex);
>             /* sometimes there are spurious wakeups */
>         }
>     }
>     .............
> }
> 
> What do you think?

It's a good idea to keep locked memory at bay, but I think flushing after N
bytes sent is too often, and will kill the performance. Even more if we assume
the non-zerocopy blocking sendmsg() is already flushing for us. 

> 
> Later, Juan.
> 

Thanks Juan!

Best regards,
Leo
Leonardo Bras June 21, 2022, 3:35 a.m. UTC | #4
On Mon, 2022-06-20 at 11:44 -0400, Peter Xu wrote:
> On Mon, Jun 20, 2022 at 11:23:53AM +0200, Juan Quintela wrote:
> > Once discussed this, what I asked in the past is that you are having too
> > much dirty memory on zero_copy.  When you have a Multiterabyte guest, in
> > a single round you have a "potentially" dirty memory on each channel of:
> > 
> >    total_amount_memory / number of channels.
> > 
> > In a Multiterabyte guest, this is going to be more that probably in the
> > dozens of gigabytes.  As far as I know there is no card/driver that will
> > benefit for so many pages in zero_copy, and kernel will move to
> > synchronous copy at some point.  (In older threads, daniel showed how to
> > test for this case).
> 
> I was wondering whether the kernel needs to cache a lot of messages for
> zero copy if we don't flush it for a long time, as recvmsg(MSG_ERRQUEUE)
> seems to be fetching one message from the kernel one at a time.  And,
> whether that queue has a limit in length or something.

IIRC, if all messages look the same, it 'merges' them in a single message, like,
'this range has these flags and output'.

So, if no issue happens, we should have a single message with the confirmation
of all sent buffers, meaning just a little memory is used for that.

> 
> Does it mean that when the kernel could have cached enough of these
> messages then it'll fallback to the no-zero-copy mode?  And probably that's
> the way how kernel protects itself from using too much buffer for the error
> msgs?

Since it merges the messages, I don't think it uses a lot of space for that.

IIRC, the kernel will fall back to copying only if the network adapter / driver
does not support MSG_ZEROCOPY, like when it does not support scatter-gather.

> 
> This reminded me - Leo, have you considered adding the patch altogether to
> detect the "fallback to non-zero-copy" condition?  Because when with it and
> when the fallback happens at some point (e.g. when the guest memory is
> larger than some value) we'll know.

I still did not consider that, but sure, how do you see that working?

We can't just disable zero-copy-send because the user actually opted in, so we
could instead add a one time error message for when it falls back to copying, as
it should happen in the first try of zero-copy send.

Or we could fail the migration, stating the interface does not support
MSG_ZEROCOPY, since it should happen in the first sendmsg().

I would personally opt for the last option.

What do you think?

> 
> Thanks,
> 

Thanks Peter!

Best regards,
Leo
Juan Quintela June 21, 2022, 2:51 p.m. UTC | #5
Leonardo Brás <leobras@redhat.com> wrote:
> On Mon, 2022-06-20 at 11:44 -0400, Peter Xu wrote:
>> On Mon, Jun 20, 2022 at 11:23:53AM +0200, Juan Quintela wrote:
>> > Once discussed this, what I asked in the past is that you are having too
>> > much dirty memory on zero_copy.  When you have a Multiterabyte guest, in
>> > a single round you have a "potentially" dirty memory on each channel of:
>> > 
>> >    total_amount_memory / number of channels.
>> > 
>> > In a Multiterabyte guest, this is going to be more that probably in the
>> > dozens of gigabytes.  As far as I know there is no card/driver that will
>> > benefit for so many pages in zero_copy, and kernel will move to
>> > synchronous copy at some point.  (In older threads, daniel showed how to
>> > test for this case).
>> 
>> I was wondering whether the kernel needs to cache a lot of messages for
>> zero copy if we don't flush it for a long time, as recvmsg(MSG_ERRQUEUE)
>> seems to be fetching one message from the kernel one at a time.  And,
>> whether that queue has a limit in length or something.
>
> IIRC, if all messages look the same, it 'merges' them in a single message, like,
> 'this range has these flags and output'.
>
> So, if no issue happens, we should have a single message with the confirmation
> of all sent buffers, meaning just a little memory is used for that.
>
>> 
>> Does it mean that when the kernel could have cached enough of these
>> messages then it'll fallback to the no-zero-copy mode?  And probably that's
>> the way how kernel protects itself from using too much buffer for the error
>> msgs?
>
> Since it merges the messages, I don't think it uses a lot of space for that.
>
> IIRC, the kernel will fall back to copying only if the network adapter / driver
> does not support MSG_ZEROCOPY, like when it does not support scatter-gather.

My understanding is that it will fallback when you have too much stuff
inflight.

>> 
>> This reminded me - Leo, have you considered adding the patch altogether to
>> detect the "fallback to non-zero-copy" condition?  Because when with it and
>> when the fallback happens at some point (e.g. when the guest memory is
>> larger than some value) we'll know.
>
> I still did not consider that, but sure, how do you see that working?

send with zero_copy(1MB)
send with zero_copy(1MB)
.... (repeat)
at some point kernel decides:
sync all queue()
send synchronously next package.

we are not wondering if the kernel does this (it does).  What we are
wondering is when it does it, i.e. after 1MB worth of writes, 2MB, 10MB
....
That is the thing that depends on kernel/network card/driver.


> We can't just disable zero-copy-send because the user actually opted in, so we
> could instead add a one time error message for when it falls back to copying, as
> it should happen in the first try of zero-copy send.

On your 1st (or second) series, Dan Berrange explained hew to use the
error message interface to detect it.

> Or we could fail the migration, stating the interface does not support
> MSG_ZEROCOPY, since it should happen in the first sendmsg().

> I would personally opt for the last option.
>
> What do you think?

Later, Juan.
Peter Xu June 21, 2022, 3:04 p.m. UTC | #6
On Tue, Jun 21, 2022 at 12:35:54AM -0300, Leonardo Brás wrote:
> > This reminded me - Leo, have you considered adding the patch altogether to
> > detect the "fallback to non-zero-copy" condition?  Because when with it and
> > when the fallback happens at some point (e.g. when the guest memory is
> > larger than some value) we'll know.
> 
> I still did not consider that, but sure, how do you see that working?
> 
> We can't just disable zero-copy-send because the user actually opted in, so we
> could instead add a one time error message for when it falls back to copying, as
> it should happen in the first try of zero-copy send.
> 
> Or we could fail the migration, stating the interface does not support
> MSG_ZEROCOPY, since it should happen in the first sendmsg().
> 
> I would personally opt for the last option.
> 
> What do you think?

I don't have a very strong feeling on it, but yes that sounds proper to me.
If one day we'll like zero-copy send be on by default then we'll consider
the other way round, but maybe not necessary for now.  Thanks.
Peter Xu June 21, 2022, 3:09 p.m. UTC | #7
On Tue, Jun 21, 2022 at 04:51:39PM +0200, Juan Quintela wrote:
> > IIRC, the kernel will fall back to copying only if the network adapter / driver
> > does not support MSG_ZEROCOPY, like when it does not support scatter-gather.
> 
> My understanding is that it will fallback when you have too much stuff
> inflight.

I think we'd better figure this out soon, because if so then IMHO we can't
simply fail the migration when the fallback happens..  If we're not sure
about that, we can always be on the safe side to dump an error only.
diff mbox series

Patch

diff --git a/migration/multifd.h b/migration/multifd.h
index 4d8d89e5e5..e7cbdf1fb4 100644
--- a/migration/multifd.h
+++ b/migration/multifd.h
@@ -22,6 +22,7 @@  bool multifd_recv_new_channel(QIOChannel *ioc, Error **errp);
 void multifd_recv_sync_main(void);
 int multifd_send_sync_main(QEMUFile *f);
 int multifd_queue_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset);
+int multifd_zero_copy_flush(void);
 
 /* Multifd Compression flags */
 #define MULTIFD_FLAG_SYNC (1 << 0)
diff --git a/migration/multifd.c b/migration/multifd.c
index 9282ab6aa4..ce4220a97d 100644
--- a/migration/multifd.c
+++ b/migration/multifd.c
@@ -566,10 +566,42 @@  void multifd_save_cleanup(void)
     multifd_send_state = NULL;
 }
 
+/*
+ * Set zero_copy_flush = true for every multifd channel
+ *
+ * When using zero-copy, it's necessary to flush the pages before any of
+ * the pages can be sent again, so we'll make sure the new version of the
+ * pages will always arrive _later_ than the old pages.
+ *
+ * Should be called only after we finished one whole scanning of
+ * all the dirty bitmaps.
+ */
+int multifd_zero_copy_flush(void)
+{
+    int i;
+    Error *local_err = NULL;
+
+    if (!migrate_use_multifd()) {
+        return 0;
+    }
+
+    for (i = 0; i < migrate_multifd_channels(); i++) {
+        MultiFDSendParams *p = &multifd_send_state->params[i];
+        int ret;
+
+        ret = qio_channel_flush(p->c, &local_err);
+        if (ret < 0) {
+            error_report_err(local_err);
+            return ret;
+        }
+    }
+
+    return 0;
+}
+
 int multifd_send_sync_main(QEMUFile *f)
 {
     int i;
-    bool flush_zero_copy;
 
     if (!migrate_use_multifd()) {
         return 0;
@@ -581,19 +613,6 @@  int multifd_send_sync_main(QEMUFile *f)
         }
     }
 
-    /*
-     * When using zero-copy, it's necessary to flush the pages before any of
-     * the pages can be sent again, so we'll make sure the new version of the
-     * pages will always arrive _later_ than the old pages.
-     *
-     * Currently we achieve this by flushing the zero-page requested writes
-     * per ram iteration, but in the future we could potentially optimize it
-     * to be less frequent, e.g. only after we finished one whole scanning of
-     * all the dirty bitmaps.
-     */
-
-    flush_zero_copy = migrate_use_zero_copy_send();
-
     for (i = 0; i < migrate_multifd_channels(); i++) {
         MultiFDSendParams *p = &multifd_send_state->params[i];
 
@@ -615,17 +634,6 @@  int multifd_send_sync_main(QEMUFile *f)
         ram_counters.transferred += p->packet_len;
         qemu_mutex_unlock(&p->mutex);
         qemu_sem_post(&p->sem);
-
-        if (flush_zero_copy && p->c) {
-            int ret;
-            Error *err = NULL;
-
-            ret = qio_channel_flush(p->c, &err);
-            if (ret < 0) {
-                error_report_err(err);
-                return -1;
-            }
-        }
     }
     for (i = 0; i < migrate_multifd_channels(); i++) {
         MultiFDSendParams *p = &multifd_send_state->params[i];
diff --git a/migration/ram.c b/migration/ram.c
index 5f5e37f64d..514584e44f 100644
--- a/migration/ram.c
+++ b/migration/ram.c
@@ -2288,6 +2288,13 @@  static int ram_find_and_save_block(RAMState *rs)
     rs->last_seen_block = pss.block;
     rs->last_page = pss.page;
 
+    if (pss.complete_round && migrate_use_zero_copy_send()) {
+        int ret = multifd_zero_copy_flush();
+        if (ret < 0) {
+            return ret;
+        }
+    }
+
     return pages;
 }