Skip to contentSkip to conversation

migration: Terminate multifd threads on yank

PATCH In review Leonardo Bras 3 participants 6 messages Jul 30, 2021 – Aug 3, 2021

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 migration: Terminate multifd threads on yank

From source host viewpoint, losing a connection during migration will cause the sockets to get stuck in sendmsg() syscall, waiting for the receiving side to reply.

In migration, yank works by shutting-down the migration QIOChannel fd. This causes a failure in the next sendmsg() for that fd, and the whole migration gets cancelled.

In multifd, due to having multiple sockets in multiple threads, on a connection loss there will be extra sockets stuck in sendmsg(), and because they will be holding their own mutex, there is good chance the main migration thread can get stuck in multifd_send_pages() waiting for one of those mutexes.

While it's waiting, the main migration thread can't run sendmsg() on it's fd, and therefore can't cause the migration to be cancelled, thus causing yank not to work.

Fixes this by shutting down all migration fds (including multifd ones), so no thread get's stuck in sendmsg() while holding a lock, and thus allowing the main migration thread to properly cancel migration when yank is used.

There is no need to do the same procedure to yank to work in the receiving host since ops->recv_pages() is kept outside the mutex protected code in multifd_recv_thread().

---
 migration/multifd.c        | 11 +++++++++++
 migration/multifd.h        |  1 +
 migration/yank_functions.c |  2 ++
 3 files changed, 14 insertions(+)
migration/multifd.c +11 -0
index 377da78f5b..744a180dfe 100644
@@ -1040,6 +1040,17 @@ void multifd_recv_sync_main(void)
     trace_multifd_recv_sync_main(multifd_recv_state->packet_num);
 }
 
+void multifd_shutdown(void)
+{
+    if (!migrate_use_multifd()) {
+        return;
+    }
+
+    if (multifd_send_state) {
+        multifd_send_terminate_threads(NULL);
+    }
+}
+
 static void *multifd_recv_thread(void *opaque)
 {
     MultiFDRecvParams *p = opaque;
migration/multifd.h +1 -0
index 8d6751f5ed..0517213bdf 100644
@@ -22,6 +22,7 @@ bool multifd_recv_new_channel(QIOChannel *ioc, Error **errp);
 void multifd_recv_sync_main(void);
 void multifd_send_sync_main(QEMUFile *f);
 int multifd_queue_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset);
+void multifd_shutdown(void);
 
 /* Multifd Compression flags */
 #define MULTIFD_FLAG_SYNC (1 << 0)
migration/yank_functions.c +2 -0
index 8c08aef14a..9335a64f00 100644
@@ -15,12 +15,14 @@
 #include "io/channel-socket.h"
 #include "io/channel-tls.h"
 #include "qemu-file.h"
+#include "multifd.h"
 
 void migration_yank_iochannel(void *opaque)
 {
     QIOChannel *ioc = QIO_CHANNEL(opaque);
 
     qio_channel_shutdown(ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
+    multifd_shutdown();
 }
 
 /* Return whether yank is supported on this ioc */
Signature · 1 line
2.32.0
#1 lore raw
Quoted 43 lines · * Leonardo Bras ([email protected]) wrote:

From source host viewpoint, losing a connection during migration will cause the sockets to get stuck in sendmsg() syscall, waiting for the receiving side to reply.

In migration, yank works by shutting-down the migration QIOChannel fd. This causes a failure in the next sendmsg() for that fd, and the whole migration gets cancelled.

In multifd, due to having multiple sockets in multiple threads, on a connection loss there will be extra sockets stuck in sendmsg(), and because they will be holding their own mutex, there is good chance the main migration thread can get stuck in multifd_send_pages() waiting for one of those mutexes.

While it's waiting, the main migration thread can't run sendmsg() on it's fd, and therefore can't cause the migration to be cancelled, thus causing yank not to work.

Fixes this by shutting down all migration fds (including multifd ones), so no thread get's stuck in sendmsg() while holding a lock, and thus allowing the main migration thread to properly cancel migration when yank is used.

There is no need to do the same procedure to yank to work in the receiving host since ops->recv_pages() is kept outside the mutex protected code in multifd_recv_thread().

---
 migration/multifd.c        | 11 +++++++++++
 migration/multifd.h        |  1 +
 migration/yank_functions.c |  2 ++
 3 files changed, 14 insertions(+)
migration/multifd.c +0 -0
index 377da78f5b..744a180dfe 100644
@@ -1040,6 +1040,17 @@ void multifd_recv_sync_main(void)
     trace_multifd_recv_sync_main(multifd_recv_state->packet_num);
 }

migration/multifd.c void multifd_recv_sync_main(void) · line 1051

Quoted 10 lines
migration/multifd.c +9 -0
 
+void multifd_shutdown(void)
+{
+    if (!migrate_use_multifd()) {
+        return;
+    }
+
+    if (multifd_send_state) {
+        multifd_send_terminate_threads(NULL);
+    }

That calls :

    for (i = 0; i < migrate_multifd_channels(); i++) {
        MultiFDSendParams *p = &multifd_send_state->params[i];

        qemu_mutex_lock(&p->mutex);
        p->quit = true;
        qemu_sem_post(&p->sem);
        qemu_mutex_unlock(&p->mutex);
    }

so why doesn't this also get stuck in the same mutex you're trying to fix?

Does the qio_channel_shutdown actually cause a shutdown on all fd's for the multifd?

(I've just seen the multifd/cancel test fail stuck in multifd_send_sync_main waiting on one of the locks).

Dave

Quoted 39 lines
migration/multifd.c +2 -0
+}
+
 static void *multifd_recv_thread(void *opaque)
 {
     MultiFDRecvParams *p = opaque;
migration/multifd.h +1 -0
index 8d6751f5ed..0517213bdf 100644
@@ -22,6 +22,7 @@ bool multifd_recv_new_channel(QIOChannel *ioc, Error **errp);
 void multifd_recv_sync_main(void);
 void multifd_send_sync_main(QEMUFile *f);
 int multifd_queue_page(QEMUFile *f, RAMBlock *block, ram_addr_t offset);
+void multifd_shutdown(void);
 
 /* Multifd Compression flags */
 #define MULTIFD_FLAG_SYNC (1 << 0)
migration/yank_functions.c +2 -0
index 8c08aef14a..9335a64f00 100644
@@ -15,12 +15,14 @@
 #include "io/channel-socket.h"
 #include "io/channel-tls.h"
 #include "qemu-file.h"
+#include "multifd.h"
 
 void migration_yank_iochannel(void *opaque)
 {
     QIOChannel *ioc = QIO_CHANNEL(opaque);
 
     qio_channel_shutdown(ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
+    multifd_shutdown();
 }
 
 /* Return whether yank is supported on this ioc */
Signature · 1 line
2.32.0
Signature · 1 line
Dr. David Alan Gilbert / [email protected] / Manchester, UK
#2 lore raw

Hello Dave,

Quoted 28 lines
Quoted 7 lines
migration/multifd.c +0 -0
index 377da78f5b..744a180dfe 100644
@@ -1040,6 +1040,17 @@ void multifd_recv_sync_main(void)
     trace_multifd_recv_sync_main(multifd_recv_state->packet_num);
 }

migration/multifd.c void multifd_recv_sync_main(void) · line 1051

Quoted 10 lines
migration/multifd.c +9 -0
 
+void multifd_shutdown(void)
+{
+    if (!migrate_use_multifd()) {
+        return;
+    }
+
+    if (multifd_send_state) {
+        multifd_send_terminate_threads(NULL);
+    }

That calls :

    for (i = 0; i < migrate_multifd_channels(); i++) {
        MultiFDSendParams *p = &multifd_send_state->params[i];

        qemu_mutex_lock(&p->mutex);
        p->quit = true;
        qemu_sem_post(&p->sem);
        qemu_mutex_unlock(&p->mutex);
    }
Quoted 2 lines

so why doesn't this also get stuck in the same mutex you're trying to fix?

You are right, I got confused over the locks.
I need to get a better look at the code, and truly understand why this patch fixes (?) the issue.

Quoted 3 lines

Does the qio_channel_shutdown actually cause a shutdown on all fd's for the multifd?

As far as I tested, it does shutdown a single fd, but whenever this fd fails in it's first sendmsg it causes migration to fail and all the other fds get shutdown as well.

Quoted 11 lines

(I've just seen the multifd/cancel test fail stuck in multifd_send_sync_main waiting on one of the locks).

Dave

Quoted 1 line

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK

I will do a little more reading / debugging in this code. Thanks Dave!

#3 lore raw
In reply to Leonardo Bras
Quoted 23 lines · On Fri, 30 Jul 2021 04:40:45 -0300 Leonardo Bras <[email protected]> wrote:

From source host viewpoint, losing a connection during migration will cause the sockets to get stuck in sendmsg() syscall, waiting for the receiving side to reply.

In migration, yank works by shutting-down the migration QIOChannel fd. This causes a failure in the next sendmsg() for that fd, and the whole migration gets cancelled.

In multifd, due to having multiple sockets in multiple threads, on a connection loss there will be extra sockets stuck in sendmsg(), and because they will be holding their own mutex, there is good chance the main migration thread can get stuck in multifd_send_pages() waiting for one of those mutexes.

While it's waiting, the main migration thread can't run sendmsg() on it's fd, and therefore can't cause the migration to be cancelled, thus causing yank not to work.

Fixes this by shutting down all migration fds (including multifd ones), so no thread get's stuck in sendmsg() while holding a lock, and thus allowing the main migration thread to properly cancel migration when yank is used.

Quoted 8 lines

There is no need to do the same procedure to yank to work in the receiving host since ops->recv_pages() is kept outside the mutex protected code in multifd_recv_thread().

---

Hi,
There is an easier explanation: I forgot the send side of multifd altogether (I thought it was covered by migration_channel_connect()). So yank won't actually shutdown() the multifd sockets on the send side.

In the bugreport you wrote

Quoted 1 line

(As a test, I called qio_channel_shutdown() in every multifd iochannel and yank worked just fine, but I could not retry migration, because it was still 'ongoing')

That sounds like a bug in the error handling for multifd. But quickly looking at the code, it should properly fail the migration.

BTW: You can shutdown outgoing sockets from outside of qemu with the 'ss' utility, like this: 'sudo ss -K dst <destination ip> dport = <destination port>'

Regards,
Lukas Straub

#4 lore raw

Hello Lukas,

Quoted 4 lines · Lukas Straub

Hi,
There is an easier explanation: I forgot the send side of multifd altogether (I thought it was covered by migration_channel_connect()). So yank won't actually shutdown() the multifd sockets on the send side.

If I could get that correctly, it seems to abort migration (and therefore close all fds) if the ft that ends up qio_channel_shutdown() get to sendmsg(), which can take a while.
But it really does not close thew fds before that.

Quoted 5 lines

In the bugreport you wrote

Quoted 1 line

(As a test, I called qio_channel_shutdown() in every multifd iochannel and yank worked just fine, but I could not retry migration, because it was still 'ongoing')

That sounds like a bug in the error handling for multifd. But quickly looking at the code, it should properly fail the migration.

In the end, just asking each thread to just exit ended up getting me a smoother migration abort.

Quoted 3 lines

BTW: You can shutdown outgoing sockets from outside of qemu with the 'ss' utility, like this: 'sudo ss -K dst <destination ip> dport = <destination port>'

Very nice tool, thanks for sharing!

Quoted 3 lines

Regards,
Lukas Straub

Best regards,
Leonardo Bras

#5 lore raw
Quoted 11 lines · On Tue, 3 Aug 2021 04:18:42 -0300 Leonardo Bras Soares Passos…

Hello Lukas,

Quoted 4 lines · Lukas Straub

Hi,
There is an easier explanation: I forgot the send side of multifd altogether (I thought it was covered by migration_channel_connect()). So yank won't actually shutdown() the multifd sockets on the send side.

If I could get that correctly, it seems to abort migration (and therefore close all fds) if the ft that ends up qio_channel_shutdown() get to sendmsg(), which can take a while.

How long is "can take a while"? Until some TCP connection times out? That would mean that it is hanging somewhere else.

I mean in precopy migration the multifd send threads should be fully utilized and always sending something until the migration finishes. In that case it is likely that all the treads become stuck in qio_channel_write_all() if the connection breaks silently (i.e. discards packets or the destination is powered off, No connection reset) since there are no TCP ACK's ariving from the destination side -> kernel tcp buffer becomes full -> qio_channel_write_all() blocks. Thus, shutdown() on the sockets should be enough to get the treads unstuck and notice that the connection broke.

If something else hangs, the question is where...

Quoted 1 line

But it really does not close thew fds before that.

Note: shutdown() is not close().

Quoted 21 lines
Quoted 5 lines

In the bugreport you wrote

Quoted 1 line

(As a test, I called qio_channel_shutdown() in every multifd iochannel and yank worked just fine, but I could not retry migration, because it was still 'ongoing')

That sounds like a bug in the error handling for multifd. But quickly looking at the code, it should properly fail the migration.

In the end, just asking each thread to just exit ended up getting me a smoother migration abort.

Quoted 3 lines

BTW: You can shutdown outgoing sockets from outside of qemu with the 'ss' utility, like this: 'sudo ss -K dst <destination ip> dport = <destination port>'

Very nice tool, thanks for sharing!

Quoted 3 lines

Regards,
Lukas Straub

Best regards,
Leonardo Bras