[PATCH v4 0/8] Fix segfault on migration return path

[PATCH v4 0/8] Fix segfault on migration return path

[Fabiano Rosas]

The major change from v3 is the dropping of shutdown() before close()
as suggested by Peter.

I also added a yank handler to rp_state.from_dst_file.

CI run: https://gitlab.com/farosas/qemu/-/pipelines/969570891

v3:
https://lore.kernel.org/r/20230811150836.2895-1-farosas@suse.de

I decided to fix the issues with the shutdown instead of complaining
about them. First 5 patches address all of the possible races I
found. The only problem left to figure out is the -EIO on shutdown
which will need more thought.

Patches 6 & 7 fix the original segfault.

Patches 8-10 make the cleanup of migration files more predictable and
centralized.

I also adjusted some small things that were mentioned by Peter:

- moved the rp.error = false outside of the thread;
- stopped checking for errors during postcopy_pause();
- dropped the tracepoint;

CI run: https://gitlab.com/farosas/qemu/-/pipelines/963407228

v2:
https://lore.kernel.org/r/20230802143644.7534-1-farosas@suse.de

- moved the await into postcopy_pause() as Peter suggested;

- brought back the mark_source_rp_bad call. Turns out that piece of
code is filled with nuance. I just moved it aside since it doesn't
make sense during pause/resume. We can tackle that when we get the
chance.

CI run: https://gitlab.com/farosas/qemu/-/pipelines/953420150
Also ran the switchover and preempt tests for 1000 times each on
x86_64.

v1:
https://lore.kernel.org/r/20230728121516.16258-1-farosas@suse.de

The /x86_64/migration/postcopy/preempt/recovery/plain test is
sometimes failing due a segmentation fault on the migration return
path. There is a race involving the retry logic of the return path and
the migration resume command.

The issue happens when the retry logic tries to cleanup the current
return path file, but ends up cleaning the new one and trying to use
it right after. Tracing shows it clearly:

open_return_path_on_source  <-- at migration start
open_return_path_on_source_continue <-- rp thread created
postcopy_pause_incoming
postcopy_pause_fast_load
qemu-system-x86_64: Detected IO failure for postcopy. Migration paused. (incoming)
postcopy_pause_fault_thread
qemu-system-x86_64: Detected IO failure for postcopy. Migration paused. (source)
postcopy_pause_incoming_continued
open_return_path_on_source   <-- NOK, too soon
postcopy_pause_continued
postcopy_pause_return_path   <-- too late, already operating on the new from_dst_file
postcopy_pause_return_path_continued <-- will continue and crash
postcopy_pause_incoming
qemu-system-x86_64: Detected IO failure for postcopy. Migration paused.
postcopy_pause_incoming_continued

We could solve this by adding some form of synchronization to ensure
that we always do the cleanup before setting up the new file, but I
find it more straight-forward to move the retry logic outside of the
thread by letting it finish and starting a new thread when resuming
the migration.

More details on the commit message.

CI run: https://gitlab.com/farosas/qemu/-/pipelines/947875609

Fabiano Rosas (8):
  migration: Fix possible race when setting rp_state.error
  migration: Fix possible races when shutting down the return path
  migration: Fix possible race when shutting down to_dst_file
  migration: Remove redundant cleanup of postcopy_qemufile_src
  migration: Consolidate return path closing code
  migration: Replace the return path retry logic
  migration: Move return path cleanup to main migration thread
  migration: Add a wrapper to cleanup migration files

 migration/migration.c | 229 +++++++++++++++---------------------------
 migration/migration.h |   1 -
 2 files changed, 80 insertions(+), 150 deletions(-)

-- 
2.35.3

[PATCH v4 1/8] migration: Fix possible race when setting rp_state.error

[Fabiano Rosas]

We don't need to set the rp_state.error right after a shutdown because
qemu_file_shutdown() always sets the QEMUFile error, so the return
path thread would have seen it and set the rp error itself.

Setting the error outside of the thread is also racy because the
thread could clear it after we set it.

Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
---
 migration/migration.c | 1 -
 1 file changed, 1 deletion(-)

diff --git a/migration/migration.c b/migration/migration.c
index 5528acb65e..f88c86079c 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -2062,7 +2062,6 @@ static int await_return_path_close_on_source(MigrationState *ms)
          * waiting for the destination.
          */
         qemu_file_shutdown(ms->rp_state.from_dst_file);
-        mark_source_rp_bad(ms);
     }
     trace_await_return_path_close_on_source_joining();
     qemu_thread_join(&ms->rp_state.rp_thread);
-- 
2.35.3

[PATCH v4 2/8] migration: Fix possible races when shutting down the return path

[Fabiano Rosas]

We cannot call qemu_file_shutdown() on the return path file without
taking the file lock. The return path thread could be running it's
cleanup code and have just cleared the from_dst_file pointer.

Checking ms->to_dst_file for errors could also race with
migrate_fd_cleanup() which clears the to_dst_file pointer.

Protect both accesses by taking the file lock.

This was caught by inspection, it should be rare, but the next patches
will start calling this code from other places, so let's do the
correct thing.

Signed-off-by: Fabiano Rosas <farosas@suse.de>
---
 migration/migration.c | 19 ++++++++++---------
 1 file changed, 10 insertions(+), 9 deletions(-)

diff --git a/migration/migration.c b/migration/migration.c
index f88c86079c..85c171f32c 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -2052,17 +2052,18 @@ static int open_return_path_on_source(MigrationState *ms,
 static int await_return_path_close_on_source(MigrationState *ms)
 {
     /*
-     * If this is a normal exit then the destination will send a SHUT and the
-     * rp_thread will exit, however if there's an error we need to cause
-     * it to exit.
+     * If this is a normal exit then the destination will send a SHUT
+     * and the rp_thread will exit, however if there's an error we
+     * need to cause it to exit. shutdown(2), if we have it, will
+     * cause it to unblock if it's stuck waiting for the destination.
      */
-    if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) {
-        /*
-         * shutdown(2), if we have it, will cause it to unblock if it's stuck
-         * waiting for the destination.
-         */
-        qemu_file_shutdown(ms->rp_state.from_dst_file);
+    WITH_QEMU_LOCK_GUARD(&ms->qemu_file_lock) {
+        if (ms->to_dst_file && ms->rp_state.from_dst_file &&
+            qemu_file_get_error(ms->to_dst_file)) {
+            qemu_file_shutdown(ms->rp_state.from_dst_file);
+        }
     }
+
     trace_await_return_path_close_on_source_joining();
     qemu_thread_join(&ms->rp_state.rp_thread);
     ms->rp_state.rp_thread_created = false;
-- 
2.35.3

[PATCH v4 3/8] migration: Fix possible race when shutting down to_dst_file

[Fabiano Rosas]

It's not safe to call qemu_file_shutdown() on the to_dst_file without
first checking for the file's presence under the lock. The cleanup of
this file happens at postcopy_pause() and migrate_fd_cleanup() which
are not necessarily running in the same thread as migrate_fd_cancel().

Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
---
 migration/migration.c | 18 +++++++++++-------
 1 file changed, 11 insertions(+), 7 deletions(-)

diff --git a/migration/migration.c b/migration/migration.c
index 85c171f32c..5e6a766235 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -1245,7 +1245,7 @@ static void migrate_fd_error(MigrationState *s, const Error *error)
 static void migrate_fd_cancel(MigrationState *s)
 {
     int old_state ;
-    QEMUFile *f = migrate_get_current()->to_dst_file;
+
     trace_migrate_fd_cancel();
 
     WITH_QEMU_LOCK_GUARD(&s->qemu_file_lock) {
@@ -1271,11 +1271,13 @@ static void migrate_fd_cancel(MigrationState *s)
      * If we're unlucky the migration code might be stuck somewhere in a
      * send/write while the network has failed and is waiting to timeout;
      * if we've got shutdown(2) available then we can force it to quit.
-     * The outgoing qemu file gets closed in migrate_fd_cleanup that is
-     * called in a bh, so there is no race against this cancel.
      */
-    if (s->state == MIGRATION_STATUS_CANCELLING && f) {
-        qemu_file_shutdown(f);
+    if (s->state == MIGRATION_STATUS_CANCELLING) {
+        WITH_QEMU_LOCK_GUARD(&s->qemu_file_lock) {
+            if (s->to_dst_file) {
+                qemu_file_shutdown(s->to_dst_file);
+            }
+        }
     }
     if (s->state == MIGRATION_STATUS_CANCELLING && s->block_inactive) {
         Error *local_err = NULL;
@@ -1519,12 +1521,14 @@ void qmp_migrate_pause(Error **errp)
 {
     MigrationState *ms = migrate_get_current();
     MigrationIncomingState *mis = migration_incoming_get_current();
-    int ret;
+    int ret = 0;
 
     if (ms->state == MIGRATION_STATUS_POSTCOPY_ACTIVE) {
         /* Source side, during postcopy */
         qemu_mutex_lock(&ms->qemu_file_lock);
-        ret = qemu_file_shutdown(ms->to_dst_file);
+        if (ms->to_dst_file) {
+            ret = qemu_file_shutdown(ms->to_dst_file);
+        }
         qemu_mutex_unlock(&ms->qemu_file_lock);
         if (ret) {
             error_setg(errp, "Failed to pause source migration");
-- 
2.35.3

[PATCH v4 4/8] migration: Remove redundant cleanup of postcopy_qemufile_src

[Fabiano Rosas]

This file is owned by the return path thread which is already doing
cleanup.

Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
---
 migration/migration.c | 6 ------
 1 file changed, 6 deletions(-)

diff --git a/migration/migration.c b/migration/migration.c
index 5e6a766235..195726eb4a 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -1177,12 +1177,6 @@ static void migrate_fd_cleanup(MigrationState *s)
         qemu_fclose(tmp);
     }
 
-    if (s->postcopy_qemufile_src) {
-        migration_ioc_unregister_yank_from_file(s->postcopy_qemufile_src);
-        qemu_fclose(s->postcopy_qemufile_src);
-        s->postcopy_qemufile_src = NULL;
-    }
-
     assert(!migration_is_active(s));
 
     if (s->state == MIGRATION_STATUS_CANCELLING) {
-- 
2.35.3

[PATCH v4 5/8] migration: Consolidate return path closing code

[Fabiano Rosas]

We'll start calling the await_return_path_close_on_source() function
from other parts of the code, so move all of the related checks and
tracepoints into it.

Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
---
 migration/migration.c | 29 ++++++++++++++---------------
 1 file changed, 14 insertions(+), 15 deletions(-)

diff --git a/migration/migration.c b/migration/migration.c
index 195726eb4a..4edbee3a5d 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -2049,6 +2049,14 @@ static int open_return_path_on_source(MigrationState *ms,
 /* Returns 0 if the RP was ok, otherwise there was an error on the RP */
 static int await_return_path_close_on_source(MigrationState *ms)
 {
+    int ret;
+
+    if (!ms->rp_state.rp_thread_created) {
+        return 0;
+    }
+
+    trace_migration_return_path_end_before();
+
     /*
      * If this is a normal exit then the destination will send a SHUT
      * and the rp_thread will exit, however if there's an error we
@@ -2066,7 +2074,10 @@ static int await_return_path_close_on_source(MigrationState *ms)
     qemu_thread_join(&ms->rp_state.rp_thread);
     ms->rp_state.rp_thread_created = false;
     trace_await_return_path_close_on_source_close();
-    return ms->rp_state.error;
+
+    ret = ms->rp_state.error;
+    trace_migration_return_path_end_after(ret);
+    return ret;
 }
 
 static inline void
@@ -2362,20 +2373,8 @@ static void migration_completion(MigrationState *s)
         goto fail;
     }
 
-    /*
-     * If rp was opened we must clean up the thread before
-     * cleaning everything else up (since if there are no failures
-     * it will wait for the destination to send it's status in
-     * a SHUT command).
-     */
-    if (s->rp_state.rp_thread_created) {
-        int rp_error;
-        trace_migration_return_path_end_before();
-        rp_error = await_return_path_close_on_source(s);
-        trace_migration_return_path_end_after(rp_error);
-        if (rp_error) {
-            goto fail;
-        }
+    if (await_return_path_close_on_source(s)) {
+        goto fail;
     }
 
     if (qemu_file_get_error(s->to_dst_file)) {
-- 
2.35.3

[PATCH v4 6/8] migration: Replace the return path retry logic

[Fabiano Rosas]

Replace the return path retry logic with finishing and restarting the
thread. This fixes a race when resuming the migration that leads to a
segfault.

Currently when doing postcopy we consider that an IO error on the
return path file could be due to a network intermittency. We then keep
the thread alive but have it do cleanup of the 'from_dst_file' and
wait on the 'postcopy_pause_rp' semaphore. When the user issues a
migrate resume, a new return path is opened and the thread is allowed
to continue.

There's a race condition in the above mechanism. It is possible for
the new return path file to be setup *before* the cleanup code in the
return path thread has had a chance to run, leading to the *new* file
being closed and the pointer set to NULL. When the thread is released
after the resume, it tries to dereference 'from_dst_file' and crashes:

Thread 7 "return path" received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0x7fffd1dbf700 (LWP 9611)]
0x00005555560e4893 in qemu_file_get_error_obj (f=0x0, errp=0x0) at ../migration/qemu-file.c:154
154         return f->last_error;

(gdb) bt
 #0  0x00005555560e4893 in qemu_file_get_error_obj (f=0x0, errp=0x0) at ../migration/qemu-file.c:154
 #1  0x00005555560e4983 in qemu_file_get_error (f=0x0) at ../migration/qemu-file.c:206
 #2  0x0000555555b9a1df in source_return_path_thread (opaque=0x555556e06000) at ../migration/migration.c:1876
 #3  0x000055555602e14f in qemu_thread_start (args=0x55555782e780) at ../util/qemu-thread-posix.c:541
 #4  0x00007ffff38d76ea in start_thread (arg=0x7fffd1dbf700) at pthread_create.c:477
 #5  0x00007ffff35efa6f in clone () at ../sysdeps/unix/sysv/linux/x86_64/clone.S:95

Here's the race (important bit is open_return_path happening before
migration_release_dst_files):

migration                 | qmp                         | return path
--------------------------+-----------------------------+---------------------------------
			    qmp_migrate_pause()
			     shutdown(ms->to_dst_file)
			      f->last_error = -EIO
migrate_detect_error()
 postcopy_pause()
  set_state(PAUSED)
  wait(postcopy_pause_sem)
			    qmp_migrate(resume)
			    migrate_fd_connect()
			     resume = state == PAUSED
			     open_return_path <-- TOO SOON!
			     set_state(RECOVER)
			     post(postcopy_pause_sem)
							(incoming closes to_src_file)
							res = qemu_file_get_error(rp)
							migration_release_dst_files()
							ms->rp_state.from_dst_file = NULL
  post(postcopy_pause_rp_sem)
							postcopy_pause_return_path_thread()
							  wait(postcopy_pause_rp_sem)
							rp = ms->rp_state.from_dst_file
							goto retry
							qemu_file_get_error(rp)
							SIGSEGV
-------------------------------------------------------------------------------------------

We can keep the retry logic without having the thread alive and
waiting. The only piece of data used by it is the 'from_dst_file' and
it is only allowed to proceed after a migrate resume is issued and the
semaphore released at migrate_fd_connect().

Move the retry logic to outside the thread by waiting for the thread
to finish before pausing the migration.

Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
---
 migration/migration.c | 60 ++++++++-----------------------------------
 migration/migration.h |  1 -
 2 files changed, 11 insertions(+), 50 deletions(-)

diff --git a/migration/migration.c b/migration/migration.c
index 4edbee3a5d..7dfcbc3634 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -1775,18 +1775,6 @@ static void migrate_handle_rp_req_pages(MigrationState *ms, const char* rbname,
     }
 }
 
-/* Return true to retry, false to quit */
-static bool postcopy_pause_return_path_thread(MigrationState *s)
-{
-    trace_postcopy_pause_return_path();
-
-    qemu_sem_wait(&s->postcopy_pause_rp_sem);
-
-    trace_postcopy_pause_return_path_continued();
-
-    return true;
-}
-
 static int migrate_handle_rp_recv_bitmap(MigrationState *s, char *block_name)
 {
     RAMBlock *block = qemu_ram_block_by_name(block_name);
@@ -1870,7 +1858,6 @@ static void *source_return_path_thread(void *opaque)
     trace_source_return_path_thread_entry();
     rcu_register_thread();
 
-retry:
     while (!ms->rp_state.error && !qemu_file_get_error(rp) &&
            migration_is_setup_or_active(ms->state)) {
         trace_source_return_path_thread_loop_top();
@@ -1992,26 +1979,7 @@ retry:
     }
 
 out:
-    res = qemu_file_get_error(rp);
-    if (res) {
-        if (res && migration_in_postcopy()) {
-            /*
-             * Maybe there is something we can do: it looks like a
-             * network down issue, and we pause for a recovery.
-             */
-            migration_release_dst_files(ms);
-            rp = NULL;
-            if (postcopy_pause_return_path_thread(ms)) {
-                /*
-                 * Reload rp, reset the rest.  Referencing it is safe since
-                 * it's reset only by us above, or when migration completes
-                 */
-                rp = ms->rp_state.from_dst_file;
-                ms->rp_state.error = false;
-                goto retry;
-            }
-        }
-
+    if (qemu_file_get_error(rp)) {
         trace_source_return_path_thread_bad_end();
         mark_source_rp_bad(ms);
     }
@@ -2022,8 +1990,7 @@ out:
     return NULL;
 }
 
-static int open_return_path_on_source(MigrationState *ms,
-                                      bool create_thread)
+static int open_return_path_on_source(MigrationState *ms)
 {
     ms->rp_state.from_dst_file = qemu_file_get_return_path(ms->to_dst_file);
     if (!ms->rp_state.from_dst_file) {
@@ -2032,11 +1999,6 @@ static int open_return_path_on_source(MigrationState *ms,
 
     trace_open_return_path_on_source();
 
-    if (!create_thread) {
-        /* We're done */
-        return 0;
-    }
-
     qemu_thread_create(&ms->rp_state.rp_thread, "return path",
                        source_return_path_thread, ms, QEMU_THREAD_JOINABLE);
     ms->rp_state.rp_thread_created = true;
@@ -2076,6 +2038,7 @@ static int await_return_path_close_on_source(MigrationState *ms)
     trace_await_return_path_close_on_source_close();
 
     ret = ms->rp_state.error;
+    ms->rp_state.error = false;
     trace_migration_return_path_end_after(ret);
     return ret;
 }
@@ -2551,6 +2514,13 @@ static MigThrError postcopy_pause(MigrationState *s)
         qemu_file_shutdown(file);
         qemu_fclose(file);
 
+        /*
+         * We're already pausing, so ignore any errors on the return
+         * path and just wait for the thread to finish. It will be
+         * re-created when we resume.
+         */
+        await_return_path_close_on_source(s);
+
         migrate_set_state(&s->state, s->state,
                           MIGRATION_STATUS_POSTCOPY_PAUSED);
 
@@ -2568,12 +2538,6 @@ static MigThrError postcopy_pause(MigrationState *s)
         if (s->state == MIGRATION_STATUS_POSTCOPY_RECOVER) {
             /* Woken up by a recover procedure. Give it a shot */
 
-            /*
-             * Firstly, let's wake up the return path now, with a new
-             * return path channel.
-             */
-            qemu_sem_post(&s->postcopy_pause_rp_sem);
-
             /* Do the resume logic */
             if (postcopy_do_resume(s) == 0) {
                 /* Let's continue! */
@@ -3263,7 +3227,7 @@ void migrate_fd_connect(MigrationState *s, Error *error_in)
      * QEMU uses the return path.
      */
     if (migrate_postcopy_ram() || migrate_return_path()) {
-        if (open_return_path_on_source(s, !resume)) {
+        if (open_return_path_on_source(s)) {
             error_setg(&local_err, "Unable to open return-path for postcopy");
             migrate_set_state(&s->state, s->state, MIGRATION_STATUS_FAILED);
             migrate_set_error(s, local_err);
@@ -3327,7 +3291,6 @@ static void migration_instance_finalize(Object *obj)
     qemu_sem_destroy(&ms->rate_limit_sem);
     qemu_sem_destroy(&ms->pause_sem);
     qemu_sem_destroy(&ms->postcopy_pause_sem);
-    qemu_sem_destroy(&ms->postcopy_pause_rp_sem);
     qemu_sem_destroy(&ms->rp_state.rp_sem);
     qemu_sem_destroy(&ms->rp_state.rp_pong_acks);
     qemu_sem_destroy(&ms->postcopy_qemufile_src_sem);
@@ -3347,7 +3310,6 @@ static void migration_instance_init(Object *obj)
     migrate_params_init(&ms->parameters);
 
     qemu_sem_init(&ms->postcopy_pause_sem, 0);
-    qemu_sem_init(&ms->postcopy_pause_rp_sem, 0);
     qemu_sem_init(&ms->rp_state.rp_sem, 0);
     qemu_sem_init(&ms->rp_state.rp_pong_acks, 0);
     qemu_sem_init(&ms->rate_limit_sem, 0);
diff --git a/migration/migration.h b/migration/migration.h
index 6eea18db36..36eb5ba70b 100644
--- a/migration/migration.h
+++ b/migration/migration.h
@@ -382,7 +382,6 @@ struct MigrationState {
 
     /* Needed by postcopy-pause state */
     QemuSemaphore postcopy_pause_sem;
-    QemuSemaphore postcopy_pause_rp_sem;
     /*
      * Whether we abort the migration if decompression errors are
      * detected at the destination. It is left at false for qemu
-- 
2.35.3

[PATCH v4 7/8] migration: Move return path cleanup to main migration thread

[Fabiano Rosas]

Now that the return path thread is allowed to finish during a paused
migration, we can move the cleanup of the QEMUFiles to the main
migration thread.

Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
---
 migration/migration.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)

diff --git a/migration/migration.c b/migration/migration.c
index 7dfcbc3634..7fec57ad7f 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -98,6 +98,7 @@ static int migration_maybe_pause(MigrationState *s,
                                  int *current_active_state,
                                  int new_state);
 static void migrate_fd_cancel(MigrationState *s);
+static int await_return_path_close_on_source(MigrationState *s);
 
 static bool migration_needs_multiple_sockets(void)
 {
@@ -1177,6 +1178,12 @@ static void migrate_fd_cleanup(MigrationState *s)
         qemu_fclose(tmp);
     }
 
+    /*
+     * We already cleaned up to_dst_file, so errors from the return
+     * path might be due to that, ignore them.
+     */
+    await_return_path_close_on_source(s);
+
     assert(!migration_is_active(s));
 
     if (s->state == MIGRATION_STATUS_CANCELLING) {
@@ -1985,7 +1992,6 @@ out:
     }
 
     trace_source_return_path_thread_end();
-    migration_release_dst_files(ms);
     rcu_unregister_thread();
     return NULL;
 }
@@ -2039,6 +2045,9 @@ static int await_return_path_close_on_source(MigrationState *ms)
 
     ret = ms->rp_state.error;
     ms->rp_state.error = false;
+
+    migration_release_dst_files(ms);
+
     trace_migration_return_path_end_after(ret);
     return ret;
 }
-- 
2.35.3

[PATCH v4 8/8] migration: Add a wrapper to cleanup migration files

[Fabiano Rosas]

We currently have a pattern for cleaning up a migration QEMUFile:

  qemu_mutex_lock(&s->qemu_file_lock);
  file = s->file_name;
  s->file_name = NULL;
  qemu_mutex_unlock(&s->qemu_file_lock);

  migration_ioc_unregister_yank_from_file(file);
  qemu_file_shutdown(file);
  qemu_fclose(file);

There are some considerations for this sequence:

- we must clear the pointer under the lock, to avoid TOC/TOU bugs;
- the close() and yank unregister expect be given a non-null parameter;
- a close() in one thread should not race with a shutdown() in another;
- we don't need to shutdown() right before close();

Create a wrapper function to make sure everything works correctly.

Note: the return path didn't have a yank handler registered, I added
      it nonetheless for uniformity.

Signed-off-by: Fabiano Rosas <farosas@suse.de>
---
 migration/migration.c | 91 +++++++++++++------------------------------
 1 file changed, 27 insertions(+), 64 deletions(-)

diff --git a/migration/migration.c b/migration/migration.c
index 7fec57ad7f..22ab7199e4 100644
--- a/migration/migration.c
+++ b/migration/migration.c
@@ -217,6 +217,26 @@ MigrationIncomingState *migration_incoming_get_current(void)
     return current_incoming;
 }
 
+static void migration_file_release(QEMUFile **file)
+{
+    MigrationState *ms = migrate_get_current();
+    QEMUFile *tmp;
+
+    /*
+     * Reset the pointer before releasing it to avoid holding the lock
+     * for too long.
+     */
+    WITH_QEMU_LOCK_GUARD(&ms->qemu_file_lock) {
+        tmp = *file;
+        *file = NULL;
+    }
+
+    if (tmp) {
+        migration_ioc_unregister_yank_from_file(tmp);
+        qemu_fclose(tmp);
+    }
+}
+
 void migration_incoming_transport_cleanup(MigrationIncomingState *mis)
 {
     if (mis->socket_address_list) {
@@ -1155,8 +1175,6 @@ static void migrate_fd_cleanup(MigrationState *s)
     qemu_savevm_state_cleanup();
 
     if (s->to_dst_file) {
-        QEMUFile *tmp;
-
         trace_migrate_fd_cleanup();
         qemu_mutex_unlock_iothread();
         if (s->migration_thread_running) {
@@ -1166,16 +1184,7 @@ static void migrate_fd_cleanup(MigrationState *s)
         qemu_mutex_lock_iothread();
 
         multifd_save_cleanup();
-        qemu_mutex_lock(&s->qemu_file_lock);
-        tmp = s->to_dst_file;
-        s->to_dst_file = NULL;
-        qemu_mutex_unlock(&s->qemu_file_lock);
-        /*
-         * Close the file handle without the lock to make sure the
-         * critical section won't block for long.
-         */
-        migration_ioc_unregister_yank_from_file(tmp);
-        qemu_fclose(tmp);
+        migration_file_release(&s->to_dst_file);
     }
 
     /*
@@ -1815,38 +1824,6 @@ static int migrate_handle_rp_resume_ack(MigrationState *s, uint32_t value)
     return 0;
 }
 
-/*
- * Release ms->rp_state.from_dst_file (and postcopy_qemufile_src if
- * existed) in a safe way.
- */
-static void migration_release_dst_files(MigrationState *ms)
-{
-    QEMUFile *file;
-
-    WITH_QEMU_LOCK_GUARD(&ms->qemu_file_lock) {
-        /*
-         * Reset the from_dst_file pointer first before releasing it, as we
-         * can't block within lock section
-         */
-        file = ms->rp_state.from_dst_file;
-        ms->rp_state.from_dst_file = NULL;
-    }
-
-    /*
-     * Do the same to postcopy fast path socket too if there is.  No
-     * locking needed because this qemufile should only be managed by
-     * return path thread.
-     */
-    if (ms->postcopy_qemufile_src) {
-        migration_ioc_unregister_yank_from_file(ms->postcopy_qemufile_src);
-        qemu_file_shutdown(ms->postcopy_qemufile_src);
-        qemu_fclose(ms->postcopy_qemufile_src);
-        ms->postcopy_qemufile_src = NULL;
-    }
-
-    qemu_fclose(file);
-}
-
 /*
  * Handles messages sent on the return path towards the source VM
  *
@@ -2003,6 +1980,8 @@ static int open_return_path_on_source(MigrationState *ms)
         return -1;
     }
 
+    migration_ioc_register_yank(qemu_file_get_ioc(ms->rp_state.from_dst_file));
+
     trace_open_return_path_on_source();
 
     qemu_thread_create(&ms->rp_state.rp_thread, "return path",
@@ -2046,7 +2025,8 @@ static int await_return_path_close_on_source(MigrationState *ms)
     ret = ms->rp_state.error;
     ms->rp_state.error = false;
 
-    migration_release_dst_files(ms);
+    migration_file_release(&ms->rp_state.from_dst_file);
+    migration_file_release(&ms->postcopy_qemufile_src);
 
     trace_migration_return_path_end_after(ret);
     return ret;
@@ -2502,26 +2482,9 @@ static MigThrError postcopy_pause(MigrationState *s)
     assert(s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE);
 
     while (true) {
-        QEMUFile *file;
-
-        /*
-         * Current channel is possibly broken. Release it.  Note that this is
-         * guaranteed even without lock because to_dst_file should only be
-         * modified by the migration thread.  That also guarantees that the
-         * unregister of yank is safe too without the lock.  It should be safe
-         * even to be within the qemu_file_lock, but we didn't do that to avoid
-         * taking more mutex (yank_lock) within qemu_file_lock.  TL;DR: we make
-         * the qemu_file_lock critical section as small as possible.
-         */
+        /* Current channel is possibly broken. Release it. */
         assert(s->to_dst_file);
-        migration_ioc_unregister_yank_from_file(s->to_dst_file);
-        qemu_mutex_lock(&s->qemu_file_lock);
-        file = s->to_dst_file;
-        s->to_dst_file = NULL;
-        qemu_mutex_unlock(&s->qemu_file_lock);
-
-        qemu_file_shutdown(file);
-        qemu_fclose(file);
+        migration_file_release(&s->to_dst_file);
 
         /*
          * We're already pausing, so ignore any errors on the return
-- 
2.35.3

Re: [PATCH v4 2/8] migration: Fix possible races when shutting down the return path

[Peter Xu]

On Wed, Aug 16, 2023 at 11:25:04AM -0300, Fabiano Rosas wrote:
> We cannot call qemu_file_shutdown() on the return path file without
> taking the file lock. The return path thread could be running it's
> cleanup code and have just cleared the from_dst_file pointer.
> 
> Checking ms->to_dst_file for errors could also race with
> migrate_fd_cleanup() which clears the to_dst_file pointer.
> 
> Protect both accesses by taking the file lock.
> 
> This was caught by inspection, it should be rare, but the next patches
> will start calling this code from other places, so let's do the
> correct thing.
> 
> Signed-off-by: Fabiano Rosas <farosas@suse.de>

Reviewed-by: Peter Xu <peterx@redhat.com>

-- 
Peter Xu

Re: [PATCH v4 8/8] migration: Add a wrapper to cleanup migration files

[Peter Xu]

On Wed, Aug 16, 2023 at 11:25:10AM -0300, Fabiano Rosas wrote:
> @@ -2003,6 +1980,8 @@ static int open_return_path_on_source(MigrationState *ms)
>          return -1;
>      }
>  
> +    migration_ioc_register_yank(qemu_file_get_ioc(ms->rp_state.from_dst_file));

I think I didn't really get why it wasn't paired before yesterday.  My
fault.

Registering from_dst_file, afaict, will register two identical yank objects
because the ioc is the same.

Should we make migration_file_release() not handle the unregister of
yank(), but leave that to callers?  Then we keep the rule of only register
yank for each ioc once.

We can also have yet another wrapper migration_file_release_with_yank() to
unregister the yank before calling migration_file_release().

-- 
Peter Xu

Re: [PATCH v4 8/8] migration: Add a wrapper to cleanup migration files

[Peter Xu]

On Wed, Aug 16, 2023 at 03:35:24PM -0400, Peter Xu wrote:
> On Wed, Aug 16, 2023 at 03:47:26PM -0300, Fabiano Rosas wrote:
> > Peter Xu <peterx@redhat.com> writes:
> > 
> > > On Wed, Aug 16, 2023 at 11:25:10AM -0300, Fabiano Rosas wrote:
> > >> @@ -2003,6 +1980,8 @@ static int open_return_path_on_source(MigrationState *ms)
> > >>          return -1;
> > >>      }
> > >>  
> > >> +    migration_ioc_register_yank(qemu_file_get_ioc(ms->rp_state.from_dst_file));
> > >
> > > I think I didn't really get why it wasn't paired before yesterday.  My
> > > fault.
> > >
> > > Registering from_dst_file, afaict, will register two identical yank objects
> > > because the ioc is the same.
> > >
> > 
> > Why do we have two QEMUFiles for the same fd again?
> 
> Because qemufile has a "direction" (either read / write)?
> 
> > 
> > We're bound to crash at some point by trying to qemu_fclose() the two
> > QEMUFiles at the same time.
> 
> Even with each qemufile holding a reference on the ioc object?  I thought
> it won't crash, but if it will please point that out; or fix it would be
> even better.
> 
> > 
> > > Should we make migration_file_release() not handle the unregister of
> > > yank(), but leave that to callers?  Then we keep the rule of only register
> > > yank for each ioc once.
> > >
> > 
> > We need the unregister to be at migration_file_release() so that it
> > takes benefit of the locking while checking the file for NULL. If it
> > moves out then the caller will have to do locking as well. Which
> > defeats the purpose of the patch.
> > 
> > I don't understand why you moved the unregister out of channel_close in
> > commit 39675ffffb ("migration: Move the yank unregister of channel_close
> > out"). You called it a "hack" at the time, but looking at the current
> > situation, it seems like a reasonable thing to do: unregister the yank
> > when the ioc refcount drops to 1.
> > 
> > I would go even further and say that qemu_fclose should also avoid
> > calling qio_channel_close if the ioc refcnt is elevated.
> 
> I'd rather not; I still think it's a hack, always open to be corrected.
> 
> I think the problem is yank can register anything so it's separate from
> iochannels.  If one would like to have ioc close() automatically
> unregister, then one should also register yank transparently without the
> ioc user even aware of yank's existance.
> 
> Now the condition is the caller register yank itself, then I think the
> caller should unreg it.. not iochannel itself, silently.

I just noticed this is not really copying the list.. let me add the cc list
back, assuming it was just forgotten.

One more thing to mention is, now I kind of agree probably we should
register yank over each qemufile, as you raised the concern in the other
thread that otherwise qmp_yank() won't set error for the qemufile, which
seems to be unexpected.

-- 
Peter Xu

Re: [PATCH v4 8/8] migration: Add a wrapper to cleanup migration files

[Fabiano Rosas]

Peter Xu <peterx@redhat.com> writes:

> On Wed, Aug 16, 2023 at 03:35:24PM -0400, Peter Xu wrote:
>> On Wed, Aug 16, 2023 at 03:47:26PM -0300, Fabiano Rosas wrote:
>> > Peter Xu <peterx@redhat.com> writes:
>> > 
>> > > On Wed, Aug 16, 2023 at 11:25:10AM -0300, Fabiano Rosas wrote:
>> > >> @@ -2003,6 +1980,8 @@ static int open_return_path_on_source(MigrationState *ms)
>> > >>          return -1;
>> > >>      }
>> > >>  
>> > >> +    migration_ioc_register_yank(qemu_file_get_ioc(ms->rp_state.from_dst_file));
>> > >
>> > > I think I didn't really get why it wasn't paired before yesterday.  My
>> > > fault.
>> > >
>> > > Registering from_dst_file, afaict, will register two identical yank objects
>> > > because the ioc is the same.
>> > >
>> > 
>> > Why do we have two QEMUFiles for the same fd again?
>> 
>> Because qemufile has a "direction" (either read / write)?
>> 
>> > 
>> > We're bound to crash at some point by trying to qemu_fclose() the two
>> > QEMUFiles at the same time.
>> 
>> Even with each qemufile holding a reference on the ioc object?  I thought
>> it won't crash, but if it will please point that out; or fix it would be
>> even better.

You're right, it wouldn't crash. But it's still the same ioc object. If
qio_channel_close() is called twice, then we could potentially close the
fd twice. Which would either error out or close a reused fd. The window
is small though, so probably unlikely to ever happen.

>> 
>> > 
>> > > Should we make migration_file_release() not handle the unregister of
>> > > yank(), but leave that to callers?  Then we keep the rule of only register
>> > > yank for each ioc once.
>> > >
>> > 
>> > We need the unregister to be at migration_file_release() so that it
>> > takes benefit of the locking while checking the file for NULL. If it
>> > moves out then the caller will have to do locking as well. Which
>> > defeats the purpose of the patch.
>> > 
>> > I don't understand why you moved the unregister out of channel_close in
>> > commit 39675ffffb ("migration: Move the yank unregister of channel_close
>> > out"). You called it a "hack" at the time, but looking at the current
>> > situation, it seems like a reasonable thing to do: unregister the yank
>> > when the ioc refcount drops to 1.
>> > 
>> > I would go even further and say that qemu_fclose should also avoid
>> > calling qio_channel_close if the ioc refcnt is elevated.
>> 
>> I'd rather not; I still think it's a hack, always open to be corrected.

It's hard to figure out what you mean by hack at times. Even more when
reading a years-old commit message.

>> 
>> I think the problem is yank can register anything so it's separate from
>> iochannels.  If one would like to have ioc close() automatically
>> unregister, then one should also register yank transparently without the
>> ioc user even aware of yank's existance.

Ok, fair point.

>> 
>> Now the condition is the caller register yank itself, then I think the
>> caller should unreg it.. not iochannel itself, silently.

I think the issue is that we're linking the yank with the QEMUFile for
no reason. The migration_yank_iochannel() performs a
qio_channel_shutdown() which is an operation on the fd. The QEMUFile
just happens to hold a pointer to the ioc.

>
> I just noticed this is not really copying the list.. let me add the cc list
> back, assuming it was just forgotten.

I'm sorry, I hit the wrong key while replying.

> One more thing to mention is, now I kind of agree probably we should
> register yank over each qemufile, as you raised the concern in the other
> thread that otherwise qmp_yank() won't set error for the qemufile, which
> seems to be unexpected.

I haven't made up my mind yet, but I think I'd rather stop setting that
error instead of doing it from other places. A shutdown() is mostly a
benign operation intended to end the connection. The fact that we use it
in some cases to kick the thread out of a possible hang doesn't seem
compelling enough to set -EIO.

Of course we currently have no other way to indicate that the file was
shutdown, so the -EIO will have to stay and that's a discussion for
another day.

Re: [PATCH v4 8/8] migration: Add a wrapper to cleanup migration files

[Peter Xu]

On Wed, Aug 16, 2023 at 06:20:58PM -0300, Fabiano Rosas wrote:
> > One more thing to mention is, now I kind of agree probably we should
> > register yank over each qemufile, as you raised the concern in the other
> > thread that otherwise qmp_yank() won't set error for the qemufile, which
> > seems to be unexpected.
> 
> I haven't made up my mind yet, but I think I'd rather stop setting that
> error instead of doing it from other places. A shutdown() is mostly a
> benign operation intended to end the connection. The fact that we use it
> in some cases to kick the thread out of a possible hang doesn't seem
> compelling enough to set -EIO.
> 
> Of course we currently have no other way to indicate that the file was
> shutdown, so the -EIO will have to stay and that's a discussion for
> another day.

Yes, if we can avoid setting -EIO at all when shutdown that'll also be
good, maybe making more sense.  Thanks,

-- 
Peter Xu