CI run: https://gitlab.com/peterx/qemu/-/pipelines/2855601505
Currently, upstream is broken with migration with qemu-7.2 machine types,
simplest reproducer:
$ QTEST_QEMU_MACHINE_TYPE=pc-q35-7.2 \
QTEST_QEMU_BINARY=./qemu-system-x86_64 \
./tests/qtest/migration-test --full \
-r /x86_64/migration/multifd+postcopy/preempt/plain
This series should fix it. Note that this solution will not work for TLS,
and due to the fact that TLS is not peekable (which the design of this
series currently rely on), TLS is not fixable. But as long as the user
doesn't enable all four features (multifd, postcopy, preempt, tls) then it
will still work. I have much more info in patch 6, feel free to read.
Patch layout:
Patch 1: Added one more test for multifd+postcopy recovery non-tls
Patch 2-5: Something I found either buggy or good to have cleanups
Patch 6: Fix the problem on incoming side
Patch 7: Fix the problem on outgoing side
Tests I've done:
- The qtest, concurrent execution of a lot of "migration-test --full",
which I normally use to catch race conditions. Nothing pops up.
- Test migration with all three features enabled (multifd, postcopy-ram,
postcopy-preempt) on different machine types and binaries combinations,
|----------+---------+-------------+--------|
| Machine | Source | Destination | Result |
|----------+---------+-------------+--------|
| q35-11.2 | master | master | PASS | --+
| | master | patched | PASS | |-----> 11.2 machine not
affected
| | patched | master | PASS | |
| | patched | patched | PASS | --+
|----------+---------+-------------+--------|
| q35-7.2 | master | master | FAIL | --------> current breakage
| | master | patched | PASS | --+
| | patched | master | PASS | |-----> patch any side should
fix
| | patched | patched | PASS | --+
|----------+---------+-------------+--------|
- Manual test of qemu-7.2 binary <-> patched new binary with either (1)
multifd migration, or (2) postcopy + preempt migration. NOTE: qemu-7.2
doesn't support enabling all.
Comments welcomed, thanks.
Peter Xu (7):
tests/migration-test: Add multifd+postcopy test for preempt recovery
migration: Add two more trace points for incoming channel mgmt
migration: Fix socket listen backlog
migration: Split migration_channel_process_incoming()
migration: Dump type of channel in set_incoming_channel tracepoint
migration: Fix incoming channel race with multifd and preempt_pre_7_2
migration: Postpone the outgoing preempt channel for preempt_pre_7_2
migration/channel.h | 6 +-
migration/migration.h | 35 ++
migration/channel.c | 453 +++++++++++++++++++++----
migration/migration.c | 58 +++-
migration/socket.c | 25 +-
tests/qtest/migration/postcopy-tests.c | 11 +
migration/trace-events | 8 +-
7 files changed, 515 insertions(+), 81 deletions(-)
--
2.55.0
