1. Add a dedicated handler for MigrationOps::ram_save_target_page in
multifd live migration.
2. Refactor ram_save_target_page_legacy so that the legacy and multifd
handlers don't have internal functions calling into each other.
Signed-off-by: Hao Xiang <hao.xiang@bytedance.com>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Message-Id: <20240226195654.934709-4-hao.xiang@bytedance.com>
Link: https://lore.kernel.org/r/20240311180015.3359271-6-hao.xiang@linux.dev
Signed-off-by: Peter Xu <peterx@redhat.com>
1. Add zero_pages field in MultiFDPacket_t.
2. Implements the zero page detection and handling on the multifd
threads for non-compression, zlib and zstd compression backends.
3. Added a new value 'multifd' in ZeroPageDetection enumeration.
4. Adds zero page counters and updates multifd send/receive tracing
format to track the newly added counters.
Signed-off-by: Hao Xiang <hao.xiang@bytedance.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240311180015.3359271-5-hao.xiang@linux.dev
Signed-off-by: Peter Xu <peterx@redhat.com>
This new parameter controls where the zero page checking is running.
1. If this parameter is set to 'legacy', zero page checking is
done in the migration main thread.
2. If this parameter is set to 'none', zero page checking is disabled.
Signed-off-by: Hao Xiang <hao.xiang@bytedance.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Link: https://lore.kernel.org/r/20240311180015.3359271-4-hao.xiang@linux.dev
Signed-off-by: Peter Xu <peterx@redhat.com>
We currently only need to clear the mapped-ram file bitmap from the
migration thread during save_zero_page.
We're about to add support for zero page detection on the multifd
thread, so allow ramblock_set_file_bmap_atomic() to also clear the
bits.
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240311180015.3359271-3-hao.xiang@linux.dev
Signed-off-by: Peter Xu <peterx@redhat.com>
Currently, it's an error to have no data pages in the multifd file
migration because zero page detection is done in the migration thread
and zero pages don't reach multifd. This is enforced with the
pages->num assert.
We're about to add zero page detection on the multifd thread. Fix the
file_write_ramblock_iov() to stop considering p->iovs_num=0 an error.
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240311180015.3359271-2-hao.xiang@linux.dev
Signed-off-by: Peter Xu <peterx@redhat.com>
Delete the MigrationState parameter from migration_is_setup_or_active
and move it to the public API in misc.h.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Link: https://lore.kernel.org/r/1710179338-294359-3-git-send-email-steven.sistare@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
A small number of migration options are accessed by migration clients,
but to see them clients must include all of options.h, which is mostly
for migration core code. migrate_mode() in particular will be needed by
multiple clients.
Refactor the option declarations so clients can see the necessary few via
misc.h, which already exports a portion of the client API.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Link: https://lore.kernel.org/r/1710179319-294320-1-git-send-email-steven.sistare@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
In file_write_ramblock_iov(), "offset" is "uintptr_t" and not
"ram_addr_t". While usually they are both equivalent, this is not the
case with CONFIG_XEN_BACKEND.
Use the right format. This will fix build on 32-bit.
Fixes: f427d90b98 ("migration/multifd: Support outgoing mapped-ram stream format")
Signed-off-by: Anthony PERARD <anthony.perard@citrix.com>
Link: https://lore.kernel.org/r/20240311123439.16844-1-anthony.perard@citrix.com
Signed-off-by: Peter Xu <peterx@redhat.com>
In commit 3fa9642ff7 change was made to convert the RDMA backend to
accept MigrateAddress struct. However, the assignment of "host" leads
to data corruption on the target host and the failure of migration.
isock->host = rdma->host;
By allocating the memory explicitly for it with g_strdup_printf(), the
issue is fixed and the migration doesn't fail any more.
Fixes: 3fa9642ff7 ("migration: convert rdma backend to accept MigrateAddress")
Cc: qemu-stable <qemu-stable@nongnu.org>
Cc: Li Zhijian <lizhijian@fujitsu.com>
Link: https://lore.kernel.org/r/CAHEcVy4L_D6tuhJ8h=xLR4WaPaprJE3nnxZAEyUnoTrxQ6CF5w@mail.gmail.com
Signed-off-by: Yu Zhang <yu.zhang@ionos.com>
[peterx: use g_strdup() instead of g_strdup_printf(), per Zhijian]
Signed-off-by: Peter Xu <peterx@redhat.com>
Commit bc38feddeb ("io: fsync before closing a file channel") added a
fsync/fdatasync at the closing point of the QIOChannelFile to ensure
integrity of the migration stream in case of QEMU crash.
The decision to do the sync at qio_channel_close() was not the best
since that function runs in the main thread and the fsync can cause
QEMU to hang for several minutes, depending on the migration size and
disk speed.
To fix the hang, remove the fsync from qio_channel_file_close().
At this moment, the migration code is the only user of the fsync and
we're taking the tradeoff of not having a sync at all, leaving the
responsibility to the upper layers.
Fixes: bc38feddeb ("io: fsync before closing a file channel")
Reviewed-by: "Daniel P. Berrangé" <berrange@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240305195629.9922-1-farosas@suse.de
Link: https://lore.kernel.org/r/20240305174332.2553-1-farosas@suse.de
[peterx: add more comment to the qio_channel_close()]
Signed-off-by: Peter Xu <peterx@redhat.com>
When commit bd2270608f ("migration/ram.c: add a notifier chain for
precopy") added PRECOPY_NOTIFY_SETUP notifiers at the end of
qemu_savevm_state_setup(), it didn't take into account a possible
error in the loop calling vmstate_save() or .save_setup() handlers.
Check ret value before calling the notifiers.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Cédric Le Goater <clg@redhat.com>
Link: https://lore.kernel.org/r/20240304122844.1888308-10-clg@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
This will help detect issues regarding I/O channels usage.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Cédric Le Goater <clg@redhat.com>
Link: https://lore.kernel.org/r/20240304122844.1888308-7-clg@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
If a migration stream is broken, the address and flag reading can return
zero. Thus, an irrelevant flag error will be returned instead of EIO.
It can be fixed by additional check after the reading.
Signed-off-by: Maksim Davydov <davydov-max@yandex-team.ru>
Link: https://lore.kernel.org/r/20240304144203.158477-1-davydov-max@yandex-team.ru
Signed-off-by: Peter Xu <peterx@redhat.com>
Commit 90697be889 ("live migration: Serialize vmstate saving in stage
2") introduced device serialization in qemu_savevm_state_iterate(). The
rationale behind it was to first complete migration of slower changing
block devices and only then migrate the RAM, to avoid sending fast
changing RAM pages over and over.
This commit was added a long time ago, and while it was useful back
then, it is not the case anymore:
1. Block migration is deprecated, see commit 66db46ca83 ("migration:
Deprecate block migration").
2. Today there are other iterative devices besides RAM and block, such
as VFIO, which are registered for migration after RAM. With current
serialization behavior, a fast changing device can block other
devices from sending their data, which may prevent migration from
converging in some cases.
The issue described in item 2 was observed in several VFIO migration
scenarios with switchover-ack capability enabled, where some workload on
the VM prevented RAM from ever reaching a hard zero, thus blocking VFIO
initial pre-copy data from being sent. Hence, destination could not ack
switchover and migration could not converge.
Fix that by not serializing iterative devices in
qemu_savevm_state_iterate().
Note that this still doesn't fully prevent device starvation. As
correctly pointed out by Peter [1], a fast changing device might
constantly consume all allocated bandwidth and block the following
devices. However, this scenario is more likely to happen only if
max-bandwidth is low.
[1] https://lore.kernel.org/qemu-devel/Zd6iw9dBhW6wKNxx@x1n/
Signed-off-by: Avihai Horon <avihaih@nvidia.com>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240304105339.20713-2-avihaih@nvidia.com
Signed-off-by: Peter Xu <peterx@redhat.com>
- Bryan's fix on multifd compression level API
- Fabiano's mapped-ram series (base + multifd only)
- Steve's amend on cpr document in qapi/
-----BEGIN PGP SIGNATURE-----
iIgEABYKADAWIQS5GE3CDMRX2s990ak7X8zN86vXBgUCZeUjKhIccGV0ZXJ4QHJl
ZGhhdC5jb20ACgkQO1/MzfOr1wbv5QD/ZexBUsmZA5qyxgGvZ2yvlUBEGNOvtmKY
kRdiYPU7khMA/0N43rn4LcqKCoq4+T+EAnYizGjIyhH/7BRUyn4DUxgO
=AeEn
-----END PGP SIGNATURE-----
Merge tag 'migration-next-pull-request' of https://gitlab.com/peterx/qemu into staging
Migartion pull request for 20240304
- Bryan's fix on multifd compression level API
- Fabiano's mapped-ram series (base + multifd only)
- Steve's amend on cpr document in qapi/
# -----BEGIN PGP SIGNATURE-----
#
# iIgEABYKADAWIQS5GE3CDMRX2s990ak7X8zN86vXBgUCZeUjKhIccGV0ZXJ4QHJl
# ZGhhdC5jb20ACgkQO1/MzfOr1wbv5QD/ZexBUsmZA5qyxgGvZ2yvlUBEGNOvtmKY
# kRdiYPU7khMA/0N43rn4LcqKCoq4+T+EAnYizGjIyhH/7BRUyn4DUxgO
# =AeEn
# -----END PGP SIGNATURE-----
# gpg: Signature made Mon 04 Mar 2024 01:26:02 GMT
# gpg: using EDDSA key B9184DC20CC457DACF7DD1A93B5FCCCDF3ABD706
# gpg: issuer "peterx@redhat.com"
# gpg: Good signature from "Peter Xu <xzpeter@gmail.com>" [marginal]
# gpg: aka "Peter Xu <peterx@redhat.com>" [marginal]
# gpg: WARNING: This key is not certified with sufficiently trusted signatures!
# gpg: It is not certain that the signature belongs to the owner.
# Primary key fingerprint: B918 4DC2 0CC4 57DA CF7D D1A9 3B5F CCCD F3AB D706
* tag 'migration-next-pull-request' of https://gitlab.com/peterx/qemu: (27 commits)
migration/multifd: Document two places for mapped-ram
tests/qtest/migration: Add a multifd + mapped-ram migration test
migration/multifd: Add mapped-ram support to fd: URI
migration/multifd: Support incoming mapped-ram stream format
migration/multifd: Support outgoing mapped-ram stream format
migration/multifd: Prepare multifd sync for mapped-ram migration
migration/multifd: Add incoming QIOChannelFile support
migration/multifd: Add outgoing QIOChannelFile support
migration/multifd: Add a wrapper for channels_created
migration/multifd: Allow receiving pages without packets
migration/multifd: Allow multifd without packets
migration/multifd: Decouple recv method from pages
migration/multifd: Rename MultiFDSend|RecvParams::data to compress_data
tests/qtest/migration: Add tests for mapped-ram file-based migration
migration/ram: Add incoming 'mapped-ram' migration
migration/ram: Add outgoing 'mapped-ram' migration
migration: Add mapped-ram URI compatibility check
migration/ram: Introduce 'mapped-ram' migration capability
migration/qemu-file: add utility methods for working with seekable channels
io: fsync before closing a file channel
...
Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
# Conflicts:
# migration/ram.c
Simplify the exec migration code by using list utility functions.
As a side effect, this also fixes a minor memory leak. On function return,
"g_auto(GStrv) argv" frees argv and each element, which is wrong, because
the function does not own the individual elements. To compensate, the code
uses g_steal_pointer which NULLs argv and prevents the destructor from
running, but argv is leaked.
Fixes: cbab4face5 ("migration: convert exec backend ...")
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Message-ID: <20240227153321.467343-4-armbru@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Add two documentations for mapped-ram migration on two spots that may not
be extremely clear.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240301091524.39900-1-peterx@redhat.com
Cc: Prasad Pandit <ppandit@redhat.com>
[peterx: fix two English errors per Prasad]
Signed-off-by: Peter Xu <peterx@redhat.com>
If we receive a file descriptor that points to a regular file, there's
nothing stopping us from doing multifd migration with mapped-ram to
that file.
Enable the fd: URI to work with multifd + mapped-ram.
Note that the fds passed into multifd are duplicated because we want
to avoid cross-thread effects when doing cleanup (i.e. close(fd)). The
original fd doesn't need to be duplicated because monitor_get_fd()
transfers ownership to the caller.
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Peter Xu <peterx@redhat.com>
Link: https://lore.kernel.org/r/20240229153017.2221-23-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
For the incoming mapped-ram migration we need to read the ramblock
headers, get the pages bitmap and send the host address of each
non-zero page to the multifd channel thread for writing.
Usage on HMP is:
(qemu) migrate_set_capability multifd on
(qemu) migrate_set_capability mapped-ram on
(qemu) migrate_incoming file:migfile
(the ram.h include needs to move because we've been previously relying
on it being included from migration.c. Now file.h will start including
multifd.h before migration.o is processed)
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-22-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
The new mapped-ram stream format uses a file transport and puts ram
pages in the migration file at their respective offsets and can be
done in parallel by using the pwritev system call which takes iovecs
and an offset.
Add support to enabling the new format along with multifd to make use
of the threading and page handling already in place.
This requires multifd to stop sending headers and leaving the stream
format to the mapped-ram code. When it comes time to write the data, we
need to call a version of qio_channel_write that can take an offset.
Usage on HMP is:
(qemu) stop
(qemu) migrate_set_capability multifd on
(qemu) migrate_set_capability mapped-ram on
(qemu) migrate_set_parameter max-bandwidth 0
(qemu) migrate_set_parameter multifd-channels 8
(qemu) migrate file:migfile
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-21-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
The mapped-ram migration can be performed live or non-live, but it is
always asynchronous, i.e. the source machine and the destination
machine are not migrating at the same time. We only need some pieces
of the multifd sync operations.
multifd_send_sync_main()
------------------------
Issued by the ram migration code on the migration thread, causes the
multifd send channels to synchronize with the migration thread and
makes the sending side emit a packet with the MULTIFD_FLUSH flag.
With mapped-ram we want to maintain the sync on the sending side
because that provides ordering between the rounds of dirty pages when
migrating live.
MULTIFD_FLUSH
-------------
On the receiving side, the presence of the MULTIFD_FLUSH flag on a
packet causes the receiving channels to start synchronizing with the
main thread.
We're not using packets with mapped-ram, so there's no MULTIFD_FLUSH
flag and therefore no channel sync on the receiving side.
multifd_recv_sync_main()
------------------------
Issued by the migration thread when the ram migration flag
RAM_SAVE_FLAG_MULTIFD_FLUSH is received, causes the migration thread
on the receiving side to start synchronizing with the recv
channels. Due to compatibility, this is also issued when
RAM_SAVE_FLAG_EOS is received.
For mapped-ram we only need to synchronize the channels at the end of
migration to avoid doing cleanup before the channels have finished
their IO.
Make sure the multifd syncs are only issued at the appropriate times.
Note that due to pre-existing backward compatibility issues, we have
the multifd_flush_after_each_section property that can cause a sync to
happen at EOS. Since the EOS flag is needed on the stream, allow
mapped-ram to just ignore it.
Also emit an error if any other unexpected flags are found on the
stream.
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Peter Xu <peterx@redhat.com>
Link: https://lore.kernel.org/r/20240229153017.2221-20-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
On the receiving side we don't need to differentiate between main
channel and threads, so whichever channel is defined first gets to be
the main one. And since there are no packets, use the atomic channel
count to index into the params array.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-19-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Allow multifd to open file-backed channels. This will be used when
enabling the mapped-ram migration stream format which expects a
seekable transport.
The QIOChannel read and write methods will use the preadv/pwritev
versions which don't update the file offset at each call so we can
reuse the fd without re-opening for every channel.
Contrary to the socket migration, the file migration doesn't need an
asynchronous channel creation process, so expose
multifd_channel_connect() and call it directly.
Note that this is just setup code and multifd cannot yet make use of
the file channels.
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Peter Xu <peterx@redhat.com>
Link: https://lore.kernel.org/r/20240229153017.2221-18-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
We'll need to access multifd_send_state->channels_created from outside
multifd.c, so introduce a helper for that.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-17-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Currently multifd does not need to have knowledge of pages on the
receiving side because all the information needed is within the
packets that come in the stream.
We're about to add support to mapped-ram migration, which cannot use
packets because it expects the ramblock section in the migration file
to contain only the guest pages data.
Add a data structure to transfer pages between the ram migration code
and the multifd receiving threads.
We don't want to reuse MultiFDPages_t for two reasons:
a) multifd threads don't really need to know about the data they're
receiving.
b) the receiving side has to be stopped to load the pages, which means
we can experiment with larger granularities than page size when
transferring data.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-16-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
For the upcoming support to the new 'mapped-ram' migration stream
format, we cannot use multifd packets because each write into the
ramblock section in the migration file is expected to contain only the
guest pages. They are written at their respective offsets relative to
the ramblock section header.
There is no space for the packet information and the expected gains
from the new approach come partly from being able to write the pages
sequentially without extraneous data in between.
The new format also simply doesn't need the packets and all necessary
information can be taken from the standard migration headers with some
(future) changes to multifd code.
Use the presence of the mapped-ram capability to decide whether to
send packets.
This only moves code under multifd_use_packets(), it has no effect for
now as mapped-ram cannot yet be enabled with multifd.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-15-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Next patches will abstract the type of data being received by the
channels, so do some cleanup now to remove references to pages and
dependency on 'normal_num'.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-14-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Use a more specific name for the compression data so we can use the
generic for the multifd core code.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-13-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Add the necessary code to parse the format changes for the
'mapped-ram' capability.
One of the more notable changes in behavior is that in the
'mapped-ram' case ram pages are restored in one go rather than
constantly looping through the migration stream.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-11-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Implement the outgoing migration side for the 'mapped-ram' capability.
A bitmap is introduced to track which pages have been written in the
migration file. Pages are written at a fixed location for every
ramblock. Zero pages are ignored as they'd be zero in the destination
migration as well.
The migration stream is altered to put the dirty pages for a ramblock
after its header instead of having a sequential stream of pages that
follow the ramblock headers.
Without mapped-ram (current): With mapped-ram (new):
--------------------- --------------------------------
| ramblock 1 header | | ramblock 1 header |
--------------------- --------------------------------
| ramblock 2 header | | ramblock 1 mapped-ram header |
--------------------- --------------------------------
| ... | | padding to next 1MB boundary |
--------------------- | ... |
| ramblock n header | --------------------------------
--------------------- | ramblock 1 pages |
| RAM_SAVE_FLAG_EOS | | ... |
--------------------- --------------------------------
| stream of pages | | ramblock 2 header |
| (iter 1) | --------------------------------
| ... | | ramblock 2 mapped-ram header |
--------------------- --------------------------------
| RAM_SAVE_FLAG_EOS | | padding to next 1MB boundary |
--------------------- | ... |
| stream of pages | --------------------------------
| (iter 2) | | ramblock 2 pages |
| ... | | ... |
--------------------- --------------------------------
| ... | | ... |
--------------------- --------------------------------
| RAM_SAVE_FLAG_EOS |
--------------------------------
| ... |
--------------------------------
where:
- ramblock header: the generic information for a ramblock, such as
idstr, used_len, etc.
- ramblock mapped-ram header: the new information added by this
feature: bitmap of pages written, bitmap size and offset of pages
in the migration file.
Signed-off-by: Nikolay Borisov <nborisov@suse.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-10-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
The mapped-ram migration format needs a channel that supports seeking
to be able to write each page to an arbitrary offset in the migration
stream.
Reviewed-by: "Daniel P. Berrangé" <berrange@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-9-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Add a new migration capability 'mapped-ram'.
The core of the feature is to ensure that RAM pages are mapped
directly to offsets in the resulting migration file instead of being
streamed at arbitrary points.
The reasons why we'd want such behavior are:
- The resulting file will have a bounded size, since pages which are
dirtied multiple times will always go to a fixed location in the
file, rather than constantly being added to a sequential
stream. This eliminates cases where a VM with, say, 1G of RAM can
result in a migration file that's 10s of GBs, provided that the
workload constantly redirties memory.
- It paves the way to implement O_DIRECT-enabled save/restore of the
migration stream as the pages are ensured to be written at aligned
offsets.
- It allows the usage of multifd so we can write RAM pages to the
migration file in parallel.
For now, enabling the capability has no effect. The next couple of
patches implement the core functionality.
Acked-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-8-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Add utility methods that will be needed when implementing 'mapped-ram'
migration capability.
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: "Daniel P. Berrangé" <berrange@redhat.com>
Link: https://lore.kernel.org/r/20240229153017.2221-7-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Some minor cleanups and documentation for multifd_recv_sync_main.
Use thread_count as done in other parts of the code. Remove p->id from
the multifd_recv_state sync, since that is global and not tied to a
channel. Add documentation for the sync steps.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240229153017.2221-2-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Some glue code was missing, so that using `qmp_migrate_set_parameters`
to set `multifd-zstd-level` or `multifd-zlib-level` did not work. This
commit adds the glue code to fix that.
Signed-off-by: Bryan Zhang <bryan.zhang@bytedance.com>
Link: https://lore.kernel.org/r/20240301035901.4006936-2-bryan.zhang@bytedance.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Replace with the maximum of the real host page size
and the target page size. This is an exact replacement.
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Ilya Leoshkevich <iii@linux.ibm.com>
Reviewed-by: Pierrick Bouvier <pierrick.bouvier@linaro.org>
Acked-by: Helge Deller <deller@gmx.de>
Message-Id: <20240102015808.132373-12-richard.henderson@linaro.org>
close_return_path_on_source() retrieves the migration error from the
the QEMUFile '->to_dst_file' to know if a shutdown is required. This
shutdown is required to exit the return-path thread.
Avoid relying on '->to_dst_file' and use migrate_has_error() instead.
(using to_dst_file is a heuristic to infer whether
rp_state.from_dst_file might be stuck on a recvmsg(). Using a generic
method for detecting errors is more reliable. We also want to reduce
dependency on QEMUFile::last_error)
Suggested-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Cédric Le Goater <clg@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
[added some words about the motivation for this patch]
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240226203122.22894-3-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
The return path thread might hang at a blocking system call. Before
joining the thread we might need to issue a shutdown() on the socket
file descriptor to release it. To determine whether the shutdown() is
necessary we look at the QEMUFile error.
Make sure we only clean up the QEMUFile after the return path has been
waited for.
This fixes a hang when qemu_savevm_state_setup() produced an error
that was detected by migration_detect_error(). That skips
migration_completion() so close_return_path_on_source() would get
stuck waiting for the RP thread to terminate.
Reported-by: Cédric Le Goater <clg@redhat.com>
Tested-by: Cédric Le Goater <clg@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240226203122.22894-2-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
The QMP command query_migrate might see incorrect throughput numbers
if it runs after we've set the migration completion status but before
migration_calculate_complete() has updated s->total_time and s->mbps.
The migration status would show COMPLETED, but the throughput value
would be the one from the last iteration and not the one from the
whole migration. This will usually be a larger value due to the time
period being smaller (one iteration).
Move migration_calculate_complete() earlier so that the status
MIGRATION_STATUS_COMPLETED is only emitted after the final counters
update. Keep everything under the BQL so the QMP thread sees the
updates as atomic.
Rename migration_calculate_complete to migration_completion_end to
reflect its new purpose of also updating s->state.
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240226143335.14282-1-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
When migration for cpr is initiated, stop the vm and set state
RUN_STATE_FINISH_MIGRATE before ram is saved. This eliminates the
possibility of ram and device state being out of sync, and guarantees
that a guest in the suspended state remains suspended, because qmp_cont
rejects a cont command in the RUN_STATE_FINISH_MIGRATE state.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Link: https://lore.kernel.org/r/1708622920-68779-11-git-send-email-steven.sistare@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Check the status returned by migration notifiers for event type
MIG_EVENT_PRECOPY_SETUP, and report errors. None of the notifiers
return an error status at this time.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Link: https://lore.kernel.org/r/1708622920-68779-10-git-send-email-steven.sistare@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Move common code for the error path in migrate_fd_connect to a shared
fail label. No functional change.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/1708622920-68779-9-git-send-email-steven.sistare@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Keep a separate list of migration notifiers for each migration mode.
Suggested-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/1708622920-68779-8-git-send-email-steven.sistare@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Define MigrationNotifyFunc to improve type safety and simplify migration
notifiers.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/1708622920-68779-7-git-send-email-steven.sistare@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Passing MigrationState to notifiers is unsound because they could access
unstable migration state internals or even modify the state. Instead, pass
the minimal info needed in a new MigrationEvent struct, which could be
extended in the future if needed.
Suggested-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/1708622920-68779-5-git-send-email-steven.sistare@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Change all migration notifiers to type NotifierWithReturn, so notifiers
can return an error status in a future patch. For now, pass NULL for the
notifier error parameter, and do not check the return value.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/1708622920-68779-4-git-send-email-steven.sistare@oracle.com
[peterx: dropped unexpected update to roms/seabios-hppa]
Signed-off-by: Peter Xu <peterx@redhat.com>
Remove the error object from opaque data passed to notifiers.
Use the new error parameter passed to the notifier instead.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/1708622920-68779-3-git-send-email-steven.sistare@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Pass an error object as the third parameter to "notifier with return"
notifiers, so clients no longer need to bundle an error object in the
opaque data. The new parameter is used in a later patch.
Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Link: https://lore.kernel.org/r/1708622920-68779-2-git-send-email-steven.sistare@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Both socket_send_channel_destroy() and multifd_send_channel_destroy() are
unnecessary wrappers to destroy an IOC, as the only thing to do is to
release the final IOC reference. We have plenty of code that destroys an
IOC using direct unref() already; keep that style.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240222095301.171137-6-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
outgoing_args is a global cache of socket address to be reused in multifd.
Freeing the cache in per-channel destructor is more or less a hack. Move
it to multifd_send_cleanup_state() so it only get checked once. Use a
small helper to do so because it's internal of socket.c.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240222095301.171137-5-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
With a clear definition of p->c protocol, where we only set it up if the
channel is fully established (TLS or non-TLS), registered_yank boolean will
have equal meaning of "p->c != NULL".
Drop registered_yank by checking p->c instead.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240222095301.171137-3-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Commit a1af605bd5 ("migration/multifd: fix hangup with TLS-Multifd due to
blocking handshake") introduced a thread for TLS channels, which will
resolve the issue on blocking the main thread. However in the same commit
p->c is slightly abused just to be able to pass over the pointer "p" into
the thread.
That's the major reason we'll need to conditionally free the io channel in
the fault paths.
To clean it up, using a separate structure to pass over both "p" and "tioc"
in the tls handshake thread. Then we can make it a rule that p->c will
never be set until the channel is completely setup. With that, we can drop
the tricky conditional unref of the io channel in the error path.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240222095301.171137-2-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Now that multifd_recv_terminate_threads() is called only once, release
the recv side sem_sync earlier like we do for the send side.
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Peter Xu <peterx@redhat.com>
Link: https://lore.kernel.org/r/20240220224138.24759-6-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Like we did on the sending side, replace the p->quit per-channel flag
with a global atomic 'exiting' flag.
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Peter Xu <peterx@redhat.com>
Link: https://lore.kernel.org/r/20240220224138.24759-5-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
It is possible that one of the multifd channels fails to be created at
multifd_new_send_channel_async() while the rest of the channel
creation tasks are still in flight.
This could lead to multifd_save_cleanup() executing the
qemu_thread_join() loop too early and not waiting for the threads
which haven't been created yet, leading to the freeing of resources
that the newly created threads will try to access and crash.
Add a synchronization point after which there will be no attempts at
thread creation and therefore calling multifd_save_cleanup() past that
point will ensure it properly waits for the threads.
A note about performance: Prior to this patch, if a channel took too
long to be established, other channels could finish connecting first
and already start taking load. Now we're bounded by the
slowest-connecting channel.
Reported-by: Avihai Horon <avihaih@nvidia.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240206215118.6171-7-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
During multifd channel creation (multifd_send_new_channel_async) when
TLS is enabled, the multifd_channel_connect function is called twice,
once to create the TLS handshake thread and another time after the
asynchrounous TLS handshake has finished.
This creates a slightly confusing call stack where
multifd_channel_connect() is called more times than the number of
channels. It also splits error handling between the two callers of
multifd_channel_connect() causing some code duplication. Lastly, it
gets in the way of having a single point to determine whether all
channel creation tasks have been initiated.
Refactor the code to move the reentrancy one level up at the
multifd_new_send_channel_async() level, de-duplicating the error
handling and allowing for the next patch to introduce a
synchronization point common to all the multifd channel creation,
regardless of TLS.
Note that the previous code would never fail once p->c had been set.
This patch changes this assumption, which affects refcounting, so add
comments around object_unref to explain the situation.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240206215118.6171-6-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
We currently have an unfavorable situation around multifd channels
creation and the migration thread execution.
We create the multifd channels with qio_channel_socket_connect_async
-> qio_task_run_in_thread, but only connect them at the
multifd_new_send_channel_async callback, called from
qio_task_complete, which is registered as a glib event.
So at multifd_send_setup() we create the channels, but they will only
be actually usable after the whole multifd_send_setup() calling stack
returns back to the main loop. Which means that the migration thread
is already up and running without any possibility for the multifd
channels to be ready on time.
We currently rely on the channels-ready semaphore blocking
multifd_send_sync_main() until channels start to come up and release
it. However there have been bugs recently found when a channel's
creation fails and multifd_send_cleanup() is allowed to run while
other channels are still being created.
Let's start to organize this situation by moving the
multifd_send_setup() call into the migration thread. That way we
unblock the main-loop to dispatch the completion callbacks and
actually have a chance of getting the multifd channels ready for when
the migration thread needs them.
The next patches will deal with the synchronization aspects.
Note that this takes multifd_send_setup() out of the BQL.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240206215118.6171-5-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
Hide the error handling inside multifd_send_setup to make it cleaner
for the next patch to move the function around.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240206215118.6171-4-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
We currently only need p->running to avoid calling qemu_thread_join()
on a non existent thread if the thread has never been created.
However, there are at least two bugs in this logic:
1) On the sending side, p->running is set too early and
qemu_thread_create() can be skipped due to an error during TLS
handshake, leaving the flag set and leading to a crash when
multifd_send_cleanup() calls qemu_thread_join().
2) During exit, the multifd thread clears the flag while holding the
channel lock. The counterpart at multifd_send_cleanup() reads the flag
outside of the lock and might free the mutex while the multifd thread
still has it locked.
Fix the first issue by setting the flag right before creating the
thread. Rename it from p->running to p->thread_created to clarify its
usage.
Fix the second issue by not clearing the flag at the multifd thread
exit. We don't have any use for that.
Note that these bugs are straight-forward logic issues and not race
conditions. There is still a gap for races to affect this code due to
multifd_send_cleanup() being allowed to run concurrently with the
thread creation loop. This issue is solved in the next patches.
Cc: qemu-stable <qemu-stable@nongnu.org>
Fixes: 2964714015 ("migration/tls: add support for multifd tls-handshake")
Reported-by: Avihai Horon <avihaih@nvidia.com>
Reported-by: chenyuhui5@huawei.com
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240206215118.6171-3-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
We're currently leaking the resources of the TLS thread by not joining
it and also overwriting the p->thread pointer altogether.
Fixes: a1af605bd5 ("migration/multifd: fix hangup with TLS-Multifd due to blocking handshake")
Cc: qemu-stable <qemu-stable@nongnu.org>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240206215118.6171-2-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
The commit in the fixes line mistakenly modified the channels and
transport compatibility check logic so it now checks multi-channel
support only for socket transport type.
Thus, running multifd migration using a transport other than socket that
is incompatible with multi-channels (such as "exec") would lead to a
segmentation fault instead of an error message.
For example:
(qemu) migrate_set_capability multifd on
(qemu) migrate -d "exec:cat > /tmp/vm_state"
Segmentation fault (core dumped)
Fix it by checking multi-channel compatibility for all transport types.
Cc: qemu-stable <qemu-stable@nongnu.org>
Fixes: d95533e1cd ("migration: modify migration_channels_and_uri_compatible() for new QAPI syntax")
Signed-off-by: Avihai Horon <avihaih@nvidia.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Link: https://lore.kernel.org/r/20240125162528.7552-2-avihaih@nvidia.com
Signed-off-by: Peter Xu <peterx@redhat.com>
When reviewing my attempt to refactor send_prepare(), Fabiano suggested we
try out with dropping the mutex in multifd code [1].
I thought about that before but I never tried to change the code. Now
maybe it's time to give it a stab. This only optimizes the sender side.
The trick here is multifd has a clear provider/consumer model, that the
migration main thread publishes requests (either pending_job/pending_sync),
while the multifd sender threads are consumers. Here we don't have a lot
of complicated data sharing, and the jobs can logically be submitted
lockless.
Arm the code with atomic weapons. Two things worth mentioning:
- For multifd_send_pages(): we can use qatomic_load_acquire() when trying
to find a free channel, but that's expensive if we attach one ACQUIRE per
channel. Instead, keep the qatomic_read() on reading the pending_job
flag as we do already, meanwhile use one smp_mb_acquire() after the loop
to guarantee the memory ordering.
- For pending_sync: it doesn't have any extra data required since now
p->flags are never touched, it should be safe to not use memory barrier.
That's different from pending_job.
Provide rich comments for all the lockless operations to state how they are
paired. With that, we can remove the mutex.
[1] https://lore.kernel.org/r/87o7d1jlu5.fsf@suse.de
Suggested-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-24-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
As reported correctly by Fabiano [1] (while per Fabiano, it sourced back to
Elena's initial report in Oct 2023), MultiFDSendParams.packet_num is buggy
to be assigned and stored. Consider two consequent operations of: (1)
queue a job into multifd send thread X, then (2) queue another sync request
to the same send thread X. Then the MultiFDSendParams.packet_num will be
assigned twice, and the first assignment can get lost already.
To avoid that, we move the packet_num assignment from p->packet_num into
where the thread will fill in the packet. Use atomic operations to protect
the field, making sure there's no race.
Note that atomic fetch_add() may not be good for scaling purposes, however
multifd should be fine as number of threads should normally not go beyond
16 threads. Let's leave that concern for later but fix the issue first.
There's also a trick on how to make it always work even on 32 bit hosts for
uint64_t packet number. Switching to uintptr_t as of now to simply the
case. It will cause packet number to overflow easier on 32 bit, but that
shouldn't be a major concern for now as 32 bit systems is not the major
audience for any performance concerns like what multifd wants to address.
We also need to move multifd_send_state definition upper, so that
multifd_send_fill_packet() can reference it.
[1] https://lore.kernel.org/r/87o7d1jlu5.fsf@suse.de
Reported-by: Elena Ufimtseva <elena.ufimtseva@oracle.com>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-23-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Most of the multifd code uses send/recv to represent the two sides, but
some rare cases use save/load.
Since send/recv is the majority, replacing the save/load use cases to use
send/recv globally. Now we reach a consensus on the naming.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-22-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Use similar logic to cleanup the recv side.
Note that multifd_recv_terminate_threads() may need some similar rework
like the sender side, but let's leave that for later.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-21-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Shrink the function by moving relevant works into helpers: move the thread
join()s into multifd_send_terminate_threads(), then create two more helpers
to cover channel/state cleanups.
Add a TODO entry for the thread terminate process because p->running is
still buggy. We need to fix it at some point but not yet covered.
Suggested-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-20-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
The current multifd_queue_page() is not easy to read and follow. It is not
good with a few reasons:
- No helper at all to show what exactly does a condition mean; in short,
readability is low.
- Rely on pages->ramblock being cleared to detect an empty queue. It's
slightly an overload of the ramblock pointer, per Fabiano [1], which I
also agree.
- Contains a self recursion, even if not necessary..
Rewrite this function. We add some comments to make it even clearer on
what it does.
[1] https://lore.kernel.org/r/87wmrpjzew.fsf@suse.de
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-19-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Split multifd_send_terminate_threads() into two functions:
- multifd_send_set_error(): used when an error happened on the sender
side, set error and quit state only
- multifd_send_terminate_threads(): used only by the main thread to kick
all multifd send threads out of sleep, for the last recycling.
Use multifd_send_set_error() in the three old call sites where only the
error will be set.
Use multifd_send_terminate_threads() in the last one where the main thread
will kick the multifd threads at last in multifd_save_cleanup().
Both helpers will need to set quitting=1.
Suggested-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-16-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Now multifd's logic is designed to have no spurious wakeup. I still
remember a talk to Juan and he seems to agree we should drop it now, and if
my memory was right it was there because multifd used to hit that when
still debugging.
Let's drop it and see what can explode; as long as it's not reaching
soft-freeze.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-15-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
This patch redefines the interfacing of ->send_prepare(). It further
simplifies multifd_send_thread() especially on zero copy.
Now with the new interface, we require the hook to do all the work for
preparing the IOVs to send. After it's completed, the IOVs should be ready
to be dumped into the specific multifd QIOChannel later.
So now the API looks like:
p->pages -----------> send_prepare() -------------> IOVs
This also prepares for the case where the input can be extended to even not
any p->pages. But that's for later.
This patch will achieve similar goal of what Fabiano used to propose here:
https://lore.kernel.org/r/20240126221943.26628-1-farosas@suse.de
However the send() interface may not be necessary. I'm boldly attaching a
"Co-developed-by" for Fabiano.
Co-developed-by: Fabiano Rosas <farosas@suse.de>
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-14-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Introduce a helper multifd_send_prepare_header() to setup the header packet
for multifd sender.
It's fine to setup the IOV[0] _before_ send_prepare() because the packet
buffer is already ready, even if the content is to be filled in.
With this helper, we can already slightly clean up the zero copy path.
Note that I explicitly put it into multifd.h, because I want it inlined
directly into multifd*.c where necessary later.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-13-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Move them into fill/unfill of packets. With that, we can further cleanup
the send/recv thread procedure, and remove one more temp var.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-12-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Just like the previous patch, move the accounting for total_normal_pages on
both src/dst sides into the packet fill/unfill procedures.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-11-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
This field, no matter whether on src or dest, is only used for debugging
purpose.
They can even be removed already, unless it still more or less provide some
accounting on "how many packets are sent/recved for this thread". The
other more important one is called packet_num, which is embeded in the
multifd packet headers (MultiFDPacket_t).
So let's keep them for now, but make them much easier to understand, by
doing below:
- Rename both of them to packets_sent / packets_recved, the old
name (num_packets) are waaay too confusing when we already have
MultiFDPacket_t.packets_num.
- Avoid worrying on the "initial packet": we know we will send it, that's
good enough. The accounting won't matter a great deal to start with 0 or
with 1.
- Move them to where we send/recv the packets. They're:
- multifd_send_fill_packet() for senders.
- multifd_recv_unfill_packet() for receivers.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-10-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
The sender thread will yield the p->mutex before IO starts, trying to not
block the requester thread. This may be unnecessary lock optimizations,
because the requester can already read pending_job safely even without the
lock, because the requester is currently the only one who can assign a
task.
Drop that lock complication on both sides:
(1) in the sender thread, always take the mutex until job done
(2) in the requester thread, check pending_job clear lockless
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-8-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Multifd provide a threaded model for processing jobs. On sender side,
there can be two kinds of job: (1) a list of pages to send, or (2) a sync
request.
The sync request is a very special kind of job. It never contains a page
array, but only a multifd packet telling the dest side to synchronize with
sent pages.
Before this patch, both requests use the pending_job field, no matter what
the request is, it will boost pending_job, while multifd sender thread will
decrement it after it finishes one job.
However this should be racy, because SYNC is special in that it needs to
set p->flags with MULTIFD_FLAG_SYNC, showing that this is a sync request.
Consider a sequence of operations where:
- migration thread enqueue a job to send some pages, pending_job++ (0->1)
- [...before the selected multifd sender thread wakes up...]
- migration thread enqueue another job to sync, pending_job++ (1->2),
setup p->flags=MULTIFD_FLAG_SYNC
- multifd sender thread wakes up, found pending_job==2
- send the 1st packet with MULTIFD_FLAG_SYNC and list of pages
- send the 2nd packet with flags==0 and no pages
This is not expected, because MULTIFD_FLAG_SYNC should hopefully be done
after all the pages are received. Meanwhile, the 2nd packet will be
completely useless, which contains zero information.
I didn't verify above, but I think this issue is still benign in that at
least on the recv side we always receive pages before handling
MULTIFD_FLAG_SYNC. However that's not always guaranteed and just tricky.
One other reason I want to separate it is using p->flags to communicate
between the two threads is also not clearly defined, it's very hard to read
and understand why accessing p->flags is always safe; see the current impl
of multifd_send_thread() where we tried to cache only p->flags. It doesn't
need to be that complicated.
This patch introduces pending_sync, a separate flag just to show that the
requester needs a sync. Alongside, we remove the tricky caching of
p->flags now because after this patch p->flags should only be used by
multifd sender thread now, which will be crystal clear. So it is always
thread safe to access p->flags.
With that, we can also safely convert the pending_job into a boolean,
because we don't support >1 pending jobs anyway.
Always use atomic ops to access both flags to make sure no cache effect.
When at it, drop the initial setting of "pending_job = 0" because it's
always allocated using g_new0().
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-7-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
This array is redundant when p->pages exists. Now we extended the life of
p->pages to the whole period where pending_job is set, it should be safe to
always use p->pages->offset[] rather than p->normal[]. Drop the array.
Alongside, the normal_num is also redundant, which is the same to
p->pages->num.
This doesn't apply to recv side, because there's no extra buffering on recv
side, so p->normal[] array is still needed.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-6-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Now we reset MultiFDPages_t object in the multifd sender thread in the
middle of the sending job. That's not necessary, because the "*pages"
struct will not be reused anyway until pending_job is cleared.
Move that to the end after the job is completed, provide a helper to reset
a "*pages" object. Use that same helper when free the object too.
This prepares us to keep using p->pages in the follow up patches, where we
may drop p->normal[].
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-5-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Multifd send side has two fields to indicate error quits:
- MultiFDSendParams.quit
- &multifd_send_state->exiting
Merge them into the global one. The replacement is done by changing all
p->quit checks into the global var check. The global check doesn't need
any lock.
A few more things done on top of this altogether:
- multifd_send_terminate_threads()
Moving the xchg() of &multifd_send_state->exiting upper, so as to cover
the tracepoint, migrate_set_error() and migrate_set_state().
- multifd_send_sync_main()
In the 2nd loop, add one more check over the global var to make sure we
don't keep the looping if QEMU already decided to quit.
- multifd_tls_outgoing_handshake()
Use multifd_send_terminate_threads() to set the error state. That has
a benefit of updating MigrationState.error to that error too, so we can
persist that 1st error we hit in that specific channel.
- multifd_new_send_channel_async()
Take similar approach like above, drop the migrate_set_error() because
multifd_send_terminate_threads() already covers that. Unwrap the helper
multifd_new_send_channel_cleanup() along the way; not really needed.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-4-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
When a multifd sender thread hit errors, it always needs to kick the main
thread by kicking all the semaphores that it can be waiting upon.
Provide a helper for it and deduplicate the code.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-3-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
We've already done that with multifd_flush_after_each_section, for multifd
in general. Drop the stale "TODO-like" comment.
Reviewed-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240202102857.110210-2-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
A memory page poisoned from the hypervisor level is no longer readable.
The migration of a VM will crash Qemu when it tries to read the
memory address space and stumbles on the poisoned page with a similar
stack trace:
Program terminated with signal SIGBUS, Bus error.
#0 _mm256_loadu_si256
#1 buffer_zero_avx2
#2 select_accel_fn
#3 buffer_is_zero
#4 save_zero_page
#5 ram_save_target_page_legacy
#6 ram_save_host_page
#7 ram_find_and_save_block
#8 ram_save_iterate
#9 qemu_savevm_state_iterate
#10 migration_iteration_run
#11 migration_thread
#12 qemu_thread_start
To avoid this VM crash during the migration, prevent the migration
when a known hardware poison exists on the VM.
Signed-off-by: William Roche <william.roche@oracle.com>
Link: https://lore.kernel.org/r/20240130190640.139364-2-william.roche@oracle.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Now that the migration state reference counting is correct, further
wrap the bottom half dispatch process to avoid future issues.
Move BH creation and scheduling together and wrap the dispatch with an
intermediary function that will ensure we always keep the ref/unref
balanced.
Also move the responsibility of deleting the BH into the wrapper and
remove the now unnecessary pointers.
Signed-off-by: Fabiano Rosas <farosas@suse.de>
Link: https://lore.kernel.org/r/20240119233922.32588-6-farosas@suse.de
Signed-off-by: Peter Xu <peterx@redhat.com>
