This is v6 of postcopy preempt series. It can also be found here: https://github.com/xzpeter/qemu/tree/postcopy-preempt
RFC: https://lore.kernel.org/qemu-devel/20220119080929.39485-1-pet...@redhat.com V1: https://lore.kernel.org/qemu-devel/20220216062809.57179-1-pet...@redhat.com V2: https://lore.kernel.org/qemu-devel/20220301083925.33483-1-pet...@redhat.com V3: https://lore.kernel.org/qemu-devel/20220330213908.26608-1-pet...@redhat.com V4: https://lore.kernel.org/qemu-devel/20220331150857.74406-1-pet...@redhat.com V5: https://lore.kernel.org/qemu-devel/20220425233847.10393-1-pet...@redhat.com The major change of this version is one more patch to handle the corner case when a postcopy request hits a preempted precopy hugepage. That's also my major test focus, to make sure that corner case will be triggered and behave as expected - as long as there's a sequential dirty thread spinning it's fairly easy to trigger the special condition. v6 changelog: - Rebased, picked up more R-bs - Replace migrate_tls_enabled() with migrate_use_tls() that Leo just introduced in the zerocopy series - Added patch "migration: Respect postcopy request order in preemption mode", per discussion with Manish. Abstract ======== This series contains two parts now: (1) Leftover patches from Dan's tls unit tests v2, which is first half (2) Leftover patches from my postcopy preempt v4, which is second half This series added a new migration capability called "postcopy-preempt". It can be enabled when postcopy is enabled, and it'll simply (but greatly) speed up postcopy page requests handling process. Below are some initial postcopy page request latency measurements after the new series applied. For each page size, I measured page request latency for three cases: (a) Vanilla: the old postcopy (b) Preempt no-break-huge: preempt enabled, x-postcopy-preempt-break-huge=off (c) Preempt full: preempt enabled, x-postcopy-preempt-break-huge=on (this is the default option when preempt enabled) Here x-postcopy-preempt-break-huge parameter is just added in v2 so as to conditionally disable the behavior to break sending a precopy huge page for debugging purpose. So when it's off, postcopy will not preempt precopy sending a huge page, but still postcopy will use its own channel. I tested it separately to give a rough idea on which part of the change helped how much of it. The overall benefit should be the comparison between case (a) and (c). |-----------+---------+-----------------------+--------------| | Page size | Vanilla | Preempt no-break-huge | Preempt full | |-----------+---------+-----------------------+--------------| | 4K | 10.68 | N/A [*] | 0.57 | | 2M | 10.58 | 5.49 | 5.02 | | 1G | 2046.65 | 933.185 | 649.445 | |-----------+---------+-----------------------+--------------| [*]: This case is N/A because 4K page does not contain huge page at all [1] https://github.com/xzpeter/small-stuffs/blob/master/tools/huge_vm/uffd-latency.bpf TODO List ========= Avoid precopy write() blocks postcopy ------------------------------------- I didn't prove this, but I always think the write() syscalls being blocked for precopy pages can affect postcopy services. If we can solve this problem then my wild guess is we can further reduce the average page latency. Two solutions at least in mind: (1) we could have made the write side of the migration channel NON_BLOCK too, or (2) multi-threads on send side, just like multifd, but we may use lock to protect which page to send too (e.g., the core idea is we should _never_ rely anything on the main thread, multifd has that dependency on queuing pages only on main thread). That can definitely be done and thought about later. Multi-channel for preemption threads ------------------------------------ Currently the postcopy preempt feature use only one extra channel and one extra thread on dest (no new thread on src QEMU). It should be mostly good enough for major use cases, but when the postcopy queue is long enough (e.g. hundreds of vCPUs faulted on different pages) logically we could still observe more delays in average. Whether growing threads/channels can solve it is debatable, but sounds worthwhile a try. That's yet another thing we can think about after this patchset lands. Logically the design provides space for that - the receiving postcopy preempt thread can understand all ram-layer migration protocol, and for multi channel and multi threads we could simply grow that into multile threads handling the same protocol (with multiple PostcopyTmpPage). The source needs more thoughts on synchronizations, though, but it shouldn't affect the whole protocol layer, so should be easy to keep compatible. Please review, thanks. Peter Xu (13): migration: Add postcopy-preempt capability migration: Postcopy preemption preparation on channel creation migration: Postcopy preemption enablement migration: Postcopy recover with preempt enabled migration: Create the postcopy preempt channel asynchronously migration: Add property x-postcopy-preempt-break-huge migration: Add helpers to detect TLS capability migration: Export tls-[creds|hostname|authz] params to cmdline too migration: Enable TLS for preempt channel migration: Respect postcopy request order in preemption mode tests: Add postcopy tls migration test tests: Add postcopy tls recovery migration test tests: Add postcopy preempt tests migration/channel.c | 9 +- migration/migration.c | 139 +++++++++++++-- migration/migration.h | 44 ++++- migration/multifd.c | 4 +- migration/postcopy-ram.c | 186 +++++++++++++++++++- migration/postcopy-ram.h | 11 ++ migration/qemu-file.c | 27 +++ migration/qemu-file.h | 1 + migration/ram.c | 330 +++++++++++++++++++++++++++++++++-- migration/ram.h | 4 +- migration/savevm.c | 46 +++-- migration/socket.c | 22 ++- migration/socket.h | 1 + migration/tls.c | 9 + migration/tls.h | 4 + migration/trace-events | 15 +- qapi/migration.json | 8 +- tests/qtest/migration-test.c | 131 ++++++++++++-- 18 files changed, 917 insertions(+), 74 deletions(-) -- 2.32.0
