Juraj Marcin <jmar...@redhat.com> writes: > Hi Daniel, > > On 2025-08-01 18:02, Daniel P. Berrangé wrote: >> This is a followup to previously merged patches that claimed to >> workaround the gnutls bug impacting migration, but in fact were >> essentially non-functional. Juraj Marcin pointed this out, and >> this new patch tweaks the workaround to make it actually do >> something useful. >> >> Daniel P. Berrangé (2): >> migration: simplify error reporting after channel read >> migration: fix workaround for gnutls thread safety >> >> crypto/tlssession.c | 16 ---------------- >> migration/qemu-file.c | 22 +++++++++++++++++----- >> 2 files changed, 17 insertions(+), 21 deletions(-) >> > > thanks for finding a fix for the workaround. I have tested it and it > resolves the issue. > > However, it significantly slows down migration, even with the workaround > disabled (and thus no locking). When benchmarking, I used the fixed > version of GNUTLS, VM with 20GB of RAM which were fully written to > before starting a normal migration with no workload during the > migration. > > Test cases: > [1]: before this patchset > [2]: with this patchset applied and GNUTLS workaround enabled > [2]: with this patchset applied and GNUTLS workaround disabled > > | Total time | Throughput | Transfered bytes | > --+------------+------------+------------------+ > 1 | 31 192 ms | 5450 mpbs | 21 230 973 763 | > 2 | 74 147 ms | 2291 mbps | 21 232 849 066 | > 3 | 72 426 ms | 2343 mbps | 21 215 009 392 |
Thanks testing this. I had just managed to convince myself that there wouldn't be any performance issues. The yield at every buffer fill on the incoming side is probably way more impactful than the poll on the RP.
