On Mon, Nov 2, 2026 at 10:42 AM Denis V. Lunev <[email protected]> wrote: > > On 11/2/26 14:41, Abhisek Panda wrote: > > When a Windows guest puts an idle SATA drive into standby, it disables > > PCI bus mastering on the AHCI controller but fails to clear the port > > engine bits in the port command register. During post-migration > > handling, QEMU’s ahci_state_post_load function checks only these port > > engine bits to determine whether to start the engines and map the FIS > > receive buffer (PxFB) and Command List buffer (PxCLB). Because PCI bus > > mastering is disabled, DMA access to guest memory is blocked. > > Consequently, the memory mapping requests fail or incorrectly route > > through bounce buffers instead of direct guest memory, causing the > > destination QEMU process to abort the migration. > The fix is right, but on current master the abort is the less common > outcome. Since 637b0aa139 ("softmmu: Support concurrent bounce > buffers") each PCI device has its own 4k bounce budget, and a started > port takes 1024 bytes for the command list plus 256 for the FIS area. > While the started ports fit into that budget, post_load succeeds with > lst and res_fis pointing at private zero-filled bounce buffers, and the > migration completes. When the guest later stops the port, the unmap > writes those buffers back, i.e. zeros land on PxCLB/PxFB in guest > memory. Only when the budget is exceeded does post_load fail. > > I have checked both on q35 with a qtest: clear bus mastering, migrate, > re-enable it, stop and restart port 0, issue IDENTIFY. With one disk > master migrates fine and stopping the port zeroes the whole command > list and FIS area. With three disks the destination fails with "bad > FIS receive buffer address". With your patch both cases migrate, > guest memory stays intact, and the restarted port works. > > Please describe both outcomes in the commit message, the silent memory > corruption is the more important one. > > > Starting DMA/FIS engines when bus mastering is disabled might lead to > > buffer mapping failures or improper bounce buffer allocations. Do not > > start the engines in the aforementioned scenario. When the guest > > re-enables bus mastering to access an idle drive, it should reset the > > port engine, safely allocating the buffers and starting the engines. > "it should" is what the patch relies on, as nothing restarts the > engines on a bus master 0->1 transition. Andy's trace in the 2019 > thread shows that Windows 10 does exactly that: it re-enables bus > mastering, rewrites PxCLB/PxFB, clears and restarts PxCMD and issues a > COMRESET. Windows 7 does not clear bus mastering at all. Please state > this instead of "should". > > > Reported-by: andychiu <[email protected]> > > Suggested-by: Michael S. Tsirkin <[email protected]> > > Suggested-by: John Snow <[email protected]> > > Link: https://lists.gnu.org/archive/html/qemu-devel/2019-09/msg01544.html > > Signed-off-by: Abhisek Panda <[email protected]> > With the commit message updated: > > Tested-by: Denis V. Lunev <[email protected]> > Reviewed-by: Denis V. Lunev <[email protected]> > > Den > > P.S. John - will you take this?
My attention is currently pulled very strongly to an in-house project; If you believe it should be merged I will trust your judgement and I wouldn't be upset if you merged it. --js
