An in-flight ATAPI buffered CD read (ide_buffered_readv(), used by both
the PIO and DMA reply paths in hw/ide/atapi.c) can collide with a
concurrent port operation, and it breaks in two distinct ways. Both are
guest-triggerable and reproduce on current master.

Neither was found by inspection: both are real issues reported against
the qemu-kvm 10.0 downstream -- the crash from production, and the
deadlock from the downstream live-migration validation suite -- on
CD-ROMs backed by high-latency network storage, where a backend read
outlives the guest's SATA command timeout.

1. Command engine restart -> NULL-deref crash (patches 1-2)

   A guest can stop and restart a port's command engine
   (PxCMD.ST 1 -> 0 -> 1) while a read is in flight. The restart re-maps
   the command list and clears AHCIDevice.cur_cmd, but nothing tears
   down the outstanding read; its completion then dereferences the
   cleared cur_cmd in ahci_pio_transfer() / ahci_dma_rw_buf(). This is
   not the COMRESET path -- an engine restart does not run ide_reset(),
   so the drive's transfer state is preserved and the read still runs.

   Patch 1 cancels the outstanding I/O when the command list is
   unmapped, reusing ide_cancel_dma_sync() exactly as ATAPI DEVICE
   RESET already does from the same MMIO context. Patch 2 is a qtest
   covering the PIO and DMA reply paths.

2. Concurrent drain -> main-loop deadlock (patch 8, repro in patch 9)

   A multi-sector PIO read whose byte-count limit spans more than one
   sector must fetch the later sectors of a DRQ burst from inside the
   completion of the first, asynchronous read. cd_read_sector_sync()
   did that with a synchronous blk_pread(), which queues behind any
   drain in progress. When a guest reset reaches
   virtio_blk_stop_ioeventfd() -> bdrv_drain_all_begin() while such a
   read is in flight, the nested read is queued while the outer
   completion still holds blk->in_flight, so the drain never finishes:
   the main loop wedges in the drain with the BQL held and every
   QMP/monitor operation blocks behind it.

   Patch 8 reads the whole elementary transfer in a single asynchronous
   request, so no read is ever issued mid-burst; cd_read_sector_sync()
   is removed and the DMA path is unchanged. Patch 9 reproduces the
   hang via x-blockdev-set-iothread, which takes the same
   bdrv_drain_all_begin() path a guest reset does.

Patches 3-7 are the test groundwork the deadlock fix needs: they fold
the IDE CD-ROM read tests into one parametrized helper and add the
previously missing coverage -- multi-sector DMA, and raw 2352-byte
READ CD on both IDE and AHCI -- so the rewritten ATAPI read path
(including its in-place 2352 unpack) is exercised on both consumers.
The test patches pass with or without patch 8, keeping the series
bisectable.

Patches 1-2 stand alone as the engine-restart crash fix; they are
included here because the deadlock work shares and extends the same
ATAPI read path and its test scaffolding.

checkpatch is clean on all nine patches; the new and existing IDE and
AHCI qtests pass.

Denis V. Lunev (9):
  hw/ide/ahci: cancel in-flight buffered reads on command engine restart
  tests/qtest/ahci: test ATAPI read completing after engine restart
  tests/qtest/ide-test: parametrize the ATAPI CD-ROM read test
  tests/qtest/ide-test: add a multi-sector ATAPI DMA read test
  tests/qtest/ide-test: cover raw (2352-byte) ATAPI CD reads
  tests/qtest/libqos/ahci: support raw (2352-byte) READ CD
  tests/qtest/ahci: cover raw (2352-byte) ATAPI CD reads
  hw/ide/atapi: read the whole elementary transfer asynchronously
  tests/qtest/ahci: regression test for ATAPI read vs. drain

 hw/ide/ahci.c             |   3 +
 hw/ide/atapi.c            | 180 +++++++++++++++-----------------
 tests/qtest/ahci-test.c   | 201 ++++++++++++++++++++++++++++++++++++
 tests/qtest/ide-test.c    | 211 ++++++++++++++++++++++++--------------
 tests/qtest/libqos/ahci.c |   8 ++
 tests/qtest/libqos/ahci.h |   2 +
 6 files changed, 433 insertions(+), 172 deletions(-)

-- 
2.53.0


Reply via email to