Cc: Bartlomiej Zolnierkiewicz <bzolnier@...>, Anders Eriksson <aeriksson@...>, Rafael J. Wysocki <rjw@...>, Jens Axboe <jens.axboe@...>, Ingo Molnar <mingo@...>, Linux Kernel Mailing List <linux-kernel@...>
Well, I admittedly haven't been involved in IDE in about a decade, so
yeah, maybe it's simply no longer true. That said, if the second bisection
was accurate (which is admittedly a rasonably big "if"), it really looks
like it would be related to the fact that we used to empty the data buffer
before handling errors for REQ_TYPE_ATA_CMD commands.
.. but as you noticed, it's almost never wrong to drain (the only chipset
it's marked for is some broken pdc202xx one), and it definitely *is* wrong
not to drain.
Also, one reason you'd not necessarily even see this is that with DMA,
this is a non-issue (since the hardware DMA engine will be doing all the
draining), so in order to ever see this you have to still use PIO _and_
you have to see IDE command errors in the first place _and_ you have to
have a device that actually keeps DRQ enabled even at an error.
All of which are hopefully fairly rare by now (and getting rarer, at last
for the PIO one).
I also wouldn't be entirely surprised if the DRQ behavior may even be
command-specific, with the regular data path for read/write quite possibly
being different from the special commands that go through some internal
drive firmware logic paths.
So I could well imagine (for example) that when a drive raises an IO error
due to a read or write fault, the DRQ line will be cleared by the drive,
but that special commands might have some firmware-directed separate FIFO
that needs draining.
Linus
--
