Re: [Qemu-devel] Bug#757927: [qemu-kvm] TRIM (discard=unmap) broken in 2.1

[Kevin Wolf]

Am 13.08.2014 um 08:41 hat Michael Tokarev geschrieben:
> [dropping debian bugreport]
> 
> 13.08.2014 10:05, Michael Tokarev wrote:
> > [This is http://bugs.debian.org/757927 -- trim stopped working in qemu 2.1]
> > 
> > 12.08.2014 18:46, Michael Tokarev wrote:
> >> 12.08.2014 17:35, Roman Mamedov wrote:
> >>> Package: qemu-kvm
> >>> Version: 2.1+dfsg-2~bpo70+2
> >>> Severity: normal
> >>>
> >>> Hello,
> >>>
> >>> I was able to successfully use the passthrough TRIM support with an IDE
> >>> interface virtual disk in Qemu-KVM version 2.0.
> >>>
> >>> However after upgrading to 2.1 and restarting my VM, TRIM now fails in it 
> >>> with
> >>> messages like those quoted below (dmesg from the guest). There are no 
> >>> messages
> >>> in dmesg related to that on the host.
> > 
> > BTW, I found a system here with an SSD which supports discard, and tried my
> > guests against it -- everything works fine when I use sata/ahci, but it 
> > breaks
> > indeed when using the default IDE interface.
> > 
> > But applying the oneliner:
> > 
> >> It is the same as 2.1+dfsg-1~bpo70+2 but with a one-line patch
> >> added:
> >>
> >> diff --git a/hw/ide/core.c b/hw/ide/core.c
> >> index db191a6..7256592 100644
> >> --- a/hw/ide/core.c
> >> +++ b/hw/ide/core.c
> >> @@ -688,7 +688,8 @@ void ide_dma_cb(void *opaque, int ret)
> >>             sector_num, n, s->dma_cmd);
> >>  #endif
> >>
> >> -    if (!ide_sect_range_ok(s, sector_num, n)) {
> >> +    if ((s->dma_cmd == IDE_DMA_READ || s->dma_cmd == IDE_DMA_WRITE) &&
> >> +        !ide_sect_range_ok(s, sector_num, n)) {
> >>          dma_buf_commit(s);
> >>          ide_dma_error(s);
> >>          return;
> > 
> > fixes the prob for me.  So apparently kwolf was right pointing to that
> > commit 58ac3211.
> 
> But this makes me wonder: why the issue only exists with ide, but not
> with ahci/sata?  The two should be sufficiently similar to trigger this
> check the same way.  Maybe we have another bug somewhere, like adding
> an extra sector in some path, which is triggered by 58ac3211?  Maybe
> even in the guest kernel (but in this case, since it works with real
> hw, we should ignore it).

I can't give you a clear answer, but it all depends on the value of
sector_num. This is the contents of the LBA registers and unused for
TRIM (spec says it's reserved, so we shouldn't be looking at it).
Whether the problematic value in the IDE case comes from a previous
command or that's just how the driver sets things and the AHCI sets
different things I can't say. But it certainly depends on the guest
behaviour whether the bug actually triggers.

Kevin