| Message ID | 1347268383-4150-1-git-send-email-deathsimple@vodafone.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Mon, 2012-09-10 at 11:13 +0200, Christian König wrote: > Only increase the higher 32bits if we really detect a wrap around. > > Fixes: > https://bugs.freedesktop.org/show_bug.cgi?id=54129 > https://bugs.freedesktop.org/show_bug.cgi?id=54662 > > Possible fixes: > https://bugzilla.redhat.com/show_bug.cgi?id=846505 > https://bugzilla.redhat.com/show_bug.cgi?id=845639 > > Signed-off-by: Christian König <deathsimple@vodafone.de> > Cc: stable@vger.kernel.org > --- > drivers/gpu/drm/radeon/radeon_fence.c | 6 +++--- > 1 file changed, 3 insertions(+), 3 deletions(-) > > diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c > index 7b737b9..4781e13 100644 > --- a/drivers/gpu/drm/radeon/radeon_fence.c > +++ b/drivers/gpu/drm/radeon/radeon_fence.c > @@ -160,7 +160,7 @@ void radeon_fence_process(struct radeon_device *rdev, int ring) > do { > seq = radeon_fence_read(rdev, ring); > seq |= last_seq & 0xffffffff00000000LL; > - if (seq < last_seq) { > + if (seq < (last_seq - 0x80000000LL)) { > seq += 0x100000000LL; > } Can you provide a bit more explanation for this change? In particular, how could the code previously detect a wraparound when there was none, and why is this the proper fix?
On 10.09.2012 13:12, Michel Dänzer wrote: > On Mon, 2012-09-10 at 11:13 +0200, Christian König wrote: >> Only increase the higher 32bits if we really detect a wrap around. >> >> Fixes: >> https://bugs.freedesktop.org/show_bug.cgi?id=54129 >> https://bugs.freedesktop.org/show_bug.cgi?id=54662 >> >> Possible fixes: >> https://bugzilla.redhat.com/show_bug.cgi?id=846505 >> https://bugzilla.redhat.com/show_bug.cgi?id=845639 >> >> Signed-off-by: Christian König <deathsimple@vodafone.de> >> Cc: stable@vger.kernel.org >> --- >> drivers/gpu/drm/radeon/radeon_fence.c | 6 +++--- >> 1 file changed, 3 insertions(+), 3 deletions(-) >> >> diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c >> index 7b737b9..4781e13 100644 >> --- a/drivers/gpu/drm/radeon/radeon_fence.c >> +++ b/drivers/gpu/drm/radeon/radeon_fence.c >> @@ -160,7 +160,7 @@ void radeon_fence_process(struct radeon_device *rdev, int ring) >> do { >> seq = radeon_fence_read(rdev, ring); >> seq |= last_seq & 0xffffffff00000000LL; >> - if (seq < last_seq) { >> + if (seq < (last_seq - 0x80000000LL)) { >> seq += 0x100000000LL; >> } > Can you provide a bit more explanation for this change? In particular, > how could the code previously detect a wraparound when there was none, > and why is this the proper fix? Honestly I also don't really understand how this bug happened in the first place. We extend the 32bit fences supported by hardware by testing if a previously read fence value is smaller than the value we read now: > if (seq < last_seq) { But the problem seems to be that on some systems we do get fence values that are decreasing, e.g. instead of 5, 6, 7, 8 we get 5, 7, 6, 8 (or maybe 5, 6, 0, 7, 8 because somebody accidentally overwrites the fence value). It might be related to a hardware bug, or the algorithm is flawed in a way I currently don't see. Anyway the old code we had wasn't so picky about such problems and the patch just tries to make the current code as robust as the old code was, which indeed seems to solve the problems we see. The wrap around detection still works (tested by setting the initial fence value to 0xfffffff0 and letting it wrap around shortly after start), so I think it we can safely commit this. Christian.
On Mon, Sep 10, 2012 at 8:02 AM, Christian König <deathsimple@vodafone.de> wrote: > On 10.09.2012 13:12, Michel Dänzer wrote: >> >> On Mon, 2012-09-10 at 11:13 +0200, Christian König wrote: >>> >>> Only increase the higher 32bits if we really detect a wrap around. >>> >>> Fixes: >>> https://bugs.freedesktop.org/show_bug.cgi?id=54129 >>> https://bugs.freedesktop.org/show_bug.cgi?id=54662 >>> >>> Possible fixes: >>> https://bugzilla.redhat.com/show_bug.cgi?id=846505 >>> https://bugzilla.redhat.com/show_bug.cgi?id=845639 >>> >>> Signed-off-by: Christian König <deathsimple@vodafone.de> >>> Cc: stable@vger.kernel.org >>> --- >>> drivers/gpu/drm/radeon/radeon_fence.c | 6 +++--- >>> 1 file changed, 3 insertions(+), 3 deletions(-) >>> >>> diff --git a/drivers/gpu/drm/radeon/radeon_fence.c >>> b/drivers/gpu/drm/radeon/radeon_fence.c >>> index 7b737b9..4781e13 100644 >>> --- a/drivers/gpu/drm/radeon/radeon_fence.c >>> +++ b/drivers/gpu/drm/radeon/radeon_fence.c >>> @@ -160,7 +160,7 @@ void radeon_fence_process(struct radeon_device *rdev, >>> int ring) >>> do { >>> seq = radeon_fence_read(rdev, ring); >>> seq |= last_seq & 0xffffffff00000000LL; >>> - if (seq < last_seq) { >>> + if (seq < (last_seq - 0x80000000LL)) { >>> seq += 0x100000000LL; >>> } >> >> Can you provide a bit more explanation for this change? In particular, >> how could the code previously detect a wraparound when there was none, >> and why is this the proper fix? > > > Honestly I also don't really understand how this bug happened in the first > place. > > We extend the 32bit fences supported by hardware by testing if a previously > read fence value is smaller than the value we read now: > >> if (seq < last_seq) { > > > But the problem seems to be that on some systems we do get fence values that > are decreasing, e.g. instead of 5, 6, 7, 8 we get 5, 7, 6, 8 (or maybe 5, 6, > 0, 7, 8 because somebody accidentally overwrites the fence value). > > It might be related to a hardware bug, or the algorithm is flawed in a way I > currently don't see. Anyway the old code we had wasn't so picky about such > problems and the patch just tries to make the current code as robust as the > old code was, which indeed seems to solve the problems we see. > > The wrap around detection still works (tested by setting the initial fence > value to 0xfffffff0 and letting it wrap around shortly after start), so I > think it we can safely commit this. > > Christian. > If fence read ever gave the value 0 then your patch is postponing the issue until last fence reach >= 0x1 8000 0001 (which will takes month/years of uptime to happen :)). Having a log of all emitted fence value and all read en fence value would probably be helpful so we have a better clue of what's going on. Sadly i don't think we don't receive fence in order but rather that we sometimes receive a 0 value fence, if we did receive fence out of order than the issue would also have happen with previous code. Otherwise patch is ack but please add a better comment about fence value most likely being 0 for unknow reasons or receive out of order (thought i don't think so). Cheers, Jerome
On Mon, 2012-09-10 at 14:02 +0200, Christian König wrote: > On 10.09.2012 13:12, Michel Dänzer wrote: > > On Mon, 2012-09-10 at 11:13 +0200, Christian König wrote: > >> Only increase the higher 32bits if we really detect a wrap around. > >> > >> Fixes: > >> https://bugs.freedesktop.org/show_bug.cgi?id=54129 > >> https://bugs.freedesktop.org/show_bug.cgi?id=54662 > >> > >> Possible fixes: > >> https://bugzilla.redhat.com/show_bug.cgi?id=846505 > >> https://bugzilla.redhat.com/show_bug.cgi?id=845639 > >> > >> Signed-off-by: Christian König <deathsimple@vodafone.de> > >> Cc: stable@vger.kernel.org > >> --- > >> drivers/gpu/drm/radeon/radeon_fence.c | 6 +++--- > >> 1 file changed, 3 insertions(+), 3 deletions(-) > >> > >> diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c > >> index 7b737b9..4781e13 100644 > >> --- a/drivers/gpu/drm/radeon/radeon_fence.c > >> +++ b/drivers/gpu/drm/radeon/radeon_fence.c > >> @@ -160,7 +160,7 @@ void radeon_fence_process(struct radeon_device *rdev, int ring) > >> do { > >> seq = radeon_fence_read(rdev, ring); > >> seq |= last_seq & 0xffffffff00000000LL; > >> - if (seq < last_seq) { > >> + if (seq < (last_seq - 0x80000000LL)) { > >> seq += 0x100000000LL; > >> } > > Can you provide a bit more explanation for this change? In particular, > > how could the code previously detect a wraparound when there was none, > > and why is this the proper fix? > > Honestly I also don't really understand how this bug happened in the > first place. > > We extend the 32bit fences supported by hardware by testing if a > previously read fence value is smaller than the value we read now: > > > if (seq < last_seq) { > > But the problem seems to be that on some systems we do get fence values > that are decreasing, e.g. instead of 5, 6, 7, 8 we get 5, 7, 6, 8 (or > maybe 5, 6, 0, 7, 8 because somebody accidentally overwrites the fence > value). Maybe some kind of race involving radeon_fence_write()? > It might be related to a hardware bug, or the algorithm is flawed in a > way I currently don't see. Anyway the old code we had wasn't so picky > about such problems and the patch just tries to make the current code as > robust as the old code was, which indeed seems to solve the problems we see. > > The wrap around detection still works (tested by setting the initial > fence value to 0xfffffff0 and letting it wrap around shortly after > start), so I think it we can safely commit this. Without knowing exactly what kind of hardware fence value pattern caused the problem, we can't be sure that the wraparound handling will work reliably, or that the values going backwards won't cause other problems. I think it would be good to get more real-world data on that.
On Mon, Sep 10, 2012 at 11:38 AM, Michel Dänzer <michel@daenzer.net> wrote: > On Mon, 2012-09-10 at 14:02 +0200, Christian König wrote: >> On 10.09.2012 13:12, Michel Dänzer wrote: >> > On Mon, 2012-09-10 at 11:13 +0200, Christian König wrote: >> >> Only increase the higher 32bits if we really detect a wrap around. >> >> >> >> Fixes: >> >> https://bugs.freedesktop.org/show_bug.cgi?id=54129 >> >> https://bugs.freedesktop.org/show_bug.cgi?id=54662 >> >> >> >> Possible fixes: >> >> https://bugzilla.redhat.com/show_bug.cgi?id=846505 >> >> https://bugzilla.redhat.com/show_bug.cgi?id=845639 >> >> >> >> Signed-off-by: Christian König <deathsimple@vodafone.de> >> >> Cc: stable@vger.kernel.org >> >> --- >> >> drivers/gpu/drm/radeon/radeon_fence.c | 6 +++--- >> >> 1 file changed, 3 insertions(+), 3 deletions(-) >> >> >> >> diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c >> >> index 7b737b9..4781e13 100644 >> >> --- a/drivers/gpu/drm/radeon/radeon_fence.c >> >> +++ b/drivers/gpu/drm/radeon/radeon_fence.c >> >> @@ -160,7 +160,7 @@ void radeon_fence_process(struct radeon_device *rdev, int ring) >> >> do { >> >> seq = radeon_fence_read(rdev, ring); >> >> seq |= last_seq & 0xffffffff00000000LL; >> >> - if (seq < last_seq) { >> >> + if (seq < (last_seq - 0x80000000LL)) { >> >> seq += 0x100000000LL; >> >> } >> > Can you provide a bit more explanation for this change? In particular, >> > how could the code previously detect a wraparound when there was none, >> > and why is this the proper fix? >> >> Honestly I also don't really understand how this bug happened in the >> first place. >> >> We extend the 32bit fences supported by hardware by testing if a >> previously read fence value is smaller than the value we read now: >> >> > if (seq < last_seq) { >> >> But the problem seems to be that on some systems we do get fence values >> that are decreasing, e.g. instead of 5, 6, 7, 8 we get 5, 7, 6, 8 (or >> maybe 5, 6, 0, 7, 8 because somebody accidentally overwrites the fence >> value). > > Maybe some kind of race involving radeon_fence_write()? > > >> It might be related to a hardware bug, or the algorithm is flawed in a >> way I currently don't see. Anyway the old code we had wasn't so picky >> about such problems and the patch just tries to make the current code as >> robust as the old code was, which indeed seems to solve the problems we see. >> >> The wrap around detection still works (tested by setting the initial >> fence value to 0xfffffff0 and letting it wrap around shortly after >> start), so I think it we can safely commit this. > > Without knowing exactly what kind of hardware fence value pattern caused > the problem, we can't be sure that the wraparound handling will work > reliably, or that the values going backwards won't cause other problems. > I think it would be good to get more real-world data on that. > As i said in my email this patch just postpone the issue to last_fence >= 0x1 8000 0001 if fence value we read back is sometimes randomly 0. If we received fence value out of order (which i highly doubt as old code would have had same issue thought on smaller scale) then if fence value 0x1 8000 0001 is received before fence value 0x1 8000 0000 we are right back to all future fence considered as signalled (again this will take month of uptime). All this probably lead to questioning the usefulness of 64bits fence. Cheers, Jerome
On Mon, Sep 10, 2012 at 11:52 AM, Jerome Glisse <j.glisse@gmail.com> wrote: > On Mon, Sep 10, 2012 at 11:38 AM, Michel Dänzer <michel@daenzer.net> wrote: >> On Mon, 2012-09-10 at 14:02 +0200, Christian König wrote: >>> On 10.09.2012 13:12, Michel Dänzer wrote: >>> > On Mon, 2012-09-10 at 11:13 +0200, Christian König wrote: >>> >> Only increase the higher 32bits if we really detect a wrap around. >>> >> >>> >> Fixes: >>> >> https://bugs.freedesktop.org/show_bug.cgi?id=54129 >>> >> https://bugs.freedesktop.org/show_bug.cgi?id=54662 >>> >> >>> >> Possible fixes: >>> >> https://bugzilla.redhat.com/show_bug.cgi?id=846505 >>> >> https://bugzilla.redhat.com/show_bug.cgi?id=845639 >>> >> >>> >> Signed-off-by: Christian König <deathsimple@vodafone.de> >>> >> Cc: stable@vger.kernel.org >>> >> --- >>> >> drivers/gpu/drm/radeon/radeon_fence.c | 6 +++--- >>> >> 1 file changed, 3 insertions(+), 3 deletions(-) >>> >> >>> >> diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c >>> >> index 7b737b9..4781e13 100644 >>> >> --- a/drivers/gpu/drm/radeon/radeon_fence.c >>> >> +++ b/drivers/gpu/drm/radeon/radeon_fence.c >>> >> @@ -160,7 +160,7 @@ void radeon_fence_process(struct radeon_device *rdev, int ring) >>> >> do { >>> >> seq = radeon_fence_read(rdev, ring); >>> >> seq |= last_seq & 0xffffffff00000000LL; >>> >> - if (seq < last_seq) { >>> >> + if (seq < (last_seq - 0x80000000LL)) { >>> >> seq += 0x100000000LL; >>> >> } >>> > Can you provide a bit more explanation for this change? In particular, >>> > how could the code previously detect a wraparound when there was none, >>> > and why is this the proper fix? >>> >>> Honestly I also don't really understand how this bug happened in the >>> first place. >>> >>> We extend the 32bit fences supported by hardware by testing if a >>> previously read fence value is smaller than the value we read now: >>> >>> > if (seq < last_seq) { >>> >>> But the problem seems to be that on some systems we do get fence values >>> that are decreasing, e.g. instead of 5, 6, 7, 8 we get 5, 7, 6, 8 (or >>> maybe 5, 6, 0, 7, 8 because somebody accidentally overwrites the fence >>> value). >> >> Maybe some kind of race involving radeon_fence_write()? >> >> >>> It might be related to a hardware bug, or the algorithm is flawed in a >>> way I currently don't see. Anyway the old code we had wasn't so picky >>> about such problems and the patch just tries to make the current code as >>> robust as the old code was, which indeed seems to solve the problems we see. >>> >>> The wrap around detection still works (tested by setting the initial >>> fence value to 0xfffffff0 and letting it wrap around shortly after >>> start), so I think it we can safely commit this. >> >> Without knowing exactly what kind of hardware fence value pattern caused >> the problem, we can't be sure that the wraparound handling will work >> reliably, or that the values going backwards won't cause other problems. >> I think it would be good to get more real-world data on that. >> > > As i said in my email this patch just postpone the issue to last_fence >>= 0x1 8000 0001 if fence value we read back is sometimes randomly 0. > If we received fence value out of order (which i highly doubt as old > code would have had same issue thought on smaller scale) then if fence > value 0x1 8000 0001 is received before fence value 0x1 8000 0000 we > are right back to all future fence considered as signalled (again this > will take month of uptime). Actually thinking back about it if fence are just received out of order then this patch corner case is if we received 0x1 ffff ffff after receiving 0x1 0000 0000, what will happen is that the 0x1 0000 0000 is the wrap over that will trigger upper 32bits to be incremented so fence become 0x2 0000 0000 then we got 0xffff ffff which with | become 0x2 ffff ffff then we get next fence value 0x0000 0001 and again we increment upper 32bits so last seq become 0x3 0000 0001. Again this will happen after month of uptime and all it does is decrement the amount of uptime for which 64bit fence are fine ie at worst we over increment by 0x2 0000 0000 instead of 0x1 0000 0000 on wrap around. Cheers, Jerome > > All this probably lead to questioning the usefulness of 64bits fence. > > Cheers, > Jerome
>>> >>> >>>> It might be related to a hardware bug, or the algorithm is flawed in a >>>> way I currently don't see. Anyway the old code we had wasn't so picky >>>> about such problems and the patch just tries to make the current code as >>>> robust as the old code was, which indeed seems to solve the problems we see. >>>> >>>> The wrap around detection still works (tested by setting the initial >>>> fence value to 0xfffffff0 and letting it wrap around shortly after >>>> start), so I think it we can safely commit this. Can we start fences off so we wrap around after say 15-20 minutes? that would ensure a) its tested b) we see failure in a lifetime. Dave.
On Mon, Sep 10, 2012 at 4:13 PM, Dave Airlie <airlied@gmail.com> wrote: >>>> >>>> >>>>> It might be related to a hardware bug, or the algorithm is flawed in a >>>>> way I currently don't see. Anyway the old code we had wasn't so picky >>>>> about such problems and the patch just tries to make the current code as >>>>> robust as the old code was, which indeed seems to solve the problems we see. >>>>> >>>>> The wrap around detection still works (tested by setting the initial >>>>> fence value to 0xfffffff0 and letting it wrap around shortly after >>>>> start), so I think it we can safely commit this. > > Can we start fences off so we wrap around after say 15-20 minutes? > that would ensure > > a) its tested > b) we see failure in a lifetime. > > Dave. IIRC normal desktop with continuous activities was around 400 fence/minutes. Anyway it all depends on what is wrong. Do we sometime get a 0 as fence value or do we get fence value in wrong order. Depending on that the wrap around is a different issue see my previous mails. Cheers, Jerome
On Mon, Sep 10, 2012 at 5:10 PM, Jerome Glisse <j.glisse@gmail.com> wrote: > On Mon, Sep 10, 2012 at 4:13 PM, Dave Airlie <airlied@gmail.com> wrote: >>>>> >>>>> >>>>>> It might be related to a hardware bug, or the algorithm is flawed in a >>>>>> way I currently don't see. Anyway the old code we had wasn't so picky >>>>>> about such problems and the patch just tries to make the current code as >>>>>> robust as the old code was, which indeed seems to solve the problems we see. >>>>>> >>>>>> The wrap around detection still works (tested by setting the initial >>>>>> fence value to 0xfffffff0 and letting it wrap around shortly after >>>>>> start), so I think it we can safely commit this. >> >> Can we start fences off so we wrap around after say 15-20 minutes? >> that would ensure >> >> a) its tested >> b) we see failure in a lifetime. >> >> Dave. > > IIRC normal desktop with continuous activities was around 400 fence/minutes. > > Anyway it all depends on what is wrong. Do we sometime get a 0 as > fence value or do we get fence value in wrong order. Depending on that > the wrap around is a different issue see my previous mails. > > Cheers, > Jerome s/minutes/seconds Jerome
diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c index 7b737b9..4781e13 100644 --- a/drivers/gpu/drm/radeon/radeon_fence.c +++ b/drivers/gpu/drm/radeon/radeon_fence.c @@ -160,7 +160,7 @@ void radeon_fence_process(struct radeon_device *rdev, int ring) do { seq = radeon_fence_read(rdev, ring); seq |= last_seq & 0xffffffff00000000LL; - if (seq < last_seq) { + if (seq < (last_seq - 0x80000000LL)) { seq += 0x100000000LL; } @@ -811,8 +811,8 @@ static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring) rdev->fence_drv[ring].cpu_addr = NULL; rdev->fence_drv[ring].gpu_addr = 0; for (i = 0; i < RADEON_NUM_RINGS; ++i) - rdev->fence_drv[ring].sync_seq[i] = 0; - atomic64_set(&rdev->fence_drv[ring].last_seq, 0); + rdev->fence_drv[ring].sync_seq[i] = 0x100000000LL; + atomic64_set(&rdev->fence_drv[ring].last_seq, 0x100000000LL); rdev->fence_drv[ring].last_activity = jiffies; rdev->fence_drv[ring].initialized = false; }
Only increase the higher 32bits if we really detect a wrap around. Fixes: https://bugs.freedesktop.org/show_bug.cgi?id=54129 https://bugs.freedesktop.org/show_bug.cgi?id=54662 Possible fixes: https://bugzilla.redhat.com/show_bug.cgi?id=846505 https://bugzilla.redhat.com/show_bug.cgi?id=845639 Signed-off-by: Christian König <deathsimple@vodafone.de> Cc: stable@vger.kernel.org --- drivers/gpu/drm/radeon/radeon_fence.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)