[Intel-gfx] [RFC PATCH] drm/nouveau: fix nested locking in mmap handler

[Maarten Lankhorst]

Op 24-09-13 12:33, Thomas Hellstrom schreef:
> On 09/24/2013 12:11 PM, Maarten Lankhorst wrote:
>> Op 24-09-13 11:36, Daniel Vetter schreef:
>>> On Tue, Sep 24, 2013 at 11:03:37AM +0200, Thomas Hellstrom wrote:
>>>> On 09/24/2013 09:34 AM, Maarten Lankhorst wrote:
>>>>> Op 24-09-13 09:22, Thomas Hellstrom schreef:
>>>>>> On 09/23/2013 05:33 PM, Maarten Lankhorst wrote:
>>>>>>> Hey,
>>>>>>>
>>>>>>> Op 13-09-13 11:00, Peter Zijlstra schreef:
>>>>>>>> On Fri, Sep 13, 2013 at 10:41:54AM +0200, Daniel Vetter wrote:
>>>>>>>>> On Fri, Sep 13, 2013 at 10:29 AM, Peter Zijlstra <peterz at infradead.org> wrote:
>>>>>>>>>> On Fri, Sep 13, 2013 at 09:46:03AM +0200, Thomas Hellstrom wrote:
>>>>>>>>>>>>> if (!bo_tryreserve()) {
>>>>>>>>>>>>>       up_read mmap_sem(); // Release the mmap_sem to avoid deadlocks.
>>>>>>>>>>>>>       bo_reserve();               // Wait for the BO to become available (interruptible)
>>>>>>>>>>>>>       bo_unreserve();           // Where is bo_wait_unreserved() when we need it, Maarten :P
>>>>>>>>>>>>>       return VM_FAULT_RETRY; // Go ahead and retry the VMA walk, after regrabbing
>>>>>>>>>>>>> }
>>>>>>>>>>> Anyway, could you describe what is wrong, with the above solution, because
>>>>>>>>>>> it seems perfectly legal to me.
>>>>>>>>>> Luckily the rule of law doesn't have anything to do with this stuff --
>>>>>>>>>> at least I sincerely hope so.
>>>>>>>>>>
>>>>>>>>>> The thing that's wrong with that pattern is that its still not
>>>>>>>>>> deterministic - although its a lot better than the pure trylock. Because
>>>>>>>>>> you have to release and re-acquire with the trylock another user might
>>>>>>>>>> have gotten in again. Its utterly prone to starvation.
>>>>>>>>>>
>>>>>>>>>> The acquire+release does remove the dead/life-lock scenario from the
>>>>>>>>>> FIFO case, since blocking on the acquire will allow the other task to
>>>>>>>>>> run (or even get boosted on -rt).
>>>>>>>>>>
>>>>>>>>>> Aside from that there's nothing particularly wrong with it and lockdep
>>>>>>>>>> should be happy afaict (but I haven't had my morning juice yet).
>>>>>>>>> bo_reserve internally maps to a ww-mutex and task can already hold
>>>>>>>>> ww-mutex (potentially even the same for especially nasty userspace).
>>>>>>>> OK, yes I wasn't aware of that. Yes in that case you're quite right.
>>>>>>>>
>>>>>>> I added a RFC patch below.  I only tested with PROVE_LOCKING, and always forced the slowpath for debugging.
>>>>>>>
>>>>>>> This fixes nouveau and core ttm to always use blocking acquisition in fastpath.
>>>>>>> Nouveau was a bit of a headache, but afaict it should work.
>>>>>>>
>>>>>>> In almost all cases relocs are not updated, so I kept intact the fastpath
>>>>>>> of not copying relocs from userspace. The slowpath tries to copy it atomically,
>>>>>>> and if that fails it will unreserve all bo's and copy everything.
>>>>>>>
>>>>>>> One thing to note is that the command submission ioctl may fail now with -EFAULT
>>>>>>> if presumed cannot be updated, while the commands are submitted succesfully.
>>>>>> I think the Nouveau guys need to comment further on this, but returning -EFAULT might break existing user-space, and that's not allowed, but IIRC the return value of "presumed" is only a hint, and if it's incorrect will only trigger future command stream patching.
>>>>>>
>>>>>> Otherwise reviewing mostly the TTM stuff. FWIW, from wat I can tell the vmwgfx driver doesn't need any fixups.
>>>>> Well because we read the list of buffer objects the presumed offsets are at least read-mapped. Although I guess in the worst case the mapping might disappear before the syscall copies back the data.
>>>>> So if -EFAULT happens here then userspace messed up in some way, either by forgetting to map the offsets read-write, which cannot happen with libdrm or free'ing the bo list before the syscall returns,
>>>>> which would probably result in libdrm crashing shortly afterwards anyway.
>>>> Hmm, is the list of buffer objects (and the "presumed" members)
>>>> really in DRM memory? Because if it resides or may reside in
>>>> anonymous system memory, it may well be paged out between reading
>>>> and writing, in which case the -EFAULT return is incorrect.
>>>>
>>>> In fact failures of pushbuf / execbuf *after* commands are
>>>> successfully submitted are generally very hard to recover from.
>>>> That's why the kernel should do whatever it takes to recover from
>>>> such failures, and user-space should do whatever it takes to recover
>>>> from copy-to-user failures of needed info from the kernel, and it
>>>> really depends on the user-space usage pattern of "presumed". IIRC
>>>> the original reason for copying it back to user-space was, that if a
>>>> relocation offsets were patched up by the kernel, and then the
>>>> process was sent a signal causing it to retry execbuf, then
>>>> "presumed" had to be updated, otherwise it would be inconsistent
>>>> with what's currently in the command stream, which is very bad. If
>>>> "presumed" is, however, only used by user-space to guess an offset,
>>>> the correct action would be to swallow the -EFAULT.
>>> In i915 we've had tons of fun with a regression in 3.7 where exactly this
>>> blows up: Some of our userspace (UXA ddx specifically) retains
>>> relocations-trees partially accross an execbuf. Which means if the kernel
>>> updates the relocations it _must_ update the presumed offset for
>>> otherwise things will go haywire on the next execbuf. So we can't return
>>> -EFAULT if the userspace memory needs to be just refaulted but still need
>>> to guarante a "correct" presumed offset.
>>>
>>> Since we didn't come up with a way to make sure this will work in all
>>> cases when we get an -EFAULT when writing back presumed offsets we have a
>>> rather tricky piece of fallback logic.
>>> - Any -EFAULT error in the fastpath will drop us into the relocation
>>>    slowpath. The slowpath completly processes relocs anew, so any updates
>>>    done by the fastpath are irrelevant.
>>>
>>> - The first thing the slowpath does is set the presumed offset in the
>>>    userspace reloc lists to an invalid address (we pick -1) to make sure
>>>    that any subsequent execbuf with the same partial reloc tree will again
>>>    go through the relocation update code.
>>>
>>> - Then we do the usual slowpath, i.e. copy relocs from userspace, re-grab
>>>    locks and then process them. The copy-back of the presumed offset
>>>    happens with an copy_to_user_inatomic, and we ignore any errors.
>>>
>>> Of course we try really hard to make sure that we never hit the reloc
>>> slowpath ;-) But nowadays this is all fully tested with some nasty
>>> testcases (and a small kernel option to disable prefaulting).
>>>
>> It seems userspace only updates offset and domain in nouveau. If it fails to update
>> it would result in the same affect as when the buffer gets moved around by TTM.
>> But hey maybe I'll have some fun, I'll lie to userspace, hardcode userspace offset
>> to 0x40000000, always force domain to be different and see what breaks.
>>
>> My guess is absolutely nothing, except it might expose some bugs where we forgot annotation..
> I think that would certainly break if your return an -ERESTARTSYS after applying relocations but
> before submitting the command stream to hardware....
The relocations are updated before submitting the command stream, but it's copied back to userspace
after submitting the command stream. I'm not sure what -ERESTARTSYS would change, the syscall
is not in an interruptible state.

Oh and after some testing with relocations it seems they always break nouveau. Originally I've set domain
and offset differently, but that crashed the card. In my second try I kept domain and offset correct, and hacked
the kernel to force apply all relocations. Unsurprisingly nouveau still breaks, even if the relocations themselves
should have been a noop. :/

Swallowing -EFAULT is fine theoretically though, I wouldn't worry about it..

~Maarten