Re: [PATCH 1/3] dma_resv: prime lockdep annotations

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 





Am 21.08.2019 20:28 schrieb "Thomas Hellström (VMware)" <thomas_os@xxxxxxxxxxxx>:
On 8/21/19 8:11 PM, Daniel Vetter wrote:
> On Wed, Aug 21, 2019 at 7:06 PM Thomas Hellström (VMware)
> <thomas_os@xxxxxxxxxxxx> wrote:
>> On 8/21/19 6:34 PM, Daniel Vetter wrote:
>>> On Wed, Aug 21, 2019 at 05:54:27PM +0200, Thomas Hellström (VMware) wrote:
>>>> On 8/20/19 4:53 PM, Daniel Vetter wrote:
>>>>> Full audit of everyone:
>>>>>
>>>>> - i915, radeon, amdgpu should be clean per their maintainers.
>>>>>
>>>>> - vram helpers should be fine, they don't do command submission, so
>>>>>      really no business holding struct_mutex while doing copy_*_user. But
>>>>>      I haven't checked them all.
>>>>>
>>>>> - panfrost seems to dma_resv_lock only in panfrost_job_push, which
>>>>>      looks clean.
>>>>>
>>>>> - v3d holds dma_resv locks in the tail of its v3d_submit_cl_ioctl(),
>>>>>      copying from/to userspace happens all in v3d_lookup_bos which is
>>>>>      outside of the critical section.
>>>>>
>>>>> - vmwgfx has a bunch of ioctls that do their own copy_*_user:
>>>>>      - vmw_execbuf_process: First this does some copies in
>>>>>        vmw_execbuf_cmdbuf() and also in the vmw_execbuf_process() itself.
>>>>>        Then comes the usual ttm reserve/validate sequence, then actual
>>>>>        submission/fencing, then unreserving, and finally some more
>>>>>        copy_to_user in vmw_execbuf_copy_fence_user. Glossing over tons of
>>>>>        details, but looks all safe.
>>>>>      - vmw_fence_event_ioctl: No ttm_reserve/dma_resv_lock anywhere to be
>>>>>        seen, seems to only create a fence and copy it out.
>>>>>      - a pile of smaller ioctl in vmwgfx_ioctl.c, no reservations to be
>>>>>        found there.
>>>>>      Summary: vmwgfx seems to be fine too.
>>>>>
>>>>> - virtio: There's virtio_gpu_execbuffer_ioctl, which does all the
>>>>>      copying from userspace before even looking up objects through their
>>>>>      handles, so safe. Plus the getparam/getcaps ioctl, also both safe.
>>>>>
>>>>> - qxl only has qxl_execbuffer_ioctl, which calls into
>>>>>      qxl_process_single_command. There's a lovely comment before the
>>>>>      __copy_from_user_inatomic that the slowpath should be copied from
>>>>>      i915, but I guess that never happened. Try not to be unlucky and get
>>>>>      your CS data evicted between when it's written and the kernel tries
>>>>>      to read it. The only other copy_from_user is for relocs, but those
>>>>>      are done before qxl_release_reserve_list(), which seems to be the
>>>>>      only thing reserving buffers (in the ttm/dma_resv sense) in that
>>>>>      code. So looks safe.
>>>>>
>>>>> - A debugfs file in nouveau_debugfs_pstate_set() and the usif ioctl in
>>>>>      usif_ioctl() look safe. nouveau_gem_ioctl_pushbuf() otoh breaks this
>>>>>      everywhere and needs to be fixed up.
>>>>>
>>>>> Cc: Alex Deucher <alexander.deucher@xxxxxxx>
>>>>> Cc: Christian König <christian.koenig@xxxxxxx>
>>>>> Cc: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx>
>>>>> Cc: Thomas Zimmermann <tzimmermann@xxxxxxx>
>>>>> Cc: Rob Herring <robh@xxxxxxxxxx>
>>>>> Cc: Tomeu Vizoso <tomeu.vizoso@xxxxxxxxxxxxx>
>>>>> Cc: Eric Anholt <eric@xxxxxxxxxx>
>>>>> Cc: Dave Airlie <airlied@xxxxxxxxxx>
>>>>> Cc: Gerd Hoffmann <kraxel@xxxxxxxxxx>
>>>>> Cc: Ben Skeggs <bskeggs@xxxxxxxxxx>
>>>>> Cc: "VMware Graphics" <linux-graphics-maintainer@xxxxxxxxxx>
>>>>> Cc: Thomas Hellstrom <thellstrom@xxxxxxxxxx>
>>>>> Signed-off-by: Daniel Vetter <daniel.vetter@xxxxxxxxx>
>>>>> ---
>>>>>     drivers/dma-buf/dma-resv.c | 12 ++++++++++++
>>>>>     1 file changed, 12 insertions(+)
>>>>>
>>>>> diff --git a/drivers/dma-buf/dma-resv.c b/drivers/dma-buf/dma-resv.c
>>>>> index 42a8f3f11681..3edca10d3faf 100644
>>>>> --- a/drivers/dma-buf/dma-resv.c
>>>>> +++ b/drivers/dma-buf/dma-resv.c
>>>>> @@ -34,6 +34,7 @@
>>>>>     #include <linux/dma-resv.h>
>>>>>     #include <linux/export.h>
>>>>> +#include <linux/sched/mm.h>
>>>>>     /**
>>>>>      * DOC: Reservation Object Overview
>>>>> @@ -107,6 +108,17 @@ void dma_resv_init(struct dma_resv *obj)
>>>>>                      &reservation_seqcount_class);
>>>>>      RCU_INIT_POINTER(obj->fence, NULL);
>>>>>      RCU_INIT_POINTER(obj->fence_excl, NULL);
>>>>> +
>>>>> +   if (IS_ENABLED(CONFIG_LOCKDEP)) {
>>>>> +           if (current->mm)
>>>>> +                   down_read(&current->mm->mmap_sem);
>>>>> +           ww_mutex_lock(&obj->lock, NULL);
>>>>> +           fs_reclaim_acquire(GFP_KERNEL);
>>>>> +           fs_reclaim_release(GFP_KERNEL);
>>>>> +           ww_mutex_unlock(&obj->lock);
>>>>> +           if (current->mm)
>>>>> +                   up_read(&current->mm->mmap_sem);
>>>>> +   }
>>>>>     }
>>>>>     EXPORT_SYMBOL(dma_resv_init);
>>>> I assume if this would have been easily done and maintainable using only
>>>> lockdep annotation instead of actually acquiring the locks, that would have
>>>> been done?
>>> There's might_lock(), plus a pile of macros, but they don't map obviuosly,
>>> so pretty good chances I accidentally end up with the wrong type of
>>> annotation. Easier to just take the locks quickly, and stuff that all into
>>> a lockdep-only section to avoid overhead.
>>>
>>>> Otherwise LGTM.
>>>>
>>>> Reviewed-by: Thomas Hellström <thellstrom@xxxxxxxxxx>
>>>>
>>>> Will test this and let you know if it trips on vmwgfx, but it really
>>>> shouldn't.
>>> Thanks, Daniel
>> One thing that strikes me is that this puts restrictions on where you
>> can actually initialize a dma_resv, even if locking orders are otherwise
>> obeyed. But that might not be a big problem.
> Hm yeah ... the trouble is a need a non-kthread thread so that I have
> a current->mm. Otherwise I'd have put it into some init section with a
> temp dma_buf. And I kinda don't want to create a fake ->mm just for
> lockdep priming. I don't expect this to be a real problem in practice,
> since before you've called dma_resv_init the reservation lock doesn't
> exist, so you can't hold it. And you've probably just allocated it, so
> fs_reclaim is going to be fine. And if you allocate dma_resv objects
> from your fault handlers I have questions anyway :-)

Coming to think of it, I think vmwgfx sometimes create bos with other
bo's reservation lock held. I guess that would trip both the mmap_sem
check the ww_mutex check?

Dito inside TTM for ghost objects.

We even used to have an dma_resv_init() in an atomic section, but that should be gone by now.

Anyway, probably have to take my review back since this will certainly go up in flames.

Need a better place for this,
Christian.


/Thomas


/Thomas




>
> So I think this should be safe.
> -Daniel



_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/dri-devel

[Index of Archives]     [Linux DRI Users]     [Linux Intel Graphics]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]
  Powered by Linux