[Intel-gfx] [PATCH 1/3] dma_resv: prime lockdep annotations

[Thomas Hellström (VMware)]

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 at amd.com>
>>> Cc: Christian König <christian.koenig at amd.com>
>>> Cc: Chris Wilson <chris at chris-wilson.co.uk>
>>> Cc: Thomas Zimmermann <tzimmermann at suse.de>
>>> Cc: Rob Herring <robh at kernel.org>
>>> Cc: Tomeu Vizoso <tomeu.vizoso at collabora.com>
>>> Cc: Eric Anholt <eric at anholt.net>
>>> Cc: Dave Airlie <airlied at redhat.com>
>>> Cc: Gerd Hoffmann <kraxel at redhat.com>
>>> Cc: Ben Skeggs <bskeggs at redhat.com>
>>> Cc: "VMware Graphics" <linux-graphics-maintainer at vmware.com>
>>> Cc: Thomas Hellstrom <thellstrom at vmware.com>
>>> Signed-off-by: Daniel Vetter <daniel.vetter at intel.com>
>>> ---
>>>    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 at vmware.com>
>>
>> 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.

/Thomas