Re: [PATCH] dma_resv: prime lockdep annotations

From: "Thomas Hellström (VMware)" <thomas_os@shipmail.org>
To: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: "Thomas Hellstrom" <thellstrom@vmware.com>,
	"Tomeu Vizoso" <tomeu.vizoso@collabora.com>,
	"DRI Development" <dri-devel@lists.freedesktop.org>,
	"VMware Graphics" <linux-graphics-maintainer@vmware.com>,
	"Gerd Hoffmann" <kraxel@redhat.com>,
	"Thomas Zimmermann" <tzimmermann@suse.de>,
	"Daniel Vetter" <daniel.vetter@intel.com>,
	"Alex Deucher" <alexander.deucher@amd.com>,
	"Dave Airlie" <airlied@redhat.com>,
	"Christian König" <christian.koenig@amd.com>,
	"Ben Skeggs" <bskeggs@redhat.com>
Subject: Re: [PATCH] dma_resv: prime lockdep annotations
Date: Thu, 22 Aug 2019 16:56:37 +0200	[thread overview]
Message-ID: <bfa6a325-6391-2c12-9b65-104193a09e91@shipmail.org> (raw)
In-Reply-To: <CAKMK7uE3K1Jh8QBv+FczXYWGei3BX7eS-sfpM5qP-nL1MSFFoQ@mail.gmail.com>

On 8/22/19 3:36 PM, Daniel Vetter wrote:
> On Thu, Aug 22, 2019 at 3:30 PM Thomas Hellström (VMware)
> <thomas_os@shipmail.org> wrote:
>> On 8/22/19 3:07 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.
>>>
>>> v2: Thomas pointed at that vmwgfx calls dma_resv_init while it holds a
>>> dma_resv lock of a different object already. Christian mentioned that
>>> ttm core does this too for ghost objects. intel-gfx-ci highlighted
>>> that i915 has similar issues.
>>>
>>> Unfortunately we can't do this in the usual module init functions,
>>> because kernel threads don't have an ->mm - we have to wait around for
>>> some user thread to do this.
>>>
>>> Solution is to spawn a worker (but only once). It's horrible, but it
>>> works.
>>>
>>> v3: We can allocate mm! (Chris). Horrible worker hack out, clean
>>> initcall solution in.
>>>
>>> v4: Annotate with __init (Rob Herring)
>>>
>>> Cc: Rob Herring <robh@kernel.org>
>>> Cc: Alex Deucher <alexander.deucher@amd.com>
>>> Cc: Christian König <christian.koenig@amd.com>
>>> Cc: Chris Wilson <chris@chris-wilson.co.uk>
>>> Cc: Thomas Zimmermann <tzimmermann@suse.de>
>>> Cc: Rob Herring <robh@kernel.org>
>>> Cc: Tomeu Vizoso <tomeu.vizoso@collabora.com>
>>> Cc: Eric Anholt <eric@anholt.net>
>>> Cc: Dave Airlie <airlied@redhat.com>
>>> Cc: Gerd Hoffmann <kraxel@redhat.com>
>>> Cc: Ben Skeggs <bskeggs@redhat.com>
>>> Cc: "VMware Graphics" <linux-graphics-maintainer@vmware.com>
>>> Cc: Thomas Hellstrom <thellstrom@vmware.com>
>>> Reviewed-by: Christian König <christian.koenig@amd.com>
>>> Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
>>> Tested-by: Chris Wilson <chris@chris-wilson.co.uk>
>>> Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
>>> ---
>>>    drivers/dma-buf/dma-resv.c | 24 ++++++++++++++++++++++++
>>>    1 file changed, 24 insertions(+)
>>>
>>> diff --git a/drivers/dma-buf/dma-resv.c b/drivers/dma-buf/dma-resv.c
>>> index 42a8f3f11681..97c4c4812d08 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
>>> @@ -95,6 +96,29 @@ static void dma_resv_list_free(struct dma_resv_list *list)
>>>        kfree_rcu(list, rcu);
>>>    }
>>>
>>> +#if IS_ENABLED(CONFIG_LOCKDEP)
>>> +static void __init dma_resv_lockdep(void)
>>> +{
>>> +     struct mm_struct *mm = mm_alloc();
>>> +     struct dma_resv obj;
>>> +
>>> +     if (!mm)
>>> +             return;
>>> +
>>> +     dma_resv_init(&obj);
>>> +
>>> +     down_read(&mm->mmap_sem);
>>
>> I took a quick look into using lockdep macros replacing the actual
>> locks: Something along the lines of
>>
>> lock_acquire(mm->mmap_sem.dep_map, 0, 0, 1, 1, NULL, _THIS_IP_);
> Yeah I'm not a fan of the magic numbers this nees :-/ And now this is
> run once at startup, so the taking the fake locks for real, once,
> shouldn't hurt. Lockdep updating it's data structures is going to be
> 100x more cpu cycles anyway :-)
>
>>> +     ww_mutex_lock(&obj.lock, NULL);
>> lock_acquire(obj.lock.dep_map, 0, 0, 0, 1, NULL, _THIS_IP_);
>>> +     fs_reclaim_acquire(GFP_KERNEL);
>>> +     fs_reclaim_release(GFP_KERNEL);
>>> +     ww_mutex_unlock(&obj.lock);
>> lock_release(obj.lock.dep_map, 0, _THIS_IP_);
>>
>>> +     up_read(&mm->mmap_sem);
>> lock_release(obj.lock.dep_map, 0, _THIS_IP_);
>>
>> Either way is fine with me, though.
>>
>> Reviewed-by: Thomas Hellström <thellstrom@vmware.com>
> Thanks for your review comments.
>
> Can you pls also run this in some test cycles, if that's easily
> possible? I'd like to have a tested-by from at least the big drivers -
> i915, amd, nouveau, vmwgfx and is definitely using ttm to its fullest
> too, so best chances for hitting an oversight.
>
> Cheers, Daniel

Tested vmwgfx with a decent OpenGL / rendercheck stress test and no 
lockdep trips.

/Thomas

Tested-by: Thomas Hellström <thellstrom@vmware.com>

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