[Intel-gfx] [PATCH 1/3] dma_resv: prime lockdep annotations
Daniel Vetter
daniel at ffwll.ch
Wed Aug 21 18:11:16 UTC 2019
On Wed, Aug 21, 2019 at 7:06 PM Thomas Hellström (VMware)
<thomas_os at shipmail.org> 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 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(¤t->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(¤t->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.
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 :-)
So I think this should be safe.
-Daniel
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
More information about the Intel-gfx
mailing list