[Nouveau] [PATCH drm-misc-next v3 6/7] drm/gpuvm: generalize dma_resv/extobj handling and GEM validation

Thomas Hellström thomas.hellstrom at linux.intel.com
Thu Sep 14 17:13:02 UTC 2023


On Thu, 2023-09-14 at 17:27 +0200, Danilo Krummrich wrote:
> On 9/14/23 13:32, Thomas Hellström wrote:
> > 
> > On 9/14/23 12:57, Danilo Krummrich wrote:
> > > On 9/13/23 14:16, Danilo Krummrich wrote:
> > > 
> > > <snip>
> > > 
> > > > > > And validate() can remove it while still holding all dma-
> > > > > > resv locks,
> > > > > > neat!
> > > > > > However, what if two tasks are trying to lock the VA space
> > > > > > concurrently? What
> > > > > > do we do when the drm_gpuvm_bo's refcount drops to zero in
> > > > > > drm_gpuva_unlink()?
> > > > > > Are we guaranteed that at this point of time the
> > > > > > drm_gpuvm_bo is not
> > > > > > on the
> > > > > > evicted list? Because otherwise we would call
> > > > > > drm_gpuvm_bo_destroy()
> > > > > > with the
> > > > > > dma-resv lock held, which wouldn't be allowed, since
> > > > > > drm_gpuvm_bo_destroy()
> > > > > > might drop the last reference to the drm_gem_object and
> > > > > > hence we'd
> > > > > > potentially
> > > > > > free the dma-resv lock while holding it, at least if it's
> > > > > > an external
> > > > > > object.
> > > > > 
> > > > > Easiest way in this scheme is to think of the lists as being
> > > > > protected
> > > > > by the vm's resv lock. That means anybody calling unlink()
> > > > > must also
> > > > > hold the vm's resv lock. (Which is OK from an UAF point of
> > > > > view, but
> > > > > perhaps not from a locking inversion POW from an async list
> > > > > update).
> > > > 
> > > > This would mean that on unlink() we'd need to hold the VM's
> > > > resv lock and the
> > > > corresponding GEM's resv lock (in case they're not the same
> > > > anyways) because the
> > > > VM's resv lock would protect the external / evicted object
> > > > lists and the GEM
> > > > objects resv lock protects the GEM's list of drm_gpuvm_bos and
> > > > the
> > > > drm_gpuvm_bo's list of drm_gpuvas.
> > > 
> > > As mentioned below the same applies for drm_gpuvm_bo_put() since
> > > it might
> > > destroy the vm_bo, which includes removing the vm_bo from
> > > external / evicted
> > > object lists and the GEMs list of vm_bos.
> > > 
> > > As mentioned, if the GEM's dma-resv is different from the VM's
> > > dma-resv we need
> > > to take both locks. Ultimately, this would mean we need a
> > > drm_exec loop, because
> > > we can't know the order in which to take these locks. Doing a
> > > full drm_exec loop
> > > just to put() a vm_bo doesn't sound reasonable to me.
> > > 
> > > Can we instead just have an internal mutex for locking the lists
> > > such that we
> > > avoid taking and dropping the spinlocks, which we use currently,
> > > in a loop?
> > 
> > You'd have the same locking inversion problem with a mutex, right?
> > Since in the eviction path you have resv->mutex, from exec you have
> > resv->mutex->resv because validate would attempt to grab resv.
> 
> Both lists, evict and extobj, would need to have a separate mutex,
> not a common one.
> We'd also need a dedicated GEM gpuva lock. Then the only rule would
> be that you can't
> hold the dma-resv lock when calling put(). Which I admit is not that
> nice.
> 
> With the current spinlock solution drivers wouldn't need to worry
> about anything locking
> related though. So maybe I come back to your proposal of having a
> switch for external
> locking with dma-resv locks entirely. Such that with external dma-
> resv locking I skip
> all the spinlocks and add lockdep checks instead.
> 
> I think that makes the most sense in terms of taking advantage of
> external dma-resv locking
> where possible and on the other hand having a self-contained solution
> if not. This should
> get all concerns out of the way, yours, Christian's and Boris'.

If we need additional locks yes, I'd prefer the opt-in/opt-out spinlock
solution, and check back after a while to see if we can remove either
option once most pitfalls are hit.

Thanks,
/Thomas


> 
> > 
> > That said, xe currently indeed does the vm+bo exec dance on vma
> > put.
> > 
> > One reason why that seemingly horrible construct is good, is that
> > when evicting an extobj and you need to access individual vmas to
> > Zap page table entries or TLB flush, those VMAs are not allowed to
> > go away (we're not refcounting them). Holding the bo resv on gpuva
> > put prevents that from happening. Possibly one could use another
> > mutex to protect the gem->vm_bo list to achieve the same, but we'd
> > need to hold it on gpuva put.
> > 
> > /Thomas
> > 
> > 
> > > 
> > > - Danilo
> > > 
> > > > 
> > > > > 
> > > > > > 
> > > > > > > > 
> > > > > > > > For extobjs an outer lock would be enough in case of
> > > > > > > > Xe, but I
> > > > > > > > really would not
> > > > > > > > like to add even more complexity just to get the
> > > > > > > > spinlock out of
> > > > > > > > the way in case
> > > > > > > > the driver already has an outer lock protecting this
> > > > > > > > path.
> > > > > > > 
> > > > > > > I must disagree here. These spinlocks and atomic
> > > > > > > operations are
> > > > > > > pretty
> > > > > > > costly and as discussed earlier this type of locking was
> > > > > > > the reason
> > > > > > > (at
> > > > > > > least according to the commit message) that made
> > > > > > > Christian drop the
> > > > > > > XArray
> > > > > > > use in drm_exec for the same set of objects: "The locking
> > > > > > > overhead
> > > > > > > is
> > > > > > > unecessary and measurable". IMHO the spinlock is the
> > > > > > > added
> > > > > > > complexity and a
> > > > > > > single wide lock following the drm locking guidelines set
> > > > > > > out by
> > > > > > > Daniel and
> > > > > > > David should really be the default choice with an opt-in
> > > > > > > for a
> > > > > > > spinlock if
> > > > > > > needed for async and pushing out to a wq is not an
> > > > > > > option.
> > > > > > 
> > > > > > For the external object list an outer lock would work as
> > > > > > long as it's
> > > > > > not the
> > > > > > dma-resv lock of the corresponding GEM object, since here
> > > > > > we actually
> > > > > > need to
> > > > > > remove the list entry from the external object list on
> > > > > > drm_gpuvm_bo_destroy().
> > > > > > It's just a bit weird design wise that drivers would need
> > > > > > to take
> > > > > > this outer
> > > > > > lock on:
> > > > > > 
> > > > > > - drm_gpuvm_bo_extobj_add()
> > > > > > - drm_gpuvm_bo_destroy()        (and hence also
> > > > > > drm_gpuvm_bo_put())
> > > > > > - drm_gpuva_unlink()            (because it needs to call
> > > > > > drm_gpuvm_bo_put())
> > > > > > - drm_gpuvm_exec_lock()
> > > > > > - drm_gpuvm_exec_lock_array()
> > > > > > - drm_gpuvm_prepare_range()
> > > > > > 
> > > > > > Given that it seems reasonable to do all the required
> > > > > > locking
> > > > > > internally.
> > > > > 
> > > > >  From a design POW, there has been a clear direction in XE to
> > > > > make
> > > > > things similar to mmap() / munmap(), so this outer lock,
> > > > > which in Xe is
> > > > > an rwsem, is used in a similar way as the mmap_lock. It's
> > > > > protecting
> > > > > the page-table structures and vma rb tree, the userptr
> > > > > structures and
> > > > > the extobj list. Basically it's taken early in the exec
> > > > > IOCTL, the
> > > > > VM_BIND ioctl, the compute rebind worker and the pagefault
> > > > > handler, so
> > > > > all of the above are just asserting that it is taken in the
> > > > > correct
> > > > > mode.
> > > > > 
> > > > > But strictly with this scheme one could also use the vm's
> > > > > dma_resv for
> > > > > the extobj list since with drm_exec, it's locked before
> > > > > traversing the
> > > > > list.
> > > > > 
> > > > > The whole point of this scheme is to rely on locks that you
> > > > > already are
> > > > > supposed to be holding for various reasons and is simple to
> > > > > comprehend.
> > > > 
> > > > I don't agree that we're supposed to hold the VM's resv lock
> > > > anyways for
> > > > functions like drm_gpuvm_bo_put() or drm_gpuva_unlink(), but
> > > > I'm fine using it
> > > > for that purpose nevertheless.
> > > > 
> > > > > 
> > > > > > 
> > > > > > In order to at least place lockdep checks, the driver would
> > > > > > need to
> > > > > > supply the
> > > > > > corresponding lock's lockdep_map, because the GPUVM
> > > > > > otherwise doesn't
> > > > > > know about
> > > > > > the lock.
> > > > > 
> > > > > Yes, that sounds reasonable. One lockdep map per list.
> > > > 
> > > > I'd really like to avoid that, especially now that everything
> > > > got simpler. We
> > > > should define the actual locks to take instead.
> > > > 
> > > > > 
> > > > > > 
> > > > > > Out of curiosity, what is the overhead of a spin_lock()
> > > > > > that doesn't
> > > > > > need to
> > > > > > spin?
> > > > > 
> > > > > I guess it's hard to tell exactly, but it is much lower on
> > > > > modern x86
> > > > > than what it used to be. Not sure about ARM, which is the
> > > > > other
> > > > > architecture important to us. I figure if there is little
> > > > > cache-line
> > > > > bouncing the main overhead comes from the implied barriers.
> > > > > 
> > > > > > 
> > > > > > > 
> > > > > > > A pretty simple way that would not add much code would be
> > > > > > > 
> > > > > > > static void gpuvm_cond_spin_lock(const struct drm_gpuvm
> > > > > > > *gpuvm,
> > > > > > > spinlock_t
> > > > > > > *lock)
> > > > > > > 
> > > > > > > {
> > > > > > > 
> > > > > > >      if (!gpuvm->resv_protected_lists)
> > > > > > >          spin_lock(lock);
> > > > > > > 
> > > > > > > }
> > > > > > > 
> > > > > > > > > For such drivers, that would require anybody calling
> > > > > > > > > unlink to
> > > > > > > > > hold the vm's
> > > > > > > > > resv, though.
> > > > > > > > In V4 I want to go back to having a dedicated lock for
> > > > > > > > the GEMs
> > > > > > > > gpuva list (or
> > > > > > > > VM_BO list to be more precise). We can't just use the
> > > > > > > > dma-resv
> > > > > > > > lock for that
> > > > > > > > with VM_BO abstractions, because on destruction of a
> > > > > > > > VM_BO we
> > > > > > > > otherwise wouldn't
> > > > > > > > be allowed to already hold the dma-resv lock. That's
> > > > > > > > the fix I
> > > > > > > > was referring to
> > > > > > > > earlier.
> > > > > > > 
> > > > > > > Yeah, I can see the need for a dedicated lock for the
> > > > > > > GEM's gpuva
> > > > > > > list, but
> > > > > > > holding the vm's dma-resv lock across the unlink
> > > > > > > shouldn't be a
> > > > > > > problem. We
> > > > > > > may free the object and a pointer to the vm's resv during
> > > > > > > unlink
> > > > > > > but we
> > > > > > > don't free the vm's resv.  It'd be a matter of ensuring
> > > > > > > that any
> > > > > > > calls to
> > > > > > > unlink from *within* drm_gpuvm allows it to be held.
> > > > > > 
> > > > > > Drivers calling unlink() from the fence signaling path
> > > > > > can't use the
> > > > > > VM's
> > > > > > dma-resv lock.
> > > > > 
> > > > > Yes, that made me a bit curious because in the current
> > > > > version the code
> > > > > required the object's dma_resv for unlink() which can't be
> > > > > grabbed
> > > > > either from the fence signaling path. So are there any
> > > > > drivers actually
> > > > > wanting to do that? If so, they will either need to resort to
> > > > > the
> > > > > current spinlock solution or they will need to call unlink
> > > > > from a
> > > > > workqueue item.
> > > > 
> > > > As Boris already mentioned we have the dma-resv lock by default
> > > > or a driver
> > > > specific GEM gpuva lock as opt-in. Now, we can get rid of the
> > > > latter.
> > > > 
> > > > > > 
> > > > > > Also, what if the object is an external object? We can't
> > > > > > use the VM's
> > > > > > dma-resv
> > > > > > lock here.
> > > > > 
> > > > > Why? Typically (sync) unlink is only ever called from an
> > > > > unbind-like
> > > > > operation where it should be trivial to grab the vm's resv.
> > > > > Or, for
> > > > > that matter any outer lock protecting the extobj list. Rule
> > > > > would be
> > > > > the drm_gpuvm_bo::entry::extobj  and
> > > > > drm_gpuvm_bo::entry::evict would
> > > > > be protected by either the vm's dma_resv (or possibly an
> > > > > outer lock in
> > > > > the case of the extobj list).
> > > > 
> > > > Outer lock wouldn't have been working for updates in the async
> > > > path, but
> > > > shouldn't be relevant anymore. We could use the VM's resv for
> > > > that.
> > > > 
> > > > > 
> > > > > >   And we can't have the GEM objs dma-resv lock held when
> > > > > > calling
> > > > > > unlink(), since unlink() calls drm_gpuvm_bo_put(), which if
> > > > > > the
> > > > > > refcount drops
> > > > > > to zero calls drm_gpuvm_bo_destroy() and
> > > > > > drm_gpuvm_bo_destroy() might
> > > > > > drop the
> > > > > > last reference of the GEM object.
> > > > > 
> > > > > Yes, but this is a different problem as to what exactly
> > > > > protects
> > > > > drm_gpuvm_bo::entry::gem. Either as you suggest an internal
> > > > > per bo list
> > > > > lock, or if we want to keep the bo's dma_resv we need to
> > > > > ensure that
> > > > > the caller of dma_resv_unlock(obj->resv) actually refcounts
> > > > > its obj
> > > > > pointer, and doesn't implicitly rely on the gpuvm_bo's
> > > > > refcount (I know
> > > > > Boris didn't like that, but requiring an explicit refcount
> > > > > for a
> > > > > pointer you dereference unless you're under a lock that
> > > > > ensures keeping
> > > > > the object alive is pretty much required?) But anyway for the
> > > > > drm_gpuvm_bo::entry::gem list protection (bo resv or internal
> > > > > spinlock)
> > > > > I don't have a strong preference.
> > > > 
> > > > We can keep the GEM objects dma-resv lock, however as mentioned
> > > > above
> > > > drm_gpuva_unlink() and drm_gpuvm_bo_put() then requires both
> > > > the VM's resv lock
> > > > and the GEM's resv lock in case they differ.
> > > > 
> > > 
> > > > > > > 
> > > 
> > 
> 



More information about the dri-devel mailing list