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.
