[Intel-gfx] [PATCH 07/10] drm/i915: Support for pread/pwrite from/to non shmem backed objects

[Tvrtko Ursulin]

On 11/01/16 17:03, Chris Wilson wrote:
> On Mon, Jan 11, 2016 at 03:11:07PM +0000, Tvrtko Ursulin wrote:
>>
>> On 11/01/16 14:45, Chris Wilson wrote:
>>> On Mon, Jan 11, 2016 at 02:21:33PM +0000, Tvrtko Ursulin wrote:
>>>>
>>>> On 22/12/15 17:40, Chris Wilson wrote:
>>>>> On Tue, Dec 22, 2015 at 11:58:33AM +0000, Tvrtko Ursulin wrote:
>>>>>> Maybe:
>>>>>>
>>>>>> 	if (!obj->base.filp || cpu_write_needs_clflush(obj))
>>>>>>     		ret = i915_gem_gtt_pwrite_fast(...);
>>>>>>
>>>>>> 	if (ret == -EFAULT && !obj->base.filp) {
>>>>>> 		ret = i915_gem_gtt_pwrite_slow(...) /* New function, doing the
>>>>>> slow_user_access loop for !filp objects, extracted from
>>>>>> gtt_pwrite_fast above. */
>>>>>
>>>>> The point is that "gtt_pwrite_slow" is going to be preferrable in the
>>>>> cases where it is possible. It just wasn't the full fallback patch for
>>>>> all objects previously, so we didn't bother to write a partial fallback
>>>>> handler.
>>>>
>>>> Maybe I don't get this - is fast_user_write expected always to fail
>>>> for non shmem backed objects? And so revert to the slow_user_path
>>>> always and immediately? Because fast_user_write is still the primary
>>>> choice for everything.
>>>
>>> If we already have a GTT mapping available, then WC writes into the
>>> object are about as fast as we can get, especially if we don't have
>>> direct page access. They also have the benefit of not polluting the
>>> cache further - though that maybe a downside as well, in which case
>>> pwrite/pread was the wrong interface to use.
>>>
>>> fast_user_write is no more likely to fail for stolen objs than for
>>> shmemfs obj, it is just that we cannot fallback to direct page access
>>> for stolen objs and so need a fallback path that writes through the GTT.
>>> That fallback path would also be preferrable to falling back from the
>>> middle of a GTT write to the direct page paths. The issue was simply
>>> that the GTT paths cannot be assumed to be universally available,
>>> whereas historically the direct page access paths were. *That* changes,
>>> and now we cannot rely on either path being universally available.
>>
>> So it sounds that we don't need to have code which falls back in the
>> middle of the write but could be written cleaner as separate
>> helpers?
>>
>> Because I really dislike that new loop...
> 
> What new loop? We don't need a new loop...
> 
> i915_gem_gtt_pwrite():
> 	/* Important and exceedingly complex setup/teardown code
> 	 * removed for brevity.
> 	 */
> 	for_each_page() {
> 		... get limits of operation in page...
> 
> 		if (fast_gtt_write(##args)) {
> 			/* Beware dragons */
> 			mutex_unlock();
> 			hit_slow_path = 1;
> 			slow_gtt_write(##args);
> 			mutex_lock();
> 		}
> 	}
> 
> 	if (hit_slow_path) {
> 		/* Beware dragons that bite */
> 		ret = i915_gem_object_set_to_gtt_domain(obj, true);
> 	}
> 
> Is that not what was written? I take it my telepathy isn't working
> again.

Sorry not a new loop, new case in a old loop. This is the hunk I think
is not helping readability:

@@ -869,11 +967,29 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_private *i915,
 		/* If we get a fault while copying data, then (presumably) our
 		 * source page isn't available.  Return the error and we'll
 		 * retry in the slow path.
+		 * If the object is non-shmem backed, we retry again with the
+		 * path that handles page fault.
 		 */
-		if (fast_user_write(i915->gtt.mappable, page_base,
-				    page_offset, user_data, page_length)) {
-			ret = -EFAULT;
-			goto out_flush;
+		if (faulted || fast_user_write(i915->gtt.mappable,
+						page_base, page_offset,
+						user_data, page_length)) {
+			if (!obj->base.filp) {
+				faulted = true;
+				mutex_unlock(&dev->struct_mutex);
+				if (slow_user_access(i915->gtt.mappable,
+						     page_base,
+						     page_offset, user_data,
+						     page_length, true)) {
+					ret = -EFAULT;
+					mutex_lock(&dev->struct_mutex);
+					goto out_flush;
+				}
+
+				mutex_lock(&dev->struct_mutex);
+			} else {
+				ret = -EFAULT;
+				goto out_flush;
+			}

Because the concept is now different for page faults on shmem based and
non-shmem based objects. Former falls out on fault and ends up in
i915_gem_shmem_pwrite, while latter keeps banging on in 
i915_gem_gtt_pwrite_fast.

I find it confusing code organization and naming. So I suggested the
new path (!shmem + fault) is added as a separate new function and called
from i915_gem_pwrite_ioctl same as i915_gem_shmem_pwrite but you
objected:

    if (!obj->base.filp || cpu_write_needs_clflush(obj))
           ret = i915_gem_gtt_pwrite_fast(...);

    if (ret == -EFAULT && !obj->base.filp) {
        ret = i915_gem_gtt_pwrite_slow(...) /* New function, doing the slow_user_access loop for !filp objects, extracted from gtt_pwrite_fast above. */
    } else if (ret == -EFAULT || ret == -ENOSPC) {
        if (obj->phys_handle)
            ...
        ...

Regards,

Tvrtko