[Intel-gfx] [PATCH 1/2] drm/i915: refactor i915_gem_object_pin_map()

[Dave Gordon]

On 19/04/16 20:50, Chris Wilson wrote:
> On Tue, Apr 19, 2016 at 06:40:07PM +0100, Dave Gordon wrote:
>> From: Alex Dai <yu.dai at intel.com>
>>
>> The recently-added i915_gem_object_pin_map() can be further optimised
>> for "small" objects. To facilitate this, and simplify the error paths
>> before adding the new code, this patch pulls out the "mapping" part of
>> the operation (involving local allocations which must be undone before
>> return) into its own subfunction.
>>
>> The next patch will then insert the new optimisation into the middle of
>> the now-separated subfunction.
>>
>> This reorganisation will probably not affect the generated code, as the
>> compiler will most likely inline it anyway, but it makes the logical
>> structure a bit clearer and easier to modify.
>>
>> Signed-off-by: Dave Gordon <david.s.gordon at intel.com>
>> Cc: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> Cc: Chris Wilson <chris at chris-wilson.co.uk>
>> ---
>>   drivers/gpu/drm/i915/i915_gem.c | 61 +++++++++++++++++++++++++++--------------
>>   1 file changed, 40 insertions(+), 21 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c
>> index 6ce2c31..fc42be0 100644
>> --- a/drivers/gpu/drm/i915/i915_gem.c
>> +++ b/drivers/gpu/drm/i915/i915_gem.c
>> @@ -2396,6 +2396,45 @@ static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj)
>>   	return 0;
>>   }
>>
>> +/* The 'mapping' part of i915_gem_object_pin_map() below */
>> +static void *i915_gem_object_map(const struct drm_i915_gem_object *obj)
>> +{
>> +	unsigned long n_pages = obj->base.size >> PAGE_SHIFT;
>> +	struct scatterlist *sg = obj->pages->sgl;
>> +	struct sg_page_iter sg_iter;
>> +	struct page **pages;
>> +	unsigned long i = 0;
>> +	void *addr = NULL;
>> +
>> +	/* A single page can always be kmapped */
>> +	if (n_pages == 1)
>> +		return kmap(sg_page(sg));
>> +
>> +	pages = drm_malloc_gfp(n_pages, sizeof(*pages), GFP_TEMPORARY);
>> +	if (pages == NULL) {
>> +		DRM_DEBUG_DRIVER("Failed to get space for pages\n");
>> +		return NULL;
>> +	}
>> +
>> +	for_each_sg_page(sg, &sg_iter, n_pages, 0) {
>> +		pages[i] = sg_page_iter_page(&sg_iter);
>
> Just pages[i++] = sg_page_iter_page(&sg_iter);
>
>> +		if (++i == n_pages) {
>> +			addr = vmap(pages, n_pages, 0, PAGE_KERNEL);
>> +			break;
>> +		}
>> +	}
>> +
>> +	/* We should have got here via the 'break' above */
>> +	WARN_ON(i != n_pages);
>> +	if (addr == NULL)
>> +		DRM_DEBUG_DRIVER("Failed to vmap pages\n");
>
> As this is a very, very confused loop.
> -Chris

I tried that approach before, but it was actually more difficult to have 
tidy error-checking that way (remembering that we must always free the 
pages array, so don't really want an early return).

Here, putting the vmap() inside the final iteration of the loop means 
that we automatically leave "addr" as NULL if we don't reach the 
expected count. The subsequent WARN_ON() tells us that this has 
happened, but we don't then have to base any further branching on this 
condition (i != n_pages) as "addr" is already right. (Obviously, we 
don't want to do the vmap() if we have exited the loop with the wrong 
page count).

I'll post the other version, but I think the post-loop checking is 
messier, to such an extent that this way round is simpler overall.

.Dave.