[Intel-gfx] [PATCH 1/3] drm/i915: Use pagecache write to prepopulate shmemfs from pwrite-ioctl
Joonas Lahtinen
joonas.lahtinen at linux.intel.com
Mon Mar 6 19:48:31 UTC 2017
On ma, 2017-03-06 at 09:29 +0000, Chris Wilson wrote:
> Before we instantiate/pin the backing store for our use, we
> can prepopulate the shmemfs filp efficiently using the a
> write into the pagecache. We avoid the penalty of instantiating
> all the pages, important if the user is just writing to a few
> and never uses the object on the GPU, and using a direct write
> into shmemfs allows it to avoid the cost of retrieving a page
> (either swapin or clearing-before-use) before it is overwritten.
>
> This can be extended later to provide additional specialisation for
> other backends (other than shmemfs).
>
> References: https://bugs.freedesktop.org/show_bug.cgi?id=99107
> Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Matthew Auld <matthew.william.auld at gmail.com>
> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> Cc: Mika Kuoppala <mika.kuoppala at linux.intel.com>
<SNIP>
> +static int
> +i915_gem_object_pwrite_gtt(struct drm_i915_gem_object *obj,
> + const struct drm_i915_gem_pwrite *arg)
> +{
> + struct address_space *mapping = obj->base.filp->f_mapping;
> + char __user *user_data = u64_to_user_ptr(arg->data_ptr);
> + u64 remain, offset;
> + unsigned int pg;
> +
> + /* Before we instantiate/pin the backing store for our use, we
> + * can prepopulate the shmemfs filp efficiently using the a
> + * write into the pagecache. We avoid the penalty of instantiating
> + * all the pages, important if the user is just writing to a few
> + * and never uses the object on the GPU, and using a direct write
> + * into shmemfs allows it to avoid the cost of retrieving a page
> + * (either swapin or clearing-before-use) before it is overwritten.
> + */
> + if (READ_ONCE(obj->mm.pages))
> + return -ENODEV;
> +
> + /* Before the pages are instantioted the object is treated as being
*instantiated
<SNIP>
> + err = pagecache_write_end(obj->base.filp, mapping, offset,
> + len, len - unwritten,
> + page, data);
Did you deliberately make the arguments different from write_begin? :)
If this solves a problem;
Reviewed-by: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
But should we instead fix libva?
Regards, Joonas
--
Joonas Lahtinen
Open Source Technology Center
Intel Corporation
More information about the Intel-gfx
mailing list