On 14/10/2016 14:53, Chris Wilson wrote:
On Fri, Oct 14, 2016 at 02:44:00PM +0100, Tvrtko Ursulin wrote:
On 14/10/2016 13:54, Chris Wilson wrote:
On Fri, Oct 14, 2016 at 01:42:35PM +0100, Tvrtko Ursulin wrote:
On 14/10/2016 13:18, Chris Wilson wrote:
+       if (IS_CRESTLINE(i915) || IS_BROADWATER(i915)) {
+               /* 965gm cannot relocate objects above 4GiB. */
+               gfp &= ~__GFP_HIGHMEM;
+               gfp |= __GFP_DMA32;
+       }
+       do {
+               int order = min(fls(npages) - 1, max_order);
I still have reservations on going back to max_order when previous
chunks required an order decrease. Size of the object is unbound
since it is indirectly controlled by userspace, correct? How about
decreasing the max_order on every repeated order decrease, following
failed order allocation?
+               struct page *page;
+               do {
+                       page = alloc_pages(gfp | (order ? QUIET : 0), order);
+                       if (page)
+                               break;
+                       if (!order--)
+                               goto err;
                        /* Limit future allocations as well */
                        max_order = order;

+               } while (1);
We do pass NORETRY | NOWARN for the higher order allocations, so it
shouldn't be as bad it seems?
I don't know for sure without looking into the implementation
details. But I assumed even with NORETRY it does some extra work to
try and free up the space. And if it fails, and we ask for it again,
it is just doing that extra work for nothing. Because within a
single allocation it sounds unlikely that something would change so
dramatically that it would start working.
iirc, NORETRY means abort after failure. In effect, it does
2 attempts from the freelist, a direct reclaim, and may then repeat
if the task's allowed set of nodes were concurrently changed.

Do you think it makes sense doing all that after it started failing, within our single get_pages allocation?



Intel-gfx mailing list

Reply via email to