When a driver specifies a beneficial order, TTM should make a reasonable effort to allocate pages at that order.
Use __GFP_RETRY_MAYFAIL instead of __GFP_NORETRY when allocating at the beneficial order. This allows reclaim to try harder before falling back to smaller orders, at the cost of longer allocation setup time. That tradeoff is acceptable for beneficial-order allocations: higher-order backing pages can improve TLB hit rates and reduce the number of TLB invalidations needed when moving memory. Cc: Christian Koenig <[email protected]> Cc: Huang Rui <[email protected]> Cc: Matthew Auld <[email protected]> Cc: Maarten Lankhorst <[email protected]> Cc: Maxime Ripard <[email protected]> Cc: Thomas Zimmermann <[email protected]> Cc: David Airlie <[email protected]> Cc: Simona Vetter <[email protected]> Cc: [email protected] Cc: [email protected] Cc: Thomas Hellström <[email protected]> Signed-off-by: Matthew Brost <[email protected]> --- Together with other Xe/core shrinker changes currently under development, this change significantly improved performance under memory fragmentation on PTL. In 3DMark Wild Life Extreme, performance improved by up to 120% compared to the baseline fragmented-memory case. GravityMark also showed approximately a 20% improvement under similar conditions. --- drivers/gpu/drm/ttm/ttm_pool.c | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/drivers/gpu/drm/ttm/ttm_pool.c b/drivers/gpu/drm/ttm/ttm_pool.c index 9f0ab99eb289..682ae4f40424 100644 --- a/drivers/gpu/drm/ttm/ttm_pool.c +++ b/drivers/gpu/drm/ttm/ttm_pool.c @@ -168,6 +168,11 @@ static struct page *ttm_pool_alloc_page(struct ttm_pool *pool, gfp_t gfp_flags, if (order && beneficial_order && order != beneficial_order) gfp_flags &= ~__GFP_RECLAIM; + if (beneficial_order && order == beneficial_order) { + gfp_flags &= ~__GFP_NORETRY; + gfp_flags |= __GFP_RETRY_MAYFAIL; + } + if (!ttm_pool_uses_dma_alloc(pool)) { p = alloc_pages_node(pool->nid, gfp_flags, order); if (p) { -- 2.34.1
