[PATCH v5 4/4] drm/xe/lnl: Offload system clear page activity to GPU

Thu Jul 4 08:18:41 UTC 2024

On LNL because of flat CCS, driver creates a migrate job to clear
CCS meta data. Extend that to also clear system pages using GPU.
Inform TTM to allocate pages without __GFP_ZERO to avoid double page
clearing by clearing out TTM_TT_FLAG_ZERO_ALLOC flag and set
TTM_TT_FLAG_CLEARED_ON_FREE while freeing to skip ttm pool's
clearn-on-free as XE now takes care of clearing pages. If a bo
is in system placement and there is a cpu map then for such BO gpu
clear will be avoided as there is no dma mapping for such BO.

To test the patch, created a small test that tries to submit a job
after binding various sizes of buffer which shows good gains for larger
buffer. For lower buffer sizes, the result is not very reliable as the
results vary a lot.

With the patch
sudo  ~/igt-gpu-tools/build/tests/xe_exec_store --run
basic-store-benchmark
IGT-Version: 1.28-g2ed908c0b (x86_64) (Linux: 6.10.0-rc2-xe+ x86_64)
Using IGT_SRANDOM=1719237905 for randomisation
Opened device: /dev/dri/card0
Starting subtest: basic-store-benchmark
Starting dynamic subtest: WC
Dynamic subtest WC: SUCCESS (0.000s)
Time taken for size SZ_4K: 9493 us
Time taken for size SZ_2M: 5503 us
Time taken for size SZ_64M: 13016 us
Time taken for size SZ_128M: 29464 us
Time taken for size SZ_256M: 38408 us
Time taken for size SZ_1G: 148758 us
Starting dynamic subtest: WB
Dynamic subtest WB: SUCCESS (0.000s)
Time taken for size SZ_4K: 3889 us
Time taken for size SZ_2M: 6091 us
Time taken for size SZ_64M: 20920 us
Time taken for size SZ_128M: 32394 us
Time taken for size SZ_256M: 61710 us
Time taken for size SZ_1G: 215437 us
Subtest basic-store-benchmark: SUCCESS (0.589s)

With the patch:
sudo  ~/igt-gpu-tools/build/tests/xe_exec_store --run
basic-store-benchmark
IGT-Version: 1.28-g2ed908c0b (x86_64) (Linux: 6.10.0-rc2-xe+ x86_64)
Using IGT_SRANDOM=1719238062 for randomisation
Opened device: /dev/dri/card0
Starting subtest: basic-store-benchmark
Starting dynamic subtest: WC
Dynamic subtest WC: SUCCESS (0.000s)
Time taken for size SZ_4K: 11803 us
Time taken for size SZ_2M: 4237 us
Time taken for size SZ_64M: 8649 us
Time taken for size SZ_128M: 14682 us
Time taken for size SZ_256M: 22156 us
Time taken for size SZ_1G: 74457 us
Starting dynamic subtest: WB
Dynamic subtest WB: SUCCESS (0.000s)
Time taken for size SZ_4K: 5129 us
Time taken for size SZ_2M: 12563 us
Time taken for size SZ_64M: 14860 us
Time taken for size SZ_128M: 26064 us
Time taken for size SZ_256M: 47167 us
Time taken for size SZ_1G: 170304 us
Subtest basic-store-benchmark: SUCCESS (0.417s)

With the patch and init_on_alloc=0
sudo  ~/igt-gpu-tools/build/tests/xe_exec_store --run
basic-store-benchmark
IGT-Version: 1.28-g2ed908c0b (x86_64) (Linux: 6.10.0-rc2-xe+ x86_64)
Using IGT_SRANDOM=1719238219 for randomisation
Opened device: /dev/dri/card0
Starting subtest: basic-store-benchmark
Starting dynamic subtest: WC
Dynamic subtest WC: SUCCESS (0.000s)
Time taken for size SZ_4K: 4803 us
Time taken for size SZ_2M: 9212 us
Time taken for size SZ_64M: 9643 us
Time taken for size SZ_128M: 13479 us
Time taken for size SZ_256M: 22429 us
Time taken for size SZ_1G: 83110 us
Starting dynamic subtest: WB
Dynamic subtest WB: SUCCESS (0.000s)
Time taken for size SZ_4K: 4003 us
Time taken for size SZ_2M: 4443 us
Time taken for size SZ_64M: 12960 us
Time taken for size SZ_128M: 13741 us
Time taken for size SZ_256M: 26841 us
Time taken for size SZ_1G: 84746 us
Subtest basic-store-benchmark: SUCCESS (0.290s)

v2: Handle regression on dgfx(Himal)
    Update commit message as no ttm API changes needed.
v3: Fix Kunit test.
v4: handle data leak on cpu mmap(Thomas)

Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray at intel.com>
Cc: Matthew Auld <matthew.auld at intel.com>
Cc: "Thomas Hellström" <thomas.hellstrom at linux.intel.com>
Signed-off-by: Nirmoy Das <nirmoy.das at intel.com>
---
 drivers/gpu/drm/xe/xe_bo.c           | 25 ++++++++++++++++++++++++-
 drivers/gpu/drm/xe/xe_device.c       |  7 +++++++
 drivers/gpu/drm/xe/xe_device_types.h |  2 ++
 3 files changed, 33 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
index 4d6315d2ae9a..b76a44fcf3b1 100644
--- a/drivers/gpu/drm/xe/xe_bo.c
+++ b/drivers/gpu/drm/xe/xe_bo.c
@@ -387,6 +387,13 @@ static struct ttm_tt *xe_ttm_tt_create(struct ttm_buffer_object *ttm_bo,
 		caching = ttm_uncached;
 	}
 
+	/* If the device can support gpu clear pages then set proper ttm
+	 * flag. Zeroed pages are only required for ttm_bo_type_device so
+	 * unwanted data is leaked to userspace.
+	 */
+	if (ttm_bo->type == ttm_bo_type_device && xe->mem.gpu_page_clear)
+		page_flags |= TTM_TT_FLAG_CLEARED_ON_FREE;
+
 	err = ttm_tt_init(&tt->ttm, &bo->ttm, page_flags, caching, extra_pages);
 	if (err) {
 		kfree(tt);
@@ -408,6 +415,10 @@ static int xe_ttm_tt_populate(struct ttm_device *ttm_dev, struct ttm_tt *tt,
 	if (tt->page_flags & TTM_TT_FLAG_EXTERNAL)
 		return 0;
 
+	/* Clear TTM_TT_FLAG_ZERO_ALLOC when GPU is set to clear pages */
+	if (tt->page_flags & TTM_TT_FLAG_CLEARED_ON_FREE)
+		tt->page_flags &= ~TTM_TT_FLAG_ZERO_ALLOC;
+
 	err = ttm_pool_alloc(&ttm_dev->pool, tt, ctx);
 	if (err)
 		return err;
@@ -653,6 +664,14 @@ static int xe_bo_move(struct ttm_buffer_object *ttm_bo, bool evict,
 
 	int ret = 0;
 
+	/*
+	 * Clear TTM_TT_FLAG_CLEARED_ON_FREE on bo creation path when
+	 * moving to system as the bo doesn't dma_mapping.
+	 */
+	if (!old_mem && ttm && !ttm_tt_is_populated(ttm)) {
+		ttm->page_flags &= ~TTM_TT_FLAG_CLEARED_ON_FREE;
+	}
+
 	/* Bo creation path, moving to system or TT. */
 	if ((!old_mem && ttm) && !handle_system_ccs) {
 		if (new_mem->mem_type == XE_PL_TT)
@@ -676,7 +695,8 @@ static int xe_bo_move(struct ttm_buffer_object *ttm_bo, bool evict,
 						(!mem_type_is_vram(old_mem_type) && !tt_has_data);
 
 	needs_clear = (ttm && ttm->page_flags & TTM_TT_FLAG_ZERO_ALLOC) ||
-		(!ttm && ttm_bo->type == ttm_bo_type_device);
+		(!ttm && ttm_bo->type == ttm_bo_type_device) ||
+		(ttm && ttm->page_flags & TTM_TT_FLAG_CLEARED_ON_FREE);
 
 	if (new_mem->mem_type == XE_PL_TT) {
 		ret = xe_tt_map_sg(ttm);
@@ -790,6 +810,9 @@ static int xe_bo_move(struct ttm_buffer_object *ttm_bo, bool evict,
 					 handle_system_ccs;
 			bool clear_bo_data = mem_type_is_vram(new_mem->mem_type);
 
+			if (ttm && (ttm->page_flags & TTM_TT_FLAG_CLEARED_ON_FREE))
+				clear_bo_data |= true;
+
 			fence = xe_migrate_clear(migrate, bo, new_mem,
 						 clear_bo_data, clear_ccs);
 		}
diff --git a/drivers/gpu/drm/xe/xe_device.c b/drivers/gpu/drm/xe/xe_device.c
index 03492fbcb8fb..7c682a53f06e 100644
--- a/drivers/gpu/drm/xe/xe_device.c
+++ b/drivers/gpu/drm/xe/xe_device.c
@@ -636,6 +636,13 @@ int xe_device_probe(struct xe_device *xe)
 	if (err)
 		goto err;
 
+	/**
+	 * On iGFX device with flat CCS, we clear CCS metadata, let's extend that
+	 * and use GPU to clear pages as well.
+	 */
+	if (xe_device_has_flat_ccs(xe) && !IS_DGFX(xe))
+		xe->mem.gpu_page_clear = true;
+
 	err = xe_vram_probe(xe);
 	if (err)
 		goto err;
diff --git a/drivers/gpu/drm/xe/xe_device_types.h b/drivers/gpu/drm/xe/xe_device_types.h
index 3bca6d344744..28eaf2ab1f25 100644
--- a/drivers/gpu/drm/xe/xe_device_types.h
+++ b/drivers/gpu/drm/xe/xe_device_types.h
@@ -325,6 +325,8 @@ struct xe_device {
 		struct xe_mem_region vram;
 		/** @mem.sys_mgr: system TTM manager */
 		struct ttm_resource_manager sys_mgr;
+		/** @gpu_page_clear: clear pages offloaded to GPU */
+		bool gpu_page_clear;
 	} mem;
 
 	/** @sriov: device level virtualization data */
-- 
2.42.0

