| Message ID | 20240816135154.19678-2-nirmoy.das@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v6,1/2] drm/ttm: Add a flag to allow drivers to skip clear-on-free | expand |
On 16/08/2024 14:51, Nirmoy Das wrote: > On LNL because of flat CCS, driver creates migrates 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 clear > on free as XE now takes care of clearing pages. If a bo is in system > placement such as BO created with DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING > and there is a cpu map then for such BO gpu clear will be avoided as > there is no dma mapping for such BO at that moment to create migration > jobs. > > Tested this patch api_overhead_benchmark_l0 from > https://github.com/intel/compute-benchmarks > > Without the patch: > api_overhead_benchmark_l0 --testFilter=UsmMemoryAllocation: > UsmMemoryAllocation(api=l0 type=Host size=4KB) 84.206 us > UsmMemoryAllocation(api=l0 type=Host size=1GB) 105775.56 us > erf tool top 5 entries: > 71.44% api_overhead_be [kernel.kallsyms] [k] clear_page_erms > 6.34% api_overhead_be [kernel.kallsyms] [k] __pageblock_pfn_to_page > 2.24% api_overhead_be [kernel.kallsyms] [k] cpa_flush > 2.15% api_overhead_be [kernel.kallsyms] [k] pages_are_mergeable > 1.94% api_overhead_be [kernel.kallsyms] [k] find_next_iomem_res > > With the patch: > api_overhead_benchmark_l0 --testFilter=UsmMemoryAllocation: > UsmMemoryAllocation(api=l0 type=Host size=4KB) 79.439 us > UsmMemoryAllocation(api=l0 type=Host size=1GB) 98677.75 us > Perf tool top 5 entries: > 11.16% api_overhead_be [kernel.kallsyms] [k] __pageblock_pfn_to_page > 7.85% api_overhead_be [kernel.kallsyms] [k] cpa_flush > 7.59% api_overhead_be [kernel.kallsyms] [k] find_next_iomem_res > 7.24% api_overhead_be [kernel.kallsyms] [k] pages_are_mergeable > 5.53% api_overhead_be [kernel.kallsyms] [k] lookup_address_in_pgd_attr > > Without this patch clear_page_erms() dominates execution time which is > also not pipelined with migration jobs. With this patch page clearing > will get pipelined with migration job and will free CPU for more work. > > 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) > v5: s/gpu_page_clear/gpu_page_clear_sys and move setting > it to xe_ttm_sys_mgr_init() and other nits (Matt Auld) > v6: Disable it when init_on_alloc and/or init_on_free is active(Matt) > Use compute-benchmarks as reporter used it to report this > allocation latency issue also a proper test application than mime. > In v5, the test showed significant reduction in alloc latency but > that is not the case any more, I think this was mostly because > previous test was done on IFWI which had low mem BW from CPU. > > Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com> > Cc: Matthew Auld <matthew.auld@intel.com> > Cc: Matthew Brost <matthew.brost@intel.com> > Cc: "Thomas Hellström" <thomas.hellstrom@linux.intel.com> > Signed-off-by: Nirmoy Das <nirmoy.das@intel.com> Reviewed-by: Matthew Auld <matthew.auld@intel.com>
On 8/19/2024 1:05 PM, Matthew Auld wrote: > On 16/08/2024 14:51, Nirmoy Das wrote: >> On LNL because of flat CCS, driver creates migrates 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 clear >> on free as XE now takes care of clearing pages. If a bo is in system >> placement such as BO created with DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING >> and there is a cpu map then for such BO gpu clear will be avoided as >> there is no dma mapping for such BO at that moment to create migration >> jobs. >> >> Tested this patch api_overhead_benchmark_l0 from >> https://github.com/intel/compute-benchmarks >> >> Without the patch: >> api_overhead_benchmark_l0 --testFilter=UsmMemoryAllocation: >> UsmMemoryAllocation(api=l0 type=Host size=4KB) 84.206 us >> UsmMemoryAllocation(api=l0 type=Host size=1GB) 105775.56 us >> erf tool top 5 entries: >> 71.44% api_overhead_be [kernel.kallsyms] [k] clear_page_erms >> 6.34% api_overhead_be [kernel.kallsyms] [k] __pageblock_pfn_to_page >> 2.24% api_overhead_be [kernel.kallsyms] [k] cpa_flush >> 2.15% api_overhead_be [kernel.kallsyms] [k] pages_are_mergeable >> 1.94% api_overhead_be [kernel.kallsyms] [k] find_next_iomem_res >> >> With the patch: >> api_overhead_benchmark_l0 --testFilter=UsmMemoryAllocation: >> UsmMemoryAllocation(api=l0 type=Host size=4KB) 79.439 us >> UsmMemoryAllocation(api=l0 type=Host size=1GB) 98677.75 us >> Perf tool top 5 entries: >> 11.16% api_overhead_be [kernel.kallsyms] [k] __pageblock_pfn_to_page >> 7.85% api_overhead_be [kernel.kallsyms] [k] cpa_flush >> 7.59% api_overhead_be [kernel.kallsyms] [k] find_next_iomem_res >> 7.24% api_overhead_be [kernel.kallsyms] [k] pages_are_mergeable >> 5.53% api_overhead_be [kernel.kallsyms] [k] >> lookup_address_in_pgd_attr >> >> Without this patch clear_page_erms() dominates execution time which is >> also not pipelined with migration jobs. With this patch page clearing >> will get pipelined with migration job and will free CPU for more work. >> >> 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) >> v5: s/gpu_page_clear/gpu_page_clear_sys and move setting >> it to xe_ttm_sys_mgr_init() and other nits (Matt Auld) >> v6: Disable it when init_on_alloc and/or init_on_free is active(Matt) >> Use compute-benchmarks as reporter used it to report this >> allocation latency issue also a proper test application than mime. >> In v5, the test showed significant reduction in alloc latency but >> that is not the case any more, I think this was mostly because >> previous test was done on IFWI which had low mem BW from CPU. >> >> Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com> >> Cc: Matthew Auld <matthew.auld@intel.com> >> Cc: Matthew Brost <matthew.brost@intel.com> >> Cc: "Thomas Hellström" <thomas.hellstrom@linux.intel.com> >> Signed-off-by: Nirmoy Das <nirmoy.das@intel.com> > > Reviewed-by: Matthew Auld <matthew.auld@intel.com> Thanks Matt. Pushed this to drm-xe-next. The series contains a ttm pool change which as agreed with Christian is small enough to not cause any issue so can be pulled though drm-xe-next. Regards, Nirmoy
Hi, Nirmoy, On Mon, 2024-08-19 at 18:01 +0200, Nirmoy Das wrote: > > On 8/19/2024 1:05 PM, Matthew Auld wrote: > > On 16/08/2024 14:51, Nirmoy Das wrote: > > > On LNL because of flat CCS, driver creates migrates 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 > > > clear > > > on free as XE now takes care of clearing pages. If a bo is in > > > system > > > placement such as BO created with > > > DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING > > > and there is a cpu map then for such BO gpu clear will be avoided > > > as > > > there is no dma mapping for such BO at that moment to create > > > migration > > > jobs. > > > > > > Tested this patch api_overhead_benchmark_l0 from > > > https://github.com/intel/compute-benchmarks > > > > > > Without the patch: > > > api_overhead_benchmark_l0 --testFilter=UsmMemoryAllocation: > > > UsmMemoryAllocation(api=l0 type=Host size=4KB) 84.206 us > > > UsmMemoryAllocation(api=l0 type=Host size=1GB) 105775.56 us > > > erf tool top 5 entries: > > > 71.44% api_overhead_be [kernel.kallsyms] [k] clear_page_erms > > > 6.34% api_overhead_be [kernel.kallsyms] [k] > > > __pageblock_pfn_to_page > > > 2.24% api_overhead_be [kernel.kallsyms] [k] cpa_flush > > > 2.15% api_overhead_be [kernel.kallsyms] [k] > > > pages_are_mergeable > > > 1.94% api_overhead_be [kernel.kallsyms] [k] > > > find_next_iomem_res > > > > > > With the patch: > > > api_overhead_benchmark_l0 --testFilter=UsmMemoryAllocation: > > > UsmMemoryAllocation(api=l0 type=Host size=4KB) 79.439 us > > > UsmMemoryAllocation(api=l0 type=Host size=1GB) 98677.75 us > > > Perf tool top 5 entries: > > > 11.16% api_overhead_be [kernel.kallsyms] [k] > > > __pageblock_pfn_to_page > > > 7.85% api_overhead_be [kernel.kallsyms] [k] cpa_flush > > > 7.59% api_overhead_be [kernel.kallsyms] [k] > > > find_next_iomem_res > > > 7.24% api_overhead_be [kernel.kallsyms] [k] > > > pages_are_mergeable > > > 5.53% api_overhead_be [kernel.kallsyms] [k] > > > lookup_address_in_pgd_attr > > > > > > Without this patch clear_page_erms() dominates execution time > > > which is > > > also not pipelined with migration jobs. With this patch page > > > clearing > > > will get pipelined with migration job and will free CPU for more > > > work. > > > > > > 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) > > > v5: s/gpu_page_clear/gpu_page_clear_sys and move setting > > > it to xe_ttm_sys_mgr_init() and other nits (Matt Auld) > > > v6: Disable it when init_on_alloc and/or init_on_free is > > > active(Matt) > > > Use compute-benchmarks as reporter used it to report this > > > allocation latency issue also a proper test application than > > > mime. > > > In v5, the test showed significant reduction in alloc > > > latency but > > > that is not the case any more, I think this was mostly > > > because > > > previous test was done on IFWI which had low mem BW from > > > CPU. > > > > > > Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com> > > > Cc: Matthew Auld <matthew.auld@intel.com> > > > Cc: Matthew Brost <matthew.brost@intel.com> > > > Cc: "Thomas Hellström" <thomas.hellstrom@linux.intel.com> > > > Signed-off-by: Nirmoy Das <nirmoy.das@intel.com> > > > > Reviewed-by: Matthew Auld <matthew.auld@intel.com> > > > Thanks Matt. > > Pushed this to drm-xe-next. The series contains a ttm pool change > which > as agreed with Christian > > is small enough to not cause any issue so can be pulled though drm- > xe-next. I have a question that was sent as a reply-to on that patch. Thanks, Thomas > > > Regards, > > Nirmoy >
diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c index 800119c8fc8d..1248acbdacfc 100644 --- a/drivers/gpu/drm/xe/xe_bo.c +++ b/drivers/gpu/drm/xe/xe_bo.c @@ -396,6 +396,14 @@ static struct ttm_tt *xe_ttm_tt_create(struct ttm_buffer_object *ttm_bo, caching = ttm_uncached; } + /* + * If the device can support gpu clear system pages then set proper ttm + * flag. Zeroed pages are only required for ttm_bo_type_device so + * unwanted data is not leaked to userspace. + */ + if (ttm_bo->type == ttm_bo_type_device && xe->mem.gpu_page_clear_sys) + 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); @@ -417,6 +425,10 @@ 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 system 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; @@ -659,8 +671,16 @@ static int xe_bo_move(struct ttm_buffer_object *ttm_bo, bool evict, bool needs_clear; bool handle_system_ccs = (!IS_DGFX(xe) && xe_bo_needs_ccs_pages(bo) && ttm && ttm_tt_is_populated(ttm)) ? true : false; + bool clear_system_pages; int ret = 0; + /* + * Clear TTM_TT_FLAG_CLEARED_ON_FREE on bo creation path when + * moving to system as the bo doesn't have 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) @@ -683,8 +703,10 @@ static int xe_bo_move(struct ttm_buffer_object *ttm_bo, bool evict, move_lacks_source = handle_system_ccs ? (!bo->ccs_cleared) : (!mem_type_is_vram(old_mem_type) && !tt_has_data); + clear_system_pages = ttm && (ttm->page_flags & TTM_TT_FLAG_CLEARED_ON_FREE); 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) || + clear_system_pages; if (new_mem->mem_type == XE_PL_TT) { ret = xe_tt_map_sg(ttm); @@ -796,7 +818,7 @@ static int xe_bo_move(struct ttm_buffer_object *ttm_bo, bool evict, if (move_lacks_source) { u32 flags = 0; - if (mem_type_is_vram(new_mem->mem_type)) + if (mem_type_is_vram(new_mem->mem_type) || clear_system_pages) flags |= XE_MIGRATE_CLEAR_FLAG_FULL; else if (handle_system_ccs) flags |= XE_MIGRATE_CLEAR_FLAG_CCS_DATA; diff --git a/drivers/gpu/drm/xe/xe_device_types.h b/drivers/gpu/drm/xe/xe_device_types.h index 16a24eadd94b..2aa18e32b32f 100644 --- a/drivers/gpu/drm/xe/xe_device_types.h +++ b/drivers/gpu/drm/xe/xe_device_types.h @@ -339,6 +339,8 @@ struct xe_device { struct xe_mem_region vram; /** @mem.sys_mgr: system TTM manager */ struct ttm_resource_manager sys_mgr; + /** @mem.gpu_page_clear_sys: clear system pages offloaded to GPU */ + bool gpu_page_clear_sys; } mem; /** @sriov: device level virtualization data */ diff --git a/drivers/gpu/drm/xe/xe_ttm_sys_mgr.c b/drivers/gpu/drm/xe/xe_ttm_sys_mgr.c index 9844a8edbfe1..e0ac20f20758 100644 --- a/drivers/gpu/drm/xe/xe_ttm_sys_mgr.c +++ b/drivers/gpu/drm/xe/xe_ttm_sys_mgr.c @@ -117,5 +117,17 @@ int xe_ttm_sys_mgr_init(struct xe_device *xe) ttm_resource_manager_init(man, &xe->ttm, gtt_size >> PAGE_SHIFT); ttm_set_driver_manager(&xe->ttm, XE_PL_TT, man); ttm_resource_manager_set_used(man, true); + + /* + * On iGFX device with flat CCS, we clear CCS metadata, let's extend that + * and use GPU to clear pages as well. + * + * Disable this when init_on_free and/or init_on_alloc is on to avoid double + * zeroing pages with CPU and GPU. + */ + if (xe_device_has_flat_ccs(xe) && !IS_DGFX(xe) && + !want_init_on_alloc(GFP_USER) && !want_init_on_free()) + xe->mem.gpu_page_clear_sys = true; + return drmm_add_action_or_reset(&xe->drm, ttm_sys_mgr_fini, xe); }
On LNL because of flat CCS, driver creates migrates 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 clear on free as XE now takes care of clearing pages. If a bo is in system placement such as BO created with DRM_XE_GEM_CREATE_FLAG_DEFER_BACKING and there is a cpu map then for such BO gpu clear will be avoided as there is no dma mapping for such BO at that moment to create migration jobs. Tested this patch api_overhead_benchmark_l0 from https://github.com/intel/compute-benchmarks Without the patch: api_overhead_benchmark_l0 --testFilter=UsmMemoryAllocation: UsmMemoryAllocation(api=l0 type=Host size=4KB) 84.206 us UsmMemoryAllocation(api=l0 type=Host size=1GB) 105775.56 us erf tool top 5 entries: 71.44% api_overhead_be [kernel.kallsyms] [k] clear_page_erms 6.34% api_overhead_be [kernel.kallsyms] [k] __pageblock_pfn_to_page 2.24% api_overhead_be [kernel.kallsyms] [k] cpa_flush 2.15% api_overhead_be [kernel.kallsyms] [k] pages_are_mergeable 1.94% api_overhead_be [kernel.kallsyms] [k] find_next_iomem_res With the patch: api_overhead_benchmark_l0 --testFilter=UsmMemoryAllocation: UsmMemoryAllocation(api=l0 type=Host size=4KB) 79.439 us UsmMemoryAllocation(api=l0 type=Host size=1GB) 98677.75 us Perf tool top 5 entries: 11.16% api_overhead_be [kernel.kallsyms] [k] __pageblock_pfn_to_page 7.85% api_overhead_be [kernel.kallsyms] [k] cpa_flush 7.59% api_overhead_be [kernel.kallsyms] [k] find_next_iomem_res 7.24% api_overhead_be [kernel.kallsyms] [k] pages_are_mergeable 5.53% api_overhead_be [kernel.kallsyms] [k] lookup_address_in_pgd_attr Without this patch clear_page_erms() dominates execution time which is also not pipelined with migration jobs. With this patch page clearing will get pipelined with migration job and will free CPU for more work. 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) v5: s/gpu_page_clear/gpu_page_clear_sys and move setting it to xe_ttm_sys_mgr_init() and other nits (Matt Auld) v6: Disable it when init_on_alloc and/or init_on_free is active(Matt) Use compute-benchmarks as reporter used it to report this allocation latency issue also a proper test application than mime. In v5, the test showed significant reduction in alloc latency but that is not the case any more, I think this was mostly because previous test was done on IFWI which had low mem BW from CPU. Cc: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com> Cc: Matthew Auld <matthew.auld@intel.com> Cc: Matthew Brost <matthew.brost@intel.com> Cc: "Thomas Hellström" <thomas.hellstrom@linux.intel.com> Signed-off-by: Nirmoy Das <nirmoy.das@intel.com> --- drivers/gpu/drm/xe/xe_bo.c | 26 ++++++++++++++++++++++++-- drivers/gpu/drm/xe/xe_device_types.h | 2 ++ drivers/gpu/drm/xe/xe_ttm_sys_mgr.c | 12 ++++++++++++ 3 files changed, 38 insertions(+), 2 deletions(-)