<html><head> <meta http-equiv="Content-Type" content="text/html; charset=utf-8"> </head> <body> <div class="moz-cite-prefix">On 2024-08-21 19:22, Ramesh Errabolu wrote: </div> <blockquote type="cite" cite="mid:20240821232252.230791-1-Ramesh.Errabolu@amd.com"> <pre class="moz-quote-pre" wrap="">KFD's design of unified memory (UM) does not allow users to configure the size of buffer used in migrating buffer either from Sysmem to VRAM or vice versa. </pre> </blockquote> This is not true, app can change range granularity attribute, to configure the buffer migration size. This module parameter is used to config the default range granularity. <blockquote type="cite" cite="mid:20240821232252.230791-1-Ramesh.Errabolu@amd.com"> <pre class="moz-quote-pre" wrap="">This patch remedies this gap, albeit at a coarse level, for workloads that deal with unregistered memory Signed-off-by: Ramesh Errabolu <a class="moz-txt-link-rfc2396E" href="mailto:Ramesh.Errabolu@amd.com"><Ramesh.Errabolu@amd.com></a> --- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 4 ++++ drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 16 ++++++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 12 ++++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 29 +++++++++++++++++-------- 4 files changed, 52 insertions(+), 9 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h index e8c284aea1f2..73dd816b01f2 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h @@ -237,6 +237,7 @@ extern int sched_policy; extern bool debug_evictions; extern bool no_system_mem_limit; extern int halt_if_hws_hang; +extern uint amdgpu_svm_attr_gobm;</pre> </blockquote> <blockquote type="cite" cite="mid:20240821232252.230791-1-Ramesh.Errabolu@amd.com"> <pre class="moz-quote-pre" wrap=""> #else static const int __maybe_unused sched_policy = KFD_SCHED_POLICY_HWS; static const bool __maybe_unused debug_evictions; /* = false */ @@ -313,6 +314,9 @@ extern int amdgpu_wbrf; /* Extra time delay(in ms) to eliminate the influence of temperature momentary fluctuation */ #define AMDGPU_SWCTF_EXTRA_DELAY 50 +/* Default size of buffer to use in migrating buffer */ +#define AMDGPU_SVM_ATTR_GOBM 9 + struct amdgpu_xcp_mgr; struct amdgpu_device; struct amdgpu_irq_src; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c index b9529948f2b2..e195e1cf0f28 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c @@ -169,6 +169,15 @@ uint amdgpu_sdma_phase_quantum = 32; char *amdgpu_disable_cu; char *amdgpu_virtual_display; bool enforce_isolation; + +/* Specifies the default size of buffer to use in + * migrating buffer from Sysmem to VRAM and vice + * versa + * + * Defined as log2(sizeof(buffer)/PAGE_SIZE) + */ +uint amdgpu_svm_attr_gobm = AMDGPU_SVM_ATTR_GOBM;</pre> </blockquote> /* add explanation, GOBM : granularity of buffer migration Use u8 type, the same type used in prange->granularity u8 amdgpu_svm_default_gobm = AMDGPU_SVM_DEFAULT_GOBM <blockquote type="cite" cite="mid:20240821232252.230791-1-Ramesh.Errabolu@amd.com"> <pre class="moz-quote-pre" wrap=""> + /* * OverDrive(bit 14) disabled by default * GFX DCS(bit 19) disabled by default @@ -320,6 +329,13 @@ module_param_named(pcie_gen2, amdgpu_pcie_gen2, int, 0444); MODULE_PARM_DESC(msi, "MSI support (1 = enable, 0 = disable, -1 = auto)"); module_param_named(msi, amdgpu_msi, int, 0444); +/** + * DOC: svm_attr_gobm (uint) + * Size of buffer to use in migrating buffer from Sysmem to VRAM and vice versa + */ +MODULE_PARM_DESC(svm_attr_gobm, "Defined as log2(sizeof(buffer)/PAGE_SIZE), e.g. 9 for 2 MiB"); +module_param_named(svm_attr_gobm, amdgpu_svm_attr_gobm, uint, 0644); + /** * DOC: lockup_timeout (string) * Set GPU scheduler timeout value in ms. diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h index 7bba6bed2f48..07b202ab008a 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h @@ -866,6 +866,18 @@ struct svm_range_list { struct delayed_work restore_work; DECLARE_BITMAP(bitmap_supported, MAX_GPU_INSTANCE); struct task_struct *faulting_task; + + /* Indicates the default size to use in migrating + * buffers of a process from Sysmem to VRAM and vice + * versa. The max legal value cannot be greater than + * 0x3F + * + * @note: A side effect of this symbol being part of + * struct svm_range_list is that it forces all buffers + * of the process of unregistered kind to use the same + * size in buffer migration + */ + uint8_t attr_gobm; }; /* Process data */ diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 10b1a1f64198..fcfe5543a3c0 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -309,12 +309,11 @@ static void svm_range_free(struct svm_range *prange, bool do_unmap) } static void -svm_range_set_default_attributes(int32_t *location, int32_t *prefetch_loc, - uint8_t *granularity, uint32_t *flags) +svm_range_set_default_attributes(int32_t *location, + int32_t *prefetch_loc, uint32_t *flags) { *location = KFD_IOCTL_SVM_LOCATION_UNDEFINED; *prefetch_loc = KFD_IOCTL_SVM_LOCATION_UNDEFINED; - *granularity = 9; *flags = KFD_IOCTL_SVM_FLAG_HOST_ACCESS | KFD_IOCTL_SVM_FLAG_COHERENT; } @@ -358,9 +357,9 @@ svm_range *svm_range_new(struct svm_range_list *svms, uint64_t start, bitmap_copy(prange->bitmap_access, svms->bitmap_supported, MAX_GPU_INSTANCE); + prange->granularity = svms->attr_gobm; svm_range_set_default_attributes(&prange->preferred_loc, - &prange->prefetch_loc, - &prange->granularity, &prange->flags); + &prange->prefetch_loc, &prange->flags); pr_debug("svms 0x%p [0x%llx 0x%llx]\n", svms, start, last); @@ -2685,10 +2684,12 @@ svm_range_get_range_boundaries(struct kfd_process *p, int64_t addr, *is_heap_stack = vma_is_initial_heap(vma) || vma_is_initial_stack(vma); + /* Determine the starting and ending page of prange */ start_limit = max(vma->vm_start >> PAGE_SHIFT, - (unsigned long)ALIGN_DOWN(addr, 2UL << 8)); + (unsigned long)ALIGN_DOWN(addr, (1 << p->svms.attr_gobm)));</pre> </blockquote> 1UL << p->svms.attr_gobm <blockquote type="cite" cite="mid:20240821232252.230791-1-Ramesh.Errabolu@amd.com"> <pre class="moz-quote-pre" wrap=""> end_limit = min(vma->vm_end >> PAGE_SHIFT, - (unsigned long)ALIGN(addr + 1, 2UL << 8)); + (unsigned long)ALIGN(addr + 1, (1 << p->svms.attr_gobm)));</pre> </blockquote> <pre class="moz-quote-pre" wrap="">1UL << p->svms.attr_gobm</pre> <blockquote type="cite" cite="mid:20240821232252.230791-1-Ramesh.Errabolu@amd.com"> <pre class="moz-quote-pre" wrap=""> + /* First range that starts after the fault address */ node = interval_tree_iter_first(&p->svms.objects, addr + 1, ULONG_MAX); if (node) { @@ -3211,6 +3212,15 @@ int svm_range_list_init(struct kfd_process *p) if (KFD_IS_SVM_API_SUPPORTED(p->pdds[i]->dev->adev)) bitmap_set(svms->bitmap_supported, i, 1); + /* Bind granularity of buffer migration, either + * the default size or one specified by the user + */ + svms->attr_gobm = amdgpu_svm_attr_gobm; + if (unlikely(amdgpu_svm_attr_gobm != AMDGPU_SVM_ATTR_GOBM)) + svms->attr_gobm = min_t(uint32_t, amdgpu_svm_attr_gobm, 0x3F);</pre> </blockquote> to simplify, remove the if check and always use this: <pre class="moz-quote-pre" wrap="">svms->attr_gobm = min_t(uint32_t, amdgpu_svm_attr_gobm, 0x3F);</pre> <blockquote type="cite" cite="mid:20240821232252.230791-1-Ramesh.Errabolu@amd.com"> <pre class="moz-quote-pre" wrap=""> + pr_debug("%s() Granularity Of Buffer Migration: %d\n", + __func__, svms->attr_gobm);</pre> </blockquote> pr_debug can output function name if adding +pf dynamically, __func__ is not needed. Regards, Philip <blockquote type="cite" cite="mid:20240821232252.230791-1-Ramesh.Errabolu@amd.com"> <pre class="moz-quote-pre" wrap=""> + return 0; } @@ -3738,8 +3748,9 @@ svm_range_get_attr(struct kfd_process *p, struct mm_struct *mm, node = interval_tree_iter_first(&svms->objects, start, last); if (!node) { pr_debug("range attrs not found return default values\n"); - svm_range_set_default_attributes(&location, &prefetch_loc, - &granularity, &flags_and); + granularity = svms->attr_gobm; + svm_range_set_default_attributes(&location, + &prefetch_loc, &flags_and); flags_or = flags_and; if (p->xnack_enabled) bitmap_copy(bitmap_access, svms->bitmap_supported, </pre> </blockquote> </body> </html>