summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--MAINTAINERS1
-rw-r--r--drivers/gpu/drm/amd/amdgpu/Makefile1
-rw-r--r--drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h15
-rw-r--r--drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c179
-rw-r--r--drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h5
-rw-r--r--drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c21
-rw-r--r--drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c18
-rw-r--r--drivers/gpu/drm/amd/include/kgd_kfd_interface.h6
8 files changed, 242 insertions, 4 deletions
diff --git a/MAINTAINERS b/MAINTAINERS
index 13c8ec11135a..9e376c7f15b1 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -766,6 +766,7 @@ F: drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.c
F: drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
F: drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v7.c
F: drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v8.c
+F: drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
F: drivers/gpu/drm/amd/amdkfd/
F: drivers/gpu/drm/amd/include/cik_structs.h
F: drivers/gpu/drm/amd/include/kgd_kfd_interface.h
diff --git a/drivers/gpu/drm/amd/amdgpu/Makefile b/drivers/gpu/drm/amd/amdgpu/Makefile
index 353c937d947d..5dd317579e77 100644
--- a/drivers/gpu/drm/amd/amdgpu/Makefile
+++ b/drivers/gpu/drm/amd/amdgpu/Makefile
@@ -129,6 +129,7 @@ amdgpu-y += \
# add amdkfd interfaces
amdgpu-y += \
amdgpu_amdkfd.o \
+ amdgpu_amdkfd_fence.o \
amdgpu_amdkfd_gfx_v8.o
# add cgs
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
index 2a519f9062ee..833dc26d402f 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.h
@@ -29,6 +29,8 @@
#include <linux/mmu_context.h>
#include <kgd_kfd_interface.h>
+extern const struct kgd2kfd_calls *kgd2kfd;
+
struct amdgpu_device;
struct kgd_mem {
@@ -37,6 +39,19 @@ struct kgd_mem {
void *cpu_ptr;
};
+/* KFD Memory Eviction */
+struct amdgpu_amdkfd_fence {
+ struct dma_fence base;
+ struct mm_struct *mm;
+ spinlock_t lock;
+ char timeline_name[TASK_COMM_LEN];
+};
+
+struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
+ struct mm_struct *mm);
+bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm);
+struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f);
+
int amdgpu_amdkfd_init(void);
void amdgpu_amdkfd_fini(void);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
new file mode 100644
index 000000000000..2c14025e5e76
--- /dev/null
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
@@ -0,0 +1,179 @@
+/*
+ * Copyright 2016-2018 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
+ * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
+ * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
+ * OTHER DEALINGS IN THE SOFTWARE.
+ */
+
+#include <linux/dma-fence.h>
+#include <linux/spinlock.h>
+#include <linux/atomic.h>
+#include <linux/stacktrace.h>
+#include <linux/sched.h>
+#include <linux/slab.h>
+#include <linux/sched/mm.h>
+#include "amdgpu_amdkfd.h"
+
+static const struct dma_fence_ops amdkfd_fence_ops;
+static atomic_t fence_seq = ATOMIC_INIT(0);
+
+/* Eviction Fence
+ * Fence helper functions to deal with KFD memory eviction.
+ * Big Idea - Since KFD submissions are done by user queues, a BO cannot be
+ * evicted unless all the user queues for that process are evicted.
+ *
+ * All the BOs in a process share an eviction fence. When process X wants
+ * to map VRAM memory but TTM can't find enough space, TTM will attempt to
+ * evict BOs from its LRU list. TTM checks if the BO is valuable to evict
+ * by calling ttm_bo_driver->eviction_valuable().
+ *
+ * ttm_bo_driver->eviction_valuable() - will return false if the BO belongs
+ * to process X. Otherwise, it will return true to indicate BO can be
+ * evicted by TTM.
+ *
+ * If ttm_bo_driver->eviction_valuable returns true, then TTM will continue
+ * the evcition process for that BO by calling ttm_bo_evict --> amdgpu_bo_move
+ * --> amdgpu_copy_buffer(). This sets up job in GPU scheduler.
+ *
+ * GPU Scheduler (amd_sched_main) - sets up a cb (fence_add_callback) to
+ * nofity when the BO is free to move. fence_add_callback --> enable_signaling
+ * --> amdgpu_amdkfd_fence.enable_signaling
+ *
+ * amdgpu_amdkfd_fence.enable_signaling - Start a work item that will quiesce
+ * user queues and signal fence. The work item will also start another delayed
+ * work item to restore BOs
+ */
+
+struct amdgpu_amdkfd_fence *amdgpu_amdkfd_fence_create(u64 context,
+ struct mm_struct *mm)
+{
+ struct amdgpu_amdkfd_fence *fence;
+
+ fence = kzalloc(sizeof(*fence), GFP_KERNEL);
+ if (fence == NULL)
+ return NULL;
+
+ /* This reference gets released in amdkfd_fence_release */
+ mmgrab(mm);
+ fence->mm = mm;
+ get_task_comm(fence->timeline_name, current);
+ spin_lock_init(&fence->lock);
+
+ dma_fence_init(&fence->base, &amdkfd_fence_ops, &fence->lock,
+ context, atomic_inc_return(&fence_seq));
+
+ return fence;
+}
+
+struct amdgpu_amdkfd_fence *to_amdgpu_amdkfd_fence(struct dma_fence *f)
+{
+ struct amdgpu_amdkfd_fence *fence;
+
+ if (!f)
+ return NULL;
+
+ fence = container_of(f, struct amdgpu_amdkfd_fence, base);
+ if (fence && f->ops == &amdkfd_fence_ops)
+ return fence;
+
+ return NULL;
+}
+
+static const char *amdkfd_fence_get_driver_name(struct dma_fence *f)
+{
+ return "amdgpu_amdkfd_fence";
+}
+
+static const char *amdkfd_fence_get_timeline_name(struct dma_fence *f)
+{
+ struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
+
+ return fence->timeline_name;
+}
+
+/**
+ * amdkfd_fence_enable_signaling - This gets called when TTM wants to evict
+ * a KFD BO and schedules a job to move the BO.
+ * If fence is already signaled return true.
+ * If fence is not signaled schedule a evict KFD process work item.
+ */
+static bool amdkfd_fence_enable_signaling(struct dma_fence *f)
+{
+ struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
+
+ if (!fence)
+ return false;
+
+ if (dma_fence_is_signaled(f))
+ return true;
+
+ if (!kgd2kfd->schedule_evict_and_restore_process(fence->mm, f))
+ return true;
+
+ return false;
+}
+
+/**
+ * amdkfd_fence_release - callback that fence can be freed
+ *
+ * @fence: fence
+ *
+ * This function is called when the reference count becomes zero.
+ * Drops the mm_struct reference and RCU schedules freeing up the fence.
+ */
+static void amdkfd_fence_release(struct dma_fence *f)
+{
+ struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
+
+ /* Unconditionally signal the fence. The process is getting
+ * terminated.
+ */
+ if (WARN_ON(!fence))
+ return; /* Not an amdgpu_amdkfd_fence */
+
+ mmdrop(fence->mm);
+ kfree_rcu(f, rcu);
+}
+
+/**
+ * amdkfd_fence_check_mm - Check if @mm is same as that of the fence @f
+ * if same return TRUE else return FALSE.
+ *
+ * @f: [IN] fence
+ * @mm: [IN] mm that needs to be verified
+ */
+bool amdkfd_fence_check_mm(struct dma_fence *f, struct mm_struct *mm)
+{
+ struct amdgpu_amdkfd_fence *fence = to_amdgpu_amdkfd_fence(f);
+
+ if (!fence)
+ return false;
+ else if (fence->mm == mm)
+ return true;
+
+ return false;
+}
+
+static const struct dma_fence_ops amdkfd_fence_ops = {
+ .get_driver_name = amdkfd_fence_get_driver_name,
+ .get_timeline_name = amdkfd_fence_get_timeline_name,
+ .enable_signaling = amdkfd_fence_enable_signaling,
+ .signaled = NULL,
+ .wait = dma_fence_default_wait,
+ .release = amdkfd_fence_release,
+};
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
index 06436c3ebd2b..1a5911882657 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
@@ -36,8 +36,9 @@
#define AMDGPU_MAX_UVD_ENC_RINGS 2
/* some special values for the owner field */
-#define AMDGPU_FENCE_OWNER_UNDEFINED ((void*)0ul)
-#define AMDGPU_FENCE_OWNER_VM ((void*)1ul)
+#define AMDGPU_FENCE_OWNER_UNDEFINED ((void *)0ul)
+#define AMDGPU_FENCE_OWNER_VM ((void *)1ul)
+#define AMDGPU_FENCE_OWNER_KFD ((void *)2ul)
#define AMDGPU_FENCE_FLAG_64BIT (1 << 0)
#define AMDGPU_FENCE_FLAG_INT (1 << 1)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
index df65c66dc956..b8d3b87fd43e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_sync.c
@@ -31,6 +31,7 @@
#include <drm/drmP.h>
#include "amdgpu.h"
#include "amdgpu_trace.h"
+#include "amdgpu_amdkfd.h"
struct amdgpu_sync_entry {
struct hlist_node node;
@@ -85,11 +86,20 @@ static bool amdgpu_sync_same_dev(struct amdgpu_device *adev,
*/
static void *amdgpu_sync_get_owner(struct dma_fence *f)
{
- struct drm_sched_fence *s_fence = to_drm_sched_fence(f);
+ struct drm_sched_fence *s_fence;
+ struct amdgpu_amdkfd_fence *kfd_fence;
+
+ if (!f)
+ return AMDGPU_FENCE_OWNER_UNDEFINED;
+ s_fence = to_drm_sched_fence(f);
if (s_fence)
return s_fence->owner;
+ kfd_fence = to_amdgpu_amdkfd_fence(f);
+ if (kfd_fence)
+ return AMDGPU_FENCE_OWNER_KFD;
+
return AMDGPU_FENCE_OWNER_UNDEFINED;
}
@@ -204,11 +214,18 @@ int amdgpu_sync_resv(struct amdgpu_device *adev,
for (i = 0; i < flist->shared_count; ++i) {
f = rcu_dereference_protected(flist->shared[i],
reservation_object_held(resv));
+ /* We only want to trigger KFD eviction fences on
+ * evict or move jobs. Skip KFD fences otherwise.
+ */
+ fence_owner = amdgpu_sync_get_owner(f);
+ if (fence_owner == AMDGPU_FENCE_OWNER_KFD &&
+ owner != AMDGPU_FENCE_OWNER_UNDEFINED)
+ continue;
+
if (amdgpu_sync_same_dev(adev, f)) {
/* VM updates are only interesting
* for other VM updates and moves.
*/
- fence_owner = amdgpu_sync_get_owner(f);
if ((owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
(fence_owner != AMDGPU_FENCE_OWNER_UNDEFINED) &&
((owner == AMDGPU_FENCE_OWNER_VM) !=
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index 28c33d711bab..5fcb3488a595 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -46,6 +46,7 @@
#include "amdgpu.h"
#include "amdgpu_object.h"
#include "amdgpu_trace.h"
+#include "amdgpu_amdkfd.h"
#include "bif/bif_4_1_d.h"
#define DRM_FILE_PAGE_OFFSET (0x100000000ULL >> PAGE_SHIFT)
@@ -1171,6 +1172,23 @@ static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
{
unsigned long num_pages = bo->mem.num_pages;
struct drm_mm_node *node = bo->mem.mm_node;
+ struct reservation_object_list *flist;
+ struct dma_fence *f;
+ int i;
+
+ /* If bo is a KFD BO, check if the bo belongs to the current process.
+ * If true, then return false as any KFD process needs all its BOs to
+ * be resident to run successfully
+ */
+ flist = reservation_object_get_list(bo->resv);
+ if (flist) {
+ for (i = 0; i < flist->shared_count; ++i) {
+ f = rcu_dereference_protected(flist->shared[i],
+ reservation_object_held(bo->resv));
+ if (amdkfd_fence_check_mm(f, current->mm))
+ return false;
+ }
+ }
switch (bo->mem.mem_type) {
case TTM_PL_TT:
diff --git a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
index 94eab54864bf..9e352499c409 100644
--- a/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
+++ b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h
@@ -30,6 +30,7 @@
#include <linux/types.h>
#include <linux/bitmap.h>
+#include <linux/dma-fence.h>
struct pci_dev;
@@ -286,6 +287,9 @@ struct kfd2kgd_calls {
*
* @resume: Notifies amdkfd about a resume action done to a kgd device
*
+ * @schedule_evict_and_restore_process: Schedules work queue that will prepare
+ * for safe eviction of KFD BOs that belong to the specified process.
+ *
* This structure contains function callback pointers so the kgd driver
* will notify to the amdkfd about certain status changes.
*
@@ -300,6 +304,8 @@ struct kgd2kfd_calls {
void (*interrupt)(struct kfd_dev *kfd, const void *ih_ring_entry);
void (*suspend)(struct kfd_dev *kfd);
int (*resume)(struct kfd_dev *kfd);
+ int (*schedule_evict_and_restore_process)(struct mm_struct *mm,
+ struct dma_fence *fence);
};
int kgd2kfd_init(unsigned interface_version,