drm/amdgpu: new VM update backends

Separate out all functions for SDMA and CPU based page table
updates into separate backends.

This way we can keep most of the complexity of those from the
core VM code.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/Makefile

@@ -53,7 +53,8 @@ amdgpu-y += amdgpu_device.o amdgpu_kms.o \
 	amdgpu_ucode.o amdgpu_bo_list.o amdgpu_ctx.o amdgpu_sync.o \
 	amdgpu_gtt_mgr.o amdgpu_vram_mgr.o amdgpu_virt.o amdgpu_atomfirmware.o \
 	amdgpu_vf_error.o amdgpu_sched.o amdgpu_debugfs.o amdgpu_ids.o \
-	amdgpu_gmc.o amdgpu_xgmi.o amdgpu_csa.o amdgpu_ras.o
+	amdgpu_gmc.o amdgpu_xgmi.o amdgpu_csa.o amdgpu_ras.o amdgpu_vm_cpu.o \
+	amdgpu_vm_sdma.o
 
 # add asic specific block
 amdgpu-$(CONFIG_DRM_AMDGPU_CIK)+= cik.o cik_ih.o kv_smc.o kv_dpm.o \

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

@@ -1242,7 +1242,7 @@ static void amdgpu_vm_do_copy_ptes(struct amdgpu_vm_update_params *params,
  * Returns:
  * The pointer for the page table entry.
  */
-static uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
+uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr)
 {
 	uint64_t result;
 
@@ -3000,6 +3000,11 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 			 vm->use_cpu_for_update ? "CPU" : "SDMA");
 	WARN_ONCE((vm->use_cpu_for_update && !amdgpu_gmc_vram_full_visible(&adev->gmc)),
 		  "CPU update of VM recommended only for large BAR system\n");
+
+	if (vm->use_cpu_for_update)
+		vm->update_funcs = &amdgpu_vm_cpu_funcs;
+	else
+		vm->update_funcs = &amdgpu_vm_sdma_funcs;
 	vm->last_update = NULL;
 
 	amdgpu_vm_bo_param(adev, vm, adev->vm_manager.root_level, &bp);

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h

@@ -203,11 +203,21 @@ struct amdgpu_vm_update_params {
 	 */
 	uint64_t src;
 
+	/**
+	 * @job: job to used for hw submission
+	 */
+	struct amdgpu_job *job;
+
 	/**
 	 * @ib: indirect buffer to fill with commands
 	 */
 	struct amdgpu_ib *ib;
 
+	/**
+	 * @num_dw_left: number of dw left for the IB
+	 */
+	unsigned int num_dw_left;
+
 	/**
 	 * @func: Function which actually does the update
 	 */
@@ -217,6 +227,17 @@ struct amdgpu_vm_update_params {
 		     uint64_t flags);
 };
 
+struct amdgpu_vm_update_funcs {
+
+	int (*prepare)(struct amdgpu_vm_update_params *p, void * owner,
+		       struct dma_fence *exclusive);
+	int (*update)(struct amdgpu_vm_update_params *p,
+		      struct amdgpu_bo *bo, uint64_t pe, uint64_t addr,
+		      unsigned count, uint32_t incr, uint64_t flags);
+	int (*commit)(struct amdgpu_vm_update_params *p,
+		      struct dma_fence **fence);
+};
+
 struct amdgpu_vm {
 	/* tree of virtual addresses mapped */
 	struct rb_root_cached	va;
@@ -252,7 +273,10 @@ struct amdgpu_vm {
 	struct amdgpu_vmid	*reserved_vmid[AMDGPU_MAX_VMHUBS];
 
 	/* Flag to indicate if VM tables are updated by CPU or GPU (SDMA) */
-	bool                    use_cpu_for_update;
+	bool					use_cpu_for_update;
+
+	/* Functions to use for VM table updates */
+	const struct amdgpu_vm_update_funcs	*update_funcs;
 
 	/* Flag to indicate ATS support from PTE for GFX9 */
 	bool			pte_support_ats;
@@ -319,6 +343,9 @@ struct amdgpu_vm_manager {
 #define amdgpu_vm_write_pte(adev, ib, pe, value, count, incr) ((adev)->vm_manager.vm_pte_funcs->write_pte((ib), (pe), (value), (count), (incr)))
 #define amdgpu_vm_set_pte_pde(adev, ib, pe, addr, count, incr, flags) ((adev)->vm_manager.vm_pte_funcs->set_pte_pde((ib), (pe), (addr), (count), (incr), (flags)))
 
+extern const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs;
+extern const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs;
+
 void amdgpu_vm_manager_init(struct amdgpu_device *adev);
 void amdgpu_vm_manager_fini(struct amdgpu_device *adev);
 
@@ -348,6 +375,7 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
 			bool clear);
 void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
 			     struct amdgpu_bo *bo, bool evicted);
+uint64_t amdgpu_vm_map_gart(const dma_addr_t *pages_addr, uint64_t addr);
 struct amdgpu_bo_va *amdgpu_vm_bo_find(struct amdgpu_vm *vm,
 				       struct amdgpu_bo *bo);
 struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,

drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c

+/*
+ * Copyright 2019 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 "amdgpu_vm.h"
+#include "amdgpu_object.h"
+#include "amdgpu_trace.h"
+
+/**
+ * amdgpu_vm_cpu_prepare - prepare page table update with the CPU
+ *
+ * @p: see amdgpu_vm_update_params definition
+ * @owner: owner we need to sync to
+ * @exclusive: exclusive move fence we need to sync to
+ *
+ * Returns:
+ * Negativ errno, 0 for success.
+ */
+static int amdgpu_vm_cpu_prepare(struct amdgpu_vm_update_params *p, void *owner,
+				 struct dma_fence *exclusive)
+{
+	int r;
+
+	/* Wait for PT BOs to be idle. PTs share the same resv. object
+	 * as the root PD BO
+	 */
+	r = amdgpu_bo_sync_wait(p->vm->root.base.bo, owner, true);
+	if (unlikely(r))
+		return r;
+
+	/* Wait for any BO move to be completed */
+	if (exclusive) {
+		r = dma_fence_wait(exclusive, true);
+		if (unlikely(r))
+			return r;
+	}
+
+	return 0;
+}
+
+/**
+ * amdgpu_vm_cpu_update - helper to update page tables via CPU
+ *
+ * @p: see amdgpu_vm_update_params definition
+ * @bo: PD/PT to update
+ * @pe: kmap addr of the page entry
+ * @addr: dst addr to write into pe
+ * @count: number of page entries to update
+ * @incr: increase next addr by incr bytes
+ * @flags: hw access flags
+ *
+ * Write count number of PT/PD entries directly.
+ */
+static int amdgpu_vm_cpu_update(struct amdgpu_vm_update_params *p,
+				struct amdgpu_bo *bo, uint64_t pe,
+				uint64_t addr, unsigned count, uint32_t incr,
+				uint64_t flags)
+{
+	unsigned int i;
+	uint64_t value;
+
+	pe += (unsigned long)amdgpu_bo_kptr(bo);
+
+	trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags);
+
+	for (i = 0; i < count; i++) {
+		value = p->pages_addr ?
+			amdgpu_vm_map_gart(p->pages_addr, addr) :
+			addr;
+		amdgpu_gmc_set_pte_pde(p->adev, (void *)(uintptr_t)pe,
+				       i, value, flags);
+		addr += incr;
+	}
+	return 0;
+}
+
+/**
+ * amdgpu_vm_cpu_commit - commit page table update to the HW
+ *
+ * @p: see amdgpu_vm_update_params definition
+ * @fence: unused
+ *
+ * Make sure that the hardware sees the page table updates.
+ */
+static int amdgpu_vm_cpu_commit(struct amdgpu_vm_update_params *p,
+				struct dma_fence **fence)
+{
+	/* Flush HDP */
+	mb();
+	amdgpu_asic_flush_hdp(p->adev, NULL);
+	return 0;
+}
+
+const struct amdgpu_vm_update_funcs amdgpu_vm_cpu_funcs = {
+	.prepare = amdgpu_vm_cpu_prepare,
+	.update = amdgpu_vm_cpu_update,
+	.commit = amdgpu_vm_cpu_commit
+};

drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c

+/*
+ * Copyright 2019 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 "amdgpu_vm.h"
+#include "amdgpu_job.h"
+#include "amdgpu_object.h"
+#include "amdgpu_trace.h"
+
+#define AMDGPU_VM_SDMA_MIN_NUM_DW	256u
+#define AMDGPU_VM_SDMA_MAX_NUM_DW	(16u * 1024u)
+
+/**
+ * amdgpu_vm_sdma_prepare - prepare SDMA command submission
+ *
+ * @p: see amdgpu_vm_update_params definition
+ * @owner: owner we need to sync to
+ * @exclusive: exclusive move fence we need to sync to
+ *
+ * Returns:
+ * Negativ errno, 0 for success.
+ */
+static int amdgpu_vm_sdma_prepare(struct amdgpu_vm_update_params *p,
+				  void *owner, struct dma_fence *exclusive)
+{
+	struct amdgpu_bo *root = p->vm->root.base.bo;
+	unsigned int ndw = AMDGPU_VM_SDMA_MIN_NUM_DW;
+	int r;
+
+	r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, &p->job);
+	if (r)
+		return r;
+
+	r = amdgpu_sync_fence(p->adev, &p->job->sync, exclusive, false);
+	if (r)
+		return r;
+
+	r = amdgpu_sync_resv(p->adev, &p->job->sync, root->tbo.resv,
+			     owner, false);
+	if (r)
+		return r;
+
+	p->num_dw_left = ndw;
+	p->ib = &p->job->ibs[0];
+	return 0;
+}
+
+/**
+ * amdgpu_vm_sdma_commit - commit SDMA command submission
+ *
+ * @p: see amdgpu_vm_update_params definition
+ * @fence: resulting fence
+ *
+ * Returns:
+ * Negativ errno, 0 for success.
+ */
+static int amdgpu_vm_sdma_commit(struct amdgpu_vm_update_params *p,
+				 struct dma_fence **fence)
+{
+	struct amdgpu_bo *root = p->vm->root.base.bo;
+	struct amdgpu_ring *ring;
+	struct dma_fence *f;
+	int r;
+
+	ring = container_of(p->vm->entity.rq->sched, struct amdgpu_ring, sched);
+
+	WARN_ON(p->ib->length_dw == 0);
+	amdgpu_ring_pad_ib(ring, p->ib);
+	WARN_ON(p->ib->length_dw > p->num_dw_left);
+	r = amdgpu_job_submit(p->job, &p->vm->entity,
+			      AMDGPU_FENCE_OWNER_VM, &f);
+	if (r)
+		goto error;
+
+	amdgpu_bo_fence(root, f, true);
+	if (fence)
+		swap(*fence, f);
+	dma_fence_put(f);
+	return 0;
+
+error:
+	amdgpu_job_free(p->job);
+	return r;
+}
+
+
+/**
+ * amdgpu_vm_sdma_copy_ptes - copy the PTEs from mapping
+ *
+ * @p: see amdgpu_vm_update_params definition
+ * @bo: PD/PT to update
+ * @pe: addr of the page entry
+ * @count: number of page entries to copy
+ *
+ * Traces the parameters and calls the DMA function to copy the PTEs.
+ */
+static void amdgpu_vm_sdma_copy_ptes(struct amdgpu_vm_update_params *p,
+				     struct amdgpu_bo *bo, uint64_t pe,
+				     unsigned count)
+{
+	uint64_t src = p->ib->gpu_addr;
+
+	src += p->num_dw_left * 4;
+
+	pe += amdgpu_bo_gpu_offset(bo);
+	trace_amdgpu_vm_copy_ptes(pe, src, count);
+
+	amdgpu_vm_copy_pte(p->adev, p->ib, pe, src, count);
+}
+
+/**
+ * amdgpu_vm_sdma_set_ptes - helper to call the right asic function
+ *
+ * @p: see amdgpu_vm_update_params definition
+ * @bo: PD/PT to update
+ * @pe: addr of the page entry
+ * @addr: dst addr to write into pe
+ * @count: number of page entries to update
+ * @incr: increase next addr by incr bytes
+ * @flags: hw access flags
+ *
+ * Traces the parameters and calls the right asic functions
+ * to setup the page table using the DMA.
+ */
+static void amdgpu_vm_sdma_set_ptes(struct amdgpu_vm_update_params *p,
+				    struct amdgpu_bo *bo, uint64_t pe,
+				    uint64_t addr, unsigned count,
+				    uint32_t incr, uint64_t flags)
+{
+	pe += amdgpu_bo_gpu_offset(bo);
+	trace_amdgpu_vm_set_ptes(pe, addr, count, incr, flags);
+	if (count < 3) {
+		amdgpu_vm_write_pte(p->adev, p->ib, pe, addr | flags,
+				    count, incr);
+	} else {
+		amdgpu_vm_set_pte_pde(p->adev, p->ib, pe, addr,
+				      count, incr, flags);
+	}
+}
+
+/**
+ * amdgpu_vm_sdma_update - execute VM update
+ *
+ * @p: see amdgpu_vm_update_params definition
+ * @bo: PD/PT to update
+ * @pe: addr of the page entry
+ * @addr: dst addr to write into pe
+ * @count: number of page entries to update
+ * @incr: increase next addr by incr bytes
+ * @flags: hw access flags
+ *
+ * Reserve space in the IB, setup mapping buffer on demand and write commands to
+ * the IB.
+ */
+static int amdgpu_vm_sdma_update(struct amdgpu_vm_update_params *p,
+				 struct amdgpu_bo *bo, uint64_t pe,
+				 uint64_t addr, unsigned count, uint32_t incr,
+				 uint64_t flags)
+{
+	unsigned int i, ndw, nptes;
+	uint64_t *pte;
+	int r;
+
+	do {
+		ndw = p->num_dw_left;
+		ndw -= p->ib->length_dw;
+
+		if (ndw < 32) {
+			r = amdgpu_vm_sdma_commit(p, NULL);
+			if (r)
+				return r;
+
+			/* estimate how many dw we need */
+			ndw = 32;
+			if (p->pages_addr)
+				ndw += count * 2;
+			ndw = max(ndw, AMDGPU_VM_SDMA_MIN_NUM_DW);
+			ndw = min(ndw, AMDGPU_VM_SDMA_MAX_NUM_DW);
+
+			r = amdgpu_job_alloc_with_ib(p->adev, ndw * 4, &p->job);
+			if (r)
+				return r;
+
+			p->num_dw_left = ndw;
+			p->ib = &p->job->ibs[0];
+		}
+
+		if (!p->pages_addr) {
+			/* set page commands needed */
+			if (bo->shadow)
+				amdgpu_vm_sdma_set_ptes(p, bo->shadow, pe, addr,
+							count, incr, flags);
+			amdgpu_vm_sdma_set_ptes(p, bo, pe, addr, count,
+						incr, flags);
+			return 0;
+		}
+
+		/* copy commands needed */
+		ndw -= p->adev->vm_manager.vm_pte_funcs->copy_pte_num_dw *
+			(bo->shadow ? 2 : 1);
+
+		/* for padding */
+		ndw -= 7;
+
+		nptes = min(count, ndw / 2);
+
+		/* Put the PTEs at the end of the IB. */
+		p->num_dw_left -= nptes * 2;
+		pte = (uint64_t *)&(p->ib->ptr[p->num_dw_left]);
+		for (i = 0; i < nptes; ++i, addr += incr) {
+			pte[i] = amdgpu_vm_map_gart(p->pages_addr, addr);
+			pte[i] |= flags;
+		}
+
+		if (bo->shadow)
+			amdgpu_vm_sdma_copy_ptes(p, bo->shadow, pe, nptes);
+		amdgpu_vm_sdma_copy_ptes(p, bo, pe, nptes);
+
+		pe += nptes * 8;
+		count -= nptes;
+	} while (count);
+
+	return 0;
+}
+
+const struct amdgpu_vm_update_funcs amdgpu_vm_sdma_funcs = {
+	.prepare = amdgpu_vm_sdma_prepare,
+	.update = amdgpu_vm_sdma_update,
+	.commit = amdgpu_vm_sdma_commit
+};