drm/msm: devcoredump iommu fault support

Wire up support to stall the SMMU on iova fault, and collect a devcore-
dump snapshot for easier debugging of faults.

Currently this is a6xx-only, but mostly only because so far it is the
only one using adreno-smmu-priv.

Signed-off-by: Rob Clark <[e-mail]>
Acked-by: Jordan Crouse <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]
Signed-off-by: Rob Clark <[e-mail]>

drivers/gpu/drm/msm/adreno/a5xx_gpu.c

···
 	struct drm_device *dev = gpu->dev;
 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
 
+	/*
+	 * If stalled on SMMU fault, we could trip the GPU's hang detection,
+	 * but the fault handler will trigger the devcore dump, and we want
+	 * to otherwise resume normally rather than killing the submit, so
+	 * just bail.
+	 */
+	if (gpu_read(gpu, REG_A5XX_RBBM_STATUS3) & BIT(24))
+		return;
+
 	DRM_DEV_ERROR(dev->dev, "gpu fault ring %d fence %x status %8.8X rb %4.4x/%4.4x ib1 %16.16llX/%4.4x ib2 %16.16llX/%4.4x\n",
 		ring ? ring->id : -1, ring ? ring->seqno : 0,
 		gpu_read(gpu, REG_A5XX_RBBM_STATUS),
···
 {
 	struct a5xx_gpu_state *a5xx_state = kzalloc(sizeof(*a5xx_state),
 			GFP_KERNEL);
+	bool stalled = !!(gpu_read(gpu, REG_A5XX_RBBM_STATUS3) & BIT(24));
 
 	if (!a5xx_state)
 		return ERR_PTR(-ENOMEM);
···
 
 	a5xx_state->base.rbbm_status = gpu_read(gpu, REG_A5XX_RBBM_STATUS);
 
-	/* Get the HLSQ regs with the help of the crashdumper */
-	a5xx_gpu_state_get_hlsq_regs(gpu, a5xx_state);
+	/*
+	 * Get the HLSQ regs with the help of the crashdumper, but only if
+	 * we are not stalled in an iommu fault (in which case the crashdumper
+	 * would not have access to memory)
+	 */
+	if (!stalled)
+		a5xx_gpu_state_get_hlsq_regs(gpu, a5xx_state);
 
 	a5xx_set_hwcg(gpu, true);
 

drivers/gpu/drm/msm/adreno/a6xx_gpu.c

···
 	struct msm_gpu *gpu = arg;
 	struct adreno_smmu_fault_info *info = data;
 	const char *type = "UNKNOWN";
+	const char *block;
+	bool do_devcoredump = info && !READ_ONCE(gpu->crashstate);
+
+	/*
+	 * If we aren't going to be resuming later from fault_worker, then do
+	 * it now.
+	 */
+	if (!do_devcoredump) {
+		gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
+	}
 
 	/*
 	 * Print a default message if we couldn't get the data from the
···
 	else if (info->fsr & ARM_SMMU_FSR_EF)
 		type = "EXTERNAL";
 
+	block = a6xx_fault_block(gpu, info->fsynr1 & 0xff);
+
 	pr_warn_ratelimited("*** gpu fault: ttbr0=%.16llx iova=%.16lx dir=%s type=%s source=%s (%u,%u,%u,%u)\n",
 			info->ttbr0, iova,
-			flags & IOMMU_FAULT_WRITE ? "WRITE" : "READ", type,
-			a6xx_fault_block(gpu, info->fsynr1 & 0xff),
+			flags & IOMMU_FAULT_WRITE ? "WRITE" : "READ",
+			type, block,
 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(4)),
 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(5)),
 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(6)),
 			gpu_read(gpu, REG_A6XX_CP_SCRATCH_REG(7)));
+
+	if (do_devcoredump) {
+		/* Turn off the hangcheck timer to keep it from bothering us */
+		del_timer(&gpu->hangcheck_timer);
+
+		gpu->fault_info.ttbr0 = info->ttbr0;
+		gpu->fault_info.iova  = iova;
+		gpu->fault_info.flags = flags;
+		gpu->fault_info.type  = type;
+		gpu->fault_info.block = block;
+
+		kthread_queue_work(gpu->worker, &gpu->fault_work);
+	}
 
 	return 0;
 }
···
 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
 	struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
+
+	/*
+	 * If stalled on SMMU fault, we could trip the GPU's hang detection,
+	 * but the fault handler will trigger the devcore dump, and we want
+	 * to otherwise resume normally rather than killing the submit, so
+	 * just bail.
+	 */
+	if (gpu_read(gpu, REG_A6XX_RBBM_STATUS3) & A6XX_RBBM_STATUS3_SMMU_STALLED_ON_FAULT)
+		return;
 
 	/*
 	 * Force the GPU to stay on until after we finish

drivers/gpu/drm/msm/adreno/a6xx_gpu_state.c

···
 		a6xx_get_ahb_gpu_registers(gpu,
 				a6xx_state, &a6xx_vbif_reglist,
 				&a6xx_state->registers[index++]);
+	if (!dumper) {
+		/*
+		 * We can't use the crashdumper when the SMMU is stalled,
+		 * because the GPU has no memory access until we resume
+		 * translation (but we don't want to do that until after
+		 * we have captured as much useful GPU state as possible).
+		 * So instead collect registers via the CPU:
+		 */
+		for (i = 0; i < ARRAY_SIZE(a6xx_reglist); i++)
+			a6xx_get_ahb_gpu_registers(gpu,
+				a6xx_state, &a6xx_reglist[i],
+				&a6xx_state->registers[index++]);
+		return;
+	}
 
 	for (i = 0; i < ARRAY_SIZE(a6xx_reglist); i++)
 		a6xx_get_crashdumper_registers(gpu,
···
 
 struct msm_gpu_state *a6xx_gpu_state_get(struct msm_gpu *gpu)
 {
-	struct a6xx_crashdumper dumper = { 0 };
+	struct a6xx_crashdumper _dumper = { 0 }, *dumper = NULL;
 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
 	struct a6xx_gpu_state *a6xx_state = kzalloc(sizeof(*a6xx_state),
 		GFP_KERNEL);
+	bool stalled = !!(gpu_read(gpu, REG_A6XX_RBBM_STATUS3) &
+			A6XX_RBBM_STATUS3_SMMU_STALLED_ON_FAULT);
 
 	if (!a6xx_state)
 		return ERR_PTR(-ENOMEM);
···
 	/* Get the banks of indexed registers */
 	a6xx_get_indexed_registers(gpu, a6xx_state);
 
-	/* Try to initialize the crashdumper */
-	if (!a6xx_crashdumper_init(gpu, &dumper)) {
-		a6xx_get_registers(gpu, a6xx_state, &dumper);
-		a6xx_get_shaders(gpu, a6xx_state, &dumper);
-		a6xx_get_clusters(gpu, a6xx_state, &dumper);
-		a6xx_get_dbgahb_clusters(gpu, a6xx_state, &dumper);
+	/*
+	 * Try to initialize the crashdumper, if we are not dumping state
+	 * with the SMMU stalled.  The crashdumper needs memory access to
+	 * write out GPU state, so we need to skip this when the SMMU is
+	 * stalled in response to an iova fault
+	 */
+	if (!stalled && !a6xx_crashdumper_init(gpu, &_dumper)) {
+		dumper = &_dumper;
+	}
 
-		msm_gem_kernel_put(dumper.bo, gpu->aspace, true);
+	a6xx_get_registers(gpu, a6xx_state, dumper);
+
+	if (dumper) {
+		a6xx_get_shaders(gpu, a6xx_state, dumper);
+		a6xx_get_clusters(gpu, a6xx_state, dumper);
+		a6xx_get_dbgahb_clusters(gpu, a6xx_state, dumper);
+
+		msm_gem_kernel_put(dumper->bo, gpu->aspace, true);
 	}
 
 	if (snapshot_debugbus)

drivers/gpu/drm/msm/adreno/adreno_gpu.c

···
 			adreno_gpu->info->revn, adreno_gpu->rev.core,
 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
 			adreno_gpu->rev.patchid);
+	/*
+	 * If this is state collected due to iova fault, so fault related info
+	 *
+	 * TTBR0 would not be zero, so this is a good way to distinguish
+	 */
+	if (state->fault_info.ttbr0) {
+		const struct msm_gpu_fault_info *info = &state->fault_info;
+
+		drm_puts(p, "fault-info:\n");
+		drm_printf(p, "  - ttbr0=%.16llx\n", info->ttbr0);
+		drm_printf(p, "  - iova=%.16lx\n", info->iova);
+		drm_printf(p, "  - dir=%s\n", info->flags & IOMMU_FAULT_WRITE ? "WRITE" : "READ");
+		drm_printf(p, "  - type=%s\n", info->type);
+		drm_printf(p, "  - source=%s\n", info->block);
+	}
 
 	drm_printf(p, "rbbm-status: 0x%08x\n", state->rbbm_status);
 

drivers/gpu/drm/msm/msm_gem.h

···
 	struct dma_fence *fence;
 	struct msm_gpu_submitqueue *queue;
 	struct pid *pid;    /* submitting process */
+	bool fault_dumped;  /* Limit devcoredump dumping to one per submit */
 	bool valid;         /* true if no cmdstream patching needed */
 	bool in_rb;         /* "sudo" mode, copy cmds into RB */
 	struct msm_ringbuffer *ring;

drivers/gpu/drm/msm/msm_gem_submit.c

···
 	submit->cmd = (void *)&submit->bos[nr_bos];
 	submit->queue = queue;
 	submit->ring = gpu->rb[queue->prio];
+	submit->fault_dumped = false;
 
 	/* initially, until copy_from_user() and bo lookup succeeds: */
 	submit->nr_bos = 0;

drivers/gpu/drm/msm/msm_gpu.c

···
 	/* Fill in the additional crash state information */
 	state->comm = kstrdup(comm, GFP_KERNEL);
 	state->cmd = kstrdup(cmd, GFP_KERNEL);
+	state->fault_info = gpu->fault_info;
 
 	if (submit) {
 		int i, nr = 0;
···
 	mutex_unlock(&dev->struct_mutex);
 
 	msm_gpu_retire(gpu);
+}
+
+static void fault_worker(struct kthread_work *work)
+{
+	struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work);
+	struct drm_device *dev = gpu->dev;
+	struct msm_gem_submit *submit;
+	struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
+	char *comm = NULL, *cmd = NULL;
+
+	mutex_lock(&dev->struct_mutex);
+
+	submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
+	if (submit && submit->fault_dumped)
+		goto resume_smmu;
+
+	if (submit) {
+		struct task_struct *task;
+
+		task = get_pid_task(submit->pid, PIDTYPE_PID);
+		if (task) {
+			comm = kstrdup(task->comm, GFP_KERNEL);
+			cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
+			put_task_struct(task);
+		}
+
+		/*
+		 * When we get GPU iova faults, we can get 1000s of them,
+		 * but we really only want to log the first one.
+		 */
+		submit->fault_dumped = true;
+	}
+
+	/* Record the crash state */
+	pm_runtime_get_sync(&gpu->pdev->dev);
+	msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
+	pm_runtime_put_sync(&gpu->pdev->dev);
+
+	kfree(cmd);
+	kfree(comm);
+
+resume_smmu:
+	memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
+	gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
+
+	mutex_unlock(&dev->struct_mutex);
 }
 
 static void hangcheck_timer_reset(struct msm_gpu *gpu)
···
 	INIT_LIST_HEAD(&gpu->active_list);
 	kthread_init_work(&gpu->retire_work, retire_worker);
 	kthread_init_work(&gpu->recover_work, recover_worker);
+	kthread_init_work(&gpu->fault_work, fault_worker);
 
 	timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
 

drivers/gpu/drm/msm/msm_gpu.h

···
 	uint32_t (*get_rptr)(struct msm_gpu *gpu, struct msm_ringbuffer *ring);
 };
 
+/* Additional state for iommu faults: */
+struct msm_gpu_fault_info {
+	u64 ttbr0;
+	unsigned long iova;
+	int flags;
+	const char *type;
+	const char *block;
+};
+
 struct msm_gpu {
 	const char *name;
 	struct drm_device *dev;
···
 
 #define DRM_MSM_HANGCHECK_DEFAULT_PERIOD 500 /* in ms */
 	struct timer_list hangcheck_timer;
+
+	/* Fault info for most recent iova fault: */
+	struct msm_gpu_fault_info fault_info;
+
+	/* work for handling GPU ioval faults: */
+	struct kthread_work fault_work;
 
 	/* work for handling GPU recovery: */
 	struct kthread_work recover_work;
···
 
 	char *comm;
 	char *cmd;
+
+	struct msm_gpu_fault_info fault_info;
 
 	int nr_bos;
 	struct msm_gpu_state_bo *bos;

drivers/gpu/drm/msm/msm_gpummu.c

···
 	return 0;
 }
 
+static void msm_gpummu_resume_translation(struct msm_mmu *mmu)
+{
+}
+
 static void msm_gpummu_destroy(struct msm_mmu *mmu)
 {
 	struct msm_gpummu *gpummu = to_msm_gpummu(mmu);
···
 		.map = msm_gpummu_map,
 		.unmap = msm_gpummu_unmap,
 		.destroy = msm_gpummu_destroy,
+		.resume_translation = msm_gpummu_resume_translation,
 };
 
 struct msm_mmu *msm_gpummu_new(struct device *dev, struct msm_gpu *gpu)

drivers/gpu/drm/msm/msm_iommu.c

···
 	 * the arm-smmu driver as a trigger to set up TTBR0
 	 */
 	if (atomic_inc_return(&iommu->pagetables) == 1) {
+		/* Enable stall on iommu fault: */
+		adreno_smmu->set_stall(adreno_smmu->cookie, true);
+
 		ret = adreno_smmu->set_ttbr0_cfg(adreno_smmu->cookie, &ttbr0_cfg);
 		if (ret) {
 			free_io_pgtable_ops(pagetable->pgtbl_ops);
···
 
 	pr_warn_ratelimited("*** fault: iova=%16lx, flags=%d\n", iova, flags);
 	return 0;
+}
+
+static void msm_iommu_resume_translation(struct msm_mmu *mmu)
+{
+	struct adreno_smmu_priv *adreno_smmu = dev_get_drvdata(mmu->dev);
+
+	adreno_smmu->resume_translation(adreno_smmu->cookie, true);
 }
 
 static void msm_iommu_detach(struct msm_mmu *mmu)
···
 		.map = msm_iommu_map,
 		.unmap = msm_iommu_unmap,
 		.destroy = msm_iommu_destroy,
+		.resume_translation = msm_iommu_resume_translation,
 };
 
 struct msm_mmu *msm_iommu_new(struct device *dev, struct iommu_domain *domain)

drivers/gpu/drm/msm/msm_mmu.h

···
 			size_t len, int prot);
 	int (*unmap)(struct msm_mmu *mmu, uint64_t iova, size_t len);
 	void (*destroy)(struct msm_mmu *mmu);
+	void (*resume_translation)(struct msm_mmu *mmu);
 };
 
 enum msm_mmu_type {