drm/amdgpu: fix and cleanup gmc_v9_0_flush_gpu_tlb_pasid

Testing for reset is pointless since the reset can start right after the
test.

The same PASID can be used by more than one VMID, invalidate each of them.

Move the KIQ and all the workaround handling into common GMC code.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

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

@@ -32,6 +32,7 @@
 #include "amdgpu.h"
 #include "amdgpu_gmc.h"
 #include "amdgpu_ras.h"
+#include "amdgpu_reset.h"
 #include "amdgpu_xgmi.h"
 
 #include <drm/drm_drv.h>
@@ -630,6 +631,65 @@ void amdgpu_gmc_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
 	dev_err(adev->dev, "Error flushing GPU TLB using the SDMA (%d)!\n", r);
 }
 
+int amdgpu_gmc_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid,
+				   uint32_t flush_type, bool all_hub,
+				   uint32_t inst)
+{
+	u32 usec_timeout = amdgpu_sriov_vf(adev) ? SRIOV_USEC_TIMEOUT :
+		adev->usec_timeout;
+	struct amdgpu_ring *ring = &adev->gfx.kiq[inst].ring;
+	struct amdgpu_kiq *kiq = &adev->gfx.kiq[inst];
+	unsigned int ndw;
+	signed long r;
+	uint32_t seq;
+
+	if (!adev->gmc.flush_pasid_uses_kiq || !ring->sched.ready ||
+	    !down_read_trylock(&adev->reset_domain->sem)) {
+		return adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid,
+								flush_type,
+								all_hub, inst);
+	}
+
+	/* 2 dwords flush + 8 dwords fence */
+	ndw = kiq->pmf->invalidate_tlbs_size + 8;
+
+	if (adev->gmc.flush_tlb_needs_extra_type_2)
+		ndw += kiq->pmf->invalidate_tlbs_size;
+
+	if (adev->gmc.flush_tlb_needs_extra_type_0)
+		ndw += kiq->pmf->invalidate_tlbs_size;
+
+	spin_lock(&adev->gfx.kiq[inst].ring_lock);
+	amdgpu_ring_alloc(ring, ndw);
+	if (adev->gmc.flush_tlb_needs_extra_type_2)
+		kiq->pmf->kiq_invalidate_tlbs(ring, pasid, 2, all_hub);
+
+	if (flush_type == 2 && adev->gmc.flush_tlb_needs_extra_type_0)
+		kiq->pmf->kiq_invalidate_tlbs(ring, pasid, 0, all_hub);
+
+	kiq->pmf->kiq_invalidate_tlbs(ring, pasid, flush_type, all_hub);
+	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
+	if (r) {
+		amdgpu_ring_undo(ring);
+		spin_unlock(&adev->gfx.kiq[inst].ring_lock);
+		goto error_unlock_reset;
+	}
+
+	amdgpu_ring_commit(ring);
+	spin_unlock(&adev->gfx.kiq[inst].ring_lock);
+	r = amdgpu_fence_wait_polling(ring, seq, usec_timeout);
+	if (r < 1) {
+		dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
+		r = -ETIME;
+		goto error_unlock_reset;
+	}
+	r = 0;
+
+error_unlock_reset:
+	up_read(&adev->reset_domain->sem);
+	return r;
+}
+
 /**
  * amdgpu_gmc_tmz_set -- check and set if a device supports TMZ
  * @adev: amdgpu_device pointer

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

@@ -335,11 +335,12 @@ struct amdgpu_gmc {
 	u64 MC_VM_MX_L1_TLB_CNTL;
 
 	u64 noretry_flags;
+
+	bool flush_tlb_needs_extra_type_0;
+	bool flush_tlb_needs_extra_type_2;
+	bool flush_pasid_uses_kiq;
 };
 
-#define amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, type, allhub, inst) \
-	((adev)->gmc.gmc_funcs->flush_gpu_tlb_pasid \
-	((adev), (pasid), (type), (allhub), (inst)))
 #define amdgpu_gmc_emit_flush_gpu_tlb(r, vmid, addr) (r)->adev->gmc.gmc_funcs->emit_flush_gpu_tlb((r), (vmid), (addr))
 #define amdgpu_gmc_emit_pasid_mapping(r, vmid, pasid) (r)->adev->gmc.gmc_funcs->emit_pasid_mapping((r), (vmid), (pasid))
 #define amdgpu_gmc_map_mtype(adev, flags) (adev)->gmc.gmc_funcs->map_mtype((adev),(flags))
@@ -404,6 +405,9 @@ void amdgpu_gmc_ras_fini(struct amdgpu_device *adev);
 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev);
 void amdgpu_gmc_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
 			      uint32_t vmhub, uint32_t flush_type);
+int amdgpu_gmc_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid,
+				   uint32_t flush_type, bool all_hub,
+				   uint32_t inst);
 
 extern void amdgpu_gmc_tmz_set(struct amdgpu_device *adev);
 extern void amdgpu_gmc_noretry_set(struct amdgpu_device *adev);

drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c

@@ -956,89 +956,30 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
 					uint16_t pasid, uint32_t flush_type,
 					bool all_hub, uint32_t inst)
 {
-	int vmid, i;
-	signed long r;
-	uint32_t seq;
-	uint16_t queried_pasid;
-	bool ret;
-	u32 usec_timeout = amdgpu_sriov_vf(adev) ? SRIOV_USEC_TIMEOUT : adev->usec_timeout;
-	struct amdgpu_ring *ring = &adev->gfx.kiq[inst].ring;
-	struct amdgpu_kiq *kiq = &adev->gfx.kiq[inst];
-
-	if (amdgpu_in_reset(adev))
-		return -EIO;
-
-	if (ring->sched.ready && down_read_trylock(&adev->reset_domain->sem)) {
-		/* Vega20+XGMI caches PTEs in TC and TLB. Add a
-		 * heavy-weight TLB flush (type 2), which flushes
-		 * both. Due to a race condition with concurrent
-		 * memory accesses using the same TLB cache line, we
-		 * still need a second TLB flush after this.
-		 */
-		bool vega20_xgmi_wa = (adev->gmc.xgmi.num_physical_nodes &&
-				       amdgpu_ip_version(adev, GC_HWIP, 0) ==
-					       IP_VERSION(9, 4, 0));
-		/* 2 dwords flush + 8 dwords fence */
-		unsigned int ndw = kiq->pmf->invalidate_tlbs_size + 8;
-
-		if (vega20_xgmi_wa)
-			ndw += kiq->pmf->invalidate_tlbs_size;
-
-		spin_lock(&adev->gfx.kiq[inst].ring_lock);
-		/* 2 dwords flush + 8 dwords fence */
-		amdgpu_ring_alloc(ring, ndw);
-		if (vega20_xgmi_wa)
-			kiq->pmf->kiq_invalidate_tlbs(ring,
-						      pasid, 2, all_hub);
-
-		if (flush_type == 2 &&
-		    amdgpu_ip_version(adev, GC_HWIP, 0) ==
-			    IP_VERSION(9, 4, 3) &&
-		    adev->rev_id == 0)
-			kiq->pmf->kiq_invalidate_tlbs(ring,
-						pasid, 0, all_hub);
-
-		kiq->pmf->kiq_invalidate_tlbs(ring,
-					pasid, flush_type, all_hub);
-		r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
-		if (r) {
-			amdgpu_ring_undo(ring);
-			spin_unlock(&adev->gfx.kiq[inst].ring_lock);
-			up_read(&adev->reset_domain->sem);
-			return -ETIME;
-		}
-
-		amdgpu_ring_commit(ring);
-		spin_unlock(&adev->gfx.kiq[inst].ring_lock);
-		r = amdgpu_fence_wait_polling(ring, seq, usec_timeout);
-		if (r < 1) {
-			dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
-			up_read(&adev->reset_domain->sem);
-			return -ETIME;
-		}
-		up_read(&adev->reset_domain->sem);
-		return 0;
-	}
+	uint16_t queried;
+	int i, vmid;
 
 	for (vmid = 1; vmid < 16; vmid++) {
+		bool valid;
 
-		ret = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
-				&queried_pasid);
-		if (ret && queried_pasid == pasid) {
-			if (all_hub) {
-				for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS)
-					gmc_v9_0_flush_gpu_tlb(adev, vmid,
-							i, flush_type);
-			} else {
-				gmc_v9_0_flush_gpu_tlb(adev, vmid,
-						AMDGPU_GFXHUB(0), flush_type);
-			}
-			break;
+		valid = gmc_v9_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
+								 &queried);
+		if (!valid || queried != pasid)
+			continue;
+
+		if (all_hub) {
+			for_each_set_bit(i, adev->vmhubs_mask,
+					 AMDGPU_MAX_VMHUBS)
+				gmc_v9_0_flush_gpu_tlb(adev, vmid, i,
+						       flush_type);
+		} else {
+			gmc_v9_0_flush_gpu_tlb(adev, vmid,
+					       AMDGPU_GFXHUB(0),
+					       flush_type);
 		}
 	}
 
 	return 0;
-
 }
 
 static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
@@ -2362,6 +2303,24 @@ static int gmc_v9_0_hw_init(void *handle)
 	bool value;
 	int i, r;
 
+	adev->gmc.flush_pasid_uses_kiq = true;
+
+	/* Vega20+XGMI caches PTEs in TC and TLB. Add a heavy-weight TLB flush
+	 * (type 2), which flushes both. Due to a race condition with
+	 * concurrent memory accesses using the same TLB cache line, we still
+	 * need a second TLB flush after this.
+	 */
+	adev->gmc.flush_tlb_needs_extra_type_2 =
+		amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 0) &&
+		adev->gmc.xgmi.num_physical_nodes;
+	/*
+	 * TODO: This workaround is badly documented and had a buggy
+	 * implementation. We should probably verify what we do here.
+	 */
+	adev->gmc.flush_tlb_needs_extra_type_0 =
+		amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(9, 4, 3) &&
+		adev->rev_id == 0;
+
 	/* The sequence of these two function calls matters.*/
 	gmc_v9_0_init_golden_registers(adev);