1 files changed, 141 insertions, 88 deletions

diff --git a/arch/powerpc/platforms/powernv/npu-dma.c b/arch/powerpc/platforms/powernv/npu-dma.c
index 4043109f4051..63f007f2de7e 100644
--- a/arch/powerpc/platforms/powernv/npu-dma.c
+++ b/arch/powerpc/platforms/powernv/npu-dma.c
@@ -413,6 +413,11 @@ struct npu_context {
 	void *priv;
 };
 
+struct mmio_atsd_reg {
+	struct npu *npu;
+	int reg;
+};
+
 /*
  * Find a free MMIO ATSD register and mark it in use. Return -ENOSPC
  * if none are available.
@@ -422,7 +427,7 @@ static int get_mmio_atsd_reg(struct npu *npu)
 	int i;
 
 	for (i = 0; i < npu->mmio_atsd_count; i++) {
-		if (!test_and_set_bit(i, &npu->mmio_atsd_usage))
+		if (!test_and_set_bit_lock(i, &npu->mmio_atsd_usage))
 			return i;
 	}
 
@@ -431,86 +436,90 @@ static int get_mmio_atsd_reg(struct npu *npu)
 
 static void put_mmio_atsd_reg(struct npu *npu, int reg)
 {
-	clear_bit(reg, &npu->mmio_atsd_usage);
+	clear_bit_unlock(reg, &npu->mmio_atsd_usage);
 }
 
 /* MMIO ATSD register offsets */
 #define XTS_ATSD_AVA  1
 #define XTS_ATSD_STAT 2
 
-static int mmio_launch_invalidate(struct npu *npu, unsigned long launch,
-				unsigned long va)
+static void mmio_launch_invalidate(struct mmio_atsd_reg *mmio_atsd_reg,
+				unsigned long launch, unsigned long va)
 {
-	int mmio_atsd_reg;
-
-	do {
-		mmio_atsd_reg = get_mmio_atsd_reg(npu);
-		cpu_relax();
-	} while (mmio_atsd_reg < 0);
+	struct npu *npu = mmio_atsd_reg->npu;
+	int reg = mmio_atsd_reg->reg;
 
 	__raw_writeq(cpu_to_be64(va),
-		npu->mmio_atsd_regs[mmio_atsd_reg] + XTS_ATSD_AVA);
+		npu->mmio_atsd_regs[reg] + XTS_ATSD_AVA);
 	eieio();
-	__raw_writeq(cpu_to_be64(launch), npu->mmio_atsd_regs[mmio_atsd_reg]);
-
-	return mmio_atsd_reg;
+	__raw_writeq(cpu_to_be64(launch), npu->mmio_atsd_regs[reg]);
 }
 
-static int mmio_invalidate_pid(struct npu *npu, unsigned long pid, bool flush)
+static void mmio_invalidate_pid(struct mmio_atsd_reg mmio_atsd_reg[NV_MAX_NPUS],
+				unsigned long pid, bool flush)
 {
+	int i;
 	unsigned long launch;
 
-	/* IS set to invalidate matching PID */
-	launch = PPC_BIT(12);
+	for (i = 0; i <= max_npu2_index; i++) {
+		if (mmio_atsd_reg[i].reg < 0)
+			continue;
+
+		/* IS set to invalidate matching PID */
+		launch = PPC_BIT(12);
 
-	/* PRS set to process-scoped */
-	launch |= PPC_BIT(13);
+		/* PRS set to process-scoped */
+		launch |= PPC_BIT(13);
 
-	/* AP */
-	launch |= (u64) mmu_get_ap(mmu_virtual_psize) << PPC_BITLSHIFT(17);
+		/* AP */
+		launch |= (u64)
+			mmu_get_ap(mmu_virtual_psize) << PPC_BITLSHIFT(17);
 
-	/* PID */
-	launch |= pid << PPC_BITLSHIFT(38);
+		/* PID */
+		launch |= pid << PPC_BITLSHIFT(38);
 
-	/* No flush */
-	launch |= !flush << PPC_BITLSHIFT(39);
+		/* No flush */
+		launch |= !flush << PPC_BITLSHIFT(39);
 
-	/* Invalidating the entire process doesn't use a va */
-	return mmio_launch_invalidate(npu, launch, 0);
+		/* Invalidating the entire process doesn't use a va */
+		mmio_launch_invalidate(&mmio_atsd_reg[i], launch, 0);
+	}
 }
 
-static int mmio_invalidate_va(struct npu *npu, unsigned long va,
-			unsigned long pid, bool flush)
+static void mmio_invalidate_va(struct mmio_atsd_reg mmio_atsd_reg[NV_MAX_NPUS],
+			unsigned long va, unsigned long pid, bool flush)
 {
+	int i;
 	unsigned long launch;
 
-	/* IS set to invalidate target VA */
-	launch = 0;
+	for (i = 0; i <= max_npu2_index; i++) {
+		if (mmio_atsd_reg[i].reg < 0)
+			continue;
+
+		/* IS set to invalidate target VA */
+		launch = 0;
 
-	/* PRS set to process scoped */
-	launch |= PPC_BIT(13);
+		/* PRS set to process scoped */
+		launch |= PPC_BIT(13);
 
-	/* AP */
-	launch |= (u64) mmu_get_ap(mmu_virtual_psize) << PPC_BITLSHIFT(17);
+		/* AP */
+		launch |= (u64)
+			mmu_get_ap(mmu_virtual_psize) << PPC_BITLSHIFT(17);
 
-	/* PID */
-	launch |= pid << PPC_BITLSHIFT(38);
+		/* PID */
+		launch |= pid << PPC_BITLSHIFT(38);
 
-	/* No flush */
-	launch |= !flush << PPC_BITLSHIFT(39);
+		/* No flush */
+		launch |= !flush << PPC_BITLSHIFT(39);
 
-	return mmio_launch_invalidate(npu, launch, va);
+		mmio_launch_invalidate(&mmio_atsd_reg[i], launch, va);
+	}
 }
 
 #define mn_to_npu_context(x) container_of(x, struct npu_context, mn)
 
-struct mmio_atsd_reg {
-	struct npu *npu;
-	int reg;
-};
-
 static void mmio_invalidate_wait(
-	struct mmio_atsd_reg mmio_atsd_reg[NV_MAX_NPUS], bool flush)
+	struct mmio_atsd_reg mmio_atsd_reg[NV_MAX_NPUS])
 {
 	struct npu *npu;
 	int i, reg;
@@ -525,16 +534,67 @@ static void mmio_invalidate_wait(
 		reg = mmio_atsd_reg[i].reg;
 		while (__raw_readq(npu->mmio_atsd_regs[reg] + XTS_ATSD_STAT))
 			cpu_relax();
+	}
+}
+
+/*
+ * Acquires all the address translation shootdown (ATSD) registers required to
+ * launch an ATSD on all links this npu_context is active on.
+ */
+static void acquire_atsd_reg(struct npu_context *npu_context,
+			struct mmio_atsd_reg mmio_atsd_reg[NV_MAX_NPUS])
+{
+	int i, j;
+	struct npu *npu;
+	struct pci_dev *npdev;
+	struct pnv_phb *nphb;
 
-		put_mmio_atsd_reg(npu, reg);
+	for (i = 0; i <= max_npu2_index; i++) {
+		mmio_atsd_reg[i].reg = -1;
+		for (j = 0; j < NV_MAX_LINKS; j++) {
+			/*
+			 * There are no ordering requirements with respect to
+			 * the setup of struct npu_context, but to ensure
+			 * consistent behaviour we need to ensure npdev[][] is
+			 * only read once.
+			 */
+			npdev = READ_ONCE(npu_context->npdev[i][j]);
+			if (!npdev)
+				continue;
 
+			nphb = pci_bus_to_host(npdev->bus)->private_data;
+			npu = &nphb->npu;
+			mmio_atsd_reg[i].npu = npu;
+			mmio_atsd_reg[i].reg = get_mmio_atsd_reg(npu);
+			while (mmio_atsd_reg[i].reg < 0) {
+				mmio_atsd_reg[i].reg = get_mmio_atsd_reg(npu);
+				cpu_relax();
+			}
+			break;
+		}
+	}
+}
+
+/*
+ * Release previously acquired ATSD registers. To avoid deadlocks the registers
+ * must be released in the same order they were acquired above in
+ * acquire_atsd_reg.
+ */
+static void release_atsd_reg(struct mmio_atsd_reg mmio_atsd_reg[NV_MAX_NPUS])
+{
+	int i;
+
+	for (i = 0; i <= max_npu2_index; i++) {
 		/*
-		 * The GPU requires two flush ATSDs to ensure all entries have
-		 * been flushed. We use PID 0 as it will never be used for a
-		 * process on the GPU.
+		 * We can't rely on npu_context->npdev[][] being the same here
+		 * as when acquire_atsd_reg() was called, hence we use the
+		 * values stored in mmio_atsd_reg during the acquire phase
+		 * rather than re-reading npdev[][].
 		 */
-		if (flush)
-			mmio_invalidate_pid(npu, 0, true);
+		if (mmio_atsd_reg[i].reg < 0)
+			continue;
+
+		put_mmio_atsd_reg(mmio_atsd_reg[i].npu, mmio_atsd_reg[i].reg);
 	}
 }
 
@@ -545,10 +605,6 @@ static void mmio_invalidate_wait(
 static void mmio_invalidate(struct npu_context *npu_context, int va,
 			unsigned long address, bool flush)
 {
-	int i, j;
-	struct npu *npu;
-	struct pnv_phb *nphb;
-	struct pci_dev *npdev;
 	struct mmio_atsd_reg mmio_atsd_reg[NV_MAX_NPUS];
 	unsigned long pid = npu_context->mm->context.id;
 
@@ -562,37 +618,25 @@ static void mmio_invalidate(struct npu_context *npu_context, int va,
 	 * Loop over all the NPUs this process is active on and launch
 	 * an invalidate.
 	 */
-	for (i = 0; i <= max_npu2_index; i++) {
-		mmio_atsd_reg[i].reg = -1;
-		for (j = 0; j < NV_MAX_LINKS; j++) {
-			npdev = npu_context->npdev[i][j];
-			if (!npdev)
-				continue;
-
-			nphb = pci_bus_to_host(npdev->bus)->private_data;
-			npu = &nphb->npu;
-			mmio_atsd_reg[i].npu = npu;
-
-			if (va)
-				mmio_atsd_reg[i].reg =
-					mmio_invalidate_va(npu, address, pid,
-							flush);
-			else
-				mmio_atsd_reg[i].reg =
-					mmio_invalidate_pid(npu, pid, flush);
-
-			/*
-			 * The NPU hardware forwards the shootdown to all GPUs
-			 * so we only have to launch one shootdown per NPU.
-			 */
-			break;
-		}
+	acquire_atsd_reg(npu_context, mmio_atsd_reg);
+	if (va)
+		mmio_invalidate_va(mmio_atsd_reg, address, pid, flush);
+	else
+		mmio_invalidate_pid(mmio_atsd_reg, pid, flush);
+
+	mmio_invalidate_wait(mmio_atsd_reg);
+	if (flush) {
+		/*
+		 * The GPU requires two flush ATSDs to ensure all entries have
+		 * been flushed. We use PID 0 as it will never be used for a
+		 * process on the GPU.
+		 */
+		mmio_invalidate_pid(mmio_atsd_reg, 0, true);
+		mmio_invalidate_wait(mmio_atsd_reg);
+		mmio_invalidate_pid(mmio_atsd_reg, 0, true);
+		mmio_invalidate_wait(mmio_atsd_reg);
 	}
-
-	mmio_invalidate_wait(mmio_atsd_reg, flush);
-	if (flush)
-		/* Wait for the flush to complete */
-		mmio_invalidate_wait(mmio_atsd_reg, false);
+	release_atsd_reg(mmio_atsd_reg);
 }
 
 static void pnv_npu2_mn_release(struct mmu_notifier *mn,
@@ -735,7 +779,16 @@ struct npu_context *pnv_npu2_init_context(struct pci_dev *gpdev,
 	if (WARN_ON(of_property_read_u32(nvlink_dn, "ibm,npu-link-index",
 							&nvlink_index)))
 		return ERR_PTR(-ENODEV);
-	npu_context->npdev[npu->index][nvlink_index] = npdev;
+
+	/*
+	 * npdev is a pci_dev pointer setup by the PCI code. We assign it to
+	 * npdev[][] to indicate to the mmu notifiers that an invalidation
+	 * should also be sent over this nvlink. The notifiers don't use any
+	 * other fields in npu_context, so we just need to ensure that when they
+	 * deference npu_context->npdev[][] it is either a valid pointer or
+	 * NULL.
+	 */
+	WRITE_ONCE(npu_context->npdev[npu->index][nvlink_index], npdev);
 
 	return npu_context;
 }
@@ -774,7 +827,7 @@ void pnv_npu2_destroy_context(struct npu_context *npu_context,
 	if (WARN_ON(of_property_read_u32(nvlink_dn, "ibm,npu-link-index",
 							&nvlink_index)))
 		return;
-	npu_context->npdev[npu->index][nvlink_index] = NULL;
+	WRITE_ONCE(npu_context->npdev[npu->index][nvlink_index], NULL);
 	opal_npu_destroy_context(nphb->opal_id, npu_context->mm->context.id,
 				PCI_DEVID(gpdev->bus->number, gpdev->devfn));
 	kref_put(&npu_context->kref, pnv_npu2_release_context);