Merge tag 'misc-habanalabs-fixes-2019-08-12' of git://people.freedesktop.org/~gabbayo/linux into char-misc-next

Oded writes:

This tag contains a couple of important fixes:

- Four fixes when running on s390 architecture (BE). With these fixes, the
  driver is fully functional on Big-endian architectures. The fixes
  include:

  - Validation/Patching of user packets
  - Completion queue handling
  - Internal H/W queues submission
  - Device IRQ unmasking operation

- Fix to double free in an error path to avoid kernel corruption

- Fix to DRAM usage accounting when a user process is terminated
  forcefully.

* tag 'misc-habanalabs-fixes-2019-08-12' of git://people.freedesktop.org/~gabbayo/linux:
  habanalabs: fix device IRQ unmasking for BE host
  habanalabs: fix endianness handling for internal QMAN submission
  habanalabs: fix completion queue handling when host is BE
  habanalabs: fix endianness handling for packets from user
  habanalabs: fix DRAM usage accounting on context tear down
  habanalabs: Avoid double free in error flow

8 files changed, 90 insertions, 54 deletions

diff --git a/drivers/misc/habanalabs/device.c b/drivers/misc/habanalabs/device.c
index 0c4894dd9c02..7a8f9d0b71b5 100644
--- a/drivers/misc/habanalabs/device.c
+++ b/drivers/misc/habanalabs/device.c
@@ -970,7 +970,8 @@ int hl_device_init(struct hl_device *hdev, struct class *hclass)
 	rc = hl_ctx_init(hdev, hdev->kernel_ctx, true);
 	if (rc) {
 		dev_err(hdev->dev, "failed to initialize kernel context\n");
-		goto free_ctx;
+		kfree(hdev->kernel_ctx);
+		goto mmu_fini;
 	}
 
 	rc = hl_cb_pool_init(hdev);
@@ -1053,8 +1054,6 @@ release_ctx:
 	if (hl_ctx_put(hdev->kernel_ctx) != 1)
 		dev_err(hdev->dev,
 			"kernel ctx is still alive on initialization failure\n");
-free_ctx:
-	kfree(hdev->kernel_ctx);
 mmu_fini:
 	hl_mmu_fini(hdev);
 eq_fini:
diff --git a/drivers/misc/habanalabs/goya/goya.c b/drivers/misc/habanalabs/goya/goya.c
index a0e181714891..271c5c8f53b4 100644
--- a/drivers/misc/habanalabs/goya/goya.c
+++ b/drivers/misc/habanalabs/goya/goya.c
@@ -2729,9 +2729,10 @@ void goya_ring_doorbell(struct hl_device *hdev, u32 hw_queue_id, u32 pi)
 				GOYA_ASYNC_EVENT_ID_PI_UPDATE);
 }
 
-void goya_flush_pq_write(struct hl_device *hdev, u64 *pq, u64 exp_val)
+void goya_pqe_write(struct hl_device *hdev, __le64 *pqe, struct hl_bd *bd)
 {
-	/* Not needed in Goya */
+	/* The QMANs are on the SRAM so need to copy to IO space */
+	memcpy_toio((void __iomem *) pqe, bd, sizeof(struct hl_bd));
 }
 
 static void *goya_dma_alloc_coherent(struct hl_device *hdev, size_t size,
@@ -3313,9 +3314,11 @@ static int goya_validate_dma_pkt_no_mmu(struct hl_device *hdev,
 	int rc;
 
 	dev_dbg(hdev->dev, "DMA packet details:\n");
-	dev_dbg(hdev->dev, "source == 0x%llx\n", user_dma_pkt->src_addr);
-	dev_dbg(hdev->dev, "destination == 0x%llx\n", user_dma_pkt->dst_addr);
-	dev_dbg(hdev->dev, "size == %u\n", user_dma_pkt->tsize);
+	dev_dbg(hdev->dev, "source == 0x%llx\n",
+		le64_to_cpu(user_dma_pkt->src_addr));
+	dev_dbg(hdev->dev, "destination == 0x%llx\n",
+		le64_to_cpu(user_dma_pkt->dst_addr));
+	dev_dbg(hdev->dev, "size == %u\n", le32_to_cpu(user_dma_pkt->tsize));
 
 	ctl = le32_to_cpu(user_dma_pkt->ctl);
 	user_dir = (ctl & GOYA_PKT_LIN_DMA_CTL_DMA_DIR_MASK) >>
@@ -3344,9 +3347,11 @@ static int goya_validate_dma_pkt_mmu(struct hl_device *hdev,
 				struct packet_lin_dma *user_dma_pkt)
 {
 	dev_dbg(hdev->dev, "DMA packet details:\n");
-	dev_dbg(hdev->dev, "source == 0x%llx\n", user_dma_pkt->src_addr);
-	dev_dbg(hdev->dev, "destination == 0x%llx\n", user_dma_pkt->dst_addr);
-	dev_dbg(hdev->dev, "size == %u\n", user_dma_pkt->tsize);
+	dev_dbg(hdev->dev, "source == 0x%llx\n",
+		le64_to_cpu(user_dma_pkt->src_addr));
+	dev_dbg(hdev->dev, "destination == 0x%llx\n",
+		le64_to_cpu(user_dma_pkt->dst_addr));
+	dev_dbg(hdev->dev, "size == %u\n", le32_to_cpu(user_dma_pkt->tsize));
 
 	/*
 	 * WA for HW-23.
@@ -3386,7 +3391,8 @@ static int goya_validate_wreg32(struct hl_device *hdev,
 
 	dev_dbg(hdev->dev, "WREG32 packet details:\n");
 	dev_dbg(hdev->dev, "reg_offset == 0x%x\n", reg_offset);
-	dev_dbg(hdev->dev, "value      == 0x%x\n", wreg_pkt->value);
+	dev_dbg(hdev->dev, "value      == 0x%x\n",
+		le32_to_cpu(wreg_pkt->value));
 
 	if (reg_offset != (mmDMA_CH_0_WR_COMP_ADDR_LO & 0x1FFF)) {
 		dev_err(hdev->dev, "WREG32 packet with illegal address 0x%x\n",
@@ -3428,12 +3434,13 @@ static int goya_validate_cb(struct hl_device *hdev,
 	while (cb_parsed_length < parser->user_cb_size) {
 		enum packet_id pkt_id;
 		u16 pkt_size;
-		void *user_pkt;
+		struct goya_packet *user_pkt;
 
-		user_pkt = (void *) (uintptr_t)
+		user_pkt = (struct goya_packet *) (uintptr_t)
 			(parser->user_cb->kernel_address + cb_parsed_length);
 
-		pkt_id = (enum packet_id) (((*(u64 *) user_pkt) &
+		pkt_id = (enum packet_id) (
+				(le64_to_cpu(user_pkt->header) &
 				PACKET_HEADER_PACKET_ID_MASK) >>
 					PACKET_HEADER_PACKET_ID_SHIFT);
 
@@ -3453,7 +3460,8 @@ static int goya_validate_cb(struct hl_device *hdev,
 			 * need to validate here as well because patch_cb() is
 			 * not called in MMU path while this function is called
 			 */
-			rc = goya_validate_wreg32(hdev, parser, user_pkt);
+			rc = goya_validate_wreg32(hdev,
+				parser, (struct packet_wreg32 *) user_pkt);
 			break;
 
 		case PACKET_WREG_BULK:
@@ -3481,10 +3489,10 @@ static int goya_validate_cb(struct hl_device *hdev,
 		case PACKET_LIN_DMA:
 			if (is_mmu)
 				rc = goya_validate_dma_pkt_mmu(hdev, parser,
-						user_pkt);
+					(struct packet_lin_dma *) user_pkt);
 			else
 				rc = goya_validate_dma_pkt_no_mmu(hdev, parser,
-						user_pkt);
+					(struct packet_lin_dma *) user_pkt);
 			break;
 
 		case PACKET_MSG_LONG:
@@ -3657,15 +3665,16 @@ static int goya_patch_cb(struct hl_device *hdev,
 		enum packet_id pkt_id;
 		u16 pkt_size;
 		u32 new_pkt_size = 0;
-		void *user_pkt, *kernel_pkt;
+		struct goya_packet *user_pkt, *kernel_pkt;
 
-		user_pkt = (void *) (uintptr_t)
+		user_pkt = (struct goya_packet *) (uintptr_t)
 			(parser->user_cb->kernel_address + cb_parsed_length);
-		kernel_pkt = (void *) (uintptr_t)
+		kernel_pkt = (struct goya_packet *) (uintptr_t)
 			(parser->patched_cb->kernel_address +
 					cb_patched_cur_length);
 
-		pkt_id = (enum packet_id) (((*(u64 *) user_pkt) &
+		pkt_id = (enum packet_id) (
+				(le64_to_cpu(user_pkt->header) &
 				PACKET_HEADER_PACKET_ID_MASK) >>
 					PACKET_HEADER_PACKET_ID_SHIFT);
 
@@ -3680,15 +3689,18 @@ static int goya_patch_cb(struct hl_device *hdev,
 
 		switch (pkt_id) {
 		case PACKET_LIN_DMA:
-			rc = goya_patch_dma_packet(hdev, parser, user_pkt,
-						kernel_pkt, &new_pkt_size);
+			rc = goya_patch_dma_packet(hdev, parser,
+					(struct packet_lin_dma *) user_pkt,
+					(struct packet_lin_dma *) kernel_pkt,
+					&new_pkt_size);
 			cb_patched_cur_length += new_pkt_size;
 			break;
 
 		case PACKET_WREG_32:
 			memcpy(kernel_pkt, user_pkt, pkt_size);
 			cb_patched_cur_length += pkt_size;
-			rc = goya_validate_wreg32(hdev, parser, kernel_pkt);
+			rc = goya_validate_wreg32(hdev, parser,
+					(struct packet_wreg32 *) kernel_pkt);
 			break;
 
 		case PACKET_WREG_BULK:
@@ -4352,6 +4364,8 @@ static int goya_unmask_irq_arr(struct hl_device *hdev, u32 *irq_arr,
 	size_t total_pkt_size;
 	long result;
 	int rc;
+	int irq_num_entries, irq_arr_index;
+	__le32 *goya_irq_arr;
 
 	total_pkt_size = sizeof(struct armcp_unmask_irq_arr_packet) +
 			irq_arr_size;
@@ -4369,8 +4383,16 @@ static int goya_unmask_irq_arr(struct hl_device *hdev, u32 *irq_arr,
 	if (!pkt)
 		return -ENOMEM;
 
-	pkt->length = cpu_to_le32(irq_arr_size / sizeof(irq_arr[0]));
-	memcpy(&pkt->irqs, irq_arr, irq_arr_size);
+	irq_num_entries = irq_arr_size / sizeof(irq_arr[0]);
+	pkt->length = cpu_to_le32(irq_num_entries);
+
+	/* We must perform any necessary endianness conversation on the irq
+	 * array being passed to the goya hardware
+	 */
+	for (irq_arr_index = 0, goya_irq_arr = (__le32 *) &pkt->irqs;
+			irq_arr_index < irq_num_entries ; irq_arr_index++)
+		goya_irq_arr[irq_arr_index] =
+				cpu_to_le32(irq_arr[irq_arr_index]);
 
 	pkt->armcp_pkt.ctl = cpu_to_le32(ARMCP_PACKET_UNMASK_RAZWI_IRQ_ARRAY <<
 						ARMCP_PKT_CTL_OPCODE_SHIFT);
@@ -5042,7 +5064,7 @@ static const struct hl_asic_funcs goya_funcs = {
 	.resume = goya_resume,
 	.cb_mmap = goya_cb_mmap,
 	.ring_doorbell = goya_ring_doorbell,
-	.flush_pq_write = goya_flush_pq_write,
+	.pqe_write = goya_pqe_write,
 	.asic_dma_alloc_coherent = goya_dma_alloc_coherent,
 	.asic_dma_free_coherent = goya_dma_free_coherent,
 	.get_int_queue_base = goya_get_int_queue_base,
diff --git a/drivers/misc/habanalabs/goya/goyaP.h b/drivers/misc/habanalabs/goya/goyaP.h
index f8c611883dc1..d7f48c9c41cd 100644
--- a/drivers/misc/habanalabs/goya/goyaP.h
+++ b/drivers/misc/habanalabs/goya/goyaP.h
@@ -177,7 +177,7 @@ int goya_late_init(struct hl_device *hdev);
 void goya_late_fini(struct hl_device *hdev);
 
 void goya_ring_doorbell(struct hl_device *hdev, u32 hw_queue_id, u32 pi);
-void goya_flush_pq_write(struct hl_device *hdev, u64 *pq, u64 exp_val);
+void goya_pqe_write(struct hl_device *hdev, __le64 *pqe, struct hl_bd *bd);
 void goya_update_eq_ci(struct hl_device *hdev, u32 val);
 void goya_restore_phase_topology(struct hl_device *hdev);
 int goya_context_switch(struct hl_device *hdev, u32 asid);
diff --git a/drivers/misc/habanalabs/habanalabs.h b/drivers/misc/habanalabs/habanalabs.h
index 6a4c64b97f38..ce83adafcf2d 100644
--- a/drivers/misc/habanalabs/habanalabs.h
+++ b/drivers/misc/habanalabs/habanalabs.h
@@ -441,7 +441,11 @@ enum hl_pll_frequency {
  * @resume: handles IP specific H/W or SW changes for resume.
  * @cb_mmap: maps a CB.
  * @ring_doorbell: increment PI on a given QMAN.
- * @flush_pq_write: flush PQ entry write if necessary, WARN if flushing failed.
+ * @pqe_write: Write the PQ entry to the PQ. This is ASIC-specific
+ *             function because the PQs are located in different memory areas
+ *             per ASIC (SRAM, DRAM, Host memory) and therefore, the method of
+ *             writing the PQE must match the destination memory area
+ *             properties.
  * @asic_dma_alloc_coherent: Allocate coherent DMA memory by calling
  *                           dma_alloc_coherent(). This is ASIC function because
  *                           its implementation is not trivial when the driver
@@ -510,7 +514,8 @@ struct hl_asic_funcs {
 	int (*cb_mmap)(struct hl_device *hdev, struct vm_area_struct *vma,
 			u64 kaddress, phys_addr_t paddress, u32 size);
 	void (*ring_doorbell)(struct hl_device *hdev, u32 hw_queue_id, u32 pi);
-	void (*flush_pq_write)(struct hl_device *hdev, u64 *pq, u64 exp_val);
+	void (*pqe_write)(struct hl_device *hdev, __le64 *pqe,
+			struct hl_bd *bd);
 	void* (*asic_dma_alloc_coherent)(struct hl_device *hdev, size_t size,
 					dma_addr_t *dma_handle, gfp_t flag);
 	void (*asic_dma_free_coherent)(struct hl_device *hdev, size_t size,
diff --git a/drivers/misc/habanalabs/hw_queue.c b/drivers/misc/habanalabs/hw_queue.c
index e3b5517897ea..5f5673b74985 100644
--- a/drivers/misc/habanalabs/hw_queue.c
+++ b/drivers/misc/habanalabs/hw_queue.c
@@ -290,23 +290,19 @@ static void int_hw_queue_schedule_job(struct hl_cs_job *job)
 	struct hl_device *hdev = job->cs->ctx->hdev;
 	struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
 	struct hl_bd bd;
-	u64 *pi, *pbd = (u64 *) &bd;
+	__le64 *pi;
 
 	bd.ctl = 0;
-	bd.len = __cpu_to_le32(job->job_cb_size);
-	bd.ptr = __cpu_to_le64((u64) (uintptr_t) job->user_cb);
+	bd.len = cpu_to_le32(job->job_cb_size);
+	bd.ptr = cpu_to_le64((u64) (uintptr_t) job->user_cb);
 
-	pi = (u64 *) (uintptr_t) (q->kernel_address +
+	pi = (__le64 *) (uintptr_t) (q->kernel_address +
 		((q->pi & (q->int_queue_len - 1)) * sizeof(bd)));
 
-	pi[0] = pbd[0];
-	pi[1] = pbd[1];
-
 	q->pi++;
 	q->pi &= ((q->int_queue_len << 1) - 1);
 
-	/* Flush PQ entry write. Relevant only for specific ASICs */
-	hdev->asic_funcs->flush_pq_write(hdev, pi, pbd[0]);
+	hdev->asic_funcs->pqe_write(hdev, pi, &bd);
 
 	hdev->asic_funcs->ring_doorbell(hdev, q->hw_queue_id, q->pi);
 }
diff --git a/drivers/misc/habanalabs/include/goya/goya_packets.h b/drivers/misc/habanalabs/include/goya/goya_packets.h
index a14407b975e4..ef54bad20509 100644
--- a/drivers/misc/habanalabs/include/goya/goya_packets.h
+++ b/drivers/misc/habanalabs/include/goya/goya_packets.h
@@ -52,6 +52,19 @@ enum goya_dma_direction {
 #define GOYA_PKT_CTL_MB_SHIFT		31
 #define GOYA_PKT_CTL_MB_MASK		0x80000000
 
+/* All packets have, at least, an 8-byte header, which contains
+ * the packet type. The kernel driver uses the packet header for packet
+ * validation and to perform any necessary required preparation before
+ * sending them off to the hardware.
+ */
+struct goya_packet {
+	__le64 header;
+	/* The rest of the packet data follows. Use the corresponding
+	 * packet_XXX struct to deference the data, based on packet type
+	 */
+	u8 contents[0];
+};
+
 struct packet_nop {
 	__le32 reserved;
 	__le32 ctl;
diff --git a/drivers/misc/habanalabs/irq.c b/drivers/misc/habanalabs/irq.c
index ea9f72ff456c..199791b57caf 100644
--- a/drivers/misc/habanalabs/irq.c
+++ b/drivers/misc/habanalabs/irq.c
@@ -80,8 +80,7 @@ irqreturn_t hl_irq_handler_cq(int irq, void *arg)
 	struct hl_cs_job *job;
 	bool shadow_index_valid;
 	u16 shadow_index;
-	u32 *cq_entry;
-	u32 *cq_base;
+	struct hl_cq_entry *cq_entry, *cq_base;
 
 	if (hdev->disabled) {
 		dev_dbg(hdev->dev,
@@ -90,29 +89,29 @@ irqreturn_t hl_irq_handler_cq(int irq, void *arg)
 		return IRQ_HANDLED;
 	}
 
-	cq_base = (u32 *) (uintptr_t) cq->kernel_address;
+	cq_base = (struct hl_cq_entry *) (uintptr_t) cq->kernel_address;
 
 	while (1) {
-		bool entry_ready = ((cq_base[cq->ci] & CQ_ENTRY_READY_MASK)
+		bool entry_ready = ((le32_to_cpu(cq_base[cq->ci].data) &
+					CQ_ENTRY_READY_MASK)
 						>> CQ_ENTRY_READY_SHIFT);
 
 		if (!entry_ready)
 			break;
 
-		cq_entry = (u32 *) &cq_base[cq->ci];
+		cq_entry = (struct hl_cq_entry *) &cq_base[cq->ci];
 
-		/*
-		 * Make sure we read CQ entry contents after we've
+		/* Make sure we read CQ entry contents after we've
 		 * checked the ownership bit.
 		 */
 		dma_rmb();
 
-		shadow_index_valid =
-			((*cq_entry & CQ_ENTRY_SHADOW_INDEX_VALID_MASK)
+		shadow_index_valid = ((le32_to_cpu(cq_entry->data) &
+					CQ_ENTRY_SHADOW_INDEX_VALID_MASK)
 					>> CQ_ENTRY_SHADOW_INDEX_VALID_SHIFT);
 
-		shadow_index = (u16)
-			((*cq_entry & CQ_ENTRY_SHADOW_INDEX_MASK)
+		shadow_index = (u16) ((le32_to_cpu(cq_entry->data) &
+					CQ_ENTRY_SHADOW_INDEX_MASK)
 					>> CQ_ENTRY_SHADOW_INDEX_SHIFT);
 
 		queue = &hdev->kernel_queues[cq->hw_queue_id];
@@ -122,8 +121,7 @@ irqreturn_t hl_irq_handler_cq(int irq, void *arg)
 			queue_work(hdev->cq_wq, &job->finish_work);
 		}
 
-		/*
-		 * Update ci of the context's queue. There is no
+		/* Update ci of the context's queue. There is no
 		 * need to protect it with spinlock because this update is
 		 * done only inside IRQ and there is a different IRQ per
 		 * queue
@@ -131,7 +129,8 @@ irqreturn_t hl_irq_handler_cq(int irq, void *arg)
 		queue->ci = hl_queue_inc_ptr(queue->ci);
 
 		/* Clear CQ entry ready bit */
-		cq_base[cq->ci] &= ~CQ_ENTRY_READY_MASK;
+		cq_entry->data = cpu_to_le32(le32_to_cpu(cq_entry->data) &
+						~CQ_ENTRY_READY_MASK);
 
 		cq->ci = hl_cq_inc_ptr(cq->ci);
 
diff --git a/drivers/misc/habanalabs/memory.c b/drivers/misc/habanalabs/memory.c
index 42d237cae1dc..365fb0cb8dff 100644
--- a/drivers/misc/habanalabs/memory.c
+++ b/drivers/misc/habanalabs/memory.c
@@ -1629,6 +1629,8 @@ void hl_vm_ctx_fini(struct hl_ctx *ctx)
 			dev_dbg(hdev->dev,
 				"page list 0x%p of asid %d is still alive\n",
 				phys_pg_list, ctx->asid);
+			atomic64_sub(phys_pg_list->total_size,
+					&hdev->dram_used_mem);
 			free_phys_pg_pack(hdev, phys_pg_list);
 			idr_remove(&vm->phys_pg_pack_handles, i);
 		}