spdk: Add block fill API to ioat driver

For those Crystal Beach DMA channels which support block fill capability,
we add a fill API here that can zero out pages or fill them with a
fixed pattern.

Change-Id: I8a57337702b951c703d494004b111f6d206279fb
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>

examples/ioat/verify/verify.c

@@ -49,6 +49,11 @@
 
 #define SRC_BUFFER_SIZE (512*1024)
 
+enum ioat_task_type {
+	IOAT_COPY_TYPE,
+	IOAT_FILL_TYPE,
+};
+
 struct user_config {
 	int queue_depth;
 	int time_in_sec;
@@ -67,6 +72,8 @@ static struct user_config g_user_config;
 struct thread_entry {
 	uint64_t xfer_completed;
 	uint64_t xfer_failed;
+	uint64_t fill_completed;
+	uint64_t fill_failed;
 	uint64_t current_queue_depth;
 	unsigned lcore_id;
 	bool is_draining;
@@ -75,9 +82,11 @@ struct thread_entry {
 };
 
 struct ioat_task {
+	enum ioat_task_type type;
 	struct thread_entry *thread_entry;
 	void *buffer;
 	int len;
+	uint64_t fill_pattern;
 	void *src;
 	void *dst;
 };
@@ -124,14 +133,29 @@ static void prepare_ioat_task(struct thread_entry *thread_entry, struct ioat_tas
 	int len;
 	int src_offset;
 	int dst_offset;
+	int num_ddwords;
+	uint64_t fill_pattern;
 
+	if (ioat_task->type == IOAT_FILL_TYPE) {
+		fill_pattern = rand_r(&seed);
+		fill_pattern = fill_pattern << 32 | rand_r(&seed);
+
+		/* ensure that the length of memset block is 8 Bytes aligned */
+		num_ddwords = (rand_r(&seed) % SRC_BUFFER_SIZE) / 8;
+		len = num_ddwords * 8;
+		if (len < 8)
+			len = 8;
+		dst_offset = rand_r(&seed) % (SRC_BUFFER_SIZE - len);
+		ioat_task->fill_pattern = fill_pattern;
+	} else {
 		src_offset = rand_r(&seed) % SRC_BUFFER_SIZE;
 		len = rand_r(&seed) % (SRC_BUFFER_SIZE - src_offset);
 		dst_offset = rand_r(&seed) % (SRC_BUFFER_SIZE - len);
 
 		memset(ioat_task->buffer, 0, SRC_BUFFER_SIZE);
-	ioat_task->len = len;
 		ioat_task->src =  g_src + src_offset;
+	}
+	ioat_task->len = len;
 	ioat_task->dst = ioat_task->buffer + dst_offset;
 	ioat_task->thread_entry = thread_entry;
 }
@@ -139,14 +163,31 @@ static void prepare_ioat_task(struct thread_entry *thread_entry, struct ioat_tas
 static void
 ioat_done(void *cb_arg)
 {
+	uint64_t *value;
+	int i, failed = 0;
 	struct ioat_task *ioat_task = (struct ioat_task *)cb_arg;
 	struct thread_entry *thread_entry = ioat_task->thread_entry;
 
+	if (ioat_task->type == IOAT_FILL_TYPE) {
+		value = (uint64_t *)ioat_task->dst;
+		for (i = 0; i < ioat_task->len / 8; i++) {
+			if (*value != ioat_task->fill_pattern) {
+				thread_entry->fill_failed++;
+				failed = 1;
+				break;
+			}
+			value++;
+		}
+		if (!failed)
+			thread_entry->fill_completed++;
+	} else {
 		if (memcmp(ioat_task->src, ioat_task->dst, ioat_task->len)) {
 			thread_entry->xfer_failed++;
 		} else {
 			thread_entry->xfer_completed++;
 		}
+	}
+
 	thread_entry->current_queue_depth--;
 	if (thread_entry->is_draining) {
 		rte_mempool_put(thread_entry->data_pool, ioat_task->buffer);
@@ -278,6 +319,9 @@ drain_xfers(struct thread_entry *thread_entry)
 static void
 submit_single_xfer(struct ioat_task *ioat_task)
 {
+	if (ioat_task->type == IOAT_FILL_TYPE)
+		ioat_submit_fill(ioat_task, ioat_done, ioat_task->dst, ioat_task->fill_pattern, ioat_task->len);
+	else
 		ioat_submit_copy(ioat_task, ioat_done, ioat_task->dst, ioat_task->src, ioat_task->len);
 	ioat_task->thread_entry->current_queue_depth++;
 }
@@ -290,6 +334,11 @@ submit_xfers(struct thread_entry *thread_entry, uint64_t queue_depth)
 		rte_mempool_get(thread_entry->task_pool, (void **)&ioat_task);
 		rte_mempool_get(thread_entry->data_pool, &(ioat_task->buffer));
 
+		ioat_task->type = IOAT_COPY_TYPE;
+		if (ioat_get_dma_capabilities() & IOAT_ENGINE_FILL_SUPPORTED) {
+			if (queue_depth % 2)
+				ioat_task->type = IOAT_FILL_TYPE;
+		}
 		prepare_ioat_task(thread_entry, ioat_task);
 		submit_single_xfer(ioat_task);
 	}
@@ -397,10 +446,12 @@ dump_result(struct thread_entry *threads, int len)
 	for (i = 0; i < len; i++) {
 		struct thread_entry *t = &threads[i];
 		total_completed += t->xfer_completed;
+		total_completed += t->fill_completed;
 		total_failed += t->xfer_failed;
+		total_failed += t->fill_failed;
 		if (t->xfer_completed || t->xfer_failed)
-			printf("lcore = %d, success = %ld, failed = %ld \n",
+			printf("lcore = %d, copy success = %ld, copy failed = %ld, fill success = %ld, fill failed = %ld \n",
-			       t->lcore_id, t->xfer_completed, t->xfer_failed);
+			       t->lcore_id, t->xfer_completed, t->xfer_failed, t->fill_completed, t->fill_failed);
 	}
 	return total_failed ? 1 : 0;
 }

include/spdk/ioat.h

@@ -90,6 +90,15 @@ void ioat_unregister_thread(void);
 int64_t ioat_submit_copy(void *cb_arg, ioat_callback_t cb_fn,
 			 void *dst, const void *src, uint64_t nbytes);
 
+/**
+ * Submit a DMA engine memory fill request.
+ *
+ * Before submitting any requests on a thread, the thread must be registered
+ * using the \ref ioat_register_thread() function.
+ */
+int64_t ioat_submit_fill(void *cb_arg, ioat_callback_t cb_fn,
+			 void *dst, uint64_t fill_pattern, uint64_t nbytes);
+
 /**
  * Check for completed requests on the current thread.
  *
@@ -100,4 +109,20 @@ int64_t ioat_submit_copy(void *cb_arg, ioat_callback_t cb_fn,
  */
 int ioat_process_events(void);
 
+/**
+ * DMA engine capability flags
+ */
+enum ioat_dma_capability_flags {
+	IOAT_ENGINE_COPY_SUPPORTED	= 0x1, /**< The memory copy is supported */
+	IOAT_ENGINE_FILL_SUPPORTED	= 0x2, /**< The memory fill is supported */
+};
+
+/**
+ * Get the DMA engine capabilities.
+ *
+ * Before submitting any requests on a thread, the thread must be registered
+ * using the \ref ioat_register_thread() function.
+ */
+uint32_t ioat_get_dma_capabilities(void);
+
 #endif

include/spdk/ioat_spec.h

@@ -53,6 +53,13 @@
 #define IOAT_CHANCTRL_ERR_COMPLETION_EN		0x0004
 #define IOAT_CHANCTRL_INT_REARM			0x0001
 
+/* DMA Channel Capabilities */
+#define	IOAT_DMACAP_PB			(1 << 0)
+#define	IOAT_DMACAP_DCA			(1 << 4)
+#define	IOAT_DMACAP_BFILL		(1 << 6)
+#define	IOAT_DMACAP_XOR			(1 << 8)
+#define	IOAT_DMACAP_PQ			(1 << 9)
+#define	IOAT_DMACAP_DMA_DIF		(1 << 10)
 
 struct ioat_registers {
 	uint8_t		chancnt;

lib/ioat/ioat.c

@@ -295,6 +295,37 @@ ioat_prep_copy(struct ioat_channel *ioat, uint64_t dst,
 	return desc;
 }
 
+static struct ioat_descriptor *
+ioat_prep_fill(struct ioat_channel *ioat, uint64_t dst,
+	       uint64_t fill_pattern, uint32_t len)
+{
+	struct ioat_descriptor *desc;
+	union ioat_hw_descriptor *hw_desc;
+
+	ioat_assert(len <= ioat->max_xfer_size);
+
+	if (ioat_get_ring_space(ioat) < 1) {
+		return NULL;
+	}
+
+	ioat_get_ring_entry(ioat, ioat->head, &desc, &hw_desc);
+
+	hw_desc->fill.u.control_raw = 0;
+	hw_desc->fill.u.control.op = IOAT_OP_FILL;
+	hw_desc->fill.u.control.completion_update = 1;
+
+	hw_desc->fill.size = len;
+	hw_desc->fill.src_data = fill_pattern;
+	hw_desc->fill.dest_addr = dst;
+
+	desc->callback_fn = NULL;
+	desc->callback_arg = NULL;
+
+	ioat_submit_single(ioat);
+
+	return desc;
+}
+
 static int ioat_reset_hw(struct ioat_channel *ioat)
 {
 	int timeout;
@@ -419,6 +450,10 @@ ioat_channel_start(struct ioat_channel *ioat)
 		return -1;
 	}
 
+	/* Always support DMA copy */
+	ioat->dma_capabilities = IOAT_ENGINE_COPY_SUPPORTED;
+	if (ioat->regs->dmacapability & IOAT_DMACAP_BFILL)
+		ioat->dma_capabilities |= IOAT_ENGINE_FILL_SUPPORTED;
 	xfercap = ioat->regs->xfercap;
 
 	/* Only bits [4:0] are valid. */
@@ -663,6 +698,75 @@ ioat_submit_copy(void *cb_arg, ioat_callback_t cb_fn,
 	return nbytes;
 }
 
+int64_t
+ioat_submit_fill(void *cb_arg, ioat_callback_t cb_fn,
+		 void *dst, uint64_t fill_pattern, uint64_t nbytes)
+{
+	struct ioat_channel	*ioat;
+	struct ioat_descriptor	*last_desc = NULL;
+	uint64_t	remaining, op_size;
+	uint64_t	vdst;
+	uint32_t	orig_head;
+
+	ioat = ioat_thread_channel;
+	if (!ioat) {
+		return -1;
+	}
+
+	if (!(ioat->dma_capabilities & IOAT_ENGINE_FILL_SUPPORTED)) {
+		ioat_printf(ioat, "Channel does not support memory fill\n");
+		return -1;
+	}
+
+	orig_head = ioat->head;
+
+	vdst = (uint64_t)dst;
+	remaining = nbytes;
+
+	while (remaining) {
+		op_size = remaining;
+		op_size = min(op_size, ioat->max_xfer_size);
+		remaining -= op_size;
+
+		last_desc = ioat_prep_fill(ioat,
+					   ioat_vtophys((void *)vdst),
+					   fill_pattern,
+					   op_size);
+
+		if (remaining == 0 || last_desc == NULL) {
+			break;
+		}
+
+		vdst += op_size;
+	}
+
+	if (last_desc) {
+		last_desc->callback_fn = cb_fn;
+		last_desc->callback_arg = cb_arg;
+	} else {
+		/*
+		 * Ran out of descriptors in the ring - reset head to leave things as they were
+		 * in case we managed to fill out any descriptors.
+		 */
+		ioat->head = orig_head;
+		return -1;
+	}
+
+	ioat_flush(ioat);
+	return nbytes;
+}
+
+uint32_t ioat_get_dma_capabilities(void)
+{
+	struct ioat_channel	*ioat;
+
+	ioat = ioat_thread_channel;
+	if (!ioat) {
+		return 0;
+	}
+	return ioat->dma_capabilities;
+}
+
 int ioat_process_events(void)
 {
 	if (!ioat_thread_channel) {

lib/ioat/ioat_internal.h

@@ -74,6 +74,7 @@ struct ioat_channel {
 	struct ioat_descriptor		*ring;
 	union ioat_hw_descriptor	*hw_ring;
 	uint64_t			hw_ring_phys_addr;
+	uint32_t			dma_capabilities;
 };
 
 static inline uint32_t