dm: storage: support discard command

Support DISCARD command is meaningful when eMMC usage is high or
there are lots of remove operations. For example, when Guest
Android is running, there will be lots of files being created and
removed. However, virtio-blk BE does not support DISCARD command,
data remove operation in UOS will not trigger erase in eMMC. After
period of time, the eMMC will be consumed out, and erase must be
done by eMMC firmware before writing any new data. This causes the
eMMC performance decrease in the whole system (SOS and UOS).
To solve the problem, DISCARD should be supported in virtio-blk BE.

Tracked-On: #2011
Signed-off-by: Conghui Chen <conghui.chen@intel.com>
Reviewed-by: Shuo A Liu <shuo.a.liu@intel.com>
Acked-by: Yu Wang <yu1.wang@intel.com>

devicemodel/hw/block_if.c

@@ -30,6 +30,7 @@
 #include <sys/queue.h>
 #include <sys/stat.h>
 #include <sys/ioctl.h>
+#include <linux/falloc.h>
 #include <linux/fs.h>
 #include <errno.h>
 #include <assert.h>
@@ -63,6 +64,7 @@
 
 #define BLOCKIF_NUMTHR	8
 #define BLOCKIF_MAXREQ	(64 + BLOCKIF_NUMTHR)
+#define MAX_DISCARD_SEGMENT	256
 
 /*
  * Debug printf
@@ -108,6 +110,9 @@ struct blockif_ctxt {
 	int			sectsz;
 	int			psectsz;
 	int			psectoff;
+	int			max_discard_sectors;
+	int			max_discard_seg;
+	int			discard_sector_alignment;
 	int			closing;
 	pthread_t		btid[BLOCKIF_NUMTHR];
 	pthread_mutex_t		mtx;
@@ -132,6 +137,12 @@ struct blockif_sig_elem {
 	struct blockif_sig_elem		*next;
 };
 
+struct discard_range {
+	uint64_t sector;
+	uint32_t num_sectors;
+	uint32_t flags;
+};
+
 static struct blockif_sig_elem *blockif_bse_head;
 
 static int
@@ -232,11 +243,92 @@ blockif_complete(struct blockif_ctxt *bc, struct blockif_elem *be)
 	TAILQ_INSERT_TAIL(&bc->freeq, be, link);
 }
 
+static int
+discard_range_validate(struct blockif_ctxt *bc, off_t start, off_t size)
+{
+	off_t start_sector = start / DEV_BSIZE;
+	off_t size_sector = size / DEV_BSIZE;
+
+	if (!size || (start + size) > (bc->size + bc->sub_file_start_lba))
+		return -1;
+
+	if ((size_sector > bc->max_discard_sectors) ||
+			(bc->discard_sector_alignment &&
+			start_sector % bc->discard_sector_alignment))
+		return -1;
+	return 0;
+}
+
+static int
+blockif_process_discard(struct blockif_ctxt *bc, struct blockif_req *br)
+{
+	int err;
+	struct discard_range *range;
+	int n_range, i, segment;
+	off_t arg[MAX_DISCARD_SEGMENT][2];
+
+	err = 0;
+	n_range = 0;
+	segment = 0;
+	if (!bc->candiscard)
+		return EOPNOTSUPP;
+
+	if (bc->rdonly)
+		return EROFS;
+
+	if (br->iovcnt == 1) {
+		/* virtio-blk use iov to transfer discard range */
+		n_range = br->iov[0].iov_len/sizeof(*range);
+		range = br->iov[0].iov_base;
+		for (i = 0; i < n_range; i++) {
+			arg[i][0] = range[i].sector * DEV_BSIZE +
+					bc->sub_file_start_lba;
+			arg[i][1] = range[i].num_sectors * DEV_BSIZE;
+			segment++;
+			if (segment > bc->max_discard_seg) {
+				WPRINTF(("segment > max_discard_seg\n"));
+				return EINVAL;
+			}
+			if (discard_range_validate(bc, arg[i][0], arg[i][1])) {
+				WPRINTF(("range [%ld: %ld] is invalid\n", arg[i][0], arg[i][1]));
+				return EINVAL;
+			}
+		}
+	} else {
+		/* ahci parse discard range to br->offset and br->reside */
+		arg[0][0] = br->offset + bc->sub_file_start_lba;
+		arg[0][1] = br->resid;
+		segment = 1;
+	}
+	for (i = 0; i < segment; i++) {
+		if (bc->isblk) {
+			err = ioctl(bc->fd, BLKDISCARD, arg[i]);
+		} else {
+			/* FALLOC_FL_PUNCH_HOLE:
+			 *	Deallocates space in the byte range starting at offset and
+			 *	continuing for length bytes.  After a successful call,
+			 *	subsequent reads from this range will return zeroes.
+			 * FALLOC_FL_KEEP_SIZE:
+			 *	Do not modify the apparent length of the file.
+			 */
+			err = fallocate(bc->fd, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
+				arg[i][0], arg[i][1]);
+		}
+		if (err) {
+			WPRINTF(("Failed to discard offset=%ld nbytes=%ld err code: %d\n",
+				 arg[i][0], arg[i][1], err));
+			return err;
+		}
+	}
+	br->resid = 0;
+
+	return 0;
+}
+
 static void
 blockif_proc(struct blockif_ctxt *bc, struct blockif_elem *be)
 {
 	struct blockif_req *br;
-	off_t arg[2];
 	ssize_t len;
 	int err;
 
@@ -271,21 +363,7 @@ blockif_proc(struct blockif_ctxt *bc, struct blockif_elem *be)
 			err = errno;
 		break;
 	case BOP_DISCARD:
-		/* only used by AHCI */
+		err = blockif_process_discard(bc, br);
-		if (!bc->candiscard)
-			err = EOPNOTSUPP;
-		else if (bc->rdonly)
-			err = EROFS;
-		else if (bc->isblk) {
-			arg[0] = br->offset;
-			arg[1] = br->resid;
-			if (ioctl(bc->fd, BLKDISCARD, arg))
-				err = errno;
-			else
-				br->resid = 0;
-		}
-		else
-			err = EOPNOTSUPP;
 		break;
 	default:
 		err = EINVAL;
@@ -425,6 +503,8 @@ blockif_open(const char *optstr, const char *ident)
 	int err_code = -1;
 	off_t sub_file_start_lba, sub_file_size;
 	int sub_file_assign;
+	int max_discard_sectors, max_discard_seg, discard_sector_alignment;
+	off_t probe_arg[] = {0, 0};
 
 	pthread_once(&blockif_once, blockif_init);
 
@@ -436,9 +516,15 @@ blockif_open(const char *optstr, const char *ident)
 	sub_file_start_lba = 0;
 	sub_file_size = 0;
 
+	max_discard_sectors = -1;
+	max_discard_seg = -1;
+	discard_sector_alignment = -1;
+
 	/* writethru is on by default */
 	writeback = 0;
 
+	candiscard = 0;
+
 	/*
 	 * The first element in the optstring is always a pathname.
 	 * Optional elements follow
@@ -458,7 +544,18 @@ blockif_open(const char *optstr, const char *ident)
 			writeback = 0;
 		else if (!strcmp(cp, "ro"))
 			ro = 1;
-		else if (!strncmp(cp, "sectorsize", strlen("sectorsize"))) {
+		else if (!strncmp(cp, "discard", strlen("discard"))) {
+			strsep(&cp, "=");
+			if (cp != NULL) {
+				if (!(!dm_strtoi(cp, &cp, 10, &max_discard_sectors) &&
+					*cp == ':' &&
+					!dm_strtoi(cp + 1, &cp, 10, &max_discard_seg) &&
+					*cp == ':' &&
+					!dm_strtoi(cp + 1, &cp, 10, &discard_sector_alignment)))
+					goto err;
+			}
+			candiscard = 1;
+		} else if (!strncmp(cp, "sectorsize", strlen("sectorsize"))) {
 			/*
 			 *  sectorsize=<sector size>
 			 * or
@@ -517,7 +614,6 @@ blockif_open(const char *optstr, const char *ident)
 	size = sbuf.st_size;
 	sectsz = DEV_BSIZE;
 	psectsz = psectoff = 0;
-	candiscard = 0;
 
 	if (S_ISBLK(sbuf.st_mode)) {
 		/* get size */
@@ -552,8 +648,22 @@ blockif_open(const char *optstr, const char *ident)
 		DPRINTF(("block partition physical sector size is 0x%lx\n",
 			 psectsz));
 
-	} else
+		if (candiscard) {
+			err_code = ioctl(fd, BLKDISCARD, probe_arg);
+			if (err_code) {
+				WPRINTF(("not support DISCARD\n"));
+				candiscard = 0;
+			}
+		}
+
+	} else {
+		if (size < DEV_BSIZE || (size & (DEV_BSIZE - 1))) {
+			WPRINTF(("%s size not corret, should be multiple of %d\n",
+						nopt, DEV_BSIZE));
+			return 0;
+		}
 		psectsz = sbuf.st_blksize;
+	}
 
 	if (ssopt != 0) {
 		if (!powerof2(ssopt) || !powerof2(pssopt) || ssopt < 512 ||
@@ -613,6 +723,15 @@ blockif_open(const char *optstr, const char *ident)
 	bc->fd = fd;
 	bc->isblk = S_ISBLK(sbuf.st_mode);
 	bc->candiscard = candiscard;
+	if (candiscard) {
+		bc->max_discard_sectors =
+			(max_discard_sectors != -1) ?
+				max_discard_sectors : (size / DEV_BSIZE);
+		bc->max_discard_seg =
+			(max_discard_seg != -1) ? max_discard_seg : 1;
+		bc->discard_sector_alignment =
+			(discard_sector_alignment != -1) ? discard_sector_alignment : 0;
+	}
 	bc->rdonly = ro;
 	bc->size = size;
 	bc->sectsz = sectsz;
@@ -906,6 +1025,27 @@ blockif_candiscard(struct blockif_ctxt *bc)
 	return bc->candiscard;
 }
 
+int
+blockif_max_discard_sectors(struct blockif_ctxt *bc)
+{
+	assert(bc->magic == BLOCKIF_SIG);
+	return bc->max_discard_sectors;
+}
+
+int
+blockif_max_discard_seg(struct blockif_ctxt *bc)
+{
+	assert(bc->magic == BLOCKIF_SIG);
+	return bc->max_discard_seg;
+}
+
+int
+blockif_discard_sector_alignment(struct blockif_ctxt *bc)
+{
+	assert(bc->magic == BLOCKIF_SIG);
+	return bc->discard_sector_alignment;
+}
+
 uint8_t
 blockif_get_wce(struct blockif_ctxt *bc)
 {

devicemodel/hw/pci/virtio/virtio_block.c

@@ -58,6 +58,8 @@
 /* Device can toggle its cache between writeback and writethrough modes */
 #define	VIRTIO_BLK_F_CONFIG_WCE	(1 << 11)
 
+#define	VIRTIO_BLK_F_DISCARD	(1 << 13)
+
 /*
  * Basic device capabilities
  */
@@ -94,6 +96,15 @@ struct virtio_blk_config {
 		uint32_t opt_io_size;
 	} topology;
 	uint8_t	writeback;
+	uint8_t unused;
+	/* Reserve for num_queues when VIRTIO_BLK_F_MQ is support*/
+	uint16_t reserve;
+	/* The maximum discard sectors (in 512-byte sectors) for one segment */
+	uint32_t max_discard_sectors;
+	/* The maximum number of discard segments */
+	uint32_t max_discard_seg;
+	/* Discard commands must be aligned to this number of sectors. */
+	uint32_t discard_sector_alignment;
 } __attribute__((packed));
 
 /*
@@ -105,6 +116,7 @@ struct virtio_blk_hdr {
 #define	VBH_OP_FLUSH		4
 #define	VBH_OP_FLUSH_OUT	5
 #define	VBH_OP_IDENT		8
+#define	VBH_OP_DISCARD		11
 #define	VBH_FLAG_BARRIER	0x80000000	/* OR'ed into type */
 	uint32_t type;
 	uint32_t ioprio;
@@ -234,12 +246,13 @@ virtio_blk_proc(struct virtio_blk *blk, struct virtio_vq_info *vq)
 	 * we don't advertise the capability.
 	 */
 	type = vbh->type & ~VBH_FLAG_BARRIER;
-	writeop = (type == VBH_OP_WRITE);
+	writeop = ((type == VBH_OP_WRITE) ||
+			(type == VBH_OP_DISCARD));
 
 	iolen = 0;
 	for (i = 1; i < n; i++) {
 		/*
-		 * - write op implies read-only descriptor,
+		 * - write/discard op implies read-only descriptor,
 		 * - read/ident op implies write-only descriptor,
 		 * therefore test the inverse of the descriptor bit
 		 * to the op.
@@ -250,7 +263,7 @@ virtio_blk_proc(struct virtio_blk *blk, struct virtio_vq_info *vq)
 	io->req.resid = iolen;
 
 	DPRINTF(("virtio_blk: %s op, %zd bytes, %d segs, offset %ld\n\r",
-		 writeop ? "write" : "read/ident", iolen, i - 1,
+		 writeop ? "write/discard" : "read/ident", iolen, i - 1,
 		 io->req.offset));
 
 	switch (type) {
@@ -279,6 +292,9 @@ virtio_blk_proc(struct virtio_blk *blk, struct virtio_vq_info *vq)
 		err = ((type == VBH_OP_READ) ? blockif_read : blockif_write)
 				(blk->bc, &io->req);
 		break;
+	case VBH_OP_DISCARD:
+		err = blockif_discard(blk->bc, &io->req);
+		break;
 	case VBH_OP_FLUSH:
 	case VBH_OP_FLUSH_OUT:
 		err = blockif_flush(blk->bc, &io->req);
@@ -315,6 +331,10 @@ virtio_blk_get_caps(struct virtio_blk *blk, bool wb)
 	caps = VIRTIO_BLK_S_HOSTCAPS;
 	if (wb)
 		caps |= VIRTIO_BLK_F_WB_BITS;
+
+	if (blockif_candiscard(blk->bc))
+		caps |= VIRTIO_BLK_F_DISCARD;
+
 	return caps;
 }
 
@@ -420,6 +440,11 @@ virtio_blk_init(struct vmctx *ctx, struct pci_vdev *dev, char *opts)
 	blk->cfg.topology.opt_io_size = 0;
 	blk->cfg.writeback = blockif_get_wce(blk->bc);
 	blk->original_wce = blk->cfg.writeback; /* save for reset */
+	if (blockif_candiscard(blk->bc)) {
+		blk->cfg.max_discard_sectors = blockif_max_discard_sectors(blk->bc);
+		blk->cfg.max_discard_seg = blockif_max_discard_seg(blk->bc);
+		blk->cfg.discard_sector_alignment = blockif_discard_sector_alignment(blk->bc);
+	}
 	blk->base.device_caps =
 		virtio_blk_get_caps(blk, !!blk->cfg.writeback);
 

devicemodel/include/block_if.h

@@ -69,5 +69,8 @@ int	blockif_close(struct blockif_ctxt *bc);
 uint8_t	blockif_get_wce(struct blockif_ctxt *bc);
 void	blockif_set_wce(struct blockif_ctxt *bc, uint8_t wce);
 int	blockif_flush_all(struct blockif_ctxt *bc);
+int	blockif_max_discard_sectors(struct blockif_ctxt *bc);
+int	blockif_max_discard_seg(struct blockif_ctxt *bc);
+int	blockif_discard_sector_alignment(struct blockif_ctxt *bc);
 
 #endif /* _BLOCK_IF_H_ */