[PATCH] Add scatter-gather support for the block layer SG_IO

Signed-off-by: Jens Axboe <axboe@suse.de>

drivers/block/ll_rw_blk.c

@@ -2148,6 +2148,50 @@ int blk_rq_map_user(request_queue_t *q, struct request *rq, void __user *ubuf,
 
 EXPORT_SYMBOL(blk_rq_map_user);
 
+/**
+ * blk_rq_map_user_iov - map user data to a request, for REQ_BLOCK_PC usage
+ * @q:		request queue where request should be inserted
+ * @rq:		request to map data to
+ * @iov:	pointer to the iovec
+ * @iov_count:	number of elements in the iovec
+ *
+ * Description:
+ *    Data will be mapped directly for zero copy io, if possible. Otherwise
+ *    a kernel bounce buffer is used.
+ *
+ *    A matching blk_rq_unmap_user() must be issued at the end of io, while
+ *    still in process context.
+ *
+ *    Note: The mapped bio may need to be bounced through blk_queue_bounce()
+ *    before being submitted to the device, as pages mapped may be out of
+ *    reach. It's the callers responsibility to make sure this happens. The
+ *    original bio must be passed back in to blk_rq_unmap_user() for proper
+ *    unmapping.
+ */
+int blk_rq_map_user_iov(request_queue_t *q, struct request *rq,
+			struct sg_iovec *iov, int iov_count)
+{
+	struct bio *bio;
+
+	if (!iov || iov_count <= 0)
+		return -EINVAL;
+
+	/* we don't allow misaligned data like bio_map_user() does.  If the
+	 * user is using sg, they're expected to know the alignment constraints
+	 * and respect them accordingly */
+	bio = bio_map_user_iov(q, NULL, iov, iov_count, rq_data_dir(rq)== READ);
+	if (IS_ERR(bio))
+		return PTR_ERR(bio);
+
+	rq->bio = rq->biotail = bio;
+	blk_rq_bio_prep(q, rq, bio);
+	rq->buffer = rq->data = NULL;
+	rq->data_len = bio->bi_size;
+	return 0;
+}
+
+EXPORT_SYMBOL(blk_rq_map_user_iov);
+
 /**
  * blk_rq_unmap_user - unmap a request with user data
  * @rq:		request to be unmapped
@@ -2207,6 +2251,19 @@ int blk_rq_map_kern(request_queue_t *q, struct request *rq, void *kbuf,
 
 EXPORT_SYMBOL(blk_rq_map_kern);
 
+void blk_execute_rq_nowait(request_queue_t *q, struct gendisk *bd_disk,
+			   struct request *rq, int at_head,
+			   void (*done)(struct request *))
+{
+	int where = at_head ? ELEVATOR_INSERT_FRONT : ELEVATOR_INSERT_BACK;
+
+	rq->rq_disk = bd_disk;
+	rq->flags |= REQ_NOMERGE;
+	rq->end_io = done;
+	elv_add_request(q, rq, where, 1);
+	generic_unplug_device(q);
+}
+
 /**
  * blk_execute_rq - insert a request into queue for execution
  * @q:		queue to insert the request in
@@ -2224,8 +2281,6 @@ int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
 	char sense[SCSI_SENSE_BUFFERSIZE];
 	int err = 0;
 
-	rq->rq_disk = bd_disk;
-
 	/*
 	 * we need an extra reference to the request, so we can look at
 	 * it after io completion
@@ -2238,11 +2293,8 @@ int blk_execute_rq(request_queue_t *q, struct gendisk *bd_disk,
 		rq->sense_len = 0;
 	}
 
-	rq->flags |= REQ_NOMERGE;
 	rq->waiting = &wait;
-	rq->end_io = blk_end_sync_rq;
+	blk_execute_rq_nowait(q, bd_disk, rq, 0, blk_end_sync_rq);
-	elv_add_request(q, rq, ELEVATOR_INSERT_BACK, 1);
-	generic_unplug_device(q);
 	wait_for_completion(&wait);
 	rq->waiting = NULL;
 

drivers/block/scsi_ioctl.c

@@ -231,17 +231,11 @@ static int sg_io(struct file *file, request_queue_t *q,
 	if (verify_command(file, cmd))
 		return -EPERM;
 
-	/*
-	 * we'll do that later
-	 */
-	if (hdr->iovec_count)
-		return -EOPNOTSUPP;
-
 	if (hdr->dxfer_len > (q->max_sectors << 9))
 		return -EIO;
 
 	reading = writing = 0;
-	if (hdr->dxfer_len) {
+	if (hdr->dxfer_len)
 		switch (hdr->dxfer_direction) {
 		default:
 			return -EINVAL;
@@ -261,11 +255,29 @@ static int sg_io(struct file *file, request_queue_t *q,
 	if (!rq)
 		return -ENOMEM;
 
-	if (reading || writing) {
+	if (hdr->iovec_count) {
-		ret = blk_rq_map_user(q, rq, hdr->dxferp, hdr->dxfer_len);
+		const int size = sizeof(struct sg_iovec) * hdr->iovec_count;
-		if (ret)
+		struct sg_iovec *iov;
+
+		iov = kmalloc(size, GFP_KERNEL);
+		if (!iov) {
+			ret = -ENOMEM;
 			goto out;
-	}
+		}
+
+		if (copy_from_user(iov, hdr->dxferp, size)) {
+			kfree(iov);
+			ret = -EFAULT;
+			goto out;
+		}
+
+		ret = blk_rq_map_user_iov(q, rq, iov, hdr->iovec_count);
+		kfree(iov);
+	} else if (hdr->dxfer_len)
+		ret = blk_rq_map_user(q, rq, hdr->dxferp, hdr->dxfer_len);
+
+	if (ret)
+		goto out;
 
 	/*
 	 * fill in request structure

fs/bio.c

@@ -25,6 +25,7 @@
 #include <linux/module.h>
 #include <linux/mempool.h>
 #include <linux/workqueue.h>
+#include <scsi/sg.h>		/* for struct sg_iovec */
 
 #define BIO_POOL_SIZE 256
 
@@ -549,22 +550,34 @@ out_bmd:
 	return ERR_PTR(ret);
 }
 
-static struct bio *__bio_map_user(request_queue_t *q, struct block_device *bdev,
+static struct bio *__bio_map_user_iov(request_queue_t *q,
-				  unsigned long uaddr, unsigned int len,
+				      struct block_device *bdev,
-				  int write_to_vm)
+				      struct sg_iovec *iov, int iov_count,
+				      int write_to_vm)
 {
-	unsigned long end = (uaddr + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
+	int i, j;
-	unsigned long start = uaddr >> PAGE_SHIFT;
+	int nr_pages = 0;
-	const int nr_pages = end - start;
-	int ret, offset, i;
 	struct page **pages;
 	struct bio *bio;
+	int cur_page = 0;
+	int ret, offset;
 
-	/*
+	for (i = 0; i < iov_count; i++) {
-	 * transfer and buffer must be aligned to at least hardsector
+		unsigned long uaddr = (unsigned long)iov[i].iov_base;
-	 * size for now, in the future we can relax this restriction
+		unsigned long len = iov[i].iov_len;
-	 */
+		unsigned long end = (uaddr + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
-	if ((uaddr & queue_dma_alignment(q)) || (len & queue_dma_alignment(q)))
+		unsigned long start = uaddr >> PAGE_SHIFT;
+
+		nr_pages += end - start;
+		/*
+		 * transfer and buffer must be aligned to at least hardsector
+		 * size for now, in the future we can relax this restriction
+		 */
+		if ((uaddr & queue_dma_alignment(q)) || (len & queue_dma_alignment(q)))
+			return ERR_PTR(-EINVAL);
+	}
+
+	if (!nr_pages)
 		return ERR_PTR(-EINVAL);
 
 	bio = bio_alloc(GFP_KERNEL, nr_pages);
@@ -576,42 +589,54 @@ static struct bio *__bio_map_user(request_queue_t *q, struct block_device *bdev,
 	if (!pages)
 		goto out;
 
-	down_read(&current->mm->mmap_sem);
+	memset(pages, 0, nr_pages * sizeof(struct page *));
-	ret = get_user_pages(current, current->mm, uaddr, nr_pages,
-						write_to_vm, 0, pages, NULL);
-	up_read(&current->mm->mmap_sem);
 
-	if (ret < nr_pages)
+	for (i = 0; i < iov_count; i++) {
-		goto out;
+		unsigned long uaddr = (unsigned long)iov[i].iov_base;
+		unsigned long len = iov[i].iov_len;
+		unsigned long end = (uaddr + len + PAGE_SIZE - 1) >> PAGE_SHIFT;
+		unsigned long start = uaddr >> PAGE_SHIFT;
+		const int local_nr_pages = end - start;
+		const int page_limit = cur_page + local_nr_pages;
+		
+		down_read(&current->mm->mmap_sem);
+		ret = get_user_pages(current, current->mm, uaddr,
+				     local_nr_pages,
+				     write_to_vm, 0, &pages[cur_page], NULL);
+		up_read(&current->mm->mmap_sem);
 
-	bio->bi_bdev = bdev;
+		if (ret < local_nr_pages)
+			goto out_unmap;
 
-	offset = uaddr & ~PAGE_MASK;
-	for (i = 0; i < nr_pages; i++) {
-		unsigned int bytes = PAGE_SIZE - offset;
 
-		if (len <= 0)
+		offset = uaddr & ~PAGE_MASK;
-			break;
+		for (j = cur_page; j < page_limit; j++) {
+			unsigned int bytes = PAGE_SIZE - offset;
 
-		if (bytes > len)
+			if (len <= 0)
-			bytes = len;
+				break;
+			
+			if (bytes > len)
+				bytes = len;
 
+			/*
+			 * sorry...
+			 */
+			if (__bio_add_page(q, bio, pages[j], bytes, offset) < bytes)
+				break;
+
+			len -= bytes;
+			offset = 0;
+		}
+
+		cur_page = j;
 		/*
-		 * sorry...
+		 * release the pages we didn't map into the bio, if any
 		 */
-		if (__bio_add_page(q, bio, pages[i], bytes, offset) < bytes)
+		while (j < page_limit)
-			break;
+			page_cache_release(pages[j++]);
-
-		len -= bytes;
-		offset = 0;
 	}
 
-	/*
-	 * release the pages we didn't map into the bio, if any
-	 */
-	while (i < nr_pages)
-		page_cache_release(pages[i++]);
-
 	kfree(pages);
 
 	/*
@@ -620,9 +645,17 @@ static struct bio *__bio_map_user(request_queue_t *q, struct block_device *bdev,
 	if (!write_to_vm)
 		bio->bi_rw |= (1 << BIO_RW);
 
+	bio->bi_bdev = bdev;
 	bio->bi_flags |= (1 << BIO_USER_MAPPED);
 	return bio;
-out:
+
+ out_unmap:
+	for (i = 0; i < nr_pages; i++) {
+		if(!pages[i])
+			break;
+		page_cache_release(pages[i]);
+	}
+ out:
 	kfree(pages);
 	bio_put(bio);
 	return ERR_PTR(ret);
@@ -642,9 +675,33 @@ out:
 struct bio *bio_map_user(request_queue_t *q, struct block_device *bdev,
 			 unsigned long uaddr, unsigned int len, int write_to_vm)
 {
-	struct bio *bio;
+	struct sg_iovec iov;
 
-	bio = __bio_map_user(q, bdev, uaddr, len, write_to_vm);
+	iov.iov_base = (__user void *)uaddr;
+	iov.iov_len = len;
+
+	return bio_map_user_iov(q, bdev, &iov, 1, write_to_vm);
+}
+
+/**
+ *	bio_map_user_iov - map user sg_iovec table into bio
+ *	@q: the request_queue_t for the bio
+ *	@bdev: destination block device
+ *	@iov:	the iovec.
+ *	@iov_count: number of elements in the iovec
+ *	@write_to_vm: bool indicating writing to pages or not
+ *
+ *	Map the user space address into a bio suitable for io to a block
+ *	device. Returns an error pointer in case of error.
+ */
+struct bio *bio_map_user_iov(request_queue_t *q, struct block_device *bdev,
+			     struct sg_iovec *iov, int iov_count,
+			     int write_to_vm)
+{
+	struct bio *bio;
+	int len = 0, i;
+
+	bio = __bio_map_user_iov(q, bdev, iov, iov_count, write_to_vm);
 
 	if (IS_ERR(bio))
 		return bio;
@@ -657,6 +714,9 @@ struct bio *bio_map_user(request_queue_t *q, struct block_device *bdev,
 	 */
 	bio_get(bio);
 
+	for (i = 0; i < iov_count; i++)
+		len += iov[i].iov_len;
+
 	if (bio->bi_size == len)
 		return bio;
 

include/linux/bio.h

@@ -281,6 +281,10 @@ extern int bio_add_page(struct bio *, struct page *, unsigned int,unsigned int);
 extern int bio_get_nr_vecs(struct block_device *);
 extern struct bio *bio_map_user(struct request_queue *, struct block_device *,
 				unsigned long, unsigned int, int);
+struct sg_iovec;
+extern struct bio *bio_map_user_iov(struct request_queue *,
+				    struct block_device *,
+				    struct sg_iovec *, int, int);
 extern void bio_unmap_user(struct bio *);
 extern struct bio *bio_map_kern(struct request_queue *, void *, unsigned int,
 				unsigned int);

include/linux/blkdev.h

@@ -561,6 +561,7 @@ extern void blk_queue_activity_fn(request_queue_t *, activity_fn *, void *);
 extern int blk_rq_map_user(request_queue_t *, struct request *, void __user *, unsigned int);
 extern int blk_rq_unmap_user(struct bio *, unsigned int);
 extern int blk_rq_map_kern(request_queue_t *, struct request *, void *, unsigned int, unsigned int);
+extern int blk_rq_map_user_iov(request_queue_t *, struct request *, struct sg_iovec *, int);
 extern int blk_execute_rq(request_queue_t *, struct gendisk *, struct request *);
 
 static inline request_queue_t *bdev_get_queue(struct block_device *bdev)