NVMe: Split non-mergeable bio requests
authorKeith Busch <keith.busch@intel.com>
Tue, 9 Apr 2013 17:59:32 +0000 (11:59 -0600)
committerMatthew Wilcox <matthew.r.wilcox@intel.com>
Thu, 2 May 2013 18:38:59 +0000 (14:38 -0400)
It is possible a bio request can not be submitted as a single NVMe IO
command if the bio_vec is not mergeable with the NVMe PRP alignement
constraints. This condition was handled by submitting an IO for the
mergeable portion then submitting a follow on IO for the remaining data
after the previous IO completes. The remainder to be sent was tracked
by manipulating the bio->bi_idx and bio->bi_sector. This patch splits
the request as many times as necessary and submits the bios together.

Since submitting the bio may cause it to be requeued on split,
nvme_resubmit_bios had to be modified to remove the wait queue when
the bio list is empty prior to submitting the bio since a split would
have added the wait queue a second time, corrupting the wait queue head
task list.

There are a few other benefits from doing this: it fixes a potential
issue with the previous handling of a non-mergeable bio as the requeuing
method could would use an unlocked nvme_queue if the callback isn't
invoked on the queue's associated cpu; it will be possible to retry a
failed bio if desired at some later time since it does not manipulate
the original bio; the bio integrity extensions require the bio to be in
its original condition for the checks to work correctly if we implement
the end-to-end data protection in the future.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Matthew Wilcox <matthew.r.wilcox@intel.com>
drivers/block/nvme-core.c

index a126c7b..f812806 100644 (file)
@@ -308,16 +308,6 @@ void nvme_free_iod(struct nvme_dev *dev, struct nvme_iod *iod)
        kfree(iod);
 }
 
-static void requeue_bio(struct nvme_dev *dev, struct bio *bio)
-{
-       struct nvme_queue *nvmeq = get_nvmeq(dev);
-       if (bio_list_empty(&nvmeq->sq_cong))
-               add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
-       bio_list_add(&nvmeq->sq_cong, bio);
-       put_nvmeq(nvmeq);
-       wake_up_process(nvme_thread);
-}
-
 static void bio_completion(struct nvme_dev *dev, void *ctx,
                                                struct nvme_completion *cqe)
 {
@@ -329,13 +319,10 @@ static void bio_completion(struct nvme_dev *dev, void *ctx,
                dma_unmap_sg(&dev->pci_dev->dev, iod->sg, iod->nents,
                        bio_data_dir(bio) ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
        nvme_free_iod(dev, iod);
-       if (status) {
+       if (status)
                bio_endio(bio, -EIO);
-       } else if (bio->bi_vcnt > bio->bi_idx) {
-               requeue_bio(dev, bio);
-       } else {
+       else
                bio_endio(bio, 0);
-       }
 }
 
 /* length is in bytes.  gfp flags indicates whether we may sleep. */
@@ -419,25 +406,130 @@ int nvme_setup_prps(struct nvme_dev *dev, struct nvme_common_command *cmd,
        return total_len;
 }
 
+struct nvme_bio_pair {
+       struct bio b1, b2, *parent;
+       struct bio_vec *bv1, *bv2;
+       int err;
+       atomic_t cnt;
+};
+
+static void nvme_bio_pair_endio(struct bio *bio, int err)
+{
+       struct nvme_bio_pair *bp = bio->bi_private;
+
+       if (err)
+               bp->err = err;
+
+       if (atomic_dec_and_test(&bp->cnt)) {
+               bio_endio(bp->parent, bp->err);
+               if (bp->bv1)
+                       kfree(bp->bv1);
+               if (bp->bv2)
+                       kfree(bp->bv2);
+               kfree(bp);
+       }
+}
+
+static struct nvme_bio_pair *nvme_bio_split(struct bio *bio, int idx,
+                                                       int len, int offset)
+{
+       struct nvme_bio_pair *bp;
+
+       BUG_ON(len > bio->bi_size);
+       BUG_ON(idx > bio->bi_vcnt);
+
+       bp = kmalloc(sizeof(*bp), GFP_ATOMIC);
+       if (!bp)
+               return NULL;
+       bp->err = 0;
+
+       bp->b1 = *bio;
+       bp->b2 = *bio;
+
+       bp->b1.bi_size = len;
+       bp->b2.bi_size -= len;
+       bp->b1.bi_vcnt = idx;
+       bp->b2.bi_idx = idx;
+       bp->b2.bi_sector += len >> 9;
+
+       if (offset) {
+               bp->bv1 = kmalloc(bio->bi_max_vecs * sizeof(struct bio_vec),
+                                                               GFP_ATOMIC);
+               if (!bp->bv1)
+                       goto split_fail_1;
+
+               bp->bv2 = kmalloc(bio->bi_max_vecs * sizeof(struct bio_vec),
+                                                               GFP_ATOMIC);
+               if (!bp->bv2)
+                       goto split_fail_2;
+
+               memcpy(bp->bv1, bio->bi_io_vec,
+                       bio->bi_max_vecs * sizeof(struct bio_vec));
+               memcpy(bp->bv2, bio->bi_io_vec,
+                       bio->bi_max_vecs * sizeof(struct bio_vec));
+
+               bp->b1.bi_io_vec = bp->bv1;
+               bp->b2.bi_io_vec = bp->bv2;
+               bp->b2.bi_io_vec[idx].bv_offset += offset;
+               bp->b2.bi_io_vec[idx].bv_len -= offset;
+               bp->b1.bi_io_vec[idx].bv_len = offset;
+               bp->b1.bi_vcnt++;
+       } else
+               bp->bv1 = bp->bv2 = NULL;
+
+       bp->b1.bi_private = bp;
+       bp->b2.bi_private = bp;
+
+       bp->b1.bi_end_io = nvme_bio_pair_endio;
+       bp->b2.bi_end_io = nvme_bio_pair_endio;
+
+       bp->parent = bio;
+       atomic_set(&bp->cnt, 2);
+
+       return bp;
+
+ split_fail_2:
+       kfree(bp->bv1);
+ split_fail_1:
+       kfree(bp);
+       return NULL;
+}
+
+static int nvme_split_and_submit(struct bio *bio, struct nvme_queue *nvmeq,
+                                               int idx, int len, int offset)
+{
+       struct nvme_bio_pair *bp = nvme_bio_split(bio, idx, len, offset);
+       if (!bp)
+               return -ENOMEM;
+
+       if (bio_list_empty(&nvmeq->sq_cong))
+               add_wait_queue(&nvmeq->sq_full, &nvmeq->sq_cong_wait);
+       bio_list_add(&nvmeq->sq_cong, &bp->b1);
+       bio_list_add(&nvmeq->sq_cong, &bp->b2);
+
+       return 0;
+}
+
 /* NVMe scatterlists require no holes in the virtual address */
 #define BIOVEC_NOT_VIRT_MERGEABLE(vec1, vec2)  ((vec2)->bv_offset || \
                        (((vec1)->bv_offset + (vec1)->bv_len) % PAGE_SIZE))
 
-static int nvme_map_bio(struct device *dev, struct nvme_iod *iod,
+static int nvme_map_bio(struct nvme_queue *nvmeq, struct nvme_iod *iod,
                struct bio *bio, enum dma_data_direction dma_dir, int psegs)
 {
        struct bio_vec *bvec, *bvprv = NULL;
        struct scatterlist *sg = NULL;
-       int i, old_idx, length = 0, nsegs = 0;
+       int i, length = 0, nsegs = 0;
 
        sg_init_table(iod->sg, psegs);
-       old_idx = bio->bi_idx;
        bio_for_each_segment(bvec, bio, i) {
                if (bvprv && BIOVEC_PHYS_MERGEABLE(bvprv, bvec)) {
                        sg->length += bvec->bv_len;
                } else {
                        if (bvprv && BIOVEC_NOT_VIRT_MERGEABLE(bvprv, bvec))
-                               break;
+                               return nvme_split_and_submit(bio, nvmeq, i,
+                                                               length, 0);
+
                        sg = sg ? sg + 1 : iod->sg;
                        sg_set_page(sg, bvec->bv_page, bvec->bv_len,
                                                        bvec->bv_offset);
@@ -446,13 +538,11 @@ static int nvme_map_bio(struct device *dev, struct nvme_iod *iod,
                length += bvec->bv_len;
                bvprv = bvec;
        }
-       bio->bi_idx = i;
        iod->nents = nsegs;
        sg_mark_end(sg);
-       if (dma_map_sg(dev, iod->sg, iod->nents, dma_dir) == 0) {
-               bio->bi_idx = old_idx;
+       if (dma_map_sg(nvmeq->q_dmadev, iod->sg, iod->nents, dma_dir) == 0)
                return -ENOMEM;
-       }
+
        return length;
 }
 
@@ -581,8 +671,8 @@ static int nvme_submit_bio_queue(struct nvme_queue *nvmeq, struct nvme_ns *ns,
                dma_dir = DMA_FROM_DEVICE;
        }
 
-       result = nvme_map_bio(nvmeq->q_dmadev, iod, bio, dma_dir, psegs);
-       if (result < 0)
+       result = nvme_map_bio(nvmeq, iod, bio, dma_dir, psegs);
+       if (result <= 0)
                goto free_cmdid;
        length = result;
 
@@ -595,8 +685,6 @@ static int nvme_submit_bio_queue(struct nvme_queue *nvmeq, struct nvme_ns *ns,
        cmnd->rw.control = cpu_to_le16(control);
        cmnd->rw.dsmgmt = cpu_to_le32(dsmgmt);
 
-       bio->bi_sector += length >> 9;
-
        if (++nvmeq->sq_tail == nvmeq->q_depth)
                nvmeq->sq_tail = 0;
        writel(nvmeq->sq_tail, nvmeq->q_db);
@@ -1281,13 +1369,17 @@ static void nvme_resubmit_bios(struct nvme_queue *nvmeq)
        while (bio_list_peek(&nvmeq->sq_cong)) {
                struct bio *bio = bio_list_pop(&nvmeq->sq_cong);
                struct nvme_ns *ns = bio->bi_bdev->bd_disk->private_data;
+
+               if (bio_list_empty(&nvmeq->sq_cong))
+                       remove_wait_queue(&nvmeq->sq_full,
+                                                       &nvmeq->sq_cong_wait);
                if (nvme_submit_bio_queue(nvmeq, ns, bio)) {
+                       if (bio_list_empty(&nvmeq->sq_cong))
+                               add_wait_queue(&nvmeq->sq_full,
+                                                       &nvmeq->sq_cong_wait);
                        bio_list_add_head(&nvmeq->sq_cong, bio);
                        break;
                }
-               if (bio_list_empty(&nvmeq->sq_cong))
-                       remove_wait_queue(&nvmeq->sq_full,
-                                                       &nvmeq->sq_cong_wait);
        }
 }