block: make generic_make_request handle arbitrarily sized bios
authorKent Overstreet <kent.overstreet@gmail.com>
Fri, 24 Apr 2015 05:37:18 +0000 (22:37 -0700)
committerJens Axboe <axboe@fb.com>
Thu, 13 Aug 2015 18:31:33 +0000 (12:31 -0600)
The way the block layer is currently written, it goes to great lengths
to avoid having to split bios; upper layer code (such as bio_add_page())
checks what the underlying device can handle and tries to always create
bios that don't need to be split.

But this approach becomes unwieldy and eventually breaks down with
stacked devices and devices with dynamic limits, and it adds a lot of
complexity. If the block layer could split bios as needed, we could
eliminate a lot of complexity elsewhere - particularly in stacked
drivers. Code that creates bios can then create whatever size bios are
convenient, and more importantly stacked drivers don't have to deal with
both their own bio size limitations and the limitations of the
(potentially multiple) devices underneath them.  In the future this will
let us delete merge_bvec_fn and a bunch of other code.

We do this by adding calls to blk_queue_split() to the various
make_request functions that need it - a few can already handle arbitrary
size bios. Note that we add the call _after_ any call to
blk_queue_bounce(); this means that blk_queue_split() and
blk_recalc_rq_segments() don't need to be concerned with bouncing
affecting segment merging.

Some make_request_fn() callbacks were simple enough to audit and verify
they don't need blk_queue_split() calls. The skipped ones are:

 * nfhd_make_request (arch/m68k/emu/nfblock.c)
 * axon_ram_make_request (arch/powerpc/sysdev/axonram.c)
 * simdisk_make_request (arch/xtensa/platforms/iss/simdisk.c)
 * brd_make_request (ramdisk - drivers/block/brd.c)
 * mtip_submit_request (drivers/block/mtip32xx/mtip32xx.c)
 * loop_make_request
 * null_queue_bio
 * bcache's make_request fns

Some others are almost certainly safe to remove now, but will be left
for future patches.

Cc: Jens Axboe <axboe@kernel.dk>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Ming Lei <ming.lei@canonical.com>
Cc: Neil Brown <neilb@suse.de>
Cc: Alasdair Kergon <agk@redhat.com>
Cc: Mike Snitzer <snitzer@redhat.com>
Cc: dm-devel@redhat.com
Cc: Lars Ellenberg <drbd-dev@lists.linbit.com>
Cc: drbd-user@lists.linbit.com
Cc: Jiri Kosina <jkosina@suse.cz>
Cc: Geoff Levand <geoff@infradead.org>
Cc: Jim Paris <jim@jtan.com>
Cc: Philip Kelleher <pjk1939@linux.vnet.ibm.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Nitin Gupta <ngupta@vflare.org>
Cc: Oleg Drokin <oleg.drokin@intel.com>
Cc: Andreas Dilger <andreas.dilger@intel.com>
Acked-by: NeilBrown <neilb@suse.de> (for the 'md/md.c' bits)
Acked-by: Mike Snitzer <snitzer@redhat.com>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Kent Overstreet <kent.overstreet@gmail.com>
[dpark: skip more mq-based drivers, resolve merge conflicts, etc.]
Signed-off-by: Dongsu Park <dpark@posteo.net>
Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
16 files changed:
block/blk-core.c
block/blk-merge.c
block/blk-mq.c
block/blk-sysfs.c
drivers/block/drbd/drbd_req.c
drivers/block/pktcdvd.c
drivers/block/ps3vram.c
drivers/block/rsxx/dev.c
drivers/block/umem.c
drivers/block/zram/zram_drv.c
drivers/md/dm.c
drivers/md/md.c
drivers/s390/block/dcssblk.c
drivers/s390/block/xpram.c
drivers/staging/lustre/lustre/llite/lloop.c
include/linux/blkdev.h

index d1796b54e97a6289fc69d073a3b02abcb207a4b7..60912e983f16a071458b25baa4eccc2e21769bd5 100644 (file)
@@ -643,6 +643,10 @@ struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
        if (q->id < 0)
                goto fail_q;
 
+       q->bio_split = bioset_create(BIO_POOL_SIZE, 0);
+       if (!q->bio_split)
+               goto fail_id;
+
        q->backing_dev_info.ra_pages =
                        (VM_MAX_READAHEAD * 1024) / PAGE_CACHE_SIZE;
        q->backing_dev_info.capabilities = BDI_CAP_CGROUP_WRITEBACK;
@@ -651,7 +655,7 @@ struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
 
        err = bdi_init(&q->backing_dev_info);
        if (err)
-               goto fail_id;
+               goto fail_split;
 
        setup_timer(&q->backing_dev_info.laptop_mode_wb_timer,
                    laptop_mode_timer_fn, (unsigned long) q);
@@ -693,6 +697,8 @@ struct request_queue *blk_alloc_queue_node(gfp_t gfp_mask, int node_id)
 
 fail_bdi:
        bdi_destroy(&q->backing_dev_info);
+fail_split:
+       bioset_free(q->bio_split);
 fail_id:
        ida_simple_remove(&blk_queue_ida, q->id);
 fail_q:
@@ -1610,6 +1616,8 @@ static void blk_queue_bio(struct request_queue *q, struct bio *bio)
        struct request *req;
        unsigned int request_count = 0;
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        /*
         * low level driver can indicate that it wants pages above a
         * certain limit bounced to low memory (ie for highmem, or even
@@ -1832,15 +1840,6 @@ generic_make_request_checks(struct bio *bio)
                goto end_io;
        }
 
-       if (likely(bio_is_rw(bio) &&
-                  nr_sectors > queue_max_hw_sectors(q))) {
-               printk(KERN_ERR "bio too big device %s (%u > %u)\n",
-                      bdevname(bio->bi_bdev, b),
-                      bio_sectors(bio),
-                      queue_max_hw_sectors(q));
-               goto end_io;
-       }
-
        part = bio->bi_bdev->bd_part;
        if (should_fail_request(part, bio->bi_iter.bi_size) ||
            should_fail_request(&part_to_disk(part)->part0,
index a455b9860143c8ee8e2ceb635cea053e3735665c..d9c3a75e4a60a677f3d0aa4c0f1df19fad029ae1 100644 (file)
 
 #include "blk.h"
 
+static struct bio *blk_bio_discard_split(struct request_queue *q,
+                                        struct bio *bio,
+                                        struct bio_set *bs)
+{
+       unsigned int max_discard_sectors, granularity;
+       int alignment;
+       sector_t tmp;
+       unsigned split_sectors;
+
+       /* Zero-sector (unknown) and one-sector granularities are the same.  */
+       granularity = max(q->limits.discard_granularity >> 9, 1U);
+
+       max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
+       max_discard_sectors -= max_discard_sectors % granularity;
+
+       if (unlikely(!max_discard_sectors)) {
+               /* XXX: warn */
+               return NULL;
+       }
+
+       if (bio_sectors(bio) <= max_discard_sectors)
+               return NULL;
+
+       split_sectors = max_discard_sectors;
+
+       /*
+        * If the next starting sector would be misaligned, stop the discard at
+        * the previous aligned sector.
+        */
+       alignment = (q->limits.discard_alignment >> 9) % granularity;
+
+       tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
+       tmp = sector_div(tmp, granularity);
+
+       if (split_sectors > tmp)
+               split_sectors -= tmp;
+
+       return bio_split(bio, split_sectors, GFP_NOIO, bs);
+}
+
+static struct bio *blk_bio_write_same_split(struct request_queue *q,
+                                           struct bio *bio,
+                                           struct bio_set *bs)
+{
+       if (!q->limits.max_write_same_sectors)
+               return NULL;
+
+       if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
+               return NULL;
+
+       return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
+}
+
+static struct bio *blk_bio_segment_split(struct request_queue *q,
+                                        struct bio *bio,
+                                        struct bio_set *bs)
+{
+       struct bio *split;
+       struct bio_vec bv, bvprv;
+       struct bvec_iter iter;
+       unsigned seg_size = 0, nsegs = 0;
+       int prev = 0;
+
+       struct bvec_merge_data bvm = {
+               .bi_bdev        = bio->bi_bdev,
+               .bi_sector      = bio->bi_iter.bi_sector,
+               .bi_size        = 0,
+               .bi_rw          = bio->bi_rw,
+       };
+
+       bio_for_each_segment(bv, bio, iter) {
+               if (q->merge_bvec_fn &&
+                   q->merge_bvec_fn(q, &bvm, &bv) < (int) bv.bv_len)
+                       goto split;
+
+               bvm.bi_size += bv.bv_len;
+
+               if (bvm.bi_size >> 9 > queue_max_sectors(q))
+                       goto split;
+
+               /*
+                * If the queue doesn't support SG gaps and adding this
+                * offset would create a gap, disallow it.
+                */
+               if (q->queue_flags & (1 << QUEUE_FLAG_SG_GAPS) &&
+                   prev && bvec_gap_to_prev(&bvprv, bv.bv_offset))
+                       goto split;
+
+               if (prev && blk_queue_cluster(q)) {
+                       if (seg_size + bv.bv_len > queue_max_segment_size(q))
+                               goto new_segment;
+                       if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
+                               goto new_segment;
+                       if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
+                               goto new_segment;
+
+                       seg_size += bv.bv_len;
+                       bvprv = bv;
+                       prev = 1;
+                       continue;
+               }
+new_segment:
+               if (nsegs == queue_max_segments(q))
+                       goto split;
+
+               nsegs++;
+               bvprv = bv;
+               prev = 1;
+               seg_size = bv.bv_len;
+       }
+
+       return NULL;
+split:
+       split = bio_clone_bioset(bio, GFP_NOIO, bs);
+
+       split->bi_iter.bi_size -= iter.bi_size;
+       bio->bi_iter = iter;
+
+       if (bio_integrity(bio)) {
+               bio_integrity_advance(bio, split->bi_iter.bi_size);
+               bio_integrity_trim(split, 0, bio_sectors(split));
+       }
+
+       return split;
+}
+
+void blk_queue_split(struct request_queue *q, struct bio **bio,
+                    struct bio_set *bs)
+{
+       struct bio *split;
+
+       if ((*bio)->bi_rw & REQ_DISCARD)
+               split = blk_bio_discard_split(q, *bio, bs);
+       else if ((*bio)->bi_rw & REQ_WRITE_SAME)
+               split = blk_bio_write_same_split(q, *bio, bs);
+       else
+               split = blk_bio_segment_split(q, *bio, q->bio_split);
+
+       if (split) {
+               bio_chain(split, *bio);
+               generic_make_request(*bio);
+               *bio = split;
+       }
+}
+EXPORT_SYMBOL(blk_queue_split);
+
 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
                                             struct bio *bio,
                                             bool no_sg_merge)
 {
        struct bio_vec bv, bvprv = { NULL };
-       int cluster, high, highprv = 1;
+       int cluster, prev = 0;
        unsigned int seg_size, nr_phys_segs;
        struct bio *fbio, *bbio;
        struct bvec_iter iter;
@@ -36,7 +182,6 @@ static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
        cluster = blk_queue_cluster(q);
        seg_size = 0;
        nr_phys_segs = 0;
-       high = 0;
        for_each_bio(bio) {
                bio_for_each_segment(bv, bio, iter) {
                        /*
@@ -46,13 +191,7 @@ static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
                        if (no_sg_merge)
                                goto new_segment;
 
-                       /*
-                        * the trick here is making sure that a high page is
-                        * never considered part of another segment, since
-                        * that might change with the bounce page.
-                        */
-                       high = page_to_pfn(bv.bv_page) > queue_bounce_pfn(q);
-                       if (!high && !highprv && cluster) {
+                       if (prev && cluster) {
                                if (seg_size + bv.bv_len
                                    > queue_max_segment_size(q))
                                        goto new_segment;
@@ -72,8 +211,8 @@ new_segment:
 
                        nr_phys_segs++;
                        bvprv = bv;
+                       prev = 1;
                        seg_size = bv.bv_len;
-                       highprv = high;
                }
                bbio = bio;
        }
index 94559025c5e6530d1a7034dc4e08d4c983e83560..81edbd95bda87eb5da6e21bfa84253daf9ca0747 100644 (file)
@@ -1287,6 +1287,8 @@ static void blk_mq_make_request(struct request_queue *q, struct bio *bio)
                return;
        }
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        if (!is_flush_fua && !blk_queue_nomerges(q) &&
            blk_attempt_plug_merge(q, bio, &request_count, &same_queue_rq))
                return;
@@ -1372,6 +1374,8 @@ static void blk_sq_make_request(struct request_queue *q, struct bio *bio)
                return;
        }
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        if (!is_flush_fua && !blk_queue_nomerges(q) &&
            blk_attempt_plug_merge(q, bio, &request_count, NULL))
                return;
index b1f34e463c0f0fd86c816bed9b7ee421d43b6b58..3e44a9da2a13579cacaee3d5e03bb3868f34d02a 100644 (file)
@@ -561,6 +561,9 @@ static void blk_release_queue(struct kobject *kobj)
 
        blk_trace_shutdown(q);
 
+       if (q->bio_split)
+               bioset_free(q->bio_split);
+
        ida_simple_remove(&blk_queue_ida, q->id);
        call_rcu(&q->rcu_head, blk_free_queue_rcu);
 }
index 9cb41166366ef93e97f37fcf4f6754f54ae88206..923c857b395b35ee4e5de2563a1f7e44781d091a 100644 (file)
@@ -1499,6 +1499,8 @@ void drbd_make_request(struct request_queue *q, struct bio *bio)
        struct drbd_device *device = (struct drbd_device *) q->queuedata;
        unsigned long start_jif;
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        start_jif = jiffies;
 
        /*
index a7a259e031da96f455eb280dd1bdfbd945697003..ee7ad5e44632ebf8a703142ab53548cb93ae76a3 100644 (file)
@@ -2447,6 +2447,10 @@ static void pkt_make_request(struct request_queue *q, struct bio *bio)
        char b[BDEVNAME_SIZE];
        struct bio *split;
 
+       blk_queue_bounce(q, &bio);
+
+       blk_queue_split(q, &bio, q->bio_split);
+
        pd = q->queuedata;
        if (!pd) {
                pr_err("%s incorrect request queue\n",
@@ -2477,8 +2481,6 @@ static void pkt_make_request(struct request_queue *q, struct bio *bio)
                goto end_io;
        }
 
-       blk_queue_bounce(q, &bio);
-
        do {
                sector_t zone = get_zone(bio->bi_iter.bi_sector, pd);
                sector_t last_zone = get_zone(bio_end_sector(bio) - 1, pd);
index 49b4706b162c851a3f6214c09e3c7a8996eda1af..d89fcac59515193c0507ab10fde01c47f3640b84 100644 (file)
@@ -606,6 +606,8 @@ static void ps3vram_make_request(struct request_queue *q, struct bio *bio)
 
        dev_dbg(&dev->core, "%s\n", __func__);
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        spin_lock_irq(&priv->lock);
        busy = !bio_list_empty(&priv->list);
        bio_list_add(&priv->list, bio);
index 63b9d2ffa8ee406628da48d60c841b6a8a65bb98..3163e4cdc2cc54a18af9ac8845311dce0440995c 100644 (file)
@@ -151,6 +151,8 @@ static void rsxx_make_request(struct request_queue *q, struct bio *bio)
        struct rsxx_bio_meta *bio_meta;
        int st = -EINVAL;
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        might_sleep();
 
        if (!card)
index 3b3afd2ec5d621c610c40b18996345730bd49405..04d65790a8862a91be1d1b40f8baef7f850824d6 100644 (file)
@@ -531,6 +531,8 @@ static void mm_make_request(struct request_queue *q, struct bio *bio)
                 (unsigned long long)bio->bi_iter.bi_sector,
                 bio->bi_iter.bi_size);
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        spin_lock_irq(&card->lock);
        *card->biotail = bio;
        bio->bi_next = NULL;
index 68c3d4800464675a99ed6dec92787a64c82630e2..aec781acee9d597f41c78cb9903adf7a609eb522 100644 (file)
@@ -900,6 +900,8 @@ static void zram_make_request(struct request_queue *queue, struct bio *bio)
        if (unlikely(!zram_meta_get(zram)))
                goto error;
 
+       blk_queue_split(queue, &bio, queue->bio_split);
+
        if (!valid_io_request(zram, bio->bi_iter.bi_sector,
                                        bio->bi_iter.bi_size)) {
                atomic64_inc(&zram->stats.invalid_io);
index 7f367fcace035a23e5da8c2ccef6b43384445221..069f8d7e890e6840c599cfafd0e246a7d057a8e1 100644 (file)
@@ -1799,6 +1799,8 @@ static void dm_make_request(struct request_queue *q, struct bio *bio)
 
        map = dm_get_live_table(md, &srcu_idx);
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        generic_start_io_acct(rw, bio_sectors(bio), &dm_disk(md)->part0);
 
        /* if we're suspended, we have to queue this io for later */
index ac4381a6625cef695857315f014b926d27d76aea..e1d8723720ccbcd72a51239927e85342d29ab4e5 100644 (file)
@@ -257,6 +257,8 @@ static void md_make_request(struct request_queue *q, struct bio *bio)
        unsigned int sectors;
        int cpu;
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        if (mddev == NULL || mddev->pers == NULL
            || !mddev->ready) {
                bio_io_error(bio);
index 8bcb822b0bac50de53873243d6a7c435314e280e..29ea2394c8966ff55380a8c5fd9414f4d32dc45e 100644 (file)
@@ -826,6 +826,8 @@ dcssblk_make_request(struct request_queue *q, struct bio *bio)
        unsigned long source_addr;
        unsigned long bytes_done;
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        bytes_done = 0;
        dev_info = bio->bi_bdev->bd_disk->private_data;
        if (dev_info == NULL)
index 93856b9b621449528cef23fe6f8efc799975c82d..02871f1db562ef42a4f6a60b50dc943a550ba251 100644 (file)
@@ -190,6 +190,8 @@ static void xpram_make_request(struct request_queue *q, struct bio *bio)
        unsigned long page_addr;
        unsigned long bytes;
 
+       blk_queue_split(q, &bio, q->bio_split);
+
        if ((bio->bi_iter.bi_sector & 7) != 0 ||
            (bio->bi_iter.bi_size & 4095) != 0)
                /* Request is not page-aligned. */
index cc00fd10fbcf77654479b2aadf6fda7c75cff91a..1e33d540b2234a6ac7953feff2150a90f963a7b0 100644 (file)
@@ -340,6 +340,8 @@ static void loop_make_request(struct request_queue *q, struct bio *old_bio)
        int rw = bio_rw(old_bio);
        int inactive;
 
+       blk_queue_split(q, &old_bio, q->bio_split);
+
        if (!lo)
                goto err;
 
index 243f29e779ec5890d40ccdad1752444ca620a098..ca778d9c7d81222f75024c93a8452d10efc49285 100644 (file)
@@ -463,6 +463,7 @@ struct request_queue {
 
        struct blk_mq_tag_set   *tag_set;
        struct list_head        tag_set_list;
+       struct bio_set          *bio_split;
 };
 
 #define QUEUE_FLAG_QUEUED      1       /* uses generic tag queueing */
@@ -783,6 +784,8 @@ extern void blk_rq_unprep_clone(struct request *rq);
 extern int blk_insert_cloned_request(struct request_queue *q,
                                     struct request *rq);
 extern void blk_delay_queue(struct request_queue *, unsigned long);
+extern void blk_queue_split(struct request_queue *, struct bio **,
+                           struct bio_set *);
 extern void blk_recount_segments(struct request_queue *, struct bio *);
 extern int scsi_verify_blk_ioctl(struct block_device *, unsigned int);
 extern int scsi_cmd_blk_ioctl(struct block_device *, fmode_t,