Preliminary increasing tag sequence implementation
Quoted 116 lines · Luben Tuikov
Signed-off-by Luben Tuikov <[email protected]>---
* This is a preliminary patch. Feel free to apply it for
inclusion and/or work/modify it to submit a more
atomically-correct version.
* Vendors: apply this to see an increasing sequence of
tags
on the wire and at your device, so that error recovery
doesn't
abort a task which isn't intended to be aborted especially
in
full duplex protocols when the command status and TMF can
cross on the wire. See this thread: http://marc.info/?l=linux-scsi&m=128902337522893&w=2
block/blk-tag.c
| 15 +++++++++++++--include/linux/blkdev.h | 1 + 2 files changed, 14 insertions(+), 2 deletions(-)block/blk-tag.c +1
-0@@ -130,6 +130,7 @@ init_tag_map(struct request_queue *q, struct blk_queue_tag *tags, int depth) tags->max_depth = depth; tags->tag_index = tag_index; tags->tag_map = tag_map; + tags->last_tag = 0; return 0;fail:
block/blk-tag.c +1
-1@@ -222,7 +223,7 @@ int blk_queue_resize_tags(struct request_queue *q, int new_depth) struct blk_queue_tag *bqt = q->queue_tags; struct request **tag_index; unsigned long *tag_map; - int max_depth, nr_ulongs; + int max_depth, nr_ulongs, last_tag;if (!bqt)
return-ENXIO;block/blk-tag.c +1
-0@@ -251,6 +252,7 @@ int blk_queue_resize_tags(struct request_queue *q, int new_depth) tag_index = bqt->tag_index; tag_map = bqt->tag_map; max_depth = bqt->real_max_depth; + last_tag = bqt->last_tag; if (init_tag_map(q, bqt, new_depth))return
-ENOMEM;block/blk-tag.c +0
-0@@ -258,6 +260,7 @@ int blk_queue_resize_tags(struct request_queue *q, int new_depth) memcpy(bqt->tag_index, tag_index, max_depth * sizeof(struct request *)); nr_ulongs = ALIGN(max_depth,BITS_PER_LONG ) /BITS_PER_LONG ; memcpy(bqt->tag_map, tag_map,
nr_ulongs* sizeof(unsigned long));
+bqt->last_tag=last_tag;
kfree(tag_index);
kfree(tag_map);block/blk-tag.c +6
-1@@ -364,12 +367,20 @@ int blk_queue_start_tag(struct request_queue *q, struct request *rq) return 1; } + if (bqt->last_tag == bqt->max_depth-1 ) + bqt->last_tag =0 ; + do { - tag = find_first_zero_bit(bqt->tag_map, max_depth); + tag = find_next_zero_bit(bqt->tag_map, + max_depth, +
bqt->last_tag+1);
if (tag >=max_depth)
return 1;} while (test_and_set_bit_lock(tag,
bqt->tag_map));block/blk-tag.c +3
-0+ + bqt->last_tag = tag; + /** We need lock ordering semantics given by
test_and_set_bit_lock.
* Seeblk_queue_end_tagfor
details.include/linux/blkdev.h +0
-0@@ -224,6 +224,7 @@ struct blk_queue_tag { int max_depth; /* what we will sendto device */
intreal_max_depth;
/* what the array can hold */
atomic_trefcnt;
/* map can be shared */
+ intlast_tag;
};
Quoted 3 lines
#define BLK_SCSI_MAX_CMDS (256)
Signature · 1 line
1.7.2.2.165.gbc382
A lot of bugs were discovered and are being discovered since this patch was applied. I'm tracking the Linux based ones at the moment as time permits (or at least trying to track them). Some of what I'm seeing is the same tag being issued for the subsequent command after an aborted command, and lookup failures occasionally (these depend on a lot of factors, for example how many LUs the device reports, how many streams are supported, whether the device runs in SS or HS mode, etc).
The purpose of the posting of this patch was to get people to either apply it or come up with a more atomically correct implementation, submit and apply it, so that everyone can 1) have correct sequence of tags on the wire and 2) discover more bugs (if any ;-) ).
Anyone?
Luben