md: add new interface 'mk_rq' in target_type
PATCH 1/2 md: add new interface 'mk_rq' in target_type
From: Robin Dong <[email protected]>
We are now trying to modify flashcache(https://github.com/facebook/flashcache) to make it request based so that
we can let cfq io-controller control the bandwidth between different io cgroups.
A search in the dm directory tells me that only multipath is a request based dm target and its functionality
is very simple and map_rq() is used to map the request to different underlying devices. We can't work in this way because:
1. the request which processed by map_rq() need to be issued to
different lower devices (disk device and cache device, in flashcache), therefore the request
can't be totally remapped by simply changing its queue and returning DM_MAPIO_REMAPPED in map_rq() like multipath_map()
2. to submit bios drectly in map_rq() (by return DM_MAPIO_SUBMITTED) will cause BUG_ON(!irqs_disabled())
in dm_request_fn() because the submit_bio()->generic_make_request()->blk_queue_bio() will definitly call spin_unlock_irq to enable the irqs
As above,the interface map_rq() provided by devcie-mapper framework is not enough for an autonomous target, like flashcache.
We propose to add a newmk_rq interface so that we can make the requests
by ourselves.
--- drivers/md/dm-io.c | 58 ++++++++++++++++++++++++++++-------------------- drivers/md/dm-log.c | 1 + include/linux/dm-io.h | 3 ++ 3 files changed, 38 insertions(+), 24 deletions(-)
drivers/md/dm-io.c +34 -24
@@ -287,8 +287,8 @@ static void km_dp_init(struct dpages *dp, void *data)/*----------------------------------------------------------------- * IO routines that accept a list of pages. *---------------------------------------------------------------*/ -static void do_region (int rw,unsigned region,struct dm_io_region *where, -struct dpages *dp,struct io *io) +static void do_region (struct dm_io_request *io_req,unsigned region, +struct dm_io_region *where,struct dpages *dp,struct io *io) {struct bio *bio;struct page *page;
@@ -298,6 +298,7 @@ static void do_region(int rw, unsigned region, struct dm_io_region *where,sector_t remaining = where->count;struct request_queue *q =bdev_get_queue (where->bdev);sector_t discard_sectors; +int rw = io_req->bi_rw;/* * where->count may be zero if rw holds a flush and we need to
@@ -339,15 +340,26 @@ static void do_region(int rw, unsigned region, struct dm_io_region *where, }atomic_inc (&io->count); -submit_bio (rw, bio); +if (!io_req->only_create_bio) +submit_bio (rw, bio); +else { + bio->bi_rw |= rw; +if (io_req->start) { + io_req->end->bi_next = bio; + io_req->end = bio; + }else + io_req->start = io_req->end = bio; + bio->bi_next =NULL ; + } }while (remaining); } -static void dispatch_io (int rw,unsigned int num_regions, +static void dispatch_io (struct dm_io_request *io_req,unsigned int num_regions,struct dm_io_region *where,struct dpages *dp,struct io *io,int sync) {int i; +int rw = io_req->bi_rw;struct dpages old_pages = *dp;BUG_ON (num_regions >DM_IO_MAX_REGIONS );
@@ -362,7 +374,7 @@ static void dispatch_io(int rw, unsigned int num_regions,for (i =0 ; i < num_regions; i++) { *dp = old_pages;if (where[i].count || (rw &REQ_FLUSH )) -do_region (rw, i, where + i, dp, io); +do_region (io_req, i, where + i, dp, io); }/*
@@ -372,8 +384,8 @@ static void dispatch_io(int rw, unsigned int num_regions,dec_count (io,0 ,0 ); } -static int sync_io (struct dm_io_client *client,unsigned int num_regions, -struct dm_io_region *where,int rw,struct dpages *dp, +static int sync_io (struct dm_io_request *io_req,unsigned int num_regions, +struct dm_io_region *where,struct dpages *dp,unsigned long *error_bits) {/*
@@ -385,7 +397,7 @@ static int sync_io(struct dm_io_client *client, unsigned int num_regions,volatile char io_[sizeof (struct io ) + __alignof__(struct io ) -1 ];struct io *io = (struct io *)PTR_ALIGN (&io_, __alignof__(struct io )); -if (num_regions >1 && (rw &RW_MASK ) !=WRITE ) { +if (num_regions >1 && (io_req->bi_rw &RW_MASK ) !=WRITE ) {WARN_ON (1 );return -EIO ; }
@@ -393,12 +405,12 @@ static int sync_io(struct dm_io_client *client, unsigned int num_regions, io->error_bits =0 ;atomic_set (&io->count,1 );/* see dispatch_io() */ io->sleeper = current; - io->client = client; + io->client = io_req->client; io->vma_invalidate_address = dp->vma_invalidate_address; io->vma_invalidate_size = dp->vma_invalidate_size; -dispatch_io (rw, num_regions, where, dp, io,1 ); +dispatch_io (io_req, num_regions, where, dp, io,1 );while (1 ) {set_current_state (TASK_UNINTERRUPTIBLE );
@@ -416,30 +428,29 @@ static int sync_io(struct dm_io_client *client, unsigned int num_regions,return io->error_bits ? -EIO :0 ; } -static int async_io (struct dm_io_client *client,unsigned int num_regions, -struct dm_io_region *where,int rw,struct dpages *dp, -io_notify_fn fn,void *context) +static int async_io (struct dm_io_request *io_req,unsigned int num_regions, +struct dm_io_region *where,struct dpages *dp) {struct io *io; -if (num_regions >1 && (rw &RW_MASK ) !=WRITE ) { +if (num_regions >1 && (io_req->bi_rw &RW_MASK ) !=WRITE ) {WARN_ON (1 ); -fn (1 , context); + io_req->notify.fn (1 , io_req->notify.context);return -EIO ; } - io =mempool_alloc (client->pool,GFP_NOIO ); + io =mempool_alloc (io_req->client->pool,GFP_NOIO ); io->error_bits =0 ;atomic_set (&io->count,1 );/* see dispatch_io() */ io->sleeper =NULL ; - io->client = client; - io->callback = fn; - io->context = context; + io->client = io_req->client; + io->callback = io_req->notify.fn; + io->context = io_req->notify.context; io->vma_invalidate_address = dp->vma_invalidate_address; io->vma_invalidate_size = dp->vma_invalidate_size; -dispatch_io (rw, num_regions, where, dp, io,0 ); +dispatch_io (io_req, num_regions, where, dp, io,0 );return 0 ; }
@@ -499,11 +510,10 @@ int dm_io(struct dm_io_request *io_req, unsigned num_regions,return r;if (!io_req->notify.fn) -return sync_io (io_req->client, num_regions, where, - io_req->bi_rw, &dp, sync_error_bits); +return sync_io (io_req, num_regions, where, + &dp, sync_error_bits); -return async_io (io_req->client, num_regions, where, io_req->bi_rw, - &dp, io_req->notify.fn, io_req->notify.context); +return async_io (io_req, num_regions, where, &dp); }EXPORT_SYMBOL (dm_io);
drivers/md/dm-log.c +1 -0
@@ -463,6 +463,7 @@ static int create_log_context(struct dm_dirty_log *log, struct dm_target *ti,kfree (lc);return r; } + lc->io_req.only_create_bio =0 ; lc->disk_header =vmalloc (buf_size);if (!lc->disk_header) {
include/linux/dm-io.h +3 -0
@@ -61,6 +61,9 @@ struct dm_io_request {struct dm_io_memory mem;/* Memory to use for io */ struct dm_io_notify notify;/* Synchronous if notify.fn is NULL */ struct dm_io_client *client;/* Client memory handler */ +int only_create_bio; +struct bio *start; +struct bio *end; };/*
Signature · 1 line
1.7.1
PATCH 2/2 md: modify dm_io() so it could return bios instead of submitting it
From: Robin Dong <[email protected]>
When trying to modify flashcache to request based (current it's bio based), we need to make request from bios by ourselves, but dm_io() will submit these bios directly, so we propose to modify the dm_io() to return bios instead of submiting it.
This could also improve the flexibility of dm_io().
--- drivers/md/dm.c | 11 +++++++++++ include/linux/device-mapper.h | 3 +++ 2 files changed, 14 insertions(+), 0 deletions(-)
drivers/md/dm.c +11 -0
@@ -1459,11 +1459,22 @@ static int dm_request_based(struct mapped_device *md)static void dm_request (struct request_queue *q,struct bio *bio) {struct mapped_device *md = q->queuedata; +struct dm_table *map =dm_get_live_table (md); +struct dm_target *ti =dm_table_find_target (map, bio->bi_sector); + +if (ti->type->mk_rq) { + ti->type->mk_rq (ti, q, bio); + goto out; + }if (dm_request_based (md))blk_queue_bio (q, bio);else _dm_request (q, bio); + +out: +dm_table_put (map); +return ; }void dm_dispatch_request (struct request *rq)
include/linux/device-mapper.h +3 -0
@@ -50,6 +50,8 @@ typedef int (*dm_map_fn) (struct dm_target *ti, struct bio *bio, union map_info *map_context);typedef int (*dm_map_request_fn ) (struct dm_target *ti,struct request *clone,union map_info *map_context); +typedef int (*dm_make_request_fn ) (struct dm_target *ti, +struct request_queue *q,struct bio *bio);/* * Returns:
@@ -136,6 +138,7 @@ struct target_type {dm_dtr_fn dtr;dm_map_fn map;dm_map_request_fn map_rq; +dm_make_request_fn mk_rq;dm_endio_fn end_io;dm_request_endio_fn rq_end_io;dm_presuspend_fn presuspend;
Signature · 1 line
1.7.1
Re: [PATCH 2/2] md: modify dm_io() so it could return bios instead of submitting it
ping
2012/9/20 Robin Dong <[email protected]>:
Quoted 65 lines
From: Robin Dong <[email protected]>
When trying to modify flashcache to request based (current it's bio based), we need to make request from bios by ourselves, but
dm_io()will submit these bios directly, so we propose to modify thedm_io()to return bios instead of submiting it.This could also improve the flexibility of
dm_io().Signed-off-by Robin Dong <[email protected]>--- drivers/md/dm.c | 11 +++++++++++ include/linux/device-mapper.h | 3 +++ 2 files changed, 14 insertions(+), 0 deletions(-)drivers/md/dm.c +11
-0@@ -1459,11 +1459,22 @@ static int dm_request_based(struct mapped_device *md)static void dm_request (struct request_queue *q,struct bio *bio) {struct mapped_device *md = q->queuedata; +struct dm_table *map =dm_get_live_table (md); +struct dm_target *ti =dm_table_find_target (map, bio->bi_sector); + +if (ti->type->mk_rq) { + ti->type->mk_rq (ti, q, bio); + goto out; + }if (dm_request_based (md))blk_queue_bio (q, bio);else _dm_request (q, bio); + +out: +dm_table_put (map); +return ; }void dm_dispatch_request (struct request *rq)include/linux/device-mapper.h +3
-1@@ -50,6 +50,8 @@ typedef int (*dm_map_fn) (struct dm_target *ti, struct bio *bio, union map_info *map_context);typedef int (*dm_map_request_fn ) (struct dm_target *ti,struct request *clone,union map_info *map_context); +typedef int (*dm_make_request_fn ) (struct dm_target *ti, +struct request_queue *q,struct bio *bio);/* * Returns: @@ -136,6 +138,7 @@ struct target_type {dm_dtr_fn dtr;dm_map_fn map;dm_map_request_fn map_rq; +dm_make_request_fn mk_rq;dm_endio_fn end_io;dm_request_endio_fn rq_end_io;dm_presuspend_fn presuspend; --1.7.1
Signature · 3 lines
-- Best Regard Robin Dong
ping
2012/9/20 Robin Dong <[email protected]>:
Quoted 212 lines
From: Robin Dong <[email protected]>
We are now trying to modify flashcache(https://github.com/facebook/flashcache) to make it request based so that
we can let cfq io-controller control the bandwidth between different io cgroups.A search in the dm directory tells me that only multipath is a request based dm target and its functionality
is very simple andmap_rq()is used to map the request to different underlying devices. We can't work in this way because:1. the request which processed by
map_rq()need to be issued to
different lower devices (disk device and cache device, in flashcache), therefore the request
can't be totally remapped by simply changing its queue and returningDM_MAPIO_REMAPPEDinmap_rq()likemultipath_map()
2. to submit bios drectly inmap_rq()(by returnDM_MAPIO_SUBMITTED) will cause BUG_ON(!irqs_disabled())in dm_request_fn() because the submit_bio()->generic_make_request()->blk_queue_bio() will definitly call spin_unlock_irq to enable the irqsAs above,the interface
map_rq()provided by devcie-mapper framework is not enough for an autonomous target, like flashcache.We propose to add a new
mk_rqinterface so that we can make the requests
by ourselves.Signed-off-by Robin Dong <[email protected]>--- drivers/md/dm-io.c | 58 ++++++++++++++++++++++++++++-------------------- drivers/md/dm-log.c | 1 + include/linux/dm-io.h | 3 ++ 3 files changed, 38 insertions(+), 24 deletions(-)drivers/md/dm-io.c +34
-24@@ -287,8 +287,8 @@ static void km_dp_init(struct dpages *dp, void *data)/*----------------------------------------------------------------- * IO routines that accept a list of pages. *---------------------------------------------------------------*/ -static void do_region (int rw,unsigned region,struct dm_io_region *where, -struct dpages *dp,struct io *io) +static void do_region (struct dm_io_request *io_req,unsigned region, +struct dm_io_region *where,struct dpages *dp,struct io *io) {struct bio *bio;struct page *page;@@ -298,6 +298,7 @@ static void do_region(int rw, unsigned region, struct dm_io_region *where,sector_t remaining = where->count;struct request_queue *q =bdev_get_queue (where->bdev);sector_t discard_sectors; +int rw = io_req->bi_rw;/* * where->count may be zero if rw holds a flush and we need to @@ -339,15 +340,26 @@ static void do_region(int rw, unsigned region, struct dm_io_region *where, }atomic_inc (&io->count); -submit_bio (rw, bio); +if (!io_req->only_create_bio) +submit_bio (rw, bio); +else { + bio->bi_rw |= rw; +if (io_req->start) { + io_req->end->bi_next = bio; + io_req->end = bio; + }else + io_req->start = io_req->end = bio; + bio->bi_next =NULL ; + } }while (remaining); } -static void dispatch_io (int rw,unsigned int num_regions, +static void dispatch_io (struct dm_io_request *io_req,unsigned int num_regions,struct dm_io_region *where,struct dpages *dp,struct io *io,int sync) {int i; +int rw = io_req->bi_rw;struct dpages old_pages = *dp;BUG_ON (num_regions >DM_IO_MAX_REGIONS );@@ -362,7 +374,7 @@ static void dispatch_io(int rw, unsigned int num_regions,for (i =0 ; i < num_regions; i++) { *dp = old_pages;if (where[i].count || (rw &REQ_FLUSH )) -do_region (rw, i, where + i, dp, io); +do_region (io_req, i, where + i, dp, io); }/* @@ -372,8 +384,8 @@ static void dispatch_io(int rw, unsigned int num_regions,dec_count (io,0 ,0 ); } -static int sync_io (struct dm_io_client *client,unsigned int num_regions, -struct dm_io_region *where,int rw,struct dpages *dp, +static int sync_io (struct dm_io_request *io_req,unsigned int num_regions, +struct dm_io_region *where,struct dpages *dp,unsigned long *error_bits) {/* @@ -385,7 +397,7 @@ static int sync_io(struct dm_io_client *client, unsigned int num_regions,volatile char io_[sizeof (struct io ) + __alignof__(struct io ) -1 ];struct io *io = (struct io *)PTR_ALIGN (&io_, __alignof__(struct io )); -if (num_regions >1 && (rw &RW_MASK ) !=WRITE ) { +if (num_regions >1 && (io_req->bi_rw &RW_MASK ) !=WRITE ) {WARN_ON (1 );return -EIO ; }@@ -393,12 +405,12 @@ static int sync_io(struct dm_io_client *client, unsigned int num_regions, io->error_bits =0 ;atomic_set (&io->count,1 );/* see dispatch_io() */ io->sleeper = current; - io->client = client; + io->client = io_req->client; io->vma_invalidate_address = dp->vma_invalidate_address; io->vma_invalidate_size = dp->vma_invalidate_size; -dispatch_io (rw, num_regions, where, dp, io,1 ); +dispatch_io (io_req, num_regions, where, dp, io,1 );while (1 ) {set_current_state (TASK_UNINTERRUPTIBLE );@@ -416,30 +428,29 @@ static int sync_io(struct dm_io_client *client, unsigned int num_regions,return io->error_bits ? -EIO :0 ; } -static int async_io (struct dm_io_client *client,unsigned int num_regions, -struct dm_io_region *where,int rw,struct dpages *dp, -io_notify_fn fn,void *context) +static int async_io (struct dm_io_request *io_req,unsigned int num_regions, +struct dm_io_region *where,struct dpages *dp) {struct io *io; -if (num_regions >1 && (rw &RW_MASK ) !=WRITE ) { +if (num_regions >1 && (io_req->bi_rw &RW_MASK ) !=WRITE ) {WARN_ON (1 ); -fn (1 , context); + io_req->notify.fn (1 , io_req->notify.context);return -EIO ; } - io =mempool_alloc (client->pool,GFP_NOIO ); + io =mempool_alloc (io_req->client->pool,GFP_NOIO ); io->error_bits =0 ;atomic_set (&io->count,1 );/* see dispatch_io() */ io->sleeper =NULL ; - io->client = client; - io->callback = fn; - io->context = context; + io->client = io_req->client; + io->callback = io_req->notify.fn; + io->context = io_req->notify.context; io->vma_invalidate_address = dp->vma_invalidate_address; io->vma_invalidate_size = dp->vma_invalidate_size; -dispatch_io (rw, num_regions, where, dp, io,0 ); +dispatch_io (io_req, num_regions, where, dp, io,0 );return 0 ; }@@ -499,11 +510,10 @@ int dm_io(struct dm_io_request *io_req, unsigned num_regions,return r;if (!io_req->notify.fn) -return sync_io (io_req->client, num_regions, where, - io_req->bi_rw, &dp, sync_error_bits); +return sync_io (io_req, num_regions, where, + &dp, sync_error_bits); -return async_io (io_req->client, num_regions, where, io_req->bi_rw, - &dp, io_req->notify.fn, io_req->notify.context); +return async_io (io_req, num_regions, where, &dp); }EXPORT_SYMBOL (dm_io);drivers/md/dm-log.c +1
-0@@ -463,6 +463,7 @@ static int create_log_context(struct dm_dirty_log *log, struct dm_target *ti,kfree (lc);return r; } + lc->io_req.only_create_bio =0 ; lc->disk_header =vmalloc (buf_size);if (!lc->disk_header) {include/linux/dm-io.h +3
-1@@ -61,6 +61,9 @@ struct dm_io_request {struct dm_io_memory mem;/* Memory to use for io */ struct dm_io_notify notify;/* Synchronous if notify.fn is NULL */ struct dm_io_client *client;/* Client memory handler */ +int only_create_bio; +struct bio *start; +struct bio *end; };/* -- 1.7.1
Signature · 3 lines
-- Best Regard Robin Dong