anticipatory I/O scheduler DESC AS: pgbench improvement EDESC From: Nick Piggin This patch improves single threaded pgbench performance from about 40 to 65 TPS. It also does a few minor tweaks here and there. Performance for all my tests is as good or better with this patch. tiobench 256 seq read performance is still bad, though not due to this patch. I am fairly sure that the problem is due to that seek distance patch. The seek distance patch also allowed any readers to seek backwards, not just the process who had last made a read. This patch nearly tripled WimMark performance, but probably does 3 times worse in tiobench 256 sequential reads. tiobench 16 seq reads also aren't as good as they should be. DESC AS: discrete read fifo batches EDESC From: Nick Piggin This patch does the following: starts a new fifo batch when changing from write to read starts from the oldest request first when chaning from write to read doesn't finish a read batch until the fifo batch is also expired fifo batches can loosely be thought of as a process IO timeslice. DESC AS sync/async batches EDESC From: Nick Piggin Performance on my tests with mm7 is good. tiobench sequential reads still isn't really good (single processor). I think this is because of the seek_mean logic. I think tiobench files aren't layed out in a very real world manner, are they? Anyway, the following patch changes the whole notion of READ/WRITE to SYNC/ASYNC. Now this is the "simple" way because it does not also keep a list in READ/WRITE order for merging. I think this is actually alright because there probably isn't a lot of wasted merging opportunity. The generic block layer shouldn't mind that we might offer a READ request as a candidate to merge with a WRITE: it checks this and will just disallow the merge. Anyway, OraSim does not change much, pgbench gains about 15%. The biggest difference is just the artificial: Bench 7 - 2 threads, 1 reading, 1 doing write+fsync 2.5.69-mm7: IO Rate: 51.21 MB/s, Sync writes per read: 0.0005 2.5.69-mm7+this: IO Rate: 36.19 MB/s Sync writes per read: 0.8674 Because we now anticipate on reads. Probably help more with ext3 + kjournald stuff. WimMark would be interesting. DESC AS: hash removal fix EDESC From Jens. DESC AS jumbo patch (for SCSI and TCQ) EDESC From: Nick Piggin It seems to go alright the basic tests. Really helps "1 reader and 1 writer" on my SCSI drive, but I'd like you to do your run of TCQ tests on it if you could. Quite a few big changes in order to get SCSI TCQ working properly. I'll give you a short: * Measure write batches by # of requests. Time has little meaning as we can fill a lot of disk cache in a small amount of time. (This option should be moved back to a time based tunable, which adjusts itself based on how long the last run took. At the moment it probably won't writeout enough on small request size devices) * When changing batch direction, don't submit any requests in the new direction until all outstanding requests have been completed. * When starting a new batch direction, make the first request a hard barrier (don't know if this helps). * Don't start timing a read (actually "SYNC") batch until the first request has been completed. This should help with writeback caches. It also includes some fixes / misc stuff done while the above was being done: * cleanup / move enums around * Now that we have sync vs async, we shouldn't need to keep track of "data dir". Get rid of it. * Thinktime is only accounted when the process has no outstanding requests. This seems pretty sane when you think about it. * Limit the rate at which seek_distance can grow. A generally nicely seeking process shouldn't suffer _too_ much if they have to seek to a fragment, or have to page in some code/library/swap, etc etc. * Properly set arq->state = AS_RQ_NEW. Now that arqs are dynamically initialised, the previous place this was done was more useless than ever. * Allow REQ_HARDBARRIER requests to be dispatched normally. No reason why they shouldn't. DESC AS: fix stupid thinko EDESC Stupid The effect is that it enables a major part of the SCSI TCQ stuff. It seems to clamp down the batches a lot more. An effect is that devices like IDE, with small request sizes get even less async writes though, so it will be important DESC AS: no batch-antic-limit EDESC From: Nick Piggin Now the current_batch_expires is set "lazily" now, and it has two meanings (for REQ_SYNC it means jiffies, REQ_ASYNC means # of requests, I really have to clean that up) this logic wasn't working properly. Remove it. Helps sync writers... DESC AS: autotune write batches EDESC From: Nick Piggin This is hopefully the last big thing AS needs. It reintroduces write_batch_expire to be a time value, and adjusts the number of requests allowed in a write batch based on this target time. DESC AS: divide by zero fix EDESC From: Robert Love Fix as-iosched divide-by-zero bug. DESC AS: more HZ != 1000 fixes EDESC From: Nick Piggin DESC AS: update_write_batch tuning EDESC From: Nick Piggin DESC AS locking EDESC From: Nick Piggin This patch locks the as_io_context when non atomic fields are modified or read->written, etc. The two places this happens are in as_update_iohist (when a request is inserted, process context), and as_completed_request (when a request is completed, interrupt context). Now the process context can't race with itself, but it can race with a completion of a request from the same process on another queue, or two completions on two queues from the same process can race. The other place which is still vulnerable is can_break_antic This only reads values, so they won't be corrupted, but may still be inaccurate. I don't know how important this is: the major stuff is done in as_update_iohist, which can't race against itself... DESC AS HZ fixes EDESC From: Nick Piggin Following patch converts sysfs input and output to ms for all HZ. drivers/block/Makefile | 3 drivers/block/as-iosched.c | 1864 +++++++++++++++++++++++++++++++++++++++++++++ drivers/block/ll_rw_blk.c | 25 fs/buffer.c | 4 fs/fs-writeback.c | 2 include/linux/elevator.h | 5 include/linux/sched.h | 5 kernel/exit.c | 2 kernel/fork.c | 1 9 files changed, 1908 insertions(+), 3 deletions(-) diff -puN /dev/null drivers/block/as-iosched.c --- /dev/null 2002-08-30 16:31:37.000000000 -0700 +++ 25-akpm/drivers/block/as-iosched.c 2003-06-21 20:47:15.000000000 -0700 @@ -0,0 +1,1864 @@ +/* + * linux/drivers/block/as-iosched.c + * + * Anticipatory & deadline i/o scheduler. + * + * Copyright (C) 2002 Jens Axboe + * Nick Piggin + * + */ +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#define REQ_SYNC 1 +#define REQ_ASYNC 0 + +/* + * See Documentation/as-iosched.txt + */ + +/* + * max time before a read is submitted. + */ +#define default_read_expire (HZ / 20) + +/* + * ditto for writes, these limits are not hard, even + * if the disk is capable of satisfying them. + */ +#define default_write_expire (HZ / 5) + +/* + * read_batch_expire describes how long we will allow a stream of reads to + * persist before looking to see whether it is time to switch over to writes. + */ +#define default_read_batch_expire (HZ / 5) + +/* + * write_batch_expire describes how long we want a stream of writes to run for. + * This is not a hard limit, but a target we set for the auto-tuning thingy. + * See, the problem is: we can send a lot of writes to disk cache / TCQ in + * a short amount of time... + */ +#define default_write_batch_expire (HZ / 20) + +/* + * max time we may wait to anticipate a read (default around 6ms) + */ +#define default_antic_expire ((HZ / 150) ? HZ / 150 : 1) + +/* + * This is the per-process anticipatory I/O scheduler state. It is refcounted + * and kmalloc'ed. + * + * There is no locking protecting the contents of this structure! Pointers + * to a single as_io_context may appear in multiple queues at once. + */ + +/* + * Keep track of up to 20ms thinktimes. We can go as big as we like here, + * however huge values tend to interfere and not decay fast enough. A program + * might be in a non-io phase of operation. Waiting on user input for example, + * or doing a lengthy computation. A small penalty can be justified there, and + * will still catch out those processes that constantly have large thinktimes. + */ +#define MAX_THINKTIME (HZ/50UL) + +/* Bits in as_io_context.state */ +enum as_io_states { + AS_TASK_RUNNING=0, /* Process has not exitted */ + AS_TASK_IORUNNING, /* Process has completed some IO */ +}; + +struct as_io_context { + atomic_t refcount; + pid_t pid; + unsigned long state; + atomic_t nr_queued; /* queued reads & sync writes */ + atomic_t nr_dispatched; /* number of requests gone to the drivers */ + + spinlock_t lock; + + /* IO History tracking */ + /* Thinktime */ + unsigned long last_end_request; + unsigned long ttime_total; + unsigned long ttime_samples; + unsigned long ttime_mean; + /* Layout pattern */ + long seek_samples; + sector_t last_request_pos; + sector_t seek_total; + sector_t seek_mean; +}; + +enum anticipation_status { + ANTIC_OFF=0, /* Not anticipating (normal operation) */ + ANTIC_WAIT_REQ, /* The last read has not yet completed */ + ANTIC_WAIT_NEXT, /* Currently anticipating a request vs + last read (which has completed) */ + ANTIC_FINISHED, /* Anticipating but have found a candidate + * or timed out */ +}; + +struct as_data { + /* + * run time data + */ + + struct request_queue *q; /* the "owner" queue */ + + /* + * requests (as_rq s) are present on both sort_list and fifo_list + */ + struct rb_root sort_list[2]; + struct list_head fifo_list[2]; + + struct as_rq *next_arq[2]; /* next in sort order */ + sector_t last_sector[2]; /* last REQ_SYNC & REQ_ASYNC sectors */ + struct list_head *dispatch; /* driver dispatch queue */ + struct list_head *hash; /* request hash */ + unsigned long hash_valid_count; /* barrier hash count */ + unsigned long current_batch_expires; + unsigned long last_check_fifo[2]; + int changed_batch; + int batch_data_dir; /* current batch REQ_SYNC / REQ_ASYNC */ + int write_batch_count; /* max # of reqs in a write batch */ + int current_write_count; /* how many requests left this batch */ + int write_batch_idled; /* has the write batch gone idle? */ + mempool_t *arq_pool; + + enum anticipation_status antic_status; + unsigned long antic_start; /* jiffies: when it started */ + struct timer_list antic_timer; /* anticipatory scheduling timer */ + struct work_struct antic_work; /* Deferred unplugging */ + struct as_io_context *as_io_context;/* Identify the expected process */ + int aic_finished; /* IO associated with as_io_context finished */ + int nr_dispatched; + + /* + * settings that change how the i/o scheduler behaves + */ + unsigned long fifo_expire[2]; + unsigned long batch_expire[2]; + unsigned long antic_expire; +}; + +#define list_entry_fifo(ptr) list_entry((ptr), struct as_rq, fifo) + +/* + * per-request data. + */ +enum arq_state { + AS_RQ_NEW=0, /* New - not referenced and not on any lists */ + AS_RQ_QUEUED, /* In the request queue. It belongs to the + scheduler */ + AS_RQ_DISPATCHED, /* On the dispatch list. It belongs to the + driver now */ +}; + +struct as_rq { + /* + * rbtree index, key is the starting offset + */ + struct rb_node rb_node; + sector_t rb_key; + + struct request *request; + + struct as_io_context *as_io_context; /* The submitting task */ + + /* + * request hash, key is the ending offset (for back merge lookup) + */ + struct list_head hash; + unsigned long hash_valid_count; + + /* + * expire fifo + */ + struct list_head fifo; + unsigned long expires; + + int is_sync; + enum arq_state state; /* debug only */ +}; + +#define RQ_DATA(rq) ((struct as_rq *) (rq)->elevator_private) + +static kmem_cache_t *arq_pool; + +/* + * IO Context helper functions + */ +/* Debug */ +static atomic_t nr_as_io_requests = ATOMIC_INIT(0); + +static void put_as_io_context(struct as_io_context **paic) +{ + struct as_io_context *aic = *paic; + + if (aic == NULL) + return; + + BUG_ON(atomic_read(&aic->refcount) == 0); + *paic = NULL; + if (atomic_dec_and_test(&aic->refcount)) { + atomic_dec(&nr_as_io_requests); + kfree(aic); + } +} + +/* Called by the exitting task */ +void exit_as_io_context(void) +{ + unsigned long flags; + struct as_io_context *aic; + + local_irq_save(flags); + aic = current->as_io_context; + if (aic) { + clear_bit(AS_TASK_RUNNING, &aic->state); + put_as_io_context(&aic); + current->as_io_context = NULL; + } + local_irq_restore(flags); +} + +/* + * If the current task has no IO context then create one and initialise it. + * If it does have a context, take a ref on it. + * + * This is always called in the context of the task which submitted the I/O. + * But weird things happen, so we disable local interrupts to ensure exclusive + * access to *current. + */ +static struct as_io_context *get_as_io_context(void) +{ + struct task_struct *tsk = current; + unsigned long flags; + struct as_io_context *ret; + + local_irq_save(flags); + ret = tsk->as_io_context; + if (ret == NULL) { + ret = kmalloc(sizeof(*ret), GFP_ATOMIC); + if (ret) { + atomic_inc(&nr_as_io_requests); + atomic_set(&ret->refcount, 1); + ret->pid = tsk->pid; + ret->state = 1 << AS_TASK_RUNNING; + atomic_set(&ret->nr_queued, 0); + atomic_set(&ret->nr_dispatched, 0); + spin_lock_init(&ret->lock); + ret->ttime_total = 0; + ret->ttime_samples = 0; + ret->ttime_mean = 0; + ret->seek_total = 0; + ret->seek_samples = 0; + ret->seek_mean = 0; + tsk->as_io_context = ret; + } + } + local_irq_restore(flags); + atomic_inc(&ret->refcount); + return ret; +} + +static void +copy_as_io_context(struct as_io_context **pdst, struct as_io_context **psrc) +{ + struct as_io_context *src = *psrc; + + if (src) { + BUG_ON(atomic_read(&src->refcount) == 0); + atomic_inc(&src->refcount); + put_as_io_context(pdst); + *pdst = src; + } +} + +static void +swap_as_io_context(struct as_io_context **aic1, struct as_io_context **aic2) +{ + struct as_io_context *temp; + temp = *aic1; + *aic1 = *aic2; + *aic2 = temp; +} + +/* + * the back merge hash support functions + */ +static const int as_hash_shift = 6; +#define AS_HASH_BLOCK(sec) ((sec) >> 3) +#define AS_HASH_FN(sec) (hash_long(AS_HASH_BLOCK((sec)), as_hash_shift)) +#define AS_HASH_ENTRIES (1 << as_hash_shift) +#define rq_hash_key(rq) ((rq)->sector + (rq)->nr_sectors) +#define list_entry_hash(ptr) list_entry((ptr), struct as_rq, hash) +#define ON_HASH(arq) (arq)->hash_valid_count + +#define AS_INVALIDATE_HASH(ad) \ + do { \ + if (!++(ad)->hash_valid_count) \ + (ad)->hash_valid_count = 1; \ + } while (0) + +static inline void __as_del_arq_hash(struct as_rq *arq) +{ + arq->hash_valid_count = 0; + list_del_init(&arq->hash); +} + +static inline void as_del_arq_hash(struct as_rq *arq) +{ + if (ON_HASH(arq)) + __as_del_arq_hash(arq); +} + +static void as_remove_merge_hints(request_queue_t *q, struct as_rq *arq) +{ + as_del_arq_hash(arq); + + if (q->last_merge == &arq->request->queuelist) + q->last_merge = NULL; +} + +static void as_add_arq_hash(struct as_data *ad, struct as_rq *arq) +{ + struct request *rq = arq->request; + + BUG_ON(ON_HASH(arq)); + + arq->hash_valid_count = ad->hash_valid_count; + list_add(&arq->hash, &ad->hash[AS_HASH_FN(rq_hash_key(rq))]); +} + +/* + * move hot entry to front of chain + */ +static inline void as_hot_arq_hash(struct as_data *ad, struct as_rq *arq) +{ + struct request *rq = arq->request; + struct list_head *head = &ad->hash[AS_HASH_FN(rq_hash_key(rq))]; + + if (ON_HASH(arq) && arq->hash.prev != head) { + list_del(&arq->hash); + list_add(&arq->hash, head); + } +} + +static struct request *as_find_arq_hash(struct as_data *ad, sector_t offset) +{ + struct list_head *hash_list = &ad->hash[AS_HASH_FN(offset)]; + struct list_head *entry, *next = hash_list->next; + + while ((entry = next) != hash_list) { + struct as_rq *arq = list_entry_hash(entry); + struct request *__rq = arq->request; + + next = entry->next; + + BUG_ON(!ON_HASH(arq)); + + if (!rq_mergeable(__rq) + || arq->hash_valid_count != ad->hash_valid_count) { + __as_del_arq_hash(arq); + continue; + } + + if (rq_hash_key(__rq) == offset) + return __rq; + } + + return NULL; +} + +/* + * rb tree support functions + */ +#define RB_NONE (2) +#define RB_EMPTY(root) ((root)->rb_node == NULL) +#define ON_RB(node) ((node)->rb_color != RB_NONE) +#define RB_CLEAR(node) ((node)->rb_color = RB_NONE) +#define rb_entry_arq(node) rb_entry((node), struct as_rq, rb_node) +#define ARQ_RB_ROOT(ad, arq) (&(ad)->sort_list[(arq)->is_sync]) +#define rq_rb_key(rq) (rq)->sector + +/* + * as_find_first_arq finds the first (lowest sector numbered) request + * for the specified data_dir. Used to sweep back to the start of the disk + * (1-way elevator) after we process the last (highest sector) request. + */ +static struct as_rq *as_find_first_arq(struct as_data *ad, int data_dir) +{ + struct rb_node *n = ad->sort_list[data_dir].rb_node; + + if (n == NULL) + return NULL; + + for (;;) { + if (n->rb_left == NULL) + return rb_entry_arq(n); + + n = n->rb_left; + } +} + +static struct as_rq *__as_add_arq_rb(struct as_data *ad, struct as_rq *arq) +{ + struct rb_node **p = &ARQ_RB_ROOT(ad, arq)->rb_node; + struct rb_node *parent = NULL; + struct as_rq *__arq; + + while (*p) { + parent = *p; + __arq = rb_entry_arq(parent); + + if (arq->rb_key < __arq->rb_key) + p = &(*p)->rb_left; + else if (arq->rb_key > __arq->rb_key) + p = &(*p)->rb_right; + else + return __arq; + } + + rb_link_node(&arq->rb_node, parent, p); + return 0; +} + +static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq); +/* + * Add the request to the rb tree if it is unique. If there is an alias (an + * existing request against the same sector), which can happen when using + * direct IO, then move the alias to the dispatch list and then add the + * request. + */ +static void as_add_arq_rb(struct as_data *ad, struct as_rq *arq) +{ + struct as_rq *alias; + struct request *rq = arq->request; + + arq->rb_key = rq_rb_key(rq); + + /* This can be caused by direct IO */ + while ((alias = __as_add_arq_rb(ad, arq))) + as_move_to_dispatch(ad, alias); + + rb_insert_color(&arq->rb_node, ARQ_RB_ROOT(ad, arq)); +} + +static inline void as_del_arq_rb(struct as_data *ad, struct as_rq *arq) +{ + if (ON_RB(&arq->rb_node)) { + rb_erase(&arq->rb_node, ARQ_RB_ROOT(ad, arq)); + RB_CLEAR(&arq->rb_node); + } +} + +static struct request * +as_find_arq_rb(struct as_data *ad, sector_t sector, int data_dir) +{ + struct rb_node *n = ad->sort_list[data_dir].rb_node; + struct as_rq *arq; + + while (n) { + arq = rb_entry_arq(n); + + if (sector < arq->rb_key) + n = n->rb_left; + else if (sector > arq->rb_key) + n = n->rb_right; + else + return arq->request; + } + + return NULL; +} + +/* + * IO Scheduler proper + */ + +#define MAXBACK (1024 * 1024) /* + * Maximum distance the disk will go backward + * for a request. + */ + +/* + * as_choose_req selects the preferred one of two requests of the same data_dir + * ignoring time - eg. timeouts, which is the job of as_dispatch_request + */ +static struct as_rq * +as_choose_req(struct as_data *ad, struct as_rq *arq1, struct as_rq *arq2) +{ + int data_dir; + sector_t last, s1, s2, d1, d2; + int r1_wrap=0, r2_wrap=0; /* requests are behind the disk head */ + const sector_t maxback = MAXBACK; + + if (arq1 == NULL || arq1 == arq2) + return arq2; + if (arq2 == NULL) + return arq1; + + data_dir = arq1->is_sync; + + last = ad->last_sector[data_dir]; + s1 = arq1->request->sector; + s2 = arq2->request->sector; + + BUG_ON(data_dir != arq2->is_sync); + + /* + * Strict one way elevator _except_ in the case where we allow + * short backward seeks which are biased as twice the cost of a + * similar forward seek. + */ + if (s1 >= last) + d1 = s1 - last; + else if (s1+maxback >= last) + d1 = (last - s1)*2; + else { + r1_wrap = 1; + d1 = 0; /* shut up, gcc */ + } + + if (s2 >= last) + d2 = s2 - last; + else if (s2+maxback >= last) + d2 = (last - s2)*2; + else { + r2_wrap = 1; + d2 = 0; + } + + /* Found required data */ + if (!r1_wrap && r2_wrap) + return arq1; + else if (!r2_wrap && r1_wrap) + return arq2; + else if (r1_wrap && r2_wrap) { + /* both behind the head */ + if (s1 <= s2) + return arq1; + else + return arq2; + } + + /* Both requests in front of the head */ + if (d1 < d2) + return arq1; + else if (d2 < d1) + return arq2; + else { + if (s1 >= s2) + return arq1; + else + return arq2; + } +} + +/* + * as_find_next_arq finds the next request after @prev in elevator order. + * this with as_choose_req form the basis for how the scheduler chooses + * what request to process next. Anticipation works on top of this. + */ +static struct as_rq *as_find_next_arq(struct as_data *ad, struct as_rq *last) +{ + const int data_dir = last->is_sync; + struct as_rq *ret; + struct rb_node *rbnext = rb_next(&last->rb_node); + struct rb_node *rbprev = rb_prev(&last->rb_node); + struct as_rq *arq_next, *arq_prev; + + BUG_ON(!ON_RB(&last->rb_node)); + + if (rbprev) + arq_prev = rb_entry_arq(rbprev); + else + arq_prev = NULL; + + if (rbnext) + arq_next = rb_entry_arq(rbnext); + else { + arq_next = as_find_first_arq(ad, data_dir); + if (arq_next == last) + arq_next = NULL; + } + + ret = as_choose_req(ad, arq_next, arq_prev); + + return ret; +} + +/* + * anticipatory scheduling functions follow + */ + +/* + * as_antic_expired tells us when we have anticipated too long. + * The funny "absolute difference" math on the elapsed time is to handle + * jiffy wraps, and disks which have been idle for 0x80000000 jiffies. + */ +static int as_antic_expired(struct as_data *ad) +{ + long delta_jif; + + delta_jif = jiffies - ad->antic_start; + if (unlikely(delta_jif < 0)) + delta_jif = -delta_jif; + if (delta_jif < ad->antic_expire) + return 0; + + return 1; +} + +/* + * as_antic_waitnext starts anticipating that a nice request will soon be + * submitted. See also as_antic_waitreq + */ +static void as_antic_waitnext(struct as_data *ad) +{ + unsigned long timeout; + + BUG_ON(ad->antic_status != ANTIC_OFF + && ad->antic_status != ANTIC_WAIT_REQ); + + timeout = ad->antic_start + ad->antic_expire; + + mod_timer(&ad->antic_timer, timeout); + + ad->antic_status = ANTIC_WAIT_NEXT; +} + +/* + * as_antic_waitreq starts anticipating. We don't start timing the anticipation + * until the request that we're anticipating on has finished. This means we + * are timing from when the candidate process wakes up hopefully. + */ +static void as_antic_waitreq(struct as_data *ad) +{ + BUG_ON(ad->antic_status == ANTIC_FINISHED); + if (ad->antic_status == ANTIC_OFF) { + if (!ad->as_io_context || ad->aic_finished) + as_antic_waitnext(ad); + else + ad->antic_status = ANTIC_WAIT_REQ; + } +} + +/* + * This is called directly by the functions in this file to stop anticipation. + * We kill the timer and schedule a call to the request_fn asap. + */ +static void as_antic_stop(struct as_data *ad) +{ + int status = ad->antic_status; + + if (status == ANTIC_WAIT_REQ || status == ANTIC_WAIT_NEXT) { + if (status == ANTIC_WAIT_NEXT) + del_timer(&ad->antic_timer); + ad->antic_status = ANTIC_FINISHED; + /* see as_work_handler */ + kblockd_schedule_work(&ad->antic_work); + } +} + +/* + * as_antic_timeout is the timer function set by as_antic_waitnext. + */ +static void as_antic_timeout(unsigned long data) +{ + struct request_queue *q = (struct request_queue *)data; + struct as_data *ad = q->elevator.elevator_data; + unsigned long flags; + + spin_lock_irqsave(q->queue_lock, flags); + if (ad->antic_status == ANTIC_WAIT_REQ + || ad->antic_status == ANTIC_WAIT_NEXT) { + ad->antic_status = ANTIC_FINISHED; + kblockd_schedule_work(&ad->antic_work); + } + spin_unlock_irqrestore(q->queue_lock, flags); +} + +/* + * as_close_req decides if one request is considered "close" to the + * previous one issued. + */ +static int as_close_req(struct as_data *ad, struct as_rq *arq) +{ + unsigned long delay; /* milliseconds */ + sector_t last = ad->last_sector[ad->batch_data_dir]; + sector_t next = arq->request->sector; + sector_t delta; /* acceptable close offset (in sectors) */ + + if (ad->antic_status == ANTIC_OFF || !ad->aic_finished) + delay = 0; + else + delay = ((jiffies - ad->antic_start) * 1000) / HZ; + + if (delay <= 1) + delta = 32; + else if (delay <= 20 && delay <= ad->antic_expire) + delta = 32 << (delay-1); + else + return 1; + + return (last - delta <= next) && (next <= last + delta); +} + +/* + * as_can_break_anticipation returns true if we have been anticipating this + * request. + * + * It also returns true if the process against which we are anticipating + * submits a write - that's presumably an fsync, O_SYNC write, etc. We want to + * dispatch it ASAP, because we know that application will not be submitting + * any new reads. + * + * If the task which has submitted the request has exitted, break anticipation. + * + * If this task has queued some other IO, do not enter enticipation. + */ +static int as_can_break_anticipation(struct as_data *ad, struct as_rq *arq) +{ + struct as_io_context *aic; + + if (arq && arq->is_sync == REQ_SYNC && as_close_req(ad, arq)) { + /* close request */ + return 1; + } + + if (ad->aic_finished && as_antic_expired(ad)) { + /* + * In this situation status should really be FINISHED, + * however the timer hasn't had the chance to run yet. + */ + return 1; + } + + aic = ad->as_io_context; + BUG_ON(!aic); + + if (arq && aic == arq->as_io_context) { + /* request from same process */ + return 1; + } + + if (!test_bit(AS_TASK_RUNNING, &aic->state)) { + /* process anticipated on has exitted */ + return 1; + } + + if (atomic_read(&aic->nr_queued) > 0) { + /* process has more requests queued */ + return 1; + } + + if (atomic_read(&aic->nr_dispatched) > 0) { + /* process has more requests dispatched */ + return 1; + } + + if (aic->ttime_mean > ad->antic_expire) { + /* the process thinks too much between requests */ + return 1; + } + + if (arq && aic->seek_samples) { + sector_t s; + if (ad->last_sector[REQ_SYNC] < arq->request->sector) + s = arq->request->sector - ad->last_sector[REQ_SYNC]; + else + s = ad->last_sector[REQ_SYNC] - arq->request->sector; + if (aic->seek_mean > s) { + /* this request is better than what we're expecting */ + return 1; + } + } + + return 0; +} + +/* + * as_can_anticipate indicates weather we should either run arq + * or keep anticipating a better request. + */ +static int as_can_anticipate(struct as_data *ad, struct as_rq *arq) +{ + if (!ad->as_io_context) + /* + * Last request submitted was a write + */ + return 0; + + if (ad->antic_status == ANTIC_FINISHED) + /* + * Don't restart if we have just finished. Run the next request + */ + return 0; + + if (as_can_break_anticipation(ad, arq)) + /* + * This request is a good candidate. Don't keep anticipating, + * run it. + */ + return 0; + + /* + * OK from here, we haven't finished, and don't have a decent request! + * Status is either ANTIC_OFF so start waiting, + * ANTIC_WAIT_REQ so continue waiting for request to finish + * or ANTIC_WAIT_NEXT so continue waiting for an acceptable request. + * + */ + + return 1; +} + +/* + * as_update_iohist keeps a decaying histogram of IO thinktimes, and + * updates @aic->ttime_mean based on that. It is called when a new + * request is queued. + */ +static void as_update_iohist(struct as_io_context *aic, struct request *rq) +{ + struct as_rq *arq = RQ_DATA(rq); + int data_dir = arq->is_sync; + unsigned long thinktime; + sector_t seek_dist; + + if (aic == NULL) + return; + + if (data_dir == REQ_SYNC) { + spin_lock(&aic->lock); + + if (test_bit(AS_TASK_IORUNNING, &aic->state) + && !atomic_read(&aic->nr_queued) + && !atomic_read(&aic->nr_dispatched)) { + /* Calculate read -> read thinktime */ + thinktime = jiffies - aic->last_end_request; + thinktime = min(thinktime, MAX_THINKTIME-1); + /* fixed point: 1.0 == 1<<8 */ + aic->ttime_samples += 256; + aic->ttime_total += 256*thinktime; + if (aic->ttime_samples) + /* fixed point factor is cancelled here */ + aic->ttime_mean = (aic->ttime_total + 128) + / aic->ttime_samples; + aic->ttime_samples = (aic->ttime_samples>>1) + + (aic->ttime_samples>>2); + aic->ttime_total = (aic->ttime_total>>1) + + (aic->ttime_total>>2); + } + + /* Calculate read -> read seek distance */ + if (aic->last_request_pos < rq->sector) + seek_dist = rq->sector - aic->last_request_pos; + else + seek_dist = aic->last_request_pos - rq->sector; + aic->last_request_pos = rq->sector + rq->nr_sectors; + + /* + * Don't allow the seek distance to get too large from the + * odd fragment, pagein, etc + */ + if (!aic->seek_samples) + seek_dist = 0; + else if (aic->seek_samples < 400) /* second&third seek */ + seek_dist = min(seek_dist, (aic->seek_mean * 4) + + 2*1024*1024); + else + seek_dist = min(seek_dist, (aic->seek_mean * 4) + + 2*1024*64); + + aic->seek_samples += 256; + aic->seek_total += 256*seek_dist; + if (aic->seek_samples) { + aic->seek_mean = aic->seek_total + 128; + do_div(aic->seek_mean, aic->seek_samples); + } + aic->seek_samples = (aic->seek_samples>>1) + + (aic->seek_samples>>2); + aic->seek_total = (aic->seek_total>>1) + + (aic->seek_total>>2); + + spin_unlock(&aic->lock); + } +} + +/* + * as_update_arq must be called whenever a request (arq) is added to + * the sort_list. This function keeps caches up to date, and checks if the + * request might be one we are "anticipating" + */ +static void as_update_arq(struct as_data *ad, struct as_rq *arq) +{ + const int data_dir = arq->is_sync; + + /* keep the next_arq cache up to date */ + ad->next_arq[data_dir] = as_choose_req(ad, arq, ad->next_arq[data_dir]); + + /* + * have we been anticipating this request? + * or does it come from the same process as the one we are anticipating + * for? + */ + if (ad->antic_status == ANTIC_WAIT_REQ + || ad->antic_status == ANTIC_WAIT_NEXT) { + if (as_can_break_anticipation(ad, arq)) + as_antic_stop(ad); + } +} + +/* + * Gathers timings and resizes the write batch automatically + */ +void update_write_batch(struct as_data *ad) +{ + unsigned long batch = ad->batch_expire[REQ_ASYNC]; + long write_time; + + write_time = (jiffies - ad->current_batch_expires) + batch; + if (write_time < 0) + write_time = 0; + + if (write_time > batch && !ad->write_batch_idled) { + if (write_time > batch * 3) + ad->write_batch_count /= 2; + else + ad->write_batch_count--; + } else if (write_time < batch && ad->current_write_count == 0) { + if (batch > write_time * 3) + ad->write_batch_count *= 2; + else + ad->write_batch_count++; + } + + if (ad->write_batch_count < 1) + ad->write_batch_count = 1; +} + +/* + * as_completed_request is to be called when a request has completed and + * returned something to the requesting process, be it an error or data. + */ +static void as_completed_request(request_queue_t *q, struct request *rq) +{ + struct as_data *ad = q->elevator.elevator_data; + struct as_rq *arq = RQ_DATA(rq); + struct as_io_context *aic = arq->as_io_context; + + if (unlikely(!blk_fs_request(rq))) { + WARN_ON(aic); + return; + } + + WARN_ON(blk_fs_request(rq) && arq->state == AS_RQ_NEW); + + if (arq->state != AS_RQ_DISPATCHED) + return; + + if (ad->changed_batch && ad->nr_dispatched == 1) { + kblockd_schedule_work(&ad->antic_work); + ad->changed_batch = 2; + } + ad->nr_dispatched--; + + /* + * Start counting the batch from when a request of that direction is + * actually serviced. This should help devices with big TCQ windows + * and writeback caches + */ + if (ad->batch_data_dir == REQ_SYNC && ad->changed_batch + && ad->batch_data_dir == arq->is_sync) { + update_write_batch(ad); + ad->current_batch_expires = jiffies + + ad->batch_expire[REQ_SYNC]; + ad->changed_batch = 0; + } + + if (!aic) + return; + + spin_lock(&aic->lock); + if (arq->is_sync == REQ_SYNC) { + set_bit(AS_TASK_IORUNNING, &aic->state); + aic->last_end_request = jiffies; + } + + if (ad->as_io_context == aic) { + ad->antic_start = jiffies; + ad->aic_finished = 1; + if (ad->antic_status == ANTIC_WAIT_REQ) { + /* + * We were waiting on this request, now anticipate + * the next one + */ + as_antic_waitnext(ad); + } + } + spin_unlock(&aic->lock); + put_as_io_context(&arq->as_io_context); +} + +/* + * as_remove_queued_request removes a request from the pre dispatch queue + * without updating refcounts. It is expected the caller will drop the + * reference unless it replaces the request at somepart of the elevator + * (ie. the dispatch queue) + */ +static void as_remove_queued_request(request_queue_t *q, struct request *rq) +{ + struct as_rq *arq = RQ_DATA(rq); + + if (!arq) + BUG(); + else { + const int data_dir = arq->is_sync; + struct as_data *ad = q->elevator.elevator_data; + + WARN_ON(arq->state != AS_RQ_QUEUED); + + if (arq->as_io_context) { + BUG_ON(!atomic_read(&arq->as_io_context->nr_queued)); + atomic_dec(&arq->as_io_context->nr_queued); + } + + /* + * Update the "next_arq" cache if we are about to remove its + * entry + */ + if (ad->next_arq[data_dir] == arq) + ad->next_arq[data_dir] = as_find_next_arq(ad, arq); + + list_del_init(&arq->fifo); + as_remove_merge_hints(q, arq); + as_del_arq_rb(ad, arq); + } + +} + +/* + * as_remove_dispatched_request is called to remove a request which has gone + * to the dispatch list. + */ +static void as_remove_dispatched_request(request_queue_t *q, struct request *rq) +{ + struct as_rq *arq = RQ_DATA(rq); + + if (arq) { + struct as_io_context *aic; + + WARN_ON(arq->state != AS_RQ_DISPATCHED); + WARN_ON(ON_RB(&arq->rb_node)); + aic = arq->as_io_context; + if (aic) { + WARN_ON(!atomic_read(&aic->nr_dispatched)); + atomic_dec(&aic->nr_dispatched); + } + } +} +/* + * as_remove_request is called when a driver has finished with a request. + * This should be only called for dispatched requests, but for some reason + * a POWER4 box running hwscan it does not. + */ +static void as_remove_request(request_queue_t *q, struct request *rq) +{ + struct as_rq *arq = RQ_DATA(rq); + + if (unlikely(!blk_fs_request(rq))) + return; + + if (arq) { + if (ON_RB(&arq->rb_node)) + as_remove_queued_request(q, rq); + else + as_remove_dispatched_request(q, rq); + } +} + +/* + * as_fifo_expired returns 0 if there are no expired reads on the fifo, + * 1 otherwise. It is ratelimited so that we only perform the check once per + * `fifo_expire' interval. Otherwise a large number of expired requests + * would create a hopeless seekstorm. + * + * See as_antic_expired comment. + */ +static int as_fifo_expired(struct as_data *ad, int adir) +{ + struct as_rq *arq; + long delta_jif; + + delta_jif = jiffies - ad->last_check_fifo[adir]; + if (unlikely(delta_jif < 0)) + delta_jif = -delta_jif; + if (delta_jif < ad->fifo_expire[adir]) + return 0; + + ad->last_check_fifo[adir] = jiffies; + + if (list_empty(&ad->fifo_list[adir])) + return 0; + + arq = list_entry_fifo(ad->fifo_list[adir].next); + + return time_after(jiffies, arq->expires); +} + +/* + * as_batch_expired returns true if the current batch has expired. A batch + * is a set of reads or a set of writes. + */ +static inline int as_batch_expired(struct as_data *ad) +{ + if (ad->changed_batch) + return 0; + + if (ad->batch_data_dir == REQ_SYNC) + /* TODO! add a check so a complete fifo gets written? */ + return time_after(jiffies, ad->current_batch_expires); + + return time_after(jiffies, ad->current_batch_expires) + || ad->current_write_count == 0; +} + +/* + * move an entry to dispatch queue + */ +static void as_move_to_dispatch(struct as_data *ad, struct as_rq *arq) +{ + const int data_dir = arq->is_sync; + + BUG_ON(!ON_RB(&arq->rb_node)); + + as_antic_stop(ad); + ad->antic_status = ANTIC_OFF; + + /* + * This has to be set in order to be correctly updated by + * as_find_next_arq + */ + ad->last_sector[data_dir] = arq->request->sector + + arq->request->nr_sectors; + + ad->nr_dispatched++; + + if (data_dir == REQ_SYNC) { + /* In case we have to anticipate after this */ + copy_as_io_context(&ad->as_io_context, &arq->as_io_context); + } else { + put_as_io_context(&ad->as_io_context); + if (ad->current_write_count != 0) + ad->current_write_count--; + } + ad->aic_finished = 0; + + ad->next_arq[data_dir] = as_find_next_arq(ad, arq); + + /* + * take it off the sort and fifo list, add to dispatch queue + */ + as_remove_queued_request(ad->q, arq->request); + list_add_tail(&arq->request->queuelist, ad->dispatch); + if (arq->as_io_context) + atomic_inc(&arq->as_io_context->nr_dispatched); + + WARN_ON(arq->state != AS_RQ_QUEUED); + arq->state = AS_RQ_DISPATCHED; +} + +/* + * as_dispatch_request selects the best request according to + * read/write expire, batch expire, etc, and moves it to the dispatch + * queue. Returns 1 if a request was found, 0 otherwise. + */ +static int as_dispatch_request(struct as_data *ad) +{ + struct as_rq *arq; + const int reads = !list_empty(&ad->fifo_list[REQ_SYNC]); + const int writes = !list_empty(&ad->fifo_list[REQ_ASYNC]); + + /* Signal that the write batch was uncontended, so we can't time it */ + if (ad->batch_data_dir == REQ_ASYNC && !reads) { + if (ad->current_write_count == 0 || !writes) + ad->write_batch_idled = 1; + } + + if (!(reads || writes) + || ad->antic_status == ANTIC_WAIT_REQ + || ad->antic_status == ANTIC_WAIT_NEXT + || ad->changed_batch == 1) + return 0; + + if (!(reads && writes && as_batch_expired(ad)) ) { + /* + * batch is still running or no reads or no writes + */ + arq = ad->next_arq[ad->batch_data_dir]; + + if (ad->batch_data_dir == REQ_SYNC && ad->antic_expire) { + if (as_fifo_expired(ad, REQ_SYNC)) + goto fifo_expired; + + if (as_can_anticipate(ad, arq)) { + as_antic_waitreq(ad); + return 0; + } + } + + if (arq) { + /* we have a "next request" */ + if (reads && !writes) + ad->current_batch_expires = + jiffies + ad->batch_expire[REQ_SYNC]; + goto dispatch_request; + } + } + + /* + * at this point we are not running a batch. select the appropriate + * data direction (read / write) + */ + + if (reads) { + BUG_ON(RB_EMPTY(&ad->sort_list[REQ_SYNC])); + + if (writes && ad->batch_data_dir == REQ_SYNC) + /* + * Last batch was a read, switch to writes + */ + goto dispatch_writes; + + if (ad->batch_data_dir == REQ_ASYNC) + ad->changed_batch = 1; + ad->batch_data_dir = REQ_SYNC; + arq = list_entry_fifo(ad->fifo_list[ad->batch_data_dir].next); + ad->last_check_fifo[ad->batch_data_dir] = jiffies; + goto dispatch_request; + } + + /* + * the last batch was a read + */ + + if (writes) { +dispatch_writes: + BUG_ON(RB_EMPTY(&ad->sort_list[REQ_ASYNC])); + + if (ad->batch_data_dir == REQ_SYNC) + ad->changed_batch = 1; + ad->batch_data_dir = REQ_ASYNC; + ad->current_write_count = ad->write_batch_count; + ad->write_batch_idled = 0; + arq = ad->next_arq[ad->batch_data_dir]; + goto dispatch_request; + } + + BUG(); + return 0; + +dispatch_request: + /* + * If a request has expired, service it. + */ + + if (as_fifo_expired(ad, ad->batch_data_dir)) { +fifo_expired: + arq = list_entry_fifo(ad->fifo_list[ad->batch_data_dir].next); + BUG_ON(arq == NULL); + } + + if (ad->changed_batch) { + if (ad->changed_batch == 1 && ad->nr_dispatched) + return 0; + if (ad->batch_data_dir == REQ_ASYNC) { + ad->current_batch_expires = jiffies + + ad->batch_expire[REQ_ASYNC]; + ad->changed_batch = 0; + } else + ad->changed_batch = 2; + arq->request->flags |= REQ_HARDBARRIER; + } + + /* + * arq is the selected appropriate request. + */ + as_move_to_dispatch(ad, arq); + + return 1; +} + +static struct request *as_next_request(request_queue_t *q) +{ + struct as_data *ad = q->elevator.elevator_data; + struct request *rq = NULL; + + /* + * if there are still requests on the dispatch queue, grab the first + */ + if (!list_empty(ad->dispatch) || as_dispatch_request(ad)) + rq = list_entry_rq(ad->dispatch->next); + + return rq; +} + +/* + * add arq to rbtree and fifo + */ +static void as_add_request(struct as_data *ad, struct as_rq *arq) +{ + int data_dir; + + if (rq_data_dir(arq->request) == READ + || current->flags&PF_SYNCWRITE) + arq->is_sync = 1; + else + arq->is_sync = 0; + data_dir = arq->is_sync; + + arq->as_io_context = get_as_io_context(); + + if (arq->as_io_context) { + atomic_inc(&arq->as_io_context->nr_queued); + as_update_iohist(arq->as_io_context, arq->request); + } + + as_add_arq_rb(ad, arq); + + /* + * set expire time (only used for reads) and add to fifo list + */ + arq->expires = jiffies + ad->fifo_expire[data_dir]; + list_add_tail(&arq->fifo, &ad->fifo_list[data_dir]); + arq->state = AS_RQ_QUEUED; + as_update_arq(ad, arq); /* keep state machine up to date */ +} + +static void +as_insert_request(request_queue_t *q, struct request *rq, + struct list_head *insert_here) +{ + struct as_data *ad = q->elevator.elevator_data; + struct as_rq *arq = RQ_DATA(rq); + + if (unlikely(rq->flags & REQ_HARDBARRIER)) { + AS_INVALIDATE_HASH(ad); + q->last_merge = NULL; + + while (ad->next_arq[REQ_SYNC]) + as_move_to_dispatch(ad, ad->next_arq[REQ_SYNC]); + + while (ad->next_arq[REQ_ASYNC]) + as_move_to_dispatch(ad, ad->next_arq[REQ_ASYNC]); + } + + if (unlikely(!blk_fs_request(rq))) { + if (!insert_here) + insert_here = ad->dispatch->prev; + + list_add(&rq->queuelist, insert_here); + + /* Stop anticipating - let this request get through */ + if (!list_empty(ad->dispatch) + && (ad->antic_status == ANTIC_WAIT_REQ + || ad->antic_status == ANTIC_WAIT_NEXT)) + as_antic_stop(ad); + + return; + } + + if (rq_mergeable(rq)) { + as_add_arq_hash(ad, arq); + + if (!q->last_merge) + q->last_merge = &rq->queuelist; + } + + as_add_request(ad, arq); +} + +/* + * as_queue_empty tells us if there are requests left in the device. It may + * not be the case that a driver can get the next request even if the queue + * is not empty - it is used in the block layer to check for plugging and + * merging opportunities + */ +static int as_queue_empty(request_queue_t *q) +{ + struct as_data *ad = q->elevator.elevator_data; + + if (!list_empty(&ad->fifo_list[REQ_ASYNC]) + || !list_empty(&ad->fifo_list[REQ_SYNC]) + || !list_empty(ad->dispatch)) + return 0; + + return 1; +} + +static struct request * +as_former_request(request_queue_t *q, struct request *rq) +{ + struct as_rq *arq = RQ_DATA(rq); + struct rb_node *rbprev = rb_prev(&arq->rb_node); + struct request *ret = NULL; + + if (rbprev) + ret = rb_entry_arq(rbprev)->request; + + return ret; +} + +static struct request * +as_latter_request(request_queue_t *q, struct request *rq) +{ + struct as_rq *arq = RQ_DATA(rq); + struct rb_node *rbnext = rb_next(&arq->rb_node); + struct request *ret = NULL; + + if (rbnext) + ret = rb_entry_arq(rbnext)->request; + + return ret; +} + +static int +as_merge(request_queue_t *q, struct list_head **insert, struct bio *bio) +{ + struct as_data *ad = q->elevator.elevator_data; + sector_t rb_key = bio->bi_sector + bio_sectors(bio); + struct request *__rq; + int ret; + + /* + * try last_merge to avoid going to hash + */ + ret = elv_try_last_merge(q, bio); + if (ret != ELEVATOR_NO_MERGE) { + __rq = list_entry_rq(q->last_merge); + goto out_insert; + } + + /* + * see if the merge hash can satisfy a back merge + */ + __rq = as_find_arq_hash(ad, bio->bi_sector); + if (__rq) { + BUG_ON(__rq->sector + __rq->nr_sectors != bio->bi_sector); + + if (elv_rq_merge_ok(__rq, bio)) { + ret = ELEVATOR_BACK_MERGE; + goto out; + } + } + + /* + * check for front merge + */ + __rq = as_find_arq_rb(ad, rb_key, bio_data_dir(bio)); + if (__rq) { + BUG_ON(rb_key != rq_rb_key(__rq)); + + if (elv_rq_merge_ok(__rq, bio)) { + ret = ELEVATOR_FRONT_MERGE; + goto out; + } + } + + return ELEVATOR_NO_MERGE; +out: + q->last_merge = &__rq->queuelist; +out_insert: + if (ret) + as_hot_arq_hash(ad, RQ_DATA(__rq)); + *insert = &__rq->queuelist; + return ret; +} + +static void as_merged_request(request_queue_t *q, struct request *req) +{ + struct as_data *ad = q->elevator.elevator_data; + struct as_rq *arq = RQ_DATA(req); + + /* + * hash always needs to be repositioned, key is end sector + */ + as_del_arq_hash(arq); + as_add_arq_hash(ad, arq); + + /* + * if the merge was a front merge, we need to reposition request + */ + if (rq_rb_key(req) != arq->rb_key) { + as_del_arq_rb(ad, arq); + as_add_arq_rb(ad, arq); + /* + * Note! At this stage of this and the next function, our next + * request may not be optimal - eg the request may have "grown" + * behind the disk head. We currently don't bother adjusting. + */ + } + + q->last_merge = &req->queuelist; +} + +static void +as_merged_requests(request_queue_t *q, struct request *req, + struct request *next) +{ + struct as_data *ad = q->elevator.elevator_data; + struct as_rq *arq = RQ_DATA(req); + struct as_rq *anext = RQ_DATA(next); + + BUG_ON(!arq); + BUG_ON(!anext); + + /* + * reposition arq (this is the merged request) in hash, and in rbtree + * in case of a front merge + */ + as_del_arq_hash(arq); + as_add_arq_hash(ad, arq); + + if (rq_rb_key(req) != arq->rb_key) { + as_del_arq_rb(ad, arq); + as_add_arq_rb(ad, arq); + } + + /* + * if anext expires before arq, assign its expire time to arq + * and move into anext position (anext will be deleted) in fifo + */ + if (!list_empty(&arq->fifo) && !list_empty(&anext->fifo)) { + if (time_before(anext->expires, arq->expires)) { + list_move(&arq->fifo, &anext->fifo); + arq->expires = anext->expires; + /* + * Don't copy here but swap, because when anext is + * removed below, it must contain the unused context + */ + swap_as_io_context(&arq->as_io_context, + &anext->as_io_context); + } + } + + /* + * kill knowledge of next, this one is a goner + */ + as_remove_queued_request(q, next); +/* put_as_io_context(&anext->as_io_context); no need to do this because as_completed_request takes care of it */ +} + +/* + * This is executed in a "deferred" process context, by kblockd. It calls the + * driver's request_fn so the driver can submit that request. + * + * IMPORTANT! This guy will reenter the elevator, so set up all queue global + * state before calling, and don't rely on any state over calls. + * + * FIXME! dispatch queue is not a queue at all! + */ +static void as_work_handler(void *data) +{ + struct request_queue *q = data; + unsigned long flags; + + spin_lock_irqsave(q->queue_lock, flags); + if (as_next_request(q)) + q->request_fn(q); + spin_unlock_irqrestore(q->queue_lock, flags); +} + +static void as_put_request(request_queue_t *q, struct request *rq) +{ + struct as_data *ad = q->elevator.elevator_data; + struct as_rq *arq = RQ_DATA(rq); + + if (arq) { + mempool_free(arq, ad->arq_pool); + rq->elevator_private = NULL; + } +} + +static int as_set_request(request_queue_t *q, struct request *rq, int gfp_mask) +{ + struct as_data *ad = q->elevator.elevator_data; + struct as_rq *arq = mempool_alloc(ad->arq_pool, gfp_mask); + + if (arq) { + RB_CLEAR(&arq->rb_node); + arq->request = rq; + arq->state = AS_RQ_NEW; + arq->as_io_context = NULL; + INIT_LIST_HEAD(&arq->hash); + arq->hash_valid_count = 0; + INIT_LIST_HEAD(&arq->fifo); + rq->elevator_private = arq; + return 0; + } + + return 1; +} + +static void as_exit(request_queue_t *q, elevator_t *e) +{ + struct as_data *ad = e->elevator_data; + + del_timer_sync(&ad->antic_timer); + kblockd_flush(); + + BUG_ON(!list_empty(&ad->fifo_list[REQ_SYNC])); + BUG_ON(!list_empty(&ad->fifo_list[REQ_ASYNC])); + + mempool_destroy(ad->arq_pool); + put_as_io_context(&ad->as_io_context); + kfree(ad->hash); + kfree(ad); +} + +/* + * initialize elevator private data (as_data), and alloc a arq for + * each request on the free lists + */ +static int as_init(request_queue_t *q, elevator_t *e) +{ + struct as_data *ad; + int i; + + if (!arq_pool) + return -ENOMEM; + + ad = kmalloc(sizeof(*ad), GFP_KERNEL); + if (!ad) + return -ENOMEM; + memset(ad, 0, sizeof(*ad)); + + ad->q = q; /* Identify what queue the data belongs to */ + + ad->hash = kmalloc(sizeof(struct list_head)*AS_HASH_ENTRIES,GFP_KERNEL); + if (!ad->hash) { + kfree(ad); + return -ENOMEM; + } + + ad->arq_pool = mempool_create(BLKDEV_MIN_RQ, mempool_alloc_slab, mempool_free_slab, arq_pool); + if (!ad->arq_pool) { + kfree(ad->hash); + kfree(ad); + return -ENOMEM; + } + + /* anticipatory scheduling helpers */ + ad->antic_timer.function = as_antic_timeout; + ad->antic_timer.data = (unsigned long)q; + init_timer(&ad->antic_timer); + INIT_WORK(&ad->antic_work, as_work_handler, q); + + for (i = 0; i < AS_HASH_ENTRIES; i++) + INIT_LIST_HEAD(&ad->hash[i]); + + INIT_LIST_HEAD(&ad->fifo_list[REQ_SYNC]); + INIT_LIST_HEAD(&ad->fifo_list[REQ_ASYNC]); + ad->sort_list[REQ_SYNC] = RB_ROOT; + ad->sort_list[REQ_ASYNC] = RB_ROOT; + ad->dispatch = &q->queue_head; + ad->fifo_expire[REQ_SYNC] = default_read_expire; + ad->fifo_expire[REQ_ASYNC] = default_write_expire; + ad->hash_valid_count = 1; + ad->antic_expire = default_antic_expire; + ad->batch_expire[REQ_SYNC] = default_read_batch_expire; + ad->batch_expire[REQ_ASYNC] = default_write_batch_expire; + e->elevator_data = ad; + + ad->current_batch_expires = jiffies + ad->batch_expire[REQ_SYNC]; + ad->write_batch_count = ad->batch_expire[REQ_ASYNC] / 10; + if (ad->write_batch_count < 2) + ad->write_batch_count = 2; + return 0; +} + +/* + * sysfs parts below + */ +struct as_fs_entry { + struct attribute attr; + ssize_t (*show)(struct as_data *, char *); + ssize_t (*store)(struct as_data *, const char *, size_t); +}; + +static ssize_t +as_var_show(unsigned int var, char *page) +{ + var = (var * 1000) / HZ; + return sprintf(page, "%d\n", var); +} + +static ssize_t +as_var_store(unsigned long *var, const char *page, size_t count) +{ + unsigned long tmp; + char *p = (char *) page; + + tmp = simple_strtoul(p, &p, 10); + if (tmp != 0) { + tmp = (tmp * HZ) / 1000; + if (tmp == 0) + tmp = 1; + } + *var = tmp; + return count; +} + +#define SHOW_FUNCTION(__FUNC, __VAR) \ +static ssize_t __FUNC(struct as_data *ad, char *page) \ +{ \ + return as_var_show(__VAR, (page)); \ +} +SHOW_FUNCTION(as_readexpire_show, ad->fifo_expire[REQ_SYNC]); +SHOW_FUNCTION(as_writeexpire_show, ad->fifo_expire[REQ_ASYNC]); +SHOW_FUNCTION(as_anticexpire_show, ad->antic_expire); +SHOW_FUNCTION(as_read_batchexpire_show, ad->batch_expire[REQ_SYNC]); +SHOW_FUNCTION(as_write_batchexpire_show, ad->batch_expire[REQ_ASYNC]); +#undef SHOW_FUNCTION + +#define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX) \ +static ssize_t __FUNC(struct as_data *ad, const char *page, size_t count) \ +{ \ + int ret = as_var_store(__PTR, (page), count); \ + if (*(__PTR) < (MIN)) \ + *(__PTR) = (MIN); \ + else if (*(__PTR) > (MAX)) \ + *(__PTR) = (MAX); \ + return ret; \ +} +STORE_FUNCTION(as_readexpire_store, &ad->fifo_expire[REQ_SYNC], 0, INT_MAX); +STORE_FUNCTION(as_writeexpire_store, &ad->fifo_expire[REQ_ASYNC], 0, INT_MAX); +STORE_FUNCTION(as_anticexpire_store, &ad->antic_expire, 0, INT_MAX); +STORE_FUNCTION(as_read_batchexpire_store, + &ad->batch_expire[REQ_SYNC], 0, INT_MAX); +STORE_FUNCTION(as_write_batchexpire_store, + &ad->batch_expire[REQ_ASYNC], 0, INT_MAX); +#undef STORE_FUNCTION + +static struct as_fs_entry as_readexpire_entry = { + .attr = {.name = "read_expire", .mode = S_IRUGO | S_IWUSR }, + .show = as_readexpire_show, + .store = as_readexpire_store, +}; +static struct as_fs_entry as_writeexpire_entry = { + .attr = {.name = "write_expire", .mode = S_IRUGO | S_IWUSR }, + .show = as_writeexpire_show, + .store = as_writeexpire_store, +}; +static struct as_fs_entry as_anticexpire_entry = { + .attr = {.name = "antic_expire", .mode = S_IRUGO | S_IWUSR }, + .show = as_anticexpire_show, + .store = as_anticexpire_store, +}; +static struct as_fs_entry as_read_batchexpire_entry = { + .attr = {.name = "read_batch_expire", .mode = S_IRUGO | S_IWUSR }, + .show = as_read_batchexpire_show, + .store = as_read_batchexpire_store, +}; +static struct as_fs_entry as_write_batchexpire_entry = { + .attr = {.name = "write_batch_expire", .mode = S_IRUGO | S_IWUSR }, + .show = as_write_batchexpire_show, + .store = as_write_batchexpire_store, +}; + +static struct attribute *default_attrs[] = { + &as_readexpire_entry.attr, + &as_writeexpire_entry.attr, + &as_anticexpire_entry.attr, + &as_read_batchexpire_entry.attr, + &as_write_batchexpire_entry.attr, + NULL, +}; + +#define to_as(atr) container_of((atr), struct as_fs_entry, attr) + +static ssize_t +as_attr_show(struct kobject *kobj, struct attribute *attr, char *page) +{ + elevator_t *e = container_of(kobj, elevator_t, kobj); + struct as_fs_entry *entry = to_as(attr); + + if (!entry->show) + return 0; + + return entry->show(e->elevator_data, page); +} + +static ssize_t +as_attr_store(struct kobject *kobj, struct attribute *attr, + const char *page, size_t length) +{ + elevator_t *e = container_of(kobj, elevator_t, kobj); + struct as_fs_entry *entry = to_as(attr); + + if (!entry->store) + return -EINVAL; + + return entry->store(e->elevator_data, page, length); +} + +static struct sysfs_ops as_sysfs_ops = { + .show = as_attr_show, + .store = as_attr_store, +}; + +struct kobj_type as_ktype = { + .sysfs_ops = &as_sysfs_ops, + .default_attrs = default_attrs, +}; + +static int __init as_slab_setup(void) +{ + arq_pool = kmem_cache_create("as_arq", sizeof(struct as_rq), + 0, 0, NULL, NULL); + + if (!arq_pool) + panic("as: can't init slab pool\n"); + + return 0; +} + +subsys_initcall(as_slab_setup); + +elevator_t iosched_as = { + .elevator_merge_fn = as_merge, + .elevator_merged_fn = as_merged_request, + .elevator_merge_req_fn = as_merged_requests, + .elevator_next_req_fn = as_next_request, + .elevator_add_req_fn = as_insert_request, + .elevator_remove_req_fn = as_remove_request, + .elevator_queue_empty_fn = as_queue_empty, + .elevator_completed_req_fn = as_completed_request, + .elevator_former_req_fn = as_former_request, + .elevator_latter_req_fn = as_latter_request, + .elevator_set_req_fn = as_set_request, + .elevator_put_req_fn = as_put_request, + .elevator_init_fn = as_init, + .elevator_exit_fn = as_exit, + + .elevator_ktype = &as_ktype, +}; + +EXPORT_SYMBOL(iosched_as); diff -puN drivers/block/ll_rw_blk.c~as-iosched drivers/block/ll_rw_blk.c --- 25/drivers/block/ll_rw_blk.c~as-iosched 2003-06-21 20:47:15.000000000 -0700 +++ 25-akpm/drivers/block/ll_rw_blk.c 2003-06-21 20:47:15.000000000 -0700 @@ -1033,7 +1033,7 @@ static inline void __generic_unplug_devi /* * was plugged, fire request_fn if queue has stuff to do */ - if (!elv_queue_empty(q)) + if (elv_next_request(q)) q->request_fn(q); } @@ -1204,6 +1204,18 @@ static int blk_init_free_list(request_qu static int __make_request(request_queue_t *, struct bio *); +static elevator_t *chosen_elevator = &iosched_as; + +static int __init elevator_setup(char *str) +{ + if (!strcmp(str, "deadline")) { + chosen_elevator = &iosched_deadline; + printk("elevator: cfq\n"); + } + return 1; +} +__setup("elevator=", elevator_setup); + /** * blk_init_queue - prepare a request queue for use with a block device * @q: The &request_queue_t to be initialised @@ -1235,11 +1247,20 @@ static int __make_request(request_queue_ int blk_init_queue(request_queue_t *q, request_fn_proc *rfn, spinlock_t *lock) { int ret; + static int printed; if (blk_init_free_list(q)) return -ENOMEM; - if ((ret = elevator_init(q, &iosched_deadline))) { + if (!printed) { + printed = 1; + if (chosen_elevator == &iosched_deadline) + printk("deadline elevator\n"); + else if (chosen_elevator == &iosched_as) + printk("anticipatory scheduling elevator\n"); + } + + if ((ret = elevator_init(q, chosen_elevator))) { blk_cleanup_queue(q); return ret; } diff -puN drivers/block/Makefile~as-iosched drivers/block/Makefile --- 25/drivers/block/Makefile~as-iosched 2003-06-21 20:47:15.000000000 -0700 +++ 25-akpm/drivers/block/Makefile 2003-06-21 20:47:15.000000000 -0700 @@ -13,7 +13,8 @@ # kblockd threads # -obj-y := elevator.o ll_rw_blk.o ioctl.o genhd.o scsi_ioctl.o deadline-iosched.o +obj-y := elevator.o ll_rw_blk.o ioctl.o genhd.o scsi_ioctl.o \ + deadline-iosched.o as-iosched.o obj-$(CONFIG_MAC_FLOPPY) += swim3.o obj-$(CONFIG_BLK_DEV_FD) += floppy.o diff -puN include/linux/elevator.h~as-iosched include/linux/elevator.h --- 25/include/linux/elevator.h~as-iosched 2003-06-21 20:47:15.000000000 -0700 +++ 25-akpm/include/linux/elevator.h 2003-06-21 20:47:15.000000000 -0700 @@ -89,6 +89,11 @@ extern elevator_t elevator_noop; */ extern elevator_t iosched_deadline; +/* + * anticipatory I/O scheduler + */ +extern elevator_t iosched_as; + extern int elevator_init(request_queue_t *, elevator_t *); extern void elevator_exit(request_queue_t *); extern inline int elv_rq_merge_ok(struct request *, struct bio *); diff -puN include/linux/sched.h~as-iosched include/linux/sched.h --- 25/include/linux/sched.h~as-iosched 2003-06-21 20:47:15.000000000 -0700 +++ 25-akpm/include/linux/sched.h 2003-06-21 20:47:15.000000000 -0700 @@ -321,6 +321,8 @@ struct k_itimer { }; +struct as_io_context; /* Anticipatory scheduler */ +void exit_as_io_context(void); struct task_struct { volatile long state; /* -1 unrunnable, 0 runnable, >0 stopped */ @@ -450,6 +452,8 @@ struct task_struct { struct dentry *proc_dentry; struct backing_dev_info *backing_dev_info; + struct as_io_context *as_io_context; + unsigned long ptrace_message; siginfo_t *last_siginfo; /* For ptrace use. */ }; @@ -481,6 +485,7 @@ do { if (atomic_dec_and_test(&(tsk)->usa #define PF_KSWAPD 0x00040000 /* I am kswapd */ #define PF_SWAPOFF 0x00080000 /* I am in swapoff */ #define PF_LESS_THROTTLE 0x01000000 /* Throttle me less: I clena memory */ +#define PF_SYNCWRITE 0x00200000 /* I am doing a sync write */ #ifdef CONFIG_SMP extern int set_cpus_allowed(task_t *p, unsigned long new_mask); diff -puN kernel/exit.c~as-iosched kernel/exit.c --- 25/kernel/exit.c~as-iosched 2003-06-21 20:47:15.000000000 -0700 +++ 25-akpm/kernel/exit.c 2003-06-21 20:47:15.000000000 -0700 @@ -680,6 +680,8 @@ NORET_TYPE void do_exit(long code) panic("Attempted to kill the idle task!"); if (unlikely(tsk->pid == 1)) panic("Attempted to kill init!"); + if (tsk->as_io_context) + exit_as_io_context(); tsk->flags |= PF_EXITING; del_timer_sync(&tsk->real_timer); diff -puN kernel/fork.c~as-iosched kernel/fork.c --- 25/kernel/fork.c~as-iosched 2003-06-21 20:47:15.000000000 -0700 +++ 25-akpm/kernel/fork.c 2003-06-21 20:47:15.000000000 -0700 @@ -864,6 +864,7 @@ struct task_struct *copy_process(unsigne p->lock_depth = -1; /* -1 = no lock */ p->start_time = get_jiffies_64(); p->security = NULL; + p->as_io_context = NULL; retval = -ENOMEM; if ((retval = security_task_alloc(p))) diff -puN fs/buffer.c~as-iosched fs/buffer.c --- 25/fs/buffer.c~as-iosched 2003-06-21 20:47:15.000000000 -0700 +++ 25-akpm/fs/buffer.c 2003-06-21 20:47:15.000000000 -0700 @@ -319,6 +319,7 @@ asmlinkage long sys_fsync(unsigned int f /* We need to protect against concurrent writers.. */ down(&inode->i_sem); + current->flags |= PF_SYNCWRITE; ret = filemap_fdatawrite(inode->i_mapping); err = file->f_op->fsync(file, dentry, 0); if (!ret) @@ -326,6 +327,7 @@ asmlinkage long sys_fsync(unsigned int f err = filemap_fdatawait(inode->i_mapping); if (!ret) ret = err; + current->flags &= ~PF_SYNCWRITE; up(&inode->i_sem); out_putf: @@ -354,6 +356,7 @@ asmlinkage long sys_fdatasync(unsigned i goto out_putf; down(&inode->i_sem); + current->flags |= PF_SYNCWRITE; ret = filemap_fdatawrite(inode->i_mapping); err = file->f_op->fsync(file, dentry, 1); if (!ret) @@ -361,6 +364,7 @@ asmlinkage long sys_fdatasync(unsigned i err = filemap_fdatawait(inode->i_mapping); if (!ret) ret = err; + current->flags &= ~PF_SYNCWRITE; up(&inode->i_sem); out_putf: diff -puN fs/fs-writeback.c~as-iosched fs/fs-writeback.c --- 25/fs/fs-writeback.c~as-iosched 2003-06-21 20:47:15.000000000 -0700 +++ 25-akpm/fs/fs-writeback.c 2003-06-21 20:47:15.000000000 -0700 @@ -516,6 +516,7 @@ int generic_osync_inode(struct inode *in int need_write_inode_now = 0; int err2; + current->flags |= PF_SYNCWRITE; if (what & OSYNC_DATA) err = filemap_fdatawrite(inode->i_mapping); if (what & (OSYNC_METADATA|OSYNC_DATA)) { @@ -528,6 +529,7 @@ int generic_osync_inode(struct inode *in if (!err) err = err2; } + current->flags &= ~PF_SYNCWRITE; spin_lock(&inode_lock); if ((inode->i_state & I_DIRTY) && _