blob: 3c7a339fe3813483155f31b05ea5aa5cd3224a36 [file] [log] [blame]
Jens Axboe86db1e22008-01-29 14:53:40 +01001/*
2 * Functions related to io context handling
3 */
4#include <linux/kernel.h>
5#include <linux/module.h>
6#include <linux/init.h>
7#include <linux/bio.h>
8#include <linux/blkdev.h>
9#include <linux/bootmem.h> /* for max_pfn/max_low_pfn */
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090010#include <linux/slab.h>
Jens Axboe86db1e22008-01-29 14:53:40 +010011
12#include "blk.h"
13
14/*
15 * For io context allocations
16 */
17static struct kmem_cache *iocontext_cachep;
18
19static void cfq_dtor(struct io_context *ioc)
20{
Jens Axboeffc4e752008-02-19 10:02:29 +010021 if (!hlist_empty(&ioc->cic_list)) {
22 struct cfq_io_context *cic;
Jens Axboe86db1e22008-01-29 14:53:40 +010023
Jens Axboeffc4e752008-02-19 10:02:29 +010024 cic = list_entry(ioc->cic_list.first, struct cfq_io_context,
25 cic_list);
26 cic->dtor(ioc);
27 }
Jens Axboe86db1e22008-01-29 14:53:40 +010028}
29
30/*
31 * IO Context helper functions. put_io_context() returns 1 if there are no
32 * more users of this io context, 0 otherwise.
33 */
34int put_io_context(struct io_context *ioc)
35{
36 if (ioc == NULL)
37 return 1;
38
Nikanth Karthikesand9c7d392009-06-10 12:57:06 -070039 BUG_ON(atomic_long_read(&ioc->refcount) == 0);
Jens Axboe86db1e22008-01-29 14:53:40 +010040
Nikanth Karthikesand9c7d392009-06-10 12:57:06 -070041 if (atomic_long_dec_and_test(&ioc->refcount)) {
Jens Axboe86db1e22008-01-29 14:53:40 +010042 rcu_read_lock();
Jens Axboe86db1e22008-01-29 14:53:40 +010043 cfq_dtor(ioc);
Jens Axboe07416d22008-05-07 09:17:12 +020044 rcu_read_unlock();
Jens Axboe86db1e22008-01-29 14:53:40 +010045
46 kmem_cache_free(iocontext_cachep, ioc);
47 return 1;
48 }
49 return 0;
50}
51EXPORT_SYMBOL(put_io_context);
52
53static void cfq_exit(struct io_context *ioc)
54{
Jens Axboe86db1e22008-01-29 14:53:40 +010055 rcu_read_lock();
Jens Axboe86db1e22008-01-29 14:53:40 +010056
Jens Axboeffc4e752008-02-19 10:02:29 +010057 if (!hlist_empty(&ioc->cic_list)) {
58 struct cfq_io_context *cic;
59
60 cic = list_entry(ioc->cic_list.first, struct cfq_io_context,
61 cic_list);
62 cic->exit(ioc);
63 }
64 rcu_read_unlock();
Jens Axboe86db1e22008-01-29 14:53:40 +010065}
66
67/* Called by the exitting task */
Louis Rillingb69f2292009-12-04 14:52:42 +010068void exit_io_context(struct task_struct *task)
Jens Axboe86db1e22008-01-29 14:53:40 +010069{
70 struct io_context *ioc;
71
Louis Rillingb69f2292009-12-04 14:52:42 +010072 task_lock(task);
73 ioc = task->io_context;
74 task->io_context = NULL;
75 task_unlock(task);
Jens Axboe86db1e22008-01-29 14:53:40 +010076
77 if (atomic_dec_and_test(&ioc->nr_tasks)) {
Jens Axboe86db1e22008-01-29 14:53:40 +010078 cfq_exit(ioc);
79
Jens Axboe86db1e22008-01-29 14:53:40 +010080 }
Louis Rilling61cc74f2009-12-04 14:52:41 +010081 put_io_context(ioc);
Jens Axboe86db1e22008-01-29 14:53:40 +010082}
83
84struct io_context *alloc_io_context(gfp_t gfp_flags, int node)
85{
86 struct io_context *ret;
87
88 ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
89 if (ret) {
Nikanth Karthikesand9c7d392009-06-10 12:57:06 -070090 atomic_long_set(&ret->refcount, 1);
Jens Axboe86db1e22008-01-29 14:53:40 +010091 atomic_set(&ret->nr_tasks, 1);
92 spin_lock_init(&ret->lock);
93 ret->ioprio_changed = 0;
94 ret->ioprio = 0;
Richard Kennedy4671a132010-03-01 10:57:22 +010095 ret->last_waited = 0; /* doesn't matter... */
Jens Axboe86db1e22008-01-29 14:53:40 +010096 ret->nr_batch_requests = 0; /* because this is 0 */
Jens Axboe86db1e22008-01-29 14:53:40 +010097 INIT_RADIX_TREE(&ret->radix_root, GFP_ATOMIC | __GFP_HIGH);
Jens Axboeffc4e752008-02-19 10:02:29 +010098 INIT_HLIST_HEAD(&ret->cic_list);
Jens Axboe86db1e22008-01-29 14:53:40 +010099 ret->ioc_data = NULL;
100 }
101
102 return ret;
103}
104
105/*
106 * If the current task has no IO context then create one and initialise it.
107 * Otherwise, return its existing IO context.
108 *
109 * This returned IO context doesn't have a specifically elevated refcount,
110 * but since the current task itself holds a reference, the context can be
111 * used in general code, so long as it stays within `current` context.
112 */
113struct io_context *current_io_context(gfp_t gfp_flags, int node)
114{
115 struct task_struct *tsk = current;
116 struct io_context *ret;
117
118 ret = tsk->io_context;
119 if (likely(ret))
120 return ret;
121
122 ret = alloc_io_context(gfp_flags, node);
123 if (ret) {
124 /* make sure set_task_ioprio() sees the settings above */
125 smp_wmb();
126 tsk->io_context = ret;
127 }
128
129 return ret;
130}
131
132/*
133 * If the current task has no IO context then create one and initialise it.
134 * If it does have a context, take a ref on it.
135 *
136 * This is always called in the context of the task which submitted the I/O.
137 */
138struct io_context *get_io_context(gfp_t gfp_flags, int node)
139{
140 struct io_context *ret = NULL;
141
142 /*
143 * Check for unlikely race with exiting task. ioc ref count is
144 * zero when ioc is being detached.
145 */
146 do {
147 ret = current_io_context(gfp_flags, node);
148 if (unlikely(!ret))
149 break;
Nikanth Karthikesand9c7d392009-06-10 12:57:06 -0700150 } while (!atomic_long_inc_not_zero(&ret->refcount));
Jens Axboe86db1e22008-01-29 14:53:40 +0100151
152 return ret;
153}
154EXPORT_SYMBOL(get_io_context);
155
Adrian Bunk13341592008-02-18 13:45:53 +0100156static int __init blk_ioc_init(void)
Jens Axboe86db1e22008-01-29 14:53:40 +0100157{
158 iocontext_cachep = kmem_cache_create("blkdev_ioc",
159 sizeof(struct io_context), 0, SLAB_PANIC, NULL);
160 return 0;
161}
162subsys_initcall(blk_ioc_init);