block/blk-ioc.c

   1 /*
   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 */
  10 #include <linux/slab.h>
  11
  12 #include "blk.h"
  13
  14 /*
  15  * For io context allocations
  16  */
  17 static struct kmem_cache *iocontext_cachep;
  18
  19 static void cfq_dtor(struct io_context *ioc)
  20 {
  21         if (!hlist_empty(&ioc->cic_list)) {
  22                 struct cfq_io_context *cic;
  23
  24                 cic = hlist_entry(ioc->cic_list.first, struct cfq_io_context,
  25                                                                 cic_list);
  26                 cic->dtor(ioc);
  27         }
  28 }
  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  */
  34 int put_io_context(struct io_context *ioc)
  35 {
  36         if (ioc == NULL)
  37                 return 1;
  38
  39         BUG_ON(atomic_long_read(&ioc->refcount) == 0);
  40
  41         if (atomic_long_dec_and_test(&ioc->refcount)) {
  42                 rcu_read_lock();
  43                 cfq_dtor(ioc);
  44                 rcu_read_unlock();
  45
  46                 kmem_cache_free(iocontext_cachep, ioc);
  47                 return 1;
  48         }
  49         return 0;
  50 }
  51 EXPORT_SYMBOL(put_io_context);
  52
  53 static void cfq_exit(struct io_context *ioc)
  54 {
  55         rcu_read_lock();
  56
  57         if (!hlist_empty(&ioc->cic_list)) {
  58                 struct cfq_io_context *cic;
  59
  60                 cic = hlist_entry(ioc->cic_list.first, struct cfq_io_context,
  61                                                                 cic_list);
  62                 cic->exit(ioc);
  63         }
  64         rcu_read_unlock();
  65 }
  66
  67 /* Called by the exiting task */
  68 void exit_io_context(struct task_struct *task)
  69 {
  70         struct io_context *ioc;
  71
  72         task_lock(task);
  73         ioc = task->io_context;
  74         task->io_context = NULL;
  75         task_unlock(task);
  76
  77         if (atomic_dec_and_test(&ioc->nr_tasks))
  78                 cfq_exit(ioc);
  79
  80         put_io_context(ioc);
  81 }
  82
  83 struct io_context *alloc_io_context(gfp_t gfp_flags, int node)
  84 {
  85         struct io_context *ret;
  86
  87         ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
  88         if (ret) {
  89                 atomic_long_set(&ret->refcount, 1);
  90                 atomic_set(&ret->nr_tasks, 1);
  91                 spin_lock_init(&ret->lock);
  92                 ret->ioprio_changed = 0;
  93                 ret->ioprio = 0;
  94                 ret->last_waited = 0; /* doesn't matter... */
  95                 ret->nr_batch_requests = 0; /* because this is 0 */
  96                 INIT_RADIX_TREE(&ret->radix_root, GFP_ATOMIC | __GFP_HIGH);
  97                 INIT_HLIST_HEAD(&ret->cic_list);
  98                 ret->ioc_data = NULL;
  99 #if defined(CONFIG_BLK_CGROUP) || defined(CONFIG_BLK_CGROUP_MODULE)
 100                 ret->cgroup_changed = 0;
 101 #endif
 102         }
 103
 104         return ret;
 105 }
 106
 107 /*
 108  * If the current task has no IO context then create one and initialise it.
 109  * Otherwise, return its existing IO context.
 110  *
 111  * This returned IO context doesn't have a specifically elevated refcount,
 112  * but since the current task itself holds a reference, the context can be
 113  * used in general code, so long as it stays within `current` context.
 114  */
 115 struct io_context *current_io_context(gfp_t gfp_flags, int node)
 116 {
 117         struct task_struct *tsk = current;
 118         struct io_context *ret;
 119
 120         ret = tsk->io_context;
 121         if (likely(ret))
 122                 return ret;
 123
 124         ret = alloc_io_context(gfp_flags, node);
 125         if (ret) {
 126                 /* make sure set_task_ioprio() sees the settings above */
 127                 smp_wmb();
 128                 tsk->io_context = ret;
 129         }
 130
 131         return ret;
 132 }
 133
 134 /*
 135  * If the current task has no IO context then create one and initialise it.
 136  * If it does have a context, take a ref on it.
 137  *
 138  * This is always called in the context of the task which submitted the I/O.
 139  */
 140 struct io_context *get_io_context(gfp_t gfp_flags, int node)
 141 {
 142         struct io_context *ret = NULL;
 143
 144         /*
 145          * Check for unlikely race with exiting task. ioc ref count is
 146          * zero when ioc is being detached.
 147          */
 148         do {
 149                 ret = current_io_context(gfp_flags, node);
 150                 if (unlikely(!ret))
 151                         break;
 152         } while (!atomic_long_inc_not_zero(&ret->refcount));
 153
 154         return ret;
 155 }
 156 EXPORT_SYMBOL(get_io_context);
 157
 158 static int __init blk_ioc_init(void)
 159 {
 160         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 161                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 162         return 0;
 163 }
 164 subsys_initcall(blk_ioc_init);