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
  11 #include "blk.h"
  12
  13 /*
  14  * For io context allocations
  15  */
  16 static struct kmem_cache *iocontext_cachep;
  17
  18 static void cfq_dtor(struct io_context *ioc)
  19 {
  20         if (!hlist_empty(&ioc->cic_list)) {
  21                 struct cfq_io_context *cic;
  22
  23                 cic = list_entry(ioc->cic_list.first, struct cfq_io_context,
  24                                                                 cic_list);
  25                 cic->dtor(ioc);
  26         }
  27 }
  28
  29 /*
  30  * IO Context helper functions. put_io_context() returns 1 if there are no
  31  * more users of this io context, 0 otherwise.
  32  */
  33 int put_io_context(struct io_context *ioc)
  34 {
  35         if (ioc == NULL)
  36                 return 1;
  37
  38         BUG_ON(atomic_long_read(&ioc->refcount) == 0);
  39
  40         if (atomic_long_dec_and_test(&ioc->refcount)) {
  41                 rcu_read_lock();
  42                 if (ioc->aic && ioc->aic->dtor)
  43                         ioc->aic->dtor(ioc->aic);
  44                 cfq_dtor(ioc);
  45                 rcu_read_unlock();
  46
  47                 kmem_cache_free(iocontext_cachep, ioc);
  48                 return 1;
  49         }
  50         return 0;
  51 }
  52 EXPORT_SYMBOL(put_io_context);
  53
  54 static void cfq_exit(struct io_context *ioc)
  55 {
  56         rcu_read_lock();
  57
  58         if (!hlist_empty(&ioc->cic_list)) {
  59                 struct cfq_io_context *cic;
  60
  61                 cic = list_entry(ioc->cic_list.first, struct cfq_io_context,
  62                                                                 cic_list);
  63                 cic->exit(ioc);
  64         }
  65         rcu_read_unlock();
  66 }
  67
  68 /* Called by the exitting task */
  69 void exit_io_context(void)
  70 {
  71         struct io_context *ioc;
  72
  73         task_lock(current);
  74         ioc = current->io_context;
  75         current->io_context = NULL;
  76         task_unlock(current);
  77
  78         if (atomic_dec_and_test(&ioc->nr_tasks)) {
  79                 if (ioc->aic && ioc->aic->exit)
  80                         ioc->aic->exit(ioc->aic);
  81                 cfq_exit(ioc);
  82
  83                 put_io_context(ioc);
  84         }
  85 }
  86
  87 struct io_context *alloc_io_context(gfp_t gfp_flags, int node)
  88 {
  89         struct io_context *ret;
  90
  91         ret = kmem_cache_alloc_node(iocontext_cachep, gfp_flags, node);
  92         if (ret) {
  93                 atomic_long_set(&ret->refcount, 1);
  94                 atomic_set(&ret->nr_tasks, 1);
  95                 spin_lock_init(&ret->lock);
  96                 ret->ioprio_changed = 0;
  97                 ret->ioprio = 0;
  98                 ret->last_waited = jiffies; /* doesn't matter... */
  99                 ret->nr_batch_requests = 0; /* because this is 0 */
 100                 ret->aic = NULL;
 101                 INIT_RADIX_TREE(&ret->radix_root, GFP_ATOMIC | __GFP_HIGH);
 102                 INIT_HLIST_HEAD(&ret->cic_list);
 103                 ret->ioc_data = NULL;
 104         }
 105
 106         return ret;
 107 }
 108
 109 /*
 110  * If the current task has no IO context then create one and initialise it.
 111  * Otherwise, return its existing IO context.
 112  *
 113  * This returned IO context doesn't have a specifically elevated refcount,
 114  * but since the current task itself holds a reference, the context can be
 115  * used in general code, so long as it stays within `current` context.
 116  */
 117 struct io_context *current_io_context(gfp_t gfp_flags, int node)
 118 {
 119         struct task_struct *tsk = current;
 120         struct io_context *ret;
 121
 122         ret = tsk->io_context;
 123         if (likely(ret))
 124                 return ret;
 125
 126         ret = alloc_io_context(gfp_flags, node);
 127         if (ret) {
 128                 /* make sure set_task_ioprio() sees the settings above */
 129                 smp_wmb();
 130                 tsk->io_context = ret;
 131         }
 132
 133         return ret;
 134 }
 135
 136 /*
 137  * If the current task has no IO context then create one and initialise it.
 138  * If it does have a context, take a ref on it.
 139  *
 140  * This is always called in the context of the task which submitted the I/O.
 141  */
 142 struct io_context *get_io_context(gfp_t gfp_flags, int node)
 143 {
 144         struct io_context *ret = NULL;
 145
 146         /*
 147          * Check for unlikely race with exiting task. ioc ref count is
 148          * zero when ioc is being detached.
 149          */
 150         do {
 151                 ret = current_io_context(gfp_flags, node);
 152                 if (unlikely(!ret))
 153                         break;
 154         } while (!atomic_long_inc_not_zero(&ret->refcount));
 155
 156         return ret;
 157 }
 158 EXPORT_SYMBOL(get_io_context);
 159
 160 void copy_io_context(struct io_context **pdst, struct io_context **psrc)
 161 {
 162         struct io_context *src = *psrc;
 163         struct io_context *dst = *pdst;
 164
 165         if (src) {
 166                 BUG_ON(atomic_long_read(&src->refcount) == 0);
 167                 atomic_long_inc(&src->refcount);
 168                 put_io_context(dst);
 169                 *pdst = src;
 170         }
 171 }
 172 EXPORT_SYMBOL(copy_io_context);
 173
 174 static int __init blk_ioc_init(void)
 175 {
 176         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 177                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 178         return 0;
 179 }
 180 subsys_initcall(blk_ioc_init);