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 /**
  20  * get_io_context - increment reference count to io_context
  21  * @ioc: io_context to get
  22  *
  23  * Increment reference count to @ioc.
  24  */
  25 void get_io_context(struct io_context *ioc)
  26 {
  27         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
  28         atomic_long_inc(&ioc->refcount);
  29 }
  30 EXPORT_SYMBOL(get_io_context);
  31
  32 /*
  33  * Releasing ioc may nest into another put_io_context() leading to nested
  34  * fast path release.  As the ioc's can't be the same, this is okay but
  35  * makes lockdep whine.  Keep track of nesting and use it as subclass.
  36  */
  37 #ifdef CONFIG_LOCKDEP
  38 #define ioc_release_depth(q)            ((q) ? (q)->ioc_release_depth : 0)
  39 #define ioc_release_depth_inc(q)        (q)->ioc_release_depth++
  40 #define ioc_release_depth_dec(q)        (q)->ioc_release_depth--
  41 #else
  42 #define ioc_release_depth(q)            0
  43 #define ioc_release_depth_inc(q)        do { } while (0)
  44 #define ioc_release_depth_dec(q)        do { } while (0)
  45 #endif
  46
  47 /*
  48  * Slow path for ioc release in put_io_context().  Performs double-lock
  49  * dancing to unlink all cic's and then frees ioc.
  50  */
  51 static void ioc_release_fn(struct work_struct *work)
  52 {
  53         struct io_context *ioc = container_of(work, struct io_context,
  54                                               release_work);
  55         struct request_queue *last_q = NULL;
  56
  57         spin_lock_irq(&ioc->lock);
  58
  59         while (!hlist_empty(&ioc->cic_list)) {
  60                 struct cfq_io_context *cic = hlist_entry(ioc->cic_list.first,
  61                                                          struct cfq_io_context,
  62                                                          cic_list);
  63                 struct request_queue *this_q = cic->q;
  64
  65                 if (this_q != last_q) {
  66                         /*
  67                          * Need to switch to @this_q.  Once we release
  68                          * @ioc->lock, it can go away along with @cic.
  69                          * Hold on to it.
  70                          */
  71                         __blk_get_queue(this_q);
  72
  73                         /*
  74                          * blk_put_queue() might sleep thanks to kobject
  75                          * idiocy.  Always release both locks, put and
  76                          * restart.
  77                          */
  78                         if (last_q) {
  79                                 spin_unlock(last_q->queue_lock);
  80                                 spin_unlock_irq(&ioc->lock);
  81                                 blk_put_queue(last_q);
  82                         } else {
  83                                 spin_unlock_irq(&ioc->lock);
  84                         }
  85
  86                         last_q = this_q;
  87                         spin_lock_irq(this_q->queue_lock);
  88                         spin_lock(&ioc->lock);
  89                         continue;
  90                 }
  91                 ioc_release_depth_inc(this_q);
  92                 cic->exit(cic);
  93                 cic->release(cic);
  94                 ioc_release_depth_dec(this_q);
  95         }
  96
  97         if (last_q) {
  98                 spin_unlock(last_q->queue_lock);
  99                 spin_unlock_irq(&ioc->lock);
 100                 blk_put_queue(last_q);
 101         } else {
 102                 spin_unlock_irq(&ioc->lock);
 103         }
 104
 105         kmem_cache_free(iocontext_cachep, ioc);
 106 }
 107
 108 /**
 109  * put_io_context - put a reference of io_context
 110  * @ioc: io_context to put
 111  * @locked_q: request_queue the caller is holding queue_lock of (hint)
 112  *
 113  * Decrement reference count of @ioc and release it if the count reaches
 114  * zero.  If the caller is holding queue_lock of a queue, it can indicate
 115  * that with @locked_q.  This is an optimization hint and the caller is
 116  * allowed to pass in %NULL even when it's holding a queue_lock.
 117  */
 118 void put_io_context(struct io_context *ioc, struct request_queue *locked_q)
 119 {
 120         struct request_queue *last_q = locked_q;
 121         unsigned long flags;
 122
 123         if (ioc == NULL)
 124                 return;
 125
 126         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
 127         if (locked_q)
 128                 lockdep_assert_held(locked_q->queue_lock);
 129
 130         if (!atomic_long_dec_and_test(&ioc->refcount))
 131                 return;
 132
 133         /*
 134          * Destroy @ioc.  This is a bit messy because cic's are chained
 135          * from both ioc and queue, and ioc->lock nests inside queue_lock.
 136          * The inner ioc->lock should be held to walk our cic_list and then
 137          * for each cic the outer matching queue_lock should be grabbed.
 138          * ie. We need to do reverse-order double lock dancing.
 139          *
 140          * Another twist is that we are often called with one of the
 141          * matching queue_locks held as indicated by @locked_q, which
 142          * prevents performing double-lock dance for other queues.
 143          *
 144          * So, we do it in two stages.  The fast path uses the queue_lock
 145          * the caller is holding and, if other queues need to be accessed,
 146          * uses trylock to avoid introducing locking dependency.  This can
 147          * handle most cases, especially if @ioc was performing IO on only
 148          * single device.
 149          *
 150          * If trylock doesn't cut it, we defer to @ioc->release_work which
 151          * can do all the double-locking dancing.
 152          */
 153         spin_lock_irqsave_nested(&ioc->lock, flags,
 154                                  ioc_release_depth(locked_q));
 155
 156         while (!hlist_empty(&ioc->cic_list)) {
 157                 struct cfq_io_context *cic = hlist_entry(ioc->cic_list.first,
 158                                                          struct cfq_io_context,
 159                                                          cic_list);
 160                 struct request_queue *this_q = cic->q;
 161
 162                 if (this_q != last_q) {
 163                         if (last_q && last_q != locked_q)
 164                                 spin_unlock(last_q->queue_lock);
 165                         last_q = NULL;
 166
 167                         if (!spin_trylock(this_q->queue_lock))
 168                                 break;
 169                         last_q = this_q;
 170                         continue;
 171                 }
 172                 ioc_release_depth_inc(this_q);
 173                 cic->exit(cic);
 174                 cic->release(cic);
 175                 ioc_release_depth_dec(this_q);
 176         }
 177
 178         if (last_q && last_q != locked_q)
 179                 spin_unlock(last_q->queue_lock);
 180
 181         spin_unlock_irqrestore(&ioc->lock, flags);
 182
 183         /* if no cic's left, we're done; otherwise, kick release_work */
 184         if (hlist_empty(&ioc->cic_list))
 185                 kmem_cache_free(iocontext_cachep, ioc);
 186         else
 187                 schedule_work(&ioc->release_work);
 188 }
 189 EXPORT_SYMBOL(put_io_context);
 190
 191 /* Called by the exiting task */
 192 void exit_io_context(struct task_struct *task)
 193 {
 194         struct io_context *ioc;
 195
 196         /* PF_EXITING prevents new io_context from being attached to @task */
 197         WARN_ON_ONCE(!(current->flags & PF_EXITING));
 198
 199         task_lock(task);
 200         ioc = task->io_context;
 201         task->io_context = NULL;
 202         task_unlock(task);
 203
 204         atomic_dec(&ioc->nr_tasks);
 205         put_io_context(ioc, NULL);
 206 }
 207
 208 static struct io_context *create_task_io_context(struct task_struct *task,
 209                                                  gfp_t gfp_flags, int node,
 210                                                  bool take_ref)
 211 {
 212         struct io_context *ioc;
 213
 214         ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
 215                                     node);
 216         if (unlikely(!ioc))
 217                 return NULL;
 218
 219         /* initialize */
 220         atomic_long_set(&ioc->refcount, 1);
 221         atomic_set(&ioc->nr_tasks, 1);
 222         spin_lock_init(&ioc->lock);
 223         INIT_RADIX_TREE(&ioc->radix_root, GFP_ATOMIC | __GFP_HIGH);
 224         INIT_HLIST_HEAD(&ioc->cic_list);
 225         INIT_WORK(&ioc->release_work, ioc_release_fn);
 226
 227         /* try to install, somebody might already have beaten us to it */
 228         task_lock(task);
 229
 230         if (!task->io_context && !(task->flags & PF_EXITING)) {
 231                 task->io_context = ioc;
 232         } else {
 233                 kmem_cache_free(iocontext_cachep, ioc);
 234                 ioc = task->io_context;
 235         }
 236
 237         if (ioc && take_ref)
 238                 get_io_context(ioc);
 239
 240         task_unlock(task);
 241         return ioc;
 242 }
 243
 244 /**
 245  * current_io_context - get io_context of %current
 246  * @gfp_flags: allocation flags, used if allocation is necessary
 247  * @node: allocation node, used if allocation is necessary
 248  *
 249  * Return io_context of %current.  If it doesn't exist, it is created with
 250  * @gfp_flags and @node.  The returned io_context does NOT have its
 251  * reference count incremented.  Because io_context is exited only on task
 252  * exit, %current can be sure that the returned io_context is valid and
 253  * alive as long as it is executing.
 254  */
 255 struct io_context *current_io_context(gfp_t gfp_flags, int node)
 256 {
 257         might_sleep_if(gfp_flags & __GFP_WAIT);
 258
 259         if (current->io_context)
 260                 return current->io_context;
 261
 262         return create_task_io_context(current, gfp_flags, node, false);
 263 }
 264 EXPORT_SYMBOL(current_io_context);
 265
 266 /**
 267  * get_task_io_context - get io_context of a task
 268  * @task: task of interest
 269  * @gfp_flags: allocation flags, used if allocation is necessary
 270  * @node: allocation node, used if allocation is necessary
 271  *
 272  * Return io_context of @task.  If it doesn't exist, it is created with
 273  * @gfp_flags and @node.  The returned io_context has its reference count
 274  * incremented.
 275  *
 276  * This function always goes through task_lock() and it's better to use
 277  * current_io_context() + get_io_context() for %current.
 278  */
 279 struct io_context *get_task_io_context(struct task_struct *task,
 280                                        gfp_t gfp_flags, int node)
 281 {
 282         struct io_context *ioc;
 283
 284         might_sleep_if(gfp_flags & __GFP_WAIT);
 285
 286         task_lock(task);
 287         ioc = task->io_context;
 288         if (likely(ioc)) {
 289                 get_io_context(ioc);
 290                 task_unlock(task);
 291                 return ioc;
 292         }
 293         task_unlock(task);
 294
 295         return create_task_io_context(task, gfp_flags, node, true);
 296 }
 297 EXPORT_SYMBOL(get_task_io_context);
 298
 299 void ioc_set_changed(struct io_context *ioc, int which)
 300 {
 301         struct cfq_io_context *cic;
 302         struct hlist_node *n;
 303
 304         hlist_for_each_entry(cic, n, &ioc->cic_list, cic_list)
 305                 set_bit(which, &cic->changed);
 306 }
 307
 308 /**
 309  * ioc_ioprio_changed - notify ioprio change
 310  * @ioc: io_context of interest
 311  * @ioprio: new ioprio
 312  *
 313  * @ioc's ioprio has changed to @ioprio.  Set %CIC_IOPRIO_CHANGED for all
 314  * cic's.  iosched is responsible for checking the bit and applying it on
 315  * request issue path.
 316  */
 317 void ioc_ioprio_changed(struct io_context *ioc, int ioprio)
 318 {
 319         unsigned long flags;
 320
 321         spin_lock_irqsave(&ioc->lock, flags);
 322         ioc->ioprio = ioprio;
 323         ioc_set_changed(ioc, CIC_IOPRIO_CHANGED);
 324         spin_unlock_irqrestore(&ioc->lock, flags);
 325 }
 326
 327 /**
 328  * ioc_cgroup_changed - notify cgroup change
 329  * @ioc: io_context of interest
 330  *
 331  * @ioc's cgroup has changed.  Set %CIC_CGROUP_CHANGED for all cic's.
 332  * iosched is responsible for checking the bit and applying it on request
 333  * issue path.
 334  */
 335 void ioc_cgroup_changed(struct io_context *ioc)
 336 {
 337         unsigned long flags;
 338
 339         spin_lock_irqsave(&ioc->lock, flags);
 340         ioc_set_changed(ioc, CIC_CGROUP_CHANGED);
 341         spin_unlock_irqrestore(&ioc->lock, flags);
 342 }
 343
 344 static int __init blk_ioc_init(void)
 345 {
 346         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 347                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 348         return 0;
 349 }
 350 subsys_initcall(blk_ioc_init);