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 icq'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->icq_list)) {
  60                 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
  61                                                 struct io_cq, ioc_node);
  62                 struct request_queue *this_q = icq->q;
  63
  64                 if (this_q != last_q) {
  65                         /*
  66                          * Need to switch to @this_q.  Once we release
  67                          * @ioc->lock, it can go away along with @cic.
  68                          * Hold on to it.
  69                          */
  70                         __blk_get_queue(this_q);
  71
  72                         /*
  73                          * blk_put_queue() might sleep thanks to kobject
  74                          * idiocy.  Always release both locks, put and
  75                          * restart.
  76                          */
  77                         if (last_q) {
  78                                 spin_unlock(last_q->queue_lock);
  79                                 spin_unlock_irq(&ioc->lock);
  80                                 blk_put_queue(last_q);
  81                         } else {
  82                                 spin_unlock_irq(&ioc->lock);
  83                         }
  84
  85                         last_q = this_q;
  86                         spin_lock_irq(this_q->queue_lock);
  87                         spin_lock(&ioc->lock);
  88                         continue;
  89                 }
  90                 ioc_release_depth_inc(this_q);
  91                 icq->exit(icq);
  92                 icq->release(icq);
  93                 ioc_release_depth_dec(this_q);
  94         }
  95
  96         if (last_q) {
  97                 spin_unlock(last_q->queue_lock);
  98                 spin_unlock_irq(&ioc->lock);
  99                 blk_put_queue(last_q);
 100         } else {
 101                 spin_unlock_irq(&ioc->lock);
 102         }
 103
 104         kmem_cache_free(iocontext_cachep, ioc);
 105 }
 106
 107 /**
 108  * put_io_context - put a reference of io_context
 109  * @ioc: io_context to put
 110  * @locked_q: request_queue the caller is holding queue_lock of (hint)
 111  *
 112  * Decrement reference count of @ioc and release it if the count reaches
 113  * zero.  If the caller is holding queue_lock of a queue, it can indicate
 114  * that with @locked_q.  This is an optimization hint and the caller is
 115  * allowed to pass in %NULL even when it's holding a queue_lock.
 116  */
 117 void put_io_context(struct io_context *ioc, struct request_queue *locked_q)
 118 {
 119         struct request_queue *last_q = locked_q;
 120         unsigned long flags;
 121
 122         if (ioc == NULL)
 123                 return;
 124
 125         BUG_ON(atomic_long_read(&ioc->refcount) <= 0);
 126         if (locked_q)
 127                 lockdep_assert_held(locked_q->queue_lock);
 128
 129         if (!atomic_long_dec_and_test(&ioc->refcount))
 130                 return;
 131
 132         /*
 133          * Destroy @ioc.  This is a bit messy because icq's are chained
 134          * from both ioc and queue, and ioc->lock nests inside queue_lock.
 135          * The inner ioc->lock should be held to walk our icq_list and then
 136          * for each icq the outer matching queue_lock should be grabbed.
 137          * ie. We need to do reverse-order double lock dancing.
 138          *
 139          * Another twist is that we are often called with one of the
 140          * matching queue_locks held as indicated by @locked_q, which
 141          * prevents performing double-lock dance for other queues.
 142          *
 143          * So, we do it in two stages.  The fast path uses the queue_lock
 144          * the caller is holding and, if other queues need to be accessed,
 145          * uses trylock to avoid introducing locking dependency.  This can
 146          * handle most cases, especially if @ioc was performing IO on only
 147          * single device.
 148          *
 149          * If trylock doesn't cut it, we defer to @ioc->release_work which
 150          * can do all the double-locking dancing.
 151          */
 152         spin_lock_irqsave_nested(&ioc->lock, flags,
 153                                  ioc_release_depth(locked_q));
 154
 155         while (!hlist_empty(&ioc->icq_list)) {
 156                 struct io_cq *icq = hlist_entry(ioc->icq_list.first,
 157                                                 struct io_cq, ioc_node);
 158                 struct request_queue *this_q = icq->q;
 159
 160                 if (this_q != last_q) {
 161                         if (last_q && last_q != locked_q)
 162                                 spin_unlock(last_q->queue_lock);
 163                         last_q = NULL;
 164
 165                         if (!spin_trylock(this_q->queue_lock))
 166                                 break;
 167                         last_q = this_q;
 168                         continue;
 169                 }
 170                 ioc_release_depth_inc(this_q);
 171                 icq->exit(icq);
 172                 icq->release(icq);
 173                 ioc_release_depth_dec(this_q);
 174         }
 175
 176         if (last_q && last_q != locked_q)
 177                 spin_unlock(last_q->queue_lock);
 178
 179         spin_unlock_irqrestore(&ioc->lock, flags);
 180
 181         /* if no icq is left, we're done; otherwise, kick release_work */
 182         if (hlist_empty(&ioc->icq_list))
 183                 kmem_cache_free(iocontext_cachep, ioc);
 184         else
 185                 schedule_work(&ioc->release_work);
 186 }
 187 EXPORT_SYMBOL(put_io_context);
 188
 189 /* Called by the exiting task */
 190 void exit_io_context(struct task_struct *task)
 191 {
 192         struct io_context *ioc;
 193
 194         /* PF_EXITING prevents new io_context from being attached to @task */
 195         WARN_ON_ONCE(!(current->flags & PF_EXITING));
 196
 197         task_lock(task);
 198         ioc = task->io_context;
 199         task->io_context = NULL;
 200         task_unlock(task);
 201
 202         atomic_dec(&ioc->nr_tasks);
 203         put_io_context(ioc, NULL);
 204 }
 205
 206 void create_io_context_slowpath(struct task_struct *task, gfp_t gfp_flags,
 207                                 int node)
 208 {
 209         struct io_context *ioc;
 210
 211         ioc = kmem_cache_alloc_node(iocontext_cachep, gfp_flags | __GFP_ZERO,
 212                                     node);
 213         if (unlikely(!ioc))
 214                 return;
 215
 216         /* initialize */
 217         atomic_long_set(&ioc->refcount, 1);
 218         atomic_set(&ioc->nr_tasks, 1);
 219         spin_lock_init(&ioc->lock);
 220         INIT_RADIX_TREE(&ioc->icq_tree, GFP_ATOMIC | __GFP_HIGH);
 221         INIT_HLIST_HEAD(&ioc->icq_list);
 222         INIT_WORK(&ioc->release_work, ioc_release_fn);
 223
 224         /* try to install, somebody might already have beaten us to it */
 225         task_lock(task);
 226         if (!task->io_context && !(task->flags & PF_EXITING))
 227                 task->io_context = ioc;
 228         else
 229                 kmem_cache_free(iocontext_cachep, ioc);
 230         task_unlock(task);
 231 }
 232 EXPORT_SYMBOL(create_io_context_slowpath);
 233
 234 /**
 235  * get_task_io_context - get io_context of a task
 236  * @task: task of interest
 237  * @gfp_flags: allocation flags, used if allocation is necessary
 238  * @node: allocation node, used if allocation is necessary
 239  *
 240  * Return io_context of @task.  If it doesn't exist, it is created with
 241  * @gfp_flags and @node.  The returned io_context has its reference count
 242  * incremented.
 243  *
 244  * This function always goes through task_lock() and it's better to use
 245  * %current->io_context + get_io_context() for %current.
 246  */
 247 struct io_context *get_task_io_context(struct task_struct *task,
 248                                        gfp_t gfp_flags, int node)
 249 {
 250         struct io_context *ioc;
 251
 252         might_sleep_if(gfp_flags & __GFP_WAIT);
 253
 254         do {
 255                 task_lock(task);
 256                 ioc = task->io_context;
 257                 if (likely(ioc)) {
 258                         get_io_context(ioc);
 259                         task_unlock(task);
 260                         return ioc;
 261                 }
 262                 task_unlock(task);
 263         } while (create_io_context(task, gfp_flags, node));
 264
 265         return NULL;
 266 }
 267 EXPORT_SYMBOL(get_task_io_context);
 268
 269 /**
 270  * ioc_lookup_icq - lookup io_cq from ioc
 271  * @ioc: the associated io_context
 272  * @q: the associated request_queue
 273  *
 274  * Look up io_cq associated with @ioc - @q pair from @ioc.  Must be called
 275  * with @q->queue_lock held.
 276  */
 277 struct io_cq *ioc_lookup_icq(struct io_context *ioc, struct request_queue *q)
 278 {
 279         struct io_cq *icq;
 280
 281         lockdep_assert_held(q->queue_lock);
 282
 283         /*
 284          * icq's are indexed from @ioc using radix tree and hint pointer,
 285          * both of which are protected with RCU.  All removals are done
 286          * holding both q and ioc locks, and we're holding q lock - if we
 287          * find a icq which points to us, it's guaranteed to be valid.
 288          */
 289         rcu_read_lock();
 290         icq = rcu_dereference(ioc->icq_hint);
 291         if (icq && icq->q == q)
 292                 goto out;
 293
 294         icq = radix_tree_lookup(&ioc->icq_tree, q->id);
 295         if (icq && icq->q == q)
 296                 rcu_assign_pointer(ioc->icq_hint, icq); /* allowed to race */
 297         else
 298                 icq = NULL;
 299 out:
 300         rcu_read_unlock();
 301         return icq;
 302 }
 303 EXPORT_SYMBOL(ioc_lookup_icq);
 304
 305 void ioc_set_changed(struct io_context *ioc, int which)
 306 {
 307         struct io_cq *icq;
 308         struct hlist_node *n;
 309
 310         hlist_for_each_entry(icq, n, &ioc->icq_list, ioc_node)
 311                 set_bit(which, &icq->changed);
 312 }
 313
 314 /**
 315  * ioc_ioprio_changed - notify ioprio change
 316  * @ioc: io_context of interest
 317  * @ioprio: new ioprio
 318  *
 319  * @ioc's ioprio has changed to @ioprio.  Set %ICQ_IOPRIO_CHANGED for all
 320  * icq's.  iosched is responsible for checking the bit and applying it on
 321  * request issue path.
 322  */
 323 void ioc_ioprio_changed(struct io_context *ioc, int ioprio)
 324 {
 325         unsigned long flags;
 326
 327         spin_lock_irqsave(&ioc->lock, flags);
 328         ioc->ioprio = ioprio;
 329         ioc_set_changed(ioc, ICQ_IOPRIO_CHANGED);
 330         spin_unlock_irqrestore(&ioc->lock, flags);
 331 }
 332
 333 /**
 334  * ioc_cgroup_changed - notify cgroup change
 335  * @ioc: io_context of interest
 336  *
 337  * @ioc's cgroup has changed.  Set %ICQ_CGROUP_CHANGED for all icq's.
 338  * iosched is responsible for checking the bit and applying it on request
 339  * issue path.
 340  */
 341 void ioc_cgroup_changed(struct io_context *ioc)
 342 {
 343         unsigned long flags;
 344
 345         spin_lock_irqsave(&ioc->lock, flags);
 346         ioc_set_changed(ioc, ICQ_CGROUP_CHANGED);
 347         spin_unlock_irqrestore(&ioc->lock, flags);
 348 }
 349
 350 static int __init blk_ioc_init(void)
 351 {
 352         iocontext_cachep = kmem_cache_create("blkdev_ioc",
 353                         sizeof(struct io_context), 0, SLAB_PANIC, NULL);
 354         return 0;
 355 }
 356 subsys_initcall(blk_ioc_init);