kernel/pwqr.c

   1 /*
   2  * Copyright (C) 2012   Pierre Habouzit <pierre.habouzit@intersec.com>
   3  * Copyright (C) 2012   Intersec SAS
   4  *
   5  * This file implements the Linux Pthread Workqueue Regulator, and is part
   6  * of the linux kernel.
   7  *
   8  * The Linux Kernel is free software: you can redistribute it and/or modify it
   9  * under the terms of the GNU General Public License version 2 as published by
  10  * the Free Software Foundation.
  11  *
  12  * The Linux Kernel is distributed in the hope that it will be useful, but
  13  * WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  14  * or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
  15  * License for more details.
  16  *
  17  * You should have received a copy of the GNU General Public License version 2
  18  * along with The Linux Kernel.  If not, see <http://www.gnu.org/licenses/>.
  19  */
  20
  21 #include <linux/cdev.h>
  22 #include <linux/device.h>
  23 #include <linux/file.h>
  24 #include <linux/fs.h>
  25 #include <linux/hash.h>
  26 #include <linux/init.h>
  27 #include <linux/kref.h>
  28 #include <linux/module.h>
  29 #include <linux/poll.h>
  30 #include <linux/sched.h>
  31 #include <linux/slab.h>
  32 #include <linux/spinlock.h>
  33 #include <linux/timer.h>
  34 #include <linux/uaccess.h>
  35 #include <linux/wait.h>
  36 #include <linux/version.h>
  37
  38 /*
  39  * The pthread workqueue regulator code is for now written as a proof of
  40  * concept module, meant to work with 2.6.23+ kernels or redhat5 ones.
  41  *
  42  * For now it uses a device /dev/pwq, which spawns magic file-descriptors
  43  * supporting a few ioctl operations (see Documentation/pwqr.adoc shipped in
  44  * the same git repository).
  45  *
  46  * This code is meant to be merged into mainline, but after the following
  47  * changes, kept here as a "todolist":
  48  *
  49  *   - get rid of the device stuff (which is 100% of the init code for 2.6.23
  50  *     kernels);
  51  *
  52  *   - resubmit the patch that makes it possible to call
  53  *     preempt_notifier_unregister from sched_in/sched_out (just a matter of a
  54  *     hlist_for_each_safe instead of hlist_for_each), and fix
  55  *     pwqr_task_release to not require RCU anymore. It makes
  56  *     pwqr_preempt_noop_ops go away.
  57  *
  58  *   - think about the possibility to add a pwq_notifier pointer directly into
  59  *     the task_struct, thought it's not *that* necessary, it grows the
  60  *     structure for a speed gain we don't really need (making pwqr_ctl
  61  *     faster). I think it's okay to crawl the preempt_notifier list instead.
  62  *     We may want to add nice "macros" for that though.
  63  *
  64  *   - replace the ioctl with a pwqr_ctl syscall
  65  *
  66  *   - create a pwqr_create() syscall to create a pwqr file-descriptor.
  67  *
  68  * Summary: most of the code should be untouched or almost not changed,
  69  * pwqr_ioctl adapted to become a syscall, and the module boilerplate replaced
  70  * with pwqr_create() and file-descriptor creation boilerplate instead. But
  71  * looking at fs/eventfd.c this looks rather simple.
  72  */
  73
  74 #ifndef CONFIG_PREEMPT_NOTIFIERS
  75 #  error PWQ module requires CONFIG_PREEMPT_NOTIFIERS
  76 #else
  77
  78 #include "pwqr.h"
  79
  80 #define PWQR_UC_DELAY           (HZ / 10)
  81 #define PWQR_OC_DELAY           (HZ / 20)
  82
  83 #define PWQR_STATE_NONE         0
  84 #define PWQR_STATE_UC           1
  85 #define PWQR_STATE_OC           2
  86 #define PWQR_STATE_DEAD         (-1)
  87
  88 /*
  89  * This is the first inclusion of CONFIG_PREEMPT_NOTIFIERS in the kernel.
  90  *
  91  * Though I want it to work on older redhat 5 kernels, that have an emulation
  92  * of the feature but not implemented the same way, and instead of linking the
  93  * preempt_notifiers from the task_struct directly, they have a private
  94  * h-table I don't have access to, so I need my own too.
  95  *
  96  * For vanilla kernels we crawl through the task_struct::preempt_notifiers
  97  * hlist until we find our entry, this list is often very short, and it's no
  98  * slower than the global h-table which also crawls a list anyway.
  99  */
 100 #define IS_PRE_2_6_23    (LINUX_VERSION_CODE < KERNEL_VERSION(2, 6, 23))
 101
 102 struct pwqr_sb {
 103         struct kref             kref;
 104         struct rcu_head         rcu;
 105         struct timer_list       timer;
 106         wait_queue_head_t       wqh;
 107         wait_queue_head_t       wqh_poll;
 108
 109         unsigned                concurrency;
 110         unsigned                registered;
 111
 112         unsigned                running;
 113         unsigned                waiting;
 114         unsigned                parked;
 115         unsigned                overcommit_wakes;
 116
 117         int                     state;
 118         unsigned                has_pollin;
 119 };
 120
 121 struct pwqr_task {
 122         struct preempt_notifier notifier;
 123         struct pwqr_sb         *sb;
 124         struct rcu_head         rcu;
 125 #if IS_PRE_2_6_23
 126         struct hlist_node       link;
 127         struct task_struct     *task;
 128 #endif
 129 };
 130
 131 #if IS_PRE_2_6_23
 132
 133 #define PWQR_HASH_BITS          5
 134 #define PWQR_HASH_SIZE          (1 << PWQR_HASH_BITS)
 135
 136 struct pwqr_task_bucket {
 137         spinlock_t              lock;
 138         struct hlist_head       tasks;
 139 };
 140
 141 static struct pwqr_task_bucket  pwqr_tasks_hash[PWQR_HASH_SIZE];
 142 #endif
 143
 144 /*
 145  * Global variables
 146  */
 147 static struct class            *pwqr_class;
 148 static int                      pwqr_major;
 149 static struct preempt_ops       pwqr_preempt_running_ops;
 150 static struct preempt_ops       pwqr_preempt_blocked_ops;
 151 static struct preempt_ops       pwqr_preempt_noop_ops;
 152
 153 /*****************************************************************************
 154  * Scoreboards
 155  */
 156
 157 #define pwqr_sb_lock_irqsave(sb, flags) \
 158         spin_lock_irqsave(&(sb)->wqh.lock, flags)
 159 #define pwqr_sb_unlock_irqrestore(sb, flags) \
 160         spin_unlock_irqrestore(&(sb)->wqh.lock, flags)
 161
 162 static inline void pwqr_arm_timer(struct pwqr_sb *sb, int how, int delay)
 163 {
 164         if (timer_pending(&sb->timer) && sb->state == how)
 165                 return;
 166         mod_timer(&sb->timer, jiffies + delay);
 167         sb->state = how;
 168 }
 169
 170 static inline void __pwqr_sb_update_state(struct pwqr_sb *sb, int running_delta)
 171 {
 172         sb->running += running_delta;
 173
 174         if (sb->running < sb->concurrency && sb->waiting == 0 && sb->parked) {
 175                 pwqr_arm_timer(sb, PWQR_STATE_UC, PWQR_UC_DELAY);
 176         } else if (sb->running > sb->concurrency) {
 177                 pwqr_arm_timer(sb, PWQR_STATE_OC, PWQR_OC_DELAY);
 178         } else {
 179                 sb->state = PWQR_STATE_NONE;
 180                 if (!timer_pending(&sb->timer))
 181                         del_timer(&sb->timer);
 182         }
 183 }
 184
 185 static void pwqr_sb_timer_cb(unsigned long arg)
 186 {
 187         struct pwqr_sb *sb = (struct pwqr_sb *)arg;
 188         unsigned long flags;
 189
 190         pwqr_sb_lock_irqsave(sb, flags);
 191         if (sb->running < sb->concurrency && sb->waiting == 0 && sb->parked) {
 192                 if (sb->overcommit_wakes == 0)
 193                         wake_up_locked(&sb->wqh);
 194         }
 195         if (sb->running > sb->concurrency) {
 196                 printk(KERN_DEBUG "wake up poll");
 197                 wake_up_poll(&sb->wqh_poll, POLLIN);
 198                 sb->has_pollin = 1;
 199         }
 200         pwqr_sb_unlock_irqrestore(sb, flags);
 201 }
 202
 203 static struct pwqr_sb *pwqr_sb_create(void)
 204 {
 205         struct pwqr_sb *sb;
 206
 207         sb = kzalloc(sizeof(struct pwqr_sb), GFP_KERNEL);
 208         if (sb == NULL)
 209                 return ERR_PTR(-ENOMEM);
 210
 211         kref_init(&sb->kref);
 212         init_waitqueue_head(&sb->wqh);
 213         init_waitqueue_head(&sb->wqh_poll);
 214         sb->concurrency    = num_online_cpus();
 215         init_timer(&sb->timer);
 216         sb->timer.function = pwqr_sb_timer_cb;
 217         sb->timer.data     = (unsigned long)sb;
 218
 219         __module_get(THIS_MODULE);
 220         return sb;
 221 }
 222 static inline void pwqr_sb_get(struct pwqr_sb *sb)
 223 {
 224         kref_get(&sb->kref);
 225 }
 226
 227 static void pwqr_sb_finalize(struct rcu_head *rcu)
 228 {
 229         struct pwqr_sb *sb = container_of(rcu, struct pwqr_sb, rcu);
 230
 231         module_put(THIS_MODULE);
 232         kfree(sb);
 233 }
 234
 235 static void pwqr_sb_release(struct kref *kref)
 236 {
 237         struct pwqr_sb *sb = container_of(kref, struct pwqr_sb, kref);
 238
 239         del_timer_sync(&sb->timer);
 240         wake_up_poll(&sb->wqh_poll, POLLHUP);
 241         call_rcu(&sb->rcu, pwqr_sb_finalize);
 242 }
 243 static inline void pwqr_sb_put(struct pwqr_sb *sb)
 244 {
 245         kref_put(&sb->kref, pwqr_sb_release);
 246 }
 247
 248 /*****************************************************************************
 249  * tasks
 250  */
 251 #if IS_PRE_2_6_23
 252 static inline struct pwqr_task_bucket *task_hbucket(struct task_struct *task)
 253 {
 254         return &pwqr_tasks_hash[hash_ptr(task, PWQR_HASH_BITS)];
 255 }
 256
 257 static struct pwqr_task *pwqr_task_find(struct task_struct *task)
 258 {
 259         struct pwqr_task_bucket *b = task_hbucket(task);
 260         struct hlist_node *node;
 261         struct pwqr_task *pwqt = NULL;
 262
 263         spin_lock(&b->lock);
 264         hlist_for_each_entry(pwqt, node, &b->tasks, link) {
 265                 if (pwqt->task == task)
 266                         break;
 267         }
 268         spin_unlock(&b->lock);
 269         return pwqt;
 270 }
 271 #else
 272 static struct pwqr_task *pwqr_task_find(struct task_struct *task)
 273 {
 274         struct hlist_node       *node;
 275         struct preempt_notifier *it;
 276         struct pwqr_task        *pwqt = NULL;
 277
 278         hlist_for_each_entry(it, node, &task->preempt_notifiers, link) {
 279                 if (it->ops == &pwqr_preempt_running_ops ||
 280                     it->ops == &pwqr_preempt_blocked_ops ||
 281                     it->ops == &pwqr_preempt_noop_ops)
 282                 {
 283                         pwqt = container_of(it, struct pwqr_task, notifier);
 284                         break;
 285                 }
 286         }
 287
 288         return pwqt;
 289 }
 290 #endif
 291
 292 static struct pwqr_task *pwqr_task_create(struct task_struct *task)
 293 {
 294         struct pwqr_task *pwqt;
 295
 296         pwqt = kmalloc(sizeof(*pwqt), GFP_KERNEL);
 297         if (pwqt == NULL)
 298                 return ERR_PTR(-ENOMEM);
 299
 300         preempt_notifier_init(&pwqt->notifier, &pwqr_preempt_running_ops);
 301         preempt_notifier_register(&pwqt->notifier);
 302 #if IS_PRE_2_6_23
 303         {
 304                 struct pwqr_task_bucket *b = task_hbucket(task);
 305
 306                 pwqt->task = task;
 307                 spin_lock(&b->lock);
 308                 hlist_add_head(&pwqt->link, &b->tasks);
 309                 spin_unlock(&b->lock);
 310         }
 311 #endif
 312         return pwqt;
 313 }
 314
 315 __cold
 316 static void pwqr_task_detach(struct pwqr_task *pwqt, struct pwqr_sb *sb)
 317 {
 318         unsigned long flags;
 319
 320         pwqr_sb_lock_irqsave(sb, flags);
 321         sb->registered--;
 322         if (pwqt->notifier.ops == &pwqr_preempt_running_ops) {
 323                 __pwqr_sb_update_state(sb, -1);
 324         } else {
 325                 __pwqr_sb_update_state(sb, 0);
 326         }
 327         pwqt->notifier.ops = &pwqr_preempt_noop_ops;
 328         pwqr_sb_unlock_irqrestore(sb, flags);
 329         pwqr_sb_put(sb);
 330         pwqt->sb = NULL;
 331 }
 332
 333 __cold
 334 static void pwqr_task_attach(struct pwqr_task *pwqt, struct pwqr_sb *sb)
 335 {
 336         unsigned long flags;
 337
 338         pwqr_sb_lock_irqsave(sb, flags);
 339         pwqr_sb_get(pwqt->sb = sb);
 340         sb->registered++;
 341         __pwqr_sb_update_state(sb, 1);
 342         pwqr_sb_unlock_irqrestore(sb, flags);
 343 }
 344
 345 __cold
 346 static void pwqr_task_release(struct pwqr_task *pwqt, bool from_notifier)
 347 {
 348 #if IS_PRE_2_6_23
 349         struct pwqr_task_bucket *b = task_hbucket(pwqt->task);
 350
 351         spin_lock(&b->lock);
 352         hlist_del(&pwqt->link);
 353         spin_unlock(&b->lock);
 354 #endif
 355
 356         if (from_notifier) {
 357                 /* When called from sched_{out,in}, it's not allowed to
 358                  * call preempt_notifier_unregister (or worse kfree())
 359                  *
 360                  * Though it's not a good idea to kfree() still registered
 361                  * callbacks if we're not dying, it'll panic on the next
 362                  * sched_{in,out} call.
 363                  */
 364                 BUG_ON(!(current->state & TASK_DEAD));
 365                 kfree_rcu(pwqt, rcu);
 366         } else {
 367                 preempt_notifier_unregister(&pwqt->notifier);
 368                 kfree(pwqt);
 369         }
 370 }
 371
 372 static void pwqr_task_noop_sched_in(struct preempt_notifier *notifier, int cpu)
 373 {
 374 }
 375
 376 static void pwqr_task_noop_sched_out(struct preempt_notifier *notifier,
 377                                      struct task_struct *next)
 378 {
 379 }
 380
 381 static void pwqr_task_blocked_sched_in(struct preempt_notifier *notifier, int cpu)
 382 {
 383         struct pwqr_task *pwqt = container_of(notifier, struct pwqr_task, notifier);
 384         struct pwqr_sb   *sb   = pwqt->sb;
 385         unsigned long flags;
 386
 387         if (unlikely(sb->state < 0)) {
 388                 pwqr_task_detach(pwqt, sb);
 389                 return;
 390         }
 391
 392         pwqt->notifier.ops = &pwqr_preempt_running_ops;
 393         pwqr_sb_lock_irqsave(sb, flags);
 394         __pwqr_sb_update_state(sb, 1);
 395         pwqr_sb_unlock_irqrestore(sb, flags);
 396 }
 397
 398 static void pwqr_task_sched_out(struct preempt_notifier *notifier,
 399                                 struct task_struct *next)
 400 {
 401         struct pwqr_task   *pwqt = container_of(notifier, struct pwqr_task, notifier);
 402         struct pwqr_sb     *sb   = pwqt->sb;
 403         struct task_struct *p    = current;
 404
 405         if (unlikely(p->state & TASK_DEAD) || unlikely(sb->state < 0)) {
 406                 pwqr_task_detach(pwqt, sb);
 407                 if (p->state & TASK_DEAD)
 408                         pwqr_task_release(pwqt, true);
 409                 return;
 410         }
 411         if (p->state == 0 || (p->state & (__TASK_STOPPED | __TASK_TRACED)))
 412                 return;
 413
 414         pwqt->notifier.ops = &pwqr_preempt_blocked_ops;
 415         /* see preempt.h: irq are disabled for sched_out */
 416         spin_lock(&sb->wqh.lock);
 417         __pwqr_sb_update_state(sb, -1);
 418         spin_unlock(&sb->wqh.lock);
 419 }
 420
 421 static struct preempt_ops __read_mostly pwqr_preempt_noop_ops = {
 422         .sched_in       = pwqr_task_noop_sched_in,
 423         .sched_out      = pwqr_task_noop_sched_out,
 424 };
 425
 426 static struct preempt_ops __read_mostly pwqr_preempt_running_ops = {
 427         .sched_in       = pwqr_task_noop_sched_in,
 428         .sched_out      = pwqr_task_sched_out,
 429 };
 430
 431 static struct preempt_ops __read_mostly pwqr_preempt_blocked_ops = {
 432         .sched_in       = pwqr_task_blocked_sched_in,
 433         .sched_out      = pwqr_task_sched_out,
 434 };
 435
 436 /*****************************************************************************
 437  * file descriptor
 438  */
 439 static int pwqr_open(struct inode *inode, struct file *filp)
 440 {
 441         struct pwqr_sb *sb;
 442
 443         sb = pwqr_sb_create();
 444         if (IS_ERR(sb))
 445                 return PTR_ERR(sb);
 446         filp->private_data = sb;
 447         return 0;
 448 }
 449
 450 static int pwqr_release(struct inode *inode, struct file *filp)
 451 {
 452         struct pwqr_sb *sb = filp->private_data;
 453         unsigned long flags;
 454
 455         pwqr_sb_lock_irqsave(sb, flags);
 456         sb->state = PWQR_STATE_DEAD;
 457         pwqr_sb_unlock_irqrestore(sb, flags);
 458         wake_up_all(&sb->wqh);
 459         pwqr_sb_put(sb);
 460         return 0;
 461 }
 462
 463 static unsigned int pwqr_poll(struct file *filp, poll_table *wait)
 464 {
 465         struct pwqr_sb *sb = filp->private_data;
 466         unsigned int events = 0;
 467         unsigned long flags;
 468
 469         poll_wait(filp, &sb->wqh_poll, wait);
 470
 471         pwqr_sb_lock_irqsave(sb, flags);
 472         if (sb->has_pollin)
 473                 events |= POLLIN;
 474         if (sb->state < 0)
 475                 events |= POLLHUP;
 476         pwqr_sb_unlock_irqrestore(sb, flags);
 477
 478         return events;
 479 }
 480
 481 static inline ssize_t pwqr_sb_read(struct pwqr_sb *sb, int no_wait, u32 *cnt)
 482 {
 483         DECLARE_WAITQUEUE(wait, current);
 484         ssize_t rc = -EAGAIN;
 485
 486         spin_lock_irq(&sb->wqh.lock);
 487         if (sb->running > sb->concurrency) {
 488                 rc = 0;
 489         } else if (!no_wait) {
 490                 add_wait_queue(&sb->wqh_poll, &wait);
 491                 for (;;) {
 492                         set_current_state(TASK_INTERRUPTIBLE);
 493                         if (sb->running > sb->concurrency) {
 494                                 rc = 0;
 495                                 break;
 496                         }
 497                         if (signal_pending(current)) {
 498                                 rc = -ERESTARTSYS;
 499                                 break;
 500                         }
 501                         spin_unlock_irq(&sb->wqh.lock);
 502                         schedule();
 503                         spin_lock_irq(&sb->wqh.lock);
 504                 }
 505                 remove_wait_queue(&sb->wqh_poll, &wait);
 506                 __set_current_state(TASK_RUNNING);
 507         }
 508         if (likely(rc == 0)) {
 509                 *cnt = sb->running - sb->concurrency;
 510                 sb->has_pollin = 0;
 511         }
 512         spin_unlock_irq(&sb->wqh.lock);
 513
 514         return rc;
 515 }
 516
 517 static ssize_t
 518 pwqr_read(struct file *filp, char __user *buf, size_t count, loff_t *ppos)
 519 {
 520         struct pwqr_sb *sb = filp->private_data;
 521         u32 cnt = 0;
 522         ssize_t rc;
 523
 524         if (count < sizeof(cnt))
 525                 return -EINVAL;
 526         rc = pwqr_sb_read(sb, filp->f_flags & O_NONBLOCK, &cnt);
 527         if (rc < 0)
 528                 return rc;
 529         return put_user(cnt, (u32 __user *)buf) ? -EFAULT : sizeof(cnt);
 530 }
 531
 532 static long
 533 do_pwqr_wait(struct pwqr_sb *sb, struct pwqr_task *pwqt,
 534              int is_wait, struct pwqr_ioc_wait __user *arg)
 535 {
 536         unsigned long flags;
 537         struct pwqr_ioc_wait wait;
 538         long rc = 0;
 539         u32 uval;
 540
 541         preempt_notifier_unregister(&pwqt->notifier);
 542
 543         if (is_wait) {
 544                 if (copy_from_user(&wait, arg, sizeof(wait))) {
 545                         rc = -EFAULT;
 546                         goto out;
 547                 }
 548                 if (unlikely((long)wait.pwqr_uaddr % sizeof(int) != 0)) {
 549                         rc = -EINVAL;
 550                         goto out;
 551                 }
 552         }
 553
 554         pwqr_sb_lock_irqsave(sb, flags);
 555         if (sb->running + sb->waiting <= sb->concurrency) {
 556                 if (is_wait) {
 557                         while (probe_kernel_address(wait.pwqr_uaddr, uval)) {
 558                                 pwqr_sb_unlock_irqrestore(sb, flags);
 559                                 rc = get_user(uval, (u32 *)wait.pwqr_uaddr);
 560                                 if (rc)
 561                                         goto out;
 562                                 pwqr_sb_lock_irqsave(sb, flags);
 563                         }
 564
 565                         if (uval != (u32)wait.pwqr_ticket) {
 566                                 rc = -EWOULDBLOCK;
 567                                 goto out_unlock;
 568                         }
 569                 } else {
 570                         goto out_unlock;
 571                 }
 572         }
 573
 574         /* @ see <wait_event_interruptible_exclusive_locked_irq> */
 575         if (likely(sb->state >= 0)) {
 576                 DEFINE_WAIT(__wait);
 577                 __wait.flags |= WQ_FLAG_EXCLUSIVE;
 578
 579                 if (is_wait) {
 580                         sb->waiting++;
 581                         __add_wait_queue(&sb->wqh, &__wait);
 582                 } else {
 583                         sb->parked++;
 584                         __add_wait_queue_tail(&sb->wqh, &__wait);
 585                 }
 586                 __pwqr_sb_update_state(sb, -1);
 587
 588                 do {
 589                         set_current_state(TASK_INTERRUPTIBLE);
 590                         if (sb->overcommit_wakes)
 591                                 break;
 592                         if (signal_pending(current)) {
 593                                 rc = -ERESTARTSYS;
 594                                 break;
 595                         }
 596                         spin_unlock_irq(&sb->wqh.lock);
 597                         schedule();
 598                         spin_lock_irq(&sb->wqh.lock);
 599                         if (is_wait)
 600                                 break;
 601                         if (sb->running + sb->waiting < sb->concurrency)
 602                                 break;
 603                 } while (likely(sb->state >= 0));
 604
 605                 __remove_wait_queue(&sb->wqh, &__wait);
 606                 __set_current_state(TASK_RUNNING);
 607                 if (is_wait) {
 608                         sb->waiting--;
 609                 } else {
 610                         sb->parked--;
 611                 }
 612                 __pwqr_sb_update_state(sb, 1);
 613
 614                 if (sb->overcommit_wakes)
 615                         sb->overcommit_wakes--;
 616                 if (sb->waiting + sb->running > sb->concurrency)
 617                         rc = -EDQUOT;
 618         }
 619
 620 out_unlock:
 621         if (unlikely(sb->state < 0))
 622                 rc = -EBADFD;
 623         pwqr_sb_unlock_irqrestore(sb, flags);
 624 out:
 625         preempt_notifier_register(&pwqt->notifier);
 626         return rc;
 627 }
 628
 629 static long do_pwqr_unregister(struct pwqr_sb *sb, struct pwqr_task *pwqt)
 630 {
 631         if (!pwqt)
 632                 return -EINVAL;
 633         if (pwqt->sb != sb)
 634                 return -ENOENT;
 635         pwqr_task_detach(pwqt, sb);
 636         pwqr_task_release(pwqt, false);
 637         return 0;
 638 }
 639
 640 static long do_pwqr_set_conc(struct pwqr_sb *sb, int conc)
 641 {
 642         long old_conc = sb->concurrency;
 643         unsigned long flags;
 644
 645         pwqr_sb_lock_irqsave(sb, flags);
 646         if (conc <= 0)
 647                 conc = num_online_cpus();
 648         if (conc != old_conc) {
 649                 sb->concurrency = conc;
 650                 __pwqr_sb_update_state(sb, 0);
 651         }
 652         pwqr_sb_unlock_irqrestore(sb, flags);
 653
 654         return old_conc;
 655 }
 656
 657 static long do_pwqr_wake(struct pwqr_sb *sb, int oc, int count)
 658 {
 659         unsigned long flags;
 660         int nwake;
 661
 662         if (count < 0)
 663                 return -EINVAL;
 664
 665         pwqr_sb_lock_irqsave(sb, flags);
 666
 667         if (oc) {
 668                 nwake = sb->waiting + sb->parked - sb->overcommit_wakes;
 669                 if (count > nwake) {
 670                         count = nwake;
 671                 } else {
 672                         nwake = count;
 673                 }
 674                 sb->overcommit_wakes += count;
 675         } else if (sb->running + sb->overcommit_wakes < sb->concurrency) {
 676                 nwake = sb->concurrency - sb->overcommit_wakes - sb->running;
 677                 if (nwake > sb->waiting + sb->parked - sb->overcommit_wakes) {
 678                         nwake = sb->waiting + sb->parked -
 679                                 sb->overcommit_wakes;
 680                 }
 681                 if (count > nwake) {
 682                         count = nwake;
 683                 } else {
 684                         nwake = count;
 685                 }
 686         } else {
 687                 /*
 688                  * This codepath deserves an explanation: waking the thread
 689                  * "for real" would overcommit, though userspace KNOWS there
 690                  * is at least one waiting thread. Such threads are threads
 691                  * that are "quarantined".
 692                  *
 693                  * Quarantined threads are woken up one by one, to allow a
 694                  * slow ramp down, trying to minimize "waiting" <-> "parked"
 695                  * flip-flops, no matter how many wakes have been asked.
 696                  *
 697                  * Since releasing one quarantined thread will wake up a
 698                  * thread that will (almost) straight go to parked mode, lie
 699                  * to userland about the fact that we unblocked that thread,
 700                  * and return 0.
 701                  *
 702                  * Though if we're already waking all waiting threads for
 703                  * overcommitting jobs, well, we don't need that.
 704                  */
 705                 count = 0;
 706                 nwake = sb->waiting > sb->overcommit_wakes;
 707         }
 708         while (nwake-- > 0)
 709                 wake_up_locked(&sb->wqh);
 710         pwqr_sb_unlock_irqrestore(sb, flags);
 711
 712         return count;
 713 }
 714
 715 static long pwqr_ioctl(struct file *filp, unsigned command, unsigned long arg)
 716 {
 717         struct pwqr_sb     *sb   = filp->private_data;
 718         struct task_struct *task = current;
 719         struct pwqr_task   *pwqt;
 720         int rc = 0;
 721
 722         switch (command) {
 723         case PWQR_CTL_GET_CONC:
 724                 return sb->concurrency;
 725         case PWQR_CTL_SET_CONC:
 726                 return do_pwqr_set_conc(sb, (int)arg);
 727
 728         case PWQR_CTL_WAKE:
 729         case PWQR_CTL_WAKE_OC:
 730                 return do_pwqr_wake(sb, command == PWQR_CTL_WAKE_OC, (int)arg);
 731
 732         case PWQR_CTL_WAIT:
 733         case PWQR_CTL_PARK:
 734         case PWQR_CTL_REGISTER:
 735         case PWQR_CTL_UNREGISTER:
 736                 break;
 737         default:
 738                 return -EINVAL;
 739         }
 740
 741         pwqt = pwqr_task_find(task);
 742         if (command == PWQR_CTL_UNREGISTER)
 743                 return do_pwqr_unregister(sb, pwqt);
 744
 745         if (pwqt == NULL) {
 746                 pwqt = pwqr_task_create(task);
 747                 if (IS_ERR(pwqt))
 748                         return PTR_ERR(pwqt);
 749                 pwqr_task_attach(pwqt, sb);
 750         } else if (unlikely(pwqt->sb != sb)) {
 751                 pwqr_task_detach(pwqt, pwqt->sb);
 752                 pwqr_task_attach(pwqt, sb);
 753         }
 754
 755         switch (command) {
 756         case PWQR_CTL_WAIT:
 757                 rc = do_pwqr_wait(sb, pwqt, true, (struct pwqr_ioc_wait __user *)arg);
 758                 break;
 759         case PWQR_CTL_PARK:
 760                 rc = do_pwqr_wait(sb, pwqt, false, NULL);
 761                 break;
 762         }
 763         return rc;
 764 }
 765
 766 static const struct file_operations pwqr_dev_fops = {
 767         .owner          = THIS_MODULE,
 768         .open           = pwqr_open,
 769         .release        = pwqr_release,
 770         .poll           = pwqr_poll,
 771         .read           = pwqr_read,
 772         .llseek         = noop_llseek,
 773         .unlocked_ioctl = pwqr_ioctl,
 774 #ifdef CONFIG_COMPAT
 775         .compat_ioctl   = pwqr_ioctl,
 776 #endif
 777 };
 778
 779 /*****************************************************************************
 780  * module
 781  */
 782 static int __init pwqr_start(void)
 783 {
 784 #if IS_PRE_2_6_23
 785         int i;
 786
 787         for (i = 0; i < PWQR_HASH_SIZE; i++) {
 788                 spin_lock_init(&pwqr_tasks_hash[i].lock);
 789                 INIT_HLIST_HEAD(&pwqr_tasks_hash[i].tasks);
 790         }
 791 #endif
 792
 793         /* Register as a character device */
 794         pwqr_major = register_chrdev(0, "pwqr", &pwqr_dev_fops);
 795         if (pwqr_major < 0) {
 796                 printk(KERN_ERR "pwqr: register_chrdev() failed\n");
 797                 return pwqr_major;
 798         }
 799
 800         /* Create a device node */
 801         pwqr_class = class_create(THIS_MODULE, PWQR_DEVICE_NAME);
 802         if (IS_ERR(pwqr_class)) {
 803                 printk(KERN_ERR "pwqr: Error creating raw class\n");
 804                 unregister_chrdev(pwqr_major, PWQR_DEVICE_NAME);
 805                 return PTR_ERR(pwqr_class);
 806         }
 807         device_create(pwqr_class, NULL, MKDEV(pwqr_major, 0), NULL, PWQR_DEVICE_NAME);
 808         printk(KERN_INFO "pwqr: PThreads Work Queues Regulator v1 loaded");
 809         return 0;
 810 }
 811
 812 static void __exit pwqr_end(void)
 813 {
 814         rcu_barrier();
 815         device_destroy(pwqr_class, MKDEV(pwqr_major, 0));
 816         class_destroy(pwqr_class);
 817         unregister_chrdev(pwqr_major, PWQR_DEVICE_NAME);
 818 }
 819
 820 module_init(pwqr_start);
 821 module_exit(pwqr_end);
 822
 823 MODULE_LICENSE("GPL");
 824 MODULE_AUTHOR("Pierre Habouzit <pierre.habouzit@intersec.com>");
 825 MODULE_DESCRIPTION("PThreads Work Queues Regulator");
 826 #endif
 827
 828 // vim:noet:sw=8:cinoptions+=\:0,L-1,=1s: