]> git.saurik.com Git - apple/xnu.git/blob - bsd/miscfs/specfs/spec_vnops.c
projects / apple / xnu.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
xnu-3789.41.3.tar.gz
[apple/xnu.git] / bsd / miscfs / specfs / spec_vnops.c
1 /*
2 * Copyright (c) 2000-2016 Apple Computer, Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28 /* Copyright (c) 1995 NeXT Computer, Inc. All Rights Reserved */
29 /*
30 * Copyright (c) 1989, 1993, 1995
31 * The Regents of the University of California. All rights reserved.
32 *
33 * Redistribution and use in source and binary forms, with or without
34 * modification, are permitted provided that the following conditions
35 * are met:
36 * 1. Redistributions of source code must retain the above copyright
37 * notice, this list of conditions and the following disclaimer.
38 * 2. Redistributions in binary form must reproduce the above copyright
39 * notice, this list of conditions and the following disclaimer in the
40 * documentation and/or other materials provided with the distribution.
41 * 3. All advertising materials mentioning features or use of this software
42 * must display the following acknowledgement:
43 * This product includes software developed by the University of
44 * California, Berkeley and its contributors.
45 * 4. Neither the name of the University nor the names of its contributors
46 * may be used to endorse or promote products derived from this software
47 * without specific prior written permission.
48 *
49 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
50 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
51 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
52 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
53 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
54 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
55 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
56 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
57 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
58 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
59 * SUCH DAMAGE.
60 *
61 * @(#)spec_vnops.c 8.14 (Berkeley) 5/21/95
62 */
63
64 #include <sys/param.h>
65 #include <sys/proc_internal.h>
66 #include <sys/kauth.h>
67 #include <sys/systm.h>
68 #include <sys/kernel.h>
69 #include <sys/conf.h>
70 #include <sys/buf_internal.h>
71 #include <sys/mount_internal.h>
72 #include <sys/vnode_internal.h>
73 #include <sys/file_internal.h>
74 #include <sys/namei.h>
75 #include <sys/stat.h>
76 #include <sys/errno.h>
77 #include <sys/ioctl.h>
78 #include <sys/file.h>
79 #include <sys/user.h>
80 #include <sys/malloc.h>
81 #include <sys/disk.h>
82 #include <sys/uio_internal.h>
83 #include <sys/resource.h>
84 #include <machine/machine_routines.h>
85 #include <miscfs/specfs/specdev.h>
86 #include <vfs/vfs_support.h>
87
88 #include <kern/assert.h>
89 #include <kern/task.h>
90 #include <kern/sched_prim.h>
91 #include <kern/thread.h>
92 #include <kern/policy_internal.h>
93 #include <kern/timer_call.h>
94
95 #include <pexpert/pexpert.h>
96
97 #include <sys/kdebug.h>
98
99 /* XXX following three prototypes should be in a header file somewhere */
100 extern dev_t chrtoblk(dev_t dev);
101 extern boolean_t iskmemdev(dev_t dev);
102 extern int bpfkqfilter(dev_t dev, struct knote *kn);
103 extern int ptsd_kqfilter(dev_t dev, struct knote *kn);
104
105 extern int ignore_is_ssd;
106
107 struct vnode *speclisth[SPECHSZ];
108
109 /* symbolic sleep message strings for devices */
110 char devopn[] = "devopn";
111 char devio[] = "devio";
112 char devwait[] = "devwait";
113 char devin[] = "devin";
114 char devout[] = "devout";
115 char devioc[] = "devioc";
116 char devcls[] = "devcls";
117
118 #define VOPFUNC int (*)(void *)
119
120 int (**spec_vnodeop_p)(void *);
121 struct vnodeopv_entry_desc spec_vnodeop_entries[] = {
122 { &vnop_default_desc, (VOPFUNC)vn_default_error },
123 { &vnop_lookup_desc, (VOPFUNC)spec_lookup }, /* lookup */
124 { &vnop_create_desc, (VOPFUNC)err_create }, /* create */
125 { &vnop_mknod_desc, (VOPFUNC)err_mknod }, /* mknod */
126 { &vnop_open_desc, (VOPFUNC)spec_open }, /* open */
127 { &vnop_close_desc, (VOPFUNC)spec_close }, /* close */
128 { &vnop_access_desc, (VOPFUNC)spec_access }, /* access */
129 { &vnop_getattr_desc, (VOPFUNC)spec_getattr }, /* getattr */
130 { &vnop_setattr_desc, (VOPFUNC)spec_setattr }, /* setattr */
131 { &vnop_read_desc, (VOPFUNC)spec_read }, /* read */
132 { &vnop_write_desc, (VOPFUNC)spec_write }, /* write */
133 { &vnop_ioctl_desc, (VOPFUNC)spec_ioctl }, /* ioctl */
134 { &vnop_select_desc, (VOPFUNC)spec_select }, /* select */
135 { &vnop_revoke_desc, (VOPFUNC)nop_revoke }, /* revoke */
136 { &vnop_mmap_desc, (VOPFUNC)err_mmap }, /* mmap */
137 { &vnop_fsync_desc, (VOPFUNC)spec_fsync }, /* fsync */
138 { &vnop_remove_desc, (VOPFUNC)err_remove }, /* remove */
139 { &vnop_link_desc, (VOPFUNC)err_link }, /* link */
140 { &vnop_rename_desc, (VOPFUNC)err_rename }, /* rename */
141 { &vnop_mkdir_desc, (VOPFUNC)err_mkdir }, /* mkdir */
142 { &vnop_rmdir_desc, (VOPFUNC)err_rmdir }, /* rmdir */
143 { &vnop_symlink_desc, (VOPFUNC)err_symlink }, /* symlink */
144 { &vnop_readdir_desc, (VOPFUNC)err_readdir }, /* readdir */
145 { &vnop_readlink_desc, (VOPFUNC)err_readlink }, /* readlink */
146 { &vnop_inactive_desc, (VOPFUNC)nop_inactive }, /* inactive */
147 { &vnop_reclaim_desc, (VOPFUNC)nop_reclaim }, /* reclaim */
148 { &vnop_strategy_desc, (VOPFUNC)spec_strategy }, /* strategy */
149 { &vnop_pathconf_desc, (VOPFUNC)spec_pathconf }, /* pathconf */
150 { &vnop_advlock_desc, (VOPFUNC)err_advlock }, /* advlock */
151 { &vnop_bwrite_desc, (VOPFUNC)spec_bwrite }, /* bwrite */
152 { &vnop_pagein_desc, (VOPFUNC)err_pagein }, /* Pagein */
153 { &vnop_pageout_desc, (VOPFUNC)err_pageout }, /* Pageout */
154 { &vnop_copyfile_desc, (VOPFUNC)err_copyfile }, /* Copyfile */
155 { &vnop_blktooff_desc, (VOPFUNC)spec_blktooff }, /* blktooff */
156 { &vnop_offtoblk_desc, (VOPFUNC)spec_offtoblk }, /* offtoblk */
157 { &vnop_blockmap_desc, (VOPFUNC)spec_blockmap }, /* blockmap */
158 { (struct vnodeop_desc*)NULL, (int(*)())NULL }
159 };
160 struct vnodeopv_desc spec_vnodeop_opv_desc =
161 { &spec_vnodeop_p, spec_vnodeop_entries };
162
163
164 static void set_blocksize(vnode_t, dev_t);
165
166 #define LOWPRI_TIER1_WINDOW_MSECS 25
167 #define LOWPRI_TIER2_WINDOW_MSECS 100
168 #define LOWPRI_TIER3_WINDOW_MSECS 500
169
170 #define LOWPRI_TIER1_IO_PERIOD_MSECS 40
171 #define LOWPRI_TIER2_IO_PERIOD_MSECS 85
172 #define LOWPRI_TIER3_IO_PERIOD_MSECS 200
173
174 #define LOWPRI_TIER1_IO_PERIOD_SSD_MSECS 5
175 #define LOWPRI_TIER2_IO_PERIOD_SSD_MSECS 15
176 #define LOWPRI_TIER3_IO_PERIOD_SSD_MSECS 25
177
178
179 int throttle_windows_msecs[THROTTLE_LEVEL_END + 1] = {
180 0,
181 LOWPRI_TIER1_WINDOW_MSECS,
182 LOWPRI_TIER2_WINDOW_MSECS,
183 LOWPRI_TIER3_WINDOW_MSECS,
184 };
185
186 int throttle_io_period_msecs[THROTTLE_LEVEL_END + 1] = {
187 0,
188 LOWPRI_TIER1_IO_PERIOD_MSECS,
189 LOWPRI_TIER2_IO_PERIOD_MSECS,
190 LOWPRI_TIER3_IO_PERIOD_MSECS,
191 };
192
193 int throttle_io_period_ssd_msecs[THROTTLE_LEVEL_END + 1] = {
194 0,
195 LOWPRI_TIER1_IO_PERIOD_SSD_MSECS,
196 LOWPRI_TIER2_IO_PERIOD_SSD_MSECS,
197 LOWPRI_TIER3_IO_PERIOD_SSD_MSECS,
198 };
199
200
201 int throttled_count[THROTTLE_LEVEL_END + 1];
202
203 struct _throttle_io_info_t {
204 lck_mtx_t throttle_lock;
205
206 struct timeval throttle_last_write_timestamp;
207 struct timeval throttle_min_timer_deadline;
208 struct timeval throttle_window_start_timestamp[THROTTLE_LEVEL_END + 1]; /* window starts at both the beginning and completion of an I/O */
209 struct timeval throttle_last_IO_timestamp[THROTTLE_LEVEL_END + 1];
210 pid_t throttle_last_IO_pid[THROTTLE_LEVEL_END + 1];
211 struct timeval throttle_start_IO_period_timestamp[THROTTLE_LEVEL_END + 1];
212 int32_t throttle_inflight_count[THROTTLE_LEVEL_END + 1];
213
214 TAILQ_HEAD( , uthread) throttle_uthlist[THROTTLE_LEVEL_END + 1]; /* Lists of throttled uthreads */
215 int throttle_next_wake_level;
216
217 thread_call_t throttle_timer_call;
218 int32_t throttle_timer_ref;
219 int32_t throttle_timer_active;
220
221 int32_t throttle_io_count;
222 int32_t throttle_io_count_begin;
223 int *throttle_io_periods;
224 uint32_t throttle_io_period_num;
225
226 int32_t throttle_refcnt;
227 int32_t throttle_alloc;
228 int32_t throttle_disabled;
229 int32_t throttle_is_fusion_with_priority;
230 };
231
232 struct _throttle_io_info_t _throttle_io_info[LOWPRI_MAX_NUM_DEV];
233
234
235 int lowpri_throttle_enabled = 1;
236
237
238 static void throttle_info_end_io_internal(struct _throttle_io_info_t *info, int throttle_level);
239 static int throttle_info_update_internal(struct _throttle_io_info_t *info, uthread_t ut, int flags, boolean_t isssd, boolean_t inflight, struct bufattr *bap);
240 static int throttle_get_thread_throttle_level(uthread_t ut);
241 static int throttle_get_thread_throttle_level_internal(uthread_t ut, int io_tier);
242
243 /*
244 * Trivial lookup routine that always fails.
245 */
246 int
247 spec_lookup(struct vnop_lookup_args *ap)
248 {
249
250 *ap->a_vpp = NULL;
251 return (ENOTDIR);
252 }
253
254 static void
255 set_blocksize(struct vnode *vp, dev_t dev)
256 {
257 int (*size)(dev_t);
258 int rsize;
259
260 if ((major(dev) < nblkdev) && (size = bdevsw[major(dev)].d_psize)) {
261 rsize = (*size)(dev);
262 if (rsize <= 0) /* did size fail? */
263 vp->v_specsize = DEV_BSIZE;
264 else
265 vp->v_specsize = rsize;
266 }
267 else
268 vp->v_specsize = DEV_BSIZE;
269 }
270
271 void
272 set_fsblocksize(struct vnode *vp)
273 {
274
275 if (vp->v_type == VBLK) {
276 dev_t dev = (dev_t)vp->v_rdev;
277 int maj = major(dev);
278
279 if ((u_int)maj >= (u_int)nblkdev)
280 return;
281
282 vnode_lock(vp);
283 set_blocksize(vp, dev);
284 vnode_unlock(vp);
285 }
286
287 }
288
289
290 /*
291 * Open a special file.
292 */
293 int
294 spec_open(struct vnop_open_args *ap)
295 {
296 struct proc *p = vfs_context_proc(ap->a_context);
297 kauth_cred_t cred = vfs_context_ucred(ap->a_context);
298 struct vnode *vp = ap->a_vp;
299 dev_t bdev, dev = (dev_t)vp->v_rdev;
300 int maj = major(dev);
301 int error;
302
303 /*
304 * Don't allow open if fs is mounted -nodev.
305 */
306 if (vp->v_mount && (vp->v_mount->mnt_flag & MNT_NODEV))
307 return (ENXIO);
308
309 switch (vp->v_type) {
310
311 case VCHR:
312 if ((u_int)maj >= (u_int)nchrdev)
313 return (ENXIO);
314 if (cred != FSCRED && (ap->a_mode & FWRITE)) {
315 /*
316 * When running in very secure mode, do not allow
317 * opens for writing of any disk character devices.
318 */
319 if (securelevel >= 2 && isdisk(dev, VCHR))
320 return (EPERM);
321
322 /* Never allow writing to /dev/mem or /dev/kmem */
323 if (iskmemdev(dev))
324 return (EPERM);
325 /*
326 * When running in secure mode, do not allow opens for
327 * writing of character devices whose corresponding block
328 * devices are currently mounted.
329 */
330 if (securelevel >= 1) {
331 if ((bdev = chrtoblk(dev)) != NODEV && check_mountedon(bdev, VBLK, &error))
332 return (error);
333 }
334 }
335
336 devsw_lock(dev, S_IFCHR);
337 error = (*cdevsw[maj].d_open)(dev, ap->a_mode, S_IFCHR, p);
338
339 if (error == 0) {
340 vp->v_specinfo->si_opencount++;
341 }
342
343 devsw_unlock(dev, S_IFCHR);
344
345 if (error == 0 && cdevsw[maj].d_type == D_DISK && !vp->v_un.vu_specinfo->si_initted) {
346 int isssd = 0;
347 uint64_t throttle_mask = 0;
348 uint32_t devbsdunit = 0;
349
350 if (VNOP_IOCTL(vp, DKIOCGETTHROTTLEMASK, (caddr_t)&throttle_mask, 0, NULL) == 0) {
351
352 if (throttle_mask != 0 &&
353 VNOP_IOCTL(vp, DKIOCISSOLIDSTATE, (caddr_t)&isssd, 0, ap->a_context) == 0) {
354 /*
355 * as a reasonable approximation, only use the lowest bit of the mask
356 * to generate a disk unit number
357 */
358 devbsdunit = num_trailing_0(throttle_mask);
359
360 vnode_lock(vp);
361
362 vp->v_un.vu_specinfo->si_isssd = isssd;
363 vp->v_un.vu_specinfo->si_devbsdunit = devbsdunit;
364 vp->v_un.vu_specinfo->si_throttle_mask = throttle_mask;
365 vp->v_un.vu_specinfo->si_throttleable = 1;
366 vp->v_un.vu_specinfo->si_initted = 1;
367
368 vnode_unlock(vp);
369 }
370 }
371 if (vp->v_un.vu_specinfo->si_initted == 0) {
372 vnode_lock(vp);
373 vp->v_un.vu_specinfo->si_initted = 1;
374 vnode_unlock(vp);
375 }
376 }
377 return (error);
378
379 case VBLK:
380 if ((u_int)maj >= (u_int)nblkdev)
381 return (ENXIO);
382 /*
383 * When running in very secure mode, do not allow
384 * opens for writing of any disk block devices.
385 */
386 if (securelevel >= 2 && cred != FSCRED &&
387 (ap->a_mode & FWRITE) && bdevsw[maj].d_type == D_DISK)
388 return (EPERM);
389 /*
390 * Do not allow opens of block devices that are
391 * currently mounted.
392 */
393 if ( (error = vfs_mountedon(vp)) )
394 return (error);
395
396 devsw_lock(dev, S_IFBLK);
397 error = (*bdevsw[maj].d_open)(dev, ap->a_mode, S_IFBLK, p);
398 if (!error) {
399 vp->v_specinfo->si_opencount++;
400 }
401 devsw_unlock(dev, S_IFBLK);
402
403 if (!error) {
404 u_int64_t blkcnt;
405 u_int32_t blksize;
406 int setsize = 0;
407 u_int32_t size512 = 512;
408
409
410 if (!VNOP_IOCTL(vp, DKIOCGETBLOCKSIZE, (caddr_t)&blksize, 0, ap->a_context)) {
411 /* Switch to 512 byte sectors (temporarily) */
412
413 if (!VNOP_IOCTL(vp, DKIOCSETBLOCKSIZE, (caddr_t)&size512, FWRITE, ap->a_context)) {
414 /* Get the number of 512 byte physical blocks. */
415 if (!VNOP_IOCTL(vp, DKIOCGETBLOCKCOUNT, (caddr_t)&blkcnt, 0, ap->a_context)) {
416 setsize = 1;
417 }
418 }
419 /* If it doesn't set back, we can't recover */
420 if (VNOP_IOCTL(vp, DKIOCSETBLOCKSIZE, (caddr_t)&blksize, FWRITE, ap->a_context))
421 error = ENXIO;
422 }
423
424
425 vnode_lock(vp);
426 set_blocksize(vp, dev);
427
428 /*
429 * Cache the size in bytes of the block device for later
430 * use by spec_write().
431 */
432 if (setsize)
433 vp->v_specdevsize = blkcnt * (u_int64_t)size512;
434 else
435 vp->v_specdevsize = (u_int64_t)0; /* Default: Can't get */
436
437 vnode_unlock(vp);
438
439 }
440 return(error);
441 default:
442 panic("spec_open type");
443 }
444 return (0);
445 }
446
447 /*
448 * Vnode op for read
449 */
450 int
451 spec_read(struct vnop_read_args *ap)
452 {
453 struct vnode *vp = ap->a_vp;
454 struct uio *uio = ap->a_uio;
455 struct buf *bp;
456 daddr64_t bn, nextbn;
457 long bsize, bscale;
458 int devBlockSize=0;
459 int n, on;
460 int error = 0;
461 dev_t dev;
462
463 #if DIAGNOSTIC
464 if (uio->uio_rw != UIO_READ)
465 panic("spec_read mode");
466 if (UIO_SEG_IS_USER_SPACE(uio->uio_segflg))
467 panic("spec_read proc");
468 #endif
469 if (uio_resid(uio) == 0)
470 return (0);
471
472 switch (vp->v_type) {
473
474 case VCHR:
475 {
476 struct _throttle_io_info_t *throttle_info = NULL;
477 int thread_throttle_level;
478 if (cdevsw[major(vp->v_rdev)].d_type == D_DISK && vp->v_un.vu_specinfo->si_throttleable) {
479 throttle_info = &_throttle_io_info[vp->v_un.vu_specinfo->si_devbsdunit];
480 thread_throttle_level = throttle_info_update_internal(throttle_info, NULL, 0, vp->v_un.vu_specinfo->si_isssd, TRUE, NULL);
481 }
482 error = (*cdevsw[major(vp->v_rdev)].d_read)
483 (vp->v_rdev, uio, ap->a_ioflag);
484
485 if (throttle_info) {
486 throttle_info_end_io_internal(throttle_info, thread_throttle_level);
487 }
488
489 return (error);
490 }
491
492 case VBLK:
493 if (uio->uio_offset < 0)
494 return (EINVAL);
495
496 dev = vp->v_rdev;
497
498 devBlockSize = vp->v_specsize;
499
500 if (devBlockSize > PAGE_SIZE)
501 return (EINVAL);
502
503 bscale = PAGE_SIZE / devBlockSize;
504 bsize = bscale * devBlockSize;
505
506 do {
507 on = uio->uio_offset % bsize;
508
509 bn = (daddr64_t)((uio->uio_offset / devBlockSize) &~ (bscale - 1));
510
511 if (vp->v_speclastr + bscale == bn) {
512 nextbn = bn + bscale;
513 error = buf_breadn(vp, bn, (int)bsize, &nextbn,
514 (int *)&bsize, 1, NOCRED, &bp);
515 } else
516 error = buf_bread(vp, bn, (int)bsize, NOCRED, &bp);
517
518 vnode_lock(vp);
519 vp->v_speclastr = bn;
520 vnode_unlock(vp);
521
522 n = bsize - buf_resid(bp);
523 if ((on > n) || error) {
524 if (!error)
525 error = EINVAL;
526 buf_brelse(bp);
527 return (error);
528 }
529 n = min((unsigned)(n - on), uio_resid(uio));
530
531 error = uiomove((char *)buf_dataptr(bp) + on, n, uio);
532 if (n + on == bsize)
533 buf_markaged(bp);
534 buf_brelse(bp);
535 } while (error == 0 && uio_resid(uio) > 0 && n != 0);
536 return (error);
537
538 default:
539 panic("spec_read type");
540 }
541 /* NOTREACHED */
542
543 return (0);
544 }
545
546 /*
547 * Vnode op for write
548 */
549 int
550 spec_write(struct vnop_write_args *ap)
551 {
552 struct vnode *vp = ap->a_vp;
553 struct uio *uio = ap->a_uio;
554 struct buf *bp;
555 daddr64_t bn;
556 int bsize, blkmask, bscale;
557 int io_sync;
558 int devBlockSize=0;
559 int n, on;
560 int error = 0;
561 dev_t dev;
562
563 #if DIAGNOSTIC
564 if (uio->uio_rw != UIO_WRITE)
565 panic("spec_write mode");
566 if (UIO_SEG_IS_USER_SPACE(uio->uio_segflg))
567 panic("spec_write proc");
568 #endif
569
570 switch (vp->v_type) {
571
572 case VCHR:
573 {
574 struct _throttle_io_info_t *throttle_info = NULL;
575 int thread_throttle_level;
576 if (cdevsw[major(vp->v_rdev)].d_type == D_DISK && vp->v_un.vu_specinfo->si_throttleable) {
577 throttle_info = &_throttle_io_info[vp->v_un.vu_specinfo->si_devbsdunit];
578
579 thread_throttle_level = throttle_info_update_internal(throttle_info, NULL, 0, vp->v_un.vu_specinfo->si_isssd, TRUE, NULL);
580
581 microuptime(&throttle_info->throttle_last_write_timestamp);
582 }
583 error = (*cdevsw[major(vp->v_rdev)].d_write)
584 (vp->v_rdev, uio, ap->a_ioflag);
585
586 if (throttle_info) {
587 throttle_info_end_io_internal(throttle_info, thread_throttle_level);
588 }
589
590 return (error);
591 }
592
593 case VBLK:
594 if (uio_resid(uio) == 0)
595 return (0);
596 if (uio->uio_offset < 0)
597 return (EINVAL);
598
599 io_sync = (ap->a_ioflag & IO_SYNC);
600
601 dev = (vp->v_rdev);
602
603 devBlockSize = vp->v_specsize;
604 if (devBlockSize > PAGE_SIZE)
605 return(EINVAL);
606
607 bscale = PAGE_SIZE / devBlockSize;
608 blkmask = bscale - 1;
609 bsize = bscale * devBlockSize;
610
611
612 do {
613 bn = (daddr64_t)((uio->uio_offset / devBlockSize) &~ blkmask);
614 on = uio->uio_offset % bsize;
615
616 n = min((unsigned)(bsize - on), uio_resid(uio));
617
618 /*
619 * Use buf_getblk() as an optimization IFF:
620 *
621 * 1) We are reading exactly a block on a block
622 * aligned boundary
623 * 2) We know the size of the device from spec_open
624 * 3) The read doesn't span the end of the device
625 *
626 * Otherwise, we fall back on buf_bread().
627 */
628 if (n == bsize &&
629 vp->v_specdevsize != (u_int64_t)0 &&
630 (uio->uio_offset + (u_int64_t)n) > vp->v_specdevsize) {
631 /* reduce the size of the read to what is there */
632 n = (uio->uio_offset + (u_int64_t)n) - vp->v_specdevsize;
633 }
634
635 if (n == bsize)
636 bp = buf_getblk(vp, bn, bsize, 0, 0, BLK_WRITE);
637 else
638 error = (int)buf_bread(vp, bn, bsize, NOCRED, &bp);
639
640 /* Translate downstream error for upstream, if needed */
641 if (!error)
642 error = (int)buf_error(bp);
643 if (error) {
644 buf_brelse(bp);
645 return (error);
646 }
647 n = min(n, bsize - buf_resid(bp));
648
649 error = uiomove((char *)buf_dataptr(bp) + on, n, uio);
650 if (error) {
651 buf_brelse(bp);
652 return (error);
653 }
654 buf_markaged(bp);
655
656 if (io_sync)
657 error = buf_bwrite(bp);
658 else {
659 if ((n + on) == bsize)
660 error = buf_bawrite(bp);
661 else
662 error = buf_bdwrite(bp);
663 }
664 } while (error == 0 && uio_resid(uio) > 0 && n != 0);
665 return (error);
666
667 default:
668 panic("spec_write type");
669 }
670 /* NOTREACHED */
671
672 return (0);
673 }
674
675 /*
676 * Device ioctl operation.
677 */
678 int
679 spec_ioctl(struct vnop_ioctl_args *ap)
680 {
681 proc_t p = vfs_context_proc(ap->a_context);
682 dev_t dev = ap->a_vp->v_rdev;
683 int retval = 0;
684
685 KERNEL_DEBUG_CONSTANT(FSDBG_CODE(DBG_IOCTL, 0) | DBG_FUNC_START,
686 dev, ap->a_command, ap->a_fflag, ap->a_vp->v_type, 0);
687
688 switch (ap->a_vp->v_type) {
689
690 case VCHR:
691 retval = (*cdevsw[major(dev)].d_ioctl)(dev, ap->a_command, ap->a_data,
692 ap->a_fflag, p);
693 break;
694
695 case VBLK:
696 retval = (*bdevsw[major(dev)].d_ioctl)(dev, ap->a_command, ap->a_data, ap->a_fflag, p);
697 if (!retval && ap->a_command == DKIOCSETBLOCKSIZE)
698 ap->a_vp->v_specsize = *(uint32_t *)ap->a_data;
699 break;
700
701 default:
702 panic("spec_ioctl");
703 /* NOTREACHED */
704 }
705 KERNEL_DEBUG_CONSTANT(FSDBG_CODE(DBG_IOCTL, 0) | DBG_FUNC_END,
706 dev, ap->a_command, ap->a_fflag, retval, 0);
707
708 return (retval);
709 }
710
711 int
712 spec_select(struct vnop_select_args *ap)
713 {
714 proc_t p = vfs_context_proc(ap->a_context);
715 dev_t dev;
716
717 switch (ap->a_vp->v_type) {
718
719 default:
720 return (1); /* XXX */
721
722 case VCHR:
723 dev = ap->a_vp->v_rdev;
724 return (*cdevsw[major(dev)].d_select)(dev, ap->a_which, ap->a_wql, p);
725 }
726 }
727
728 static int filt_specattach(struct knote *kn);
729
730 int
731 spec_kqfilter(vnode_t vp, struct knote *kn)
732 {
733 dev_t dev;
734
735 assert(vnode_ischr(vp));
736
737 dev = vnode_specrdev(vp);
738
739 #if NETWORKING
740 /*
741 * Try a bpf device, as defined in bsd/net/bpf.c
742 * If it doesn't error out the attach, then it
743 * claimed it. Otherwise, fall through and try
744 * a regular spec attach.
745 */
746 int32_t tmp_flags = kn->kn_flags;
747 int64_t tmp_data = kn->kn_data;
748 int res;
749
750 res = bpfkqfilter(dev, kn);
751 if ((kn->kn_flags & EV_ERROR) == 0) {
752 return res;
753 }
754 kn->kn_flags = tmp_flags;
755 kn->kn_data = tmp_data;
756 #endif
757
758 /* Try to attach to other char special devices */
759 return filt_specattach(kn);
760 }
761
762 /*
763 * Synch buffers associated with a block device
764 */
765 int
766 spec_fsync_internal(vnode_t vp, int waitfor, __unused vfs_context_t context)
767 {
768 if (vp->v_type == VCHR)
769 return (0);
770 /*
771 * Flush all dirty buffers associated with a block device.
772 */
773 buf_flushdirtyblks(vp, (waitfor == MNT_WAIT || waitfor == MNT_DWAIT), 0, "spec_fsync");
774
775 return (0);
776 }
777
778 int
779 spec_fsync(struct vnop_fsync_args *ap)
780 {
781 return spec_fsync_internal(ap->a_vp, ap->a_waitfor, ap->a_context);
782 }
783
784
785 /*
786 * Just call the device strategy routine
787 */
788 void throttle_init(void);
789
790
791 #if 0
792 #define DEBUG_ALLOC_THROTTLE_INFO(format, debug_info, args...) \
793 do { \
794 if ((debug_info)->alloc) \
795 printf("%s: "format, __FUNCTION__, ## args); \
796 } while(0)
797
798 #else
799 #define DEBUG_ALLOC_THROTTLE_INFO(format, debug_info, args...)
800 #endif
801
802
803 SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier1_window_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_windows_msecs[THROTTLE_LEVEL_TIER1], 0, "");
804 SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier2_window_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_windows_msecs[THROTTLE_LEVEL_TIER2], 0, "");
805 SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier3_window_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_windows_msecs[THROTTLE_LEVEL_TIER3], 0, "");
806
807 SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier1_io_period_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_msecs[THROTTLE_LEVEL_TIER1], 0, "");
808 SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier2_io_period_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_msecs[THROTTLE_LEVEL_TIER2], 0, "");
809 SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier3_io_period_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_msecs[THROTTLE_LEVEL_TIER3], 0, "");
810
811 SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier1_io_period_ssd_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_ssd_msecs[THROTTLE_LEVEL_TIER1], 0, "");
812 SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier2_io_period_ssd_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_ssd_msecs[THROTTLE_LEVEL_TIER2], 0, "");
813 SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_tier3_io_period_ssd_msecs, CTLFLAG_RW | CTLFLAG_LOCKED, &throttle_io_period_ssd_msecs[THROTTLE_LEVEL_TIER3], 0, "");
814
815 SYSCTL_INT(_debug, OID_AUTO, lowpri_throttle_enabled, CTLFLAG_RW | CTLFLAG_LOCKED, &lowpri_throttle_enabled, 0, "");
816
817
818 static lck_grp_t *throttle_lock_grp;
819 static lck_attr_t *throttle_lock_attr;
820 static lck_grp_attr_t *throttle_lock_grp_attr;
821
822
823 /*
824 * throttled I/O helper function
825 * convert the index of the lowest set bit to a device index
826 */
827 int
828 num_trailing_0(uint64_t n)
829 {
830 /*
831 * since in most cases the number of trailing 0s is very small,
832 * we simply counting sequentially from the lowest bit
833 */
834 if (n == 0)
835 return sizeof(n) * 8;
836 int count = 0;
837 while (!ISSET(n, 1)) {
838 n >>= 1;
839 ++count;
840 }
841 return count;
842 }
843
844
845 /*
846 * Release the reference and if the item was allocated and this is the last
847 * reference then free it.
848 *
849 * This routine always returns the old value.
850 */
851 static int
852 throttle_info_rel(struct _throttle_io_info_t *info)
853 {
854 SInt32 oldValue = OSDecrementAtomic(&info->throttle_refcnt);
855
856 DEBUG_ALLOC_THROTTLE_INFO("refcnt = %d info = %p\n",
857 info, (int)(oldValue -1), info );
858
859 /* The reference count just went negative, very bad */
860 if (oldValue == 0)
861 panic("throttle info ref cnt went negative!");
862
863 /*
864 * Once reference count is zero, no one else should be able to take a
865 * reference
866 */
867 if ((info->throttle_refcnt == 0) && (info->throttle_alloc)) {
868 DEBUG_ALLOC_THROTTLE_INFO("Freeing info = %p\n", info);
869
870 lck_mtx_destroy(&info->throttle_lock, throttle_lock_grp);
871 FREE(info, M_TEMP);
872 }
873 return oldValue;
874 }
875
876
877 /*
878 * Just take a reference on the throttle info structure.
879 *
880 * This routine always returns the old value.
881 */
882 static SInt32
883 throttle_info_ref(struct _throttle_io_info_t *info)
884 {
885 SInt32 oldValue = OSIncrementAtomic(&info->throttle_refcnt);
886
887 DEBUG_ALLOC_THROTTLE_INFO("refcnt = %d info = %p\n",
888 info, (int)(oldValue -1), info );
889 /* Allocated items should never have a reference of zero */
890 if (info->throttle_alloc && (oldValue == 0))
891 panic("Taking a reference without calling create throttle info!\n");
892
893 return oldValue;
894 }
895
896 /*
897 * on entry the throttle_lock is held...
898 * this function is responsible for taking
899 * and dropping the reference on the info
900 * structure which will keep it from going
901 * away while the timer is running if it
902 * happens to have been dynamically allocated by
903 * a network fileystem kext which is now trying
904 * to free it
905 */
906 static uint32_t
907 throttle_timer_start(struct _throttle_io_info_t *info, boolean_t update_io_count, int wakelevel)
908 {
909 struct timeval elapsed;
910 struct timeval now;
911 struct timeval period;
912 uint64_t elapsed_msecs;
913 int throttle_level;
914 int level;
915 int msecs;
916 boolean_t throttled = FALSE;
917 boolean_t need_timer = FALSE;
918
919 microuptime(&now);
920
921 if (update_io_count == TRUE) {
922 info->throttle_io_count_begin = info->throttle_io_count;
923 info->throttle_io_period_num++;
924
925 while (wakelevel >= THROTTLE_LEVEL_THROTTLED)
926 info->throttle_start_IO_period_timestamp[wakelevel--] = now;
927
928 info->throttle_min_timer_deadline = now;
929
930 msecs = info->throttle_io_periods[THROTTLE_LEVEL_THROTTLED];
931 period.tv_sec = msecs / 1000;
932 period.tv_usec = (msecs % 1000) * 1000;
933
934 timevaladd(&info->throttle_min_timer_deadline, &period);
935 }
936 for (throttle_level = THROTTLE_LEVEL_START; throttle_level < THROTTLE_LEVEL_END; throttle_level++) {
937
938 elapsed = now;
939 timevalsub(&elapsed, &info->throttle_window_start_timestamp[throttle_level]);
940 elapsed_msecs = (uint64_t)elapsed.tv_sec * (uint64_t)1000 + (elapsed.tv_usec / 1000);
941
942 for (level = throttle_level + 1; level <= THROTTLE_LEVEL_END; level++) {
943
944 if (!TAILQ_EMPTY(&info->throttle_uthlist[level])) {
945
946 if (elapsed_msecs < (uint64_t)throttle_windows_msecs[level] || info->throttle_inflight_count[throttle_level]) {
947 /*
948 * we had an I/O occur at a higher priority tier within
949 * this tier's throttle window
950 */
951 throttled = TRUE;
952 }
953 /*
954 * we assume that the windows are the same or longer
955 * as we drop through the throttling tiers... thus
956 * we can stop looking once we run into a tier with
957 * threads to schedule regardless of whether it's
958 * still in its throttling window or not
959 */
960 break;
961 }
962 }
963 if (throttled == TRUE)
964 break;
965 }
966 if (throttled == TRUE) {
967 uint64_t deadline = 0;
968 struct timeval target;
969 struct timeval min_target;
970
971 /*
972 * we've got at least one tier still in a throttled window
973 * so we need a timer running... compute the next deadline
974 * and schedule it
975 */
976 for (level = throttle_level+1; level <= THROTTLE_LEVEL_END; level++) {
977
978 if (TAILQ_EMPTY(&info->throttle_uthlist[level]))
979 continue;
980
981 target = info->throttle_start_IO_period_timestamp[level];
982
983 msecs = info->throttle_io_periods[level];
984 period.tv_sec = msecs / 1000;
985 period.tv_usec = (msecs % 1000) * 1000;
986
987 timevaladd(&target, &period);
988
989 if (need_timer == FALSE || timevalcmp(&target, &min_target, <)) {
990 min_target = target;
991 need_timer = TRUE;
992 }
993 }
994 if (timevalcmp(&info->throttle_min_timer_deadline, &now, >)) {
995 if (timevalcmp(&info->throttle_min_timer_deadline, &min_target, >))
996 min_target = info->throttle_min_timer_deadline;
997 }
998
999 if (info->throttle_timer_active) {
1000 if (thread_call_cancel(info->throttle_timer_call) == FALSE) {
1001 /*
1002 * couldn't kill the timer because it's already
1003 * been dispatched, so don't try to start a new
1004 * one... once we drop the lock, the timer will
1005 * proceed and eventually re-run this function
1006 */
1007 need_timer = FALSE;
1008 } else
1009 info->throttle_timer_active = 0;
1010 }
1011 if (need_timer == TRUE) {
1012 /*
1013 * This is defined as an int (32-bit) rather than a 64-bit
1014 * value because it would need a really big period in the
1015 * order of ~500 days to overflow this. So, we let this be
1016 * 32-bit which allows us to use the clock_interval_to_deadline()
1017 * routine.
1018 */
1019 int target_msecs;
1020
1021 if (info->throttle_timer_ref == 0) {
1022 /*
1023 * take a reference for the timer
1024 */
1025 throttle_info_ref(info);
1026
1027 info->throttle_timer_ref = 1;
1028 }
1029 elapsed = min_target;
1030 timevalsub(&elapsed, &now);
1031 target_msecs = elapsed.tv_sec * 1000 + elapsed.tv_usec / 1000;
1032
1033 if (target_msecs <= 0) {
1034 /*
1035 * we may have computed a deadline slightly in the past
1036 * due to various factors... if so, just set the timer
1037 * to go off in the near future (we don't need to be precise)
1038 */
1039 target_msecs = 1;
1040 }
1041 clock_interval_to_deadline(target_msecs, 1000000, &deadline);
1042
1043 thread_call_enter_delayed(info->throttle_timer_call, deadline);
1044 info->throttle_timer_active = 1;
1045 }
1046 }
1047 return (throttle_level);
1048 }
1049
1050
1051 static void
1052 throttle_timer(struct _throttle_io_info_t *info)
1053 {
1054 uthread_t ut, utlist;
1055 struct timeval elapsed;
1056 struct timeval now;
1057 uint64_t elapsed_msecs;
1058 int throttle_level;
1059 int level;
1060 int wake_level;
1061 caddr_t wake_address = NULL;
1062 boolean_t update_io_count = FALSE;
1063 boolean_t need_wakeup = FALSE;
1064 boolean_t need_release = FALSE;
1065
1066 ut = NULL;
1067 lck_mtx_lock(&info->throttle_lock);
1068
1069 info->throttle_timer_active = 0;
1070 microuptime(&now);
1071
1072 elapsed = now;
1073 timevalsub(&elapsed, &info->throttle_start_IO_period_timestamp[THROTTLE_LEVEL_THROTTLED]);
1074 elapsed_msecs = (uint64_t)elapsed.tv_sec * (uint64_t)1000 + (elapsed.tv_usec / 1000);
1075
1076 if (elapsed_msecs >= (uint64_t)info->throttle_io_periods[THROTTLE_LEVEL_THROTTLED]) {
1077
1078 wake_level = info->throttle_next_wake_level;
1079
1080 for (level = THROTTLE_LEVEL_START; level < THROTTLE_LEVEL_END; level++) {
1081
1082 elapsed = now;
1083 timevalsub(&elapsed, &info->throttle_start_IO_period_timestamp[wake_level]);
1084 elapsed_msecs = (uint64_t)elapsed.tv_sec * (uint64_t)1000 + (elapsed.tv_usec / 1000);
1085
1086 if (elapsed_msecs >= (uint64_t)info->throttle_io_periods[wake_level] && !TAILQ_EMPTY(&info->throttle_uthlist[wake_level])) {
1087 /*
1088 * we're closing out the current IO period...
1089 * if we have a waiting thread, wake it up
1090 * after we have reset the I/O window info
1091 */
1092 need_wakeup = TRUE;
1093 update_io_count = TRUE;
1094
1095 info->throttle_next_wake_level = wake_level - 1;
1096
1097 if (info->throttle_next_wake_level == THROTTLE_LEVEL_START)
1098 info->throttle_next_wake_level = THROTTLE_LEVEL_END;
1099
1100 break;
1101 }
1102 wake_level--;
1103
1104 if (wake_level == THROTTLE_LEVEL_START)
1105 wake_level = THROTTLE_LEVEL_END;
1106 }
1107 }
1108 if (need_wakeup == TRUE) {
1109 if (!TAILQ_EMPTY(&info->throttle_uthlist[wake_level])) {
1110
1111 ut = (uthread_t)TAILQ_FIRST(&info->throttle_uthlist[wake_level]);
1112 TAILQ_REMOVE(&info->throttle_uthlist[wake_level], ut, uu_throttlelist);
1113 ut->uu_on_throttlelist = THROTTLE_LEVEL_NONE;
1114 ut->uu_is_throttled = FALSE;
1115
1116 wake_address = (caddr_t)&ut->uu_on_throttlelist;
1117 }
1118 } else
1119 wake_level = THROTTLE_LEVEL_START;
1120
1121 throttle_level = throttle_timer_start(info, update_io_count, wake_level);
1122
1123 if (wake_address != NULL)
1124 wakeup(wake_address);
1125
1126 for (level = THROTTLE_LEVEL_THROTTLED; level <= throttle_level; level++) {
1127
1128 TAILQ_FOREACH_SAFE(ut, &info->throttle_uthlist[level], uu_throttlelist, utlist) {
1129
1130 TAILQ_REMOVE(&info->throttle_uthlist[level], ut, uu_throttlelist);
1131 ut->uu_on_throttlelist = THROTTLE_LEVEL_NONE;
1132 ut->uu_is_throttled = FALSE;
1133
1134 wakeup(&ut->uu_on_throttlelist);
1135 }
1136 }
1137 if (info->throttle_timer_active == 0 && info->throttle_timer_ref) {
1138 info->throttle_timer_ref = 0;
1139 need_release = TRUE;
1140 }
1141 lck_mtx_unlock(&info->throttle_lock);
1142
1143 if (need_release == TRUE)
1144 throttle_info_rel(info);
1145 }
1146
1147
1148 static int
1149 throttle_add_to_list(struct _throttle_io_info_t *info, uthread_t ut, int mylevel, boolean_t insert_tail)
1150 {
1151 boolean_t start_timer = FALSE;
1152 int level = THROTTLE_LEVEL_START;
1153
1154 if (TAILQ_EMPTY(&info->throttle_uthlist[mylevel])) {
1155 info->throttle_start_IO_period_timestamp[mylevel] = info->throttle_last_IO_timestamp[mylevel];
1156 start_timer = TRUE;
1157 }
1158
1159 if (insert_tail == TRUE)
1160 TAILQ_INSERT_TAIL(&info->throttle_uthlist[mylevel], ut, uu_throttlelist);
1161 else
1162 TAILQ_INSERT_HEAD(&info->throttle_uthlist[mylevel], ut, uu_throttlelist);
1163
1164 ut->uu_on_throttlelist = mylevel;
1165
1166 if (start_timer == TRUE) {
1167 /* we may need to start or rearm the timer */
1168 level = throttle_timer_start(info, FALSE, THROTTLE_LEVEL_START);
1169
1170 if (level == THROTTLE_LEVEL_END) {
1171 if (ut->uu_on_throttlelist >= THROTTLE_LEVEL_THROTTLED) {
1172 TAILQ_REMOVE(&info->throttle_uthlist[ut->uu_on_throttlelist], ut, uu_throttlelist);
1173
1174 ut->uu_on_throttlelist = THROTTLE_LEVEL_NONE;
1175 }
1176 }
1177 }
1178 return (level);
1179 }
1180
1181 static void
1182 throttle_init_throttle_window(void)
1183 {
1184 int throttle_window_size;
1185
1186 /*
1187 * The hierarchy of throttle window values is as follows:
1188 * - Global defaults
1189 * - Device tree properties
1190 * - Boot-args
1191 * All values are specified in msecs.
1192 */
1193
1194 /* Override global values with device-tree properties */
1195 if (PE_get_default("kern.io_throttle_window_tier1", &throttle_window_size, sizeof(throttle_window_size)))
1196 throttle_windows_msecs[THROTTLE_LEVEL_TIER1] = throttle_window_size;
1197
1198 if (PE_get_default("kern.io_throttle_window_tier2", &throttle_window_size, sizeof(throttle_window_size)))
1199 throttle_windows_msecs[THROTTLE_LEVEL_TIER2] = throttle_window_size;
1200
1201 if (PE_get_default("kern.io_throttle_window_tier3", &throttle_window_size, sizeof(throttle_window_size)))
1202 throttle_windows_msecs[THROTTLE_LEVEL_TIER3] = throttle_window_size;
1203
1204 /* Override with boot-args */
1205 if (PE_parse_boot_argn("io_throttle_window_tier1", &throttle_window_size, sizeof(throttle_window_size)))
1206 throttle_windows_msecs[THROTTLE_LEVEL_TIER1] = throttle_window_size;
1207
1208 if (PE_parse_boot_argn("io_throttle_window_tier2", &throttle_window_size, sizeof(throttle_window_size)))
1209 throttle_windows_msecs[THROTTLE_LEVEL_TIER2] = throttle_window_size;
1210
1211 if (PE_parse_boot_argn("io_throttle_window_tier3", &throttle_window_size, sizeof(throttle_window_size)))
1212 throttle_windows_msecs[THROTTLE_LEVEL_TIER3] = throttle_window_size;
1213 }
1214
1215 static void
1216 throttle_init_throttle_period(struct _throttle_io_info_t *info, boolean_t isssd)
1217 {
1218 int throttle_period_size;
1219
1220 /*
1221 * The hierarchy of throttle period values is as follows:
1222 * - Global defaults
1223 * - Device tree properties
1224 * - Boot-args
1225 * All values are specified in msecs.
1226 */
1227
1228 /* Assign global defaults */
1229 if ((isssd == TRUE) && (info->throttle_is_fusion_with_priority == 0))
1230 info->throttle_io_periods = &throttle_io_period_ssd_msecs[0];
1231 else
1232 info->throttle_io_periods = &throttle_io_period_msecs[0];
1233
1234 /* Override global values with device-tree properties */
1235 if (PE_get_default("kern.io_throttle_period_tier1", &throttle_period_size, sizeof(throttle_period_size)))
1236 info->throttle_io_periods[THROTTLE_LEVEL_TIER1] = throttle_period_size;
1237
1238 if (PE_get_default("kern.io_throttle_period_tier2", &throttle_period_size, sizeof(throttle_period_size)))
1239 info->throttle_io_periods[THROTTLE_LEVEL_TIER2] = throttle_period_size;
1240
1241 if (PE_get_default("kern.io_throttle_period_tier3", &throttle_period_size, sizeof(throttle_period_size)))
1242 info->throttle_io_periods[THROTTLE_LEVEL_TIER3] = throttle_period_size;
1243
1244 /* Override with boot-args */
1245 if (PE_parse_boot_argn("io_throttle_period_tier1", &throttle_period_size, sizeof(throttle_period_size)))
1246 info->throttle_io_periods[THROTTLE_LEVEL_TIER1] = throttle_period_size;
1247
1248 if (PE_parse_boot_argn("io_throttle_period_tier2", &throttle_period_size, sizeof(throttle_period_size)))
1249 info->throttle_io_periods[THROTTLE_LEVEL_TIER2] = throttle_period_size;
1250
1251 if (PE_parse_boot_argn("io_throttle_period_tier3", &throttle_period_size, sizeof(throttle_period_size)))
1252 info->throttle_io_periods[THROTTLE_LEVEL_TIER3] = throttle_period_size;
1253
1254 }
1255
1256 #if CONFIG_IOSCHED
1257 extern void vm_io_reprioritize_init(void);
1258 int iosched_enabled = 1;
1259 #endif
1260
1261 void
1262 throttle_init(void)
1263 {
1264 struct _throttle_io_info_t *info;
1265 int i;
1266 int level;
1267 #if CONFIG_IOSCHED
1268 int iosched;
1269 #endif
1270 /*
1271 * allocate lock group attribute and group
1272 */
1273 throttle_lock_grp_attr = lck_grp_attr_alloc_init();
1274 throttle_lock_grp = lck_grp_alloc_init("throttle I/O", throttle_lock_grp_attr);
1275
1276 /* Update throttle parameters based on device tree configuration */
1277 throttle_init_throttle_window();
1278
1279 /*
1280 * allocate the lock attribute
1281 */
1282 throttle_lock_attr = lck_attr_alloc_init();
1283
1284 for (i = 0; i < LOWPRI_MAX_NUM_DEV; i++) {
1285 info = &_throttle_io_info[i];
1286
1287 lck_mtx_init(&info->throttle_lock, throttle_lock_grp, throttle_lock_attr);
1288 info->throttle_timer_call = thread_call_allocate((thread_call_func_t)throttle_timer, (thread_call_param_t)info);
1289
1290 for (level = 0; level <= THROTTLE_LEVEL_END; level++) {
1291 TAILQ_INIT(&info->throttle_uthlist[level]);
1292 info->throttle_last_IO_pid[level] = 0;
1293 info->throttle_inflight_count[level] = 0;
1294 }
1295 info->throttle_next_wake_level = THROTTLE_LEVEL_END;
1296 info->throttle_disabled = 0;
1297 info->throttle_is_fusion_with_priority = 0;
1298 }
1299 #if CONFIG_IOSCHED
1300 if (PE_parse_boot_argn("iosched", &iosched, sizeof(iosched))) {
1301 iosched_enabled = iosched;
1302 }
1303 if (iosched_enabled) {
1304 /* Initialize I/O Reprioritization mechanism */
1305 vm_io_reprioritize_init();
1306 }
1307 #endif
1308 }
1309
1310 void
1311 sys_override_io_throttle(int flag)
1312 {
1313 if (flag == THROTTLE_IO_ENABLE)
1314 lowpri_throttle_enabled = 1;
1315
1316 if (flag == THROTTLE_IO_DISABLE)
1317 lowpri_throttle_enabled = 0;
1318 }
1319
1320 int rethrottle_wakeups = 0;
1321
1322 /*
1323 * the uu_rethrottle_lock is used to synchronize this function
1324 * with "throttle_lowpri_io" which is where a throttled thread
1325 * will block... that function will grab this lock before beginning
1326 * it's decision making process concerning the need to block, and
1327 * hold it through the assert_wait. When that thread is awakened
1328 * for any reason (timer or rethrottle), it will reacquire the
1329 * uu_rethrottle_lock before determining if it really is ok for
1330 * it to now run. This is the point at which the thread could
1331 * enter a different throttling queue and reblock or return from
1332 * the throttle w/o having waited out it's entire throttle if
1333 * the rethrottle has now moved it out of any currently
1334 * active throttle window.
1335 *
1336 *
1337 * NOTES:
1338 * 1 - This may be called with the task lock held.
1339 * 2 - This may be called with preemption and interrupts disabled
1340 * in the kqueue wakeup path so we can't take the throttle_lock which is a mutex
1341 * 3 - This cannot safely dereference uu_throttle_info, as it may
1342 * get deallocated out from under us
1343 */
1344
1345 void
1346 rethrottle_thread(uthread_t ut)
1347 {
1348 /*
1349 * If uthread doesn't have throttle state, then there's no chance
1350 * of it needing a rethrottle.
1351 */
1352 if (ut->uu_throttle_info == NULL)
1353 return;
1354
1355 boolean_t s = ml_set_interrupts_enabled(FALSE);
1356 lck_spin_lock(&ut->uu_rethrottle_lock);
1357
1358 if (ut->uu_is_throttled == FALSE)
1359 ut->uu_was_rethrottled = TRUE;
1360 else {
1361 int my_new_level = throttle_get_thread_throttle_level(ut);
1362
1363 if (my_new_level != ut->uu_on_throttlelist) {
1364 /*
1365 * ut is currently blocked (as indicated by
1366 * ut->uu_is_throttled == TRUE)
1367 * and we're changing it's throttle level, so
1368 * we need to wake it up.
1369 */
1370 ut->uu_is_throttled = FALSE;
1371 wakeup(&ut->uu_on_throttlelist);
1372
1373 rethrottle_wakeups++;
1374 KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_FSRW, 102)), thread_tid(ut->uu_thread), ut->uu_on_throttlelist, my_new_level, 0, 0);
1375 }
1376 }
1377 lck_spin_unlock(&ut->uu_rethrottle_lock);
1378 ml_set_interrupts_enabled(s);
1379 }
1380
1381
1382 /*
1383 * KPI routine
1384 *
1385 * Create and take a reference on a throttle info structure and return a
1386 * pointer for the file system to use when calling throttle_info_update.
1387 * Calling file system must have a matching release for every create.
1388 */
1389 void *
1390 throttle_info_create(void)
1391 {
1392 struct _throttle_io_info_t *info;
1393 int level;
1394
1395 MALLOC(info, struct _throttle_io_info_t *, sizeof(*info), M_TEMP, M_ZERO | M_WAITOK);
1396 /* Should never happen but just in case */
1397 if (info == NULL)
1398 return NULL;
1399 /* Mark that this one was allocated and needs to be freed */
1400 DEBUG_ALLOC_THROTTLE_INFO("Creating info = %p\n", info, info );
1401 info->throttle_alloc = TRUE;
1402
1403 lck_mtx_init(&info->throttle_lock, throttle_lock_grp, throttle_lock_attr);
1404 info->throttle_timer_call = thread_call_allocate((thread_call_func_t)throttle_timer, (thread_call_param_t)info);
1405
1406 for (level = 0; level <= THROTTLE_LEVEL_END; level++) {
1407 TAILQ_INIT(&info->throttle_uthlist[level]);
1408 }
1409 info->throttle_next_wake_level = THROTTLE_LEVEL_END;
1410
1411 /* Take a reference */
1412 OSIncrementAtomic(&info->throttle_refcnt);
1413 return info;
1414 }
1415
1416 /*
1417 * KPI routine
1418 *
1419 * Release the throttle info pointer if all the reference are gone. Should be
1420 * called to release reference taken by throttle_info_create
1421 */
1422 void
1423 throttle_info_release(void *throttle_info)
1424 {
1425 DEBUG_ALLOC_THROTTLE_INFO("Releaseing info = %p\n",
1426 (struct _throttle_io_info_t *)throttle_info,
1427 (struct _throttle_io_info_t *)throttle_info);
1428 if (throttle_info) /* Just to be careful */
1429 throttle_info_rel(throttle_info);
1430 }
1431
1432 /*
1433 * KPI routine
1434 *
1435 * File Systems that create an info structure, need to call this routine in
1436 * their mount routine (used by cluster code). File Systems that call this in
1437 * their mount routines must call throttle_info_mount_rel in their unmount
1438 * routines.
1439 */
1440 void
1441 throttle_info_mount_ref(mount_t mp, void *throttle_info)
1442 {
1443 if ((throttle_info == NULL) || (mp == NULL))
1444 return;
1445 throttle_info_ref(throttle_info);
1446
1447 /*
1448 * We already have a reference release it before adding the new one
1449 */
1450 if (mp->mnt_throttle_info)
1451 throttle_info_rel(mp->mnt_throttle_info);
1452 mp->mnt_throttle_info = throttle_info;
1453 }
1454
1455 /*
1456 * Private KPI routine
1457 *
1458 * return a handle for accessing throttle_info given a throttle_mask. The
1459 * handle must be released by throttle_info_rel_by_mask
1460 */
1461 int
1462 throttle_info_ref_by_mask(uint64_t throttle_mask, throttle_info_handle_t *throttle_info_handle)
1463 {
1464 int dev_index;
1465 struct _throttle_io_info_t *info;
1466
1467 if (throttle_info_handle == NULL)
1468 return EINVAL;
1469
1470 dev_index = num_trailing_0(throttle_mask);
1471 info = &_throttle_io_info[dev_index];
1472 throttle_info_ref(info);
1473 *(struct _throttle_io_info_t**)throttle_info_handle = info;
1474
1475 return 0;
1476 }
1477
1478 /*
1479 * Private KPI routine
1480 *
1481 * release the handle obtained by throttle_info_ref_by_mask
1482 */
1483 void
1484 throttle_info_rel_by_mask(throttle_info_handle_t throttle_info_handle)
1485 {
1486 /*
1487 * for now the handle is just a pointer to _throttle_io_info_t
1488 */
1489 throttle_info_rel((struct _throttle_io_info_t*)throttle_info_handle);
1490 }
1491
1492 /*
1493 * KPI routine
1494 *
1495 * File Systems that throttle_info_mount_ref, must call this routine in their
1496 * umount routine.
1497 */
1498 void
1499 throttle_info_mount_rel(mount_t mp)
1500 {
1501 if (mp->mnt_throttle_info)
1502 throttle_info_rel(mp->mnt_throttle_info);
1503 mp->mnt_throttle_info = NULL;
1504 }
1505
1506 void
1507 throttle_info_get_last_io_time(mount_t mp, struct timeval *tv)
1508 {
1509 struct _throttle_io_info_t *info;
1510
1511 if (mp == NULL)
1512 info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
1513 else if (mp->mnt_throttle_info == NULL)
1514 info = &_throttle_io_info[mp->mnt_devbsdunit];
1515 else
1516 info = mp->mnt_throttle_info;
1517
1518 *tv = info->throttle_last_write_timestamp;
1519 }
1520
1521 void
1522 update_last_io_time(mount_t mp)
1523 {
1524 struct _throttle_io_info_t *info;
1525
1526 if (mp == NULL)
1527 info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
1528 else if (mp->mnt_throttle_info == NULL)
1529 info = &_throttle_io_info[mp->mnt_devbsdunit];
1530 else
1531 info = mp->mnt_throttle_info;
1532
1533 microuptime(&info->throttle_last_write_timestamp);
1534 if (mp != NULL)
1535 mp->mnt_last_write_completed_timestamp = info->throttle_last_write_timestamp;
1536 }
1537
1538
1539 int
1540 throttle_get_io_policy(uthread_t *ut)
1541 {
1542 if (ut != NULL)
1543 *ut = get_bsdthread_info(current_thread());
1544
1545 return (proc_get_effective_thread_policy(current_thread(), TASK_POLICY_IO));
1546 }
1547
1548 int
1549 throttle_get_passive_io_policy(uthread_t *ut)
1550 {
1551 if (ut != NULL)
1552 *ut = get_bsdthread_info(current_thread());
1553
1554 return (proc_get_effective_thread_policy(current_thread(), TASK_POLICY_PASSIVE_IO));
1555 }
1556
1557
1558 static int
1559 throttle_get_thread_throttle_level(uthread_t ut)
1560 {
1561 uthread_t *ut_p = (ut == NULL) ? &ut : NULL;
1562 int io_tier = throttle_get_io_policy(ut_p);
1563
1564 return throttle_get_thread_throttle_level_internal(ut, io_tier);
1565 }
1566
1567 /*
1568 * Return a throttle level given an existing I/O tier (such as returned by throttle_get_io_policy)
1569 */
1570 static int
1571 throttle_get_thread_throttle_level_internal(uthread_t ut, int io_tier) {
1572 int thread_throttle_level = io_tier;
1573 int user_idle_level;
1574
1575 assert(ut != NULL);
1576
1577 /* Bootcache misses should always be throttled */
1578 if (ut->uu_throttle_bc == TRUE)
1579 thread_throttle_level = THROTTLE_LEVEL_TIER3;
1580
1581 /*
1582 * Issue tier3 I/O as tier2 when the user is idle
1583 * to allow maintenance tasks to make more progress.
1584 *
1585 * Assume any positive idle level is enough... for now it's
1586 * only ever 0 or 128 but this is not defined anywhere.
1587 */
1588 if (thread_throttle_level >= THROTTLE_LEVEL_TIER3) {
1589 user_idle_level = timer_get_user_idle_level();
1590 if (user_idle_level > 0) {
1591 thread_throttle_level--;
1592 }
1593 }
1594
1595 return (thread_throttle_level);
1596 }
1597
1598 /*
1599 * I/O will be throttled if either of the following are true:
1600 * - Higher tiers have in-flight I/O
1601 * - The time delta since the last start/completion of a higher tier is within the throttle window interval
1602 *
1603 * In-flight I/O is bookended by throttle_info_update_internal/throttle_info_end_io_internal
1604 */
1605 static int
1606 throttle_io_will_be_throttled_internal(void * throttle_info, int * mylevel, int * throttling_level)
1607 {
1608 struct _throttle_io_info_t *info = throttle_info;
1609 struct timeval elapsed;
1610 struct timeval now;
1611 uint64_t elapsed_msecs;
1612 int thread_throttle_level;
1613 int throttle_level;
1614
1615 if ((thread_throttle_level = throttle_get_thread_throttle_level(NULL)) < THROTTLE_LEVEL_THROTTLED)
1616 return (THROTTLE_DISENGAGED);
1617
1618 microuptime(&now);
1619
1620 for (throttle_level = THROTTLE_LEVEL_START; throttle_level < thread_throttle_level; throttle_level++) {
1621 if (info->throttle_inflight_count[throttle_level]) {
1622 break;
1623 }
1624 elapsed = now;
1625 timevalsub(&elapsed, &info->throttle_window_start_timestamp[throttle_level]);
1626 elapsed_msecs = (uint64_t)elapsed.tv_sec * (uint64_t)1000 + (elapsed.tv_usec / 1000);
1627
1628 if (elapsed_msecs < (uint64_t)throttle_windows_msecs[thread_throttle_level])
1629 break;
1630 }
1631 if (throttle_level >= thread_throttle_level) {
1632 /*
1633 * we're beyond all of the throttle windows
1634 * that affect the throttle level of this thread,
1635 * so go ahead and treat as normal I/O
1636 */
1637 return (THROTTLE_DISENGAGED);
1638 }
1639 if (mylevel)
1640 *mylevel = thread_throttle_level;
1641 if (throttling_level)
1642 *throttling_level = throttle_level;
1643
1644 if (info->throttle_io_count != info->throttle_io_count_begin) {
1645 /*
1646 * we've already issued at least one throttleable I/O
1647 * in the current I/O window, so avoid issuing another one
1648 */
1649 return (THROTTLE_NOW);
1650 }
1651 /*
1652 * we're in the throttle window, so
1653 * cut the I/O size back
1654 */
1655 return (THROTTLE_ENGAGED);
1656 }
1657
1658 /*
1659 * If we have a mount point and it has a throttle info pointer then
1660 * use it to do the check, otherwise use the device unit number to find
1661 * the correct throttle info array element.
1662 */
1663 int
1664 throttle_io_will_be_throttled(__unused int lowpri_window_msecs, mount_t mp)
1665 {
1666 struct _throttle_io_info_t *info;
1667
1668 /*
1669 * Should we just return zero if no mount point
1670 */
1671 if (mp == NULL)
1672 info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
1673 else if (mp->mnt_throttle_info == NULL)
1674 info = &_throttle_io_info[mp->mnt_devbsdunit];
1675 else
1676 info = mp->mnt_throttle_info;
1677
1678 if (info->throttle_is_fusion_with_priority) {
1679 uthread_t ut = get_bsdthread_info(current_thread());
1680 if (ut->uu_lowpri_window == 0)
1681 return (THROTTLE_DISENGAGED);
1682 }
1683
1684 if (info->throttle_disabled)
1685 return (THROTTLE_DISENGAGED);
1686 else
1687 return throttle_io_will_be_throttled_internal(info, NULL, NULL);
1688 }
1689
1690 /*
1691 * Routine to increment I/O throttling counters maintained in the proc
1692 */
1693
1694 static void
1695 throttle_update_proc_stats(pid_t throttling_pid, int count)
1696 {
1697 proc_t throttling_proc;
1698 proc_t throttled_proc = current_proc();
1699
1700 /* The throttled_proc is always the current proc; so we are not concerned with refs */
1701 OSAddAtomic64(count, &(throttled_proc->was_throttled));
1702
1703 /* The throttling pid might have exited by now */
1704 throttling_proc = proc_find(throttling_pid);
1705 if (throttling_proc != PROC_NULL) {
1706 OSAddAtomic64(count, &(throttling_proc->did_throttle));
1707 proc_rele(throttling_proc);
1708 }
1709 }
1710
1711 /*
1712 * Block until woken up by the throttle timer or by a rethrottle call.
1713 * As long as we hold the throttle_lock while querying the throttle tier, we're
1714 * safe against seeing an old throttle tier after a rethrottle.
1715 */
1716 uint32_t
1717 throttle_lowpri_io(int sleep_amount)
1718 {
1719 uthread_t ut;
1720 struct _throttle_io_info_t *info;
1721 int throttle_type = 0;
1722 int mylevel = 0;
1723 int throttling_level = THROTTLE_LEVEL_NONE;
1724 int sleep_cnt = 0;
1725 uint32_t throttle_io_period_num = 0;
1726 boolean_t insert_tail = TRUE;
1727 boolean_t s;
1728
1729 ut = get_bsdthread_info(current_thread());
1730
1731 if (ut->uu_lowpri_window == 0)
1732 return (0);
1733
1734 info = ut->uu_throttle_info;
1735
1736 if (info == NULL) {
1737 ut->uu_throttle_bc = FALSE;
1738 ut->uu_lowpri_window = 0;
1739 return (0);
1740 }
1741 lck_mtx_lock(&info->throttle_lock);
1742 assert(ut->uu_on_throttlelist < THROTTLE_LEVEL_THROTTLED);
1743
1744 if (sleep_amount == 0)
1745 goto done;
1746
1747 if (sleep_amount == 1 && ut->uu_throttle_bc == FALSE)
1748 sleep_amount = 0;
1749
1750 throttle_io_period_num = info->throttle_io_period_num;
1751
1752 ut->uu_was_rethrottled = FALSE;
1753
1754 while ( (throttle_type = throttle_io_will_be_throttled_internal(info, &mylevel, &throttling_level)) ) {
1755
1756 if (throttle_type == THROTTLE_ENGAGED) {
1757 if (sleep_amount == 0)
1758 break;
1759 if (info->throttle_io_period_num < throttle_io_period_num)
1760 break;
1761 if ((info->throttle_io_period_num - throttle_io_period_num) >= (uint32_t)sleep_amount)
1762 break;
1763 }
1764 /*
1765 * keep the same position in the list if "rethrottle_thread" changes our throttle level and
1766 * then puts us back to the original level before we get a chance to run
1767 */
1768 if (ut->uu_on_throttlelist >= THROTTLE_LEVEL_THROTTLED && ut->uu_on_throttlelist != mylevel) {
1769 /*
1770 * must have been awakened via "rethrottle_thread" (the timer pulls us off the list)
1771 * and we've changed our throttling level, so pull ourselves off of the appropriate list
1772 * and make sure we get put on the tail of the new list since we're starting anew w/r to
1773 * the throttling engine
1774 */
1775 TAILQ_REMOVE(&info->throttle_uthlist[ut->uu_on_throttlelist], ut, uu_throttlelist);
1776 ut->uu_on_throttlelist = THROTTLE_LEVEL_NONE;
1777 insert_tail = TRUE;
1778 }
1779 if (ut->uu_on_throttlelist < THROTTLE_LEVEL_THROTTLED) {
1780 if (throttle_add_to_list(info, ut, mylevel, insert_tail) == THROTTLE_LEVEL_END)
1781 goto done;
1782 }
1783 assert(throttling_level >= THROTTLE_LEVEL_START && throttling_level <= THROTTLE_LEVEL_END);
1784
1785 s = ml_set_interrupts_enabled(FALSE);
1786 lck_spin_lock(&ut->uu_rethrottle_lock);
1787
1788 /*
1789 * this is the critical section w/r to our interaction
1790 * with "rethrottle_thread"
1791 */
1792 if (ut->uu_was_rethrottled == TRUE) {
1793
1794 lck_spin_unlock(&ut->uu_rethrottle_lock);
1795 ml_set_interrupts_enabled(s);
1796 lck_mtx_yield(&info->throttle_lock);
1797
1798 KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_FSRW, 103)), thread_tid(ut->uu_thread), ut->uu_on_throttlelist, 0, 0, 0);
1799
1800 ut->uu_was_rethrottled = FALSE;
1801 continue;
1802 }
1803 KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_THROTTLE, PROCESS_THROTTLED)) | DBG_FUNC_NONE,
1804 info->throttle_last_IO_pid[throttling_level], throttling_level, proc_selfpid(), mylevel, 0);
1805
1806 if (sleep_cnt == 0) {
1807 KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_START,
1808 throttle_windows_msecs[mylevel], info->throttle_io_periods[mylevel], info->throttle_io_count, 0, 0);
1809 throttled_count[mylevel]++;
1810 }
1811 ut->uu_wmesg = "throttle_lowpri_io";
1812
1813 assert_wait((caddr_t)&ut->uu_on_throttlelist, THREAD_UNINT);
1814
1815 ut->uu_is_throttled = TRUE;
1816 lck_spin_unlock(&ut->uu_rethrottle_lock);
1817 ml_set_interrupts_enabled(s);
1818
1819 lck_mtx_unlock(&info->throttle_lock);
1820
1821 thread_block(THREAD_CONTINUE_NULL);
1822
1823 ut->uu_wmesg = NULL;
1824
1825 ut->uu_is_throttled = FALSE;
1826 ut->uu_was_rethrottled = FALSE;
1827
1828 lck_mtx_lock(&info->throttle_lock);
1829
1830 sleep_cnt++;
1831
1832 if (sleep_amount == 0)
1833 insert_tail = FALSE;
1834 else if (info->throttle_io_period_num < throttle_io_period_num ||
1835 (info->throttle_io_period_num - throttle_io_period_num) >= (uint32_t)sleep_amount) {
1836 insert_tail = FALSE;
1837 sleep_amount = 0;
1838 }
1839 }
1840 done:
1841 if (ut->uu_on_throttlelist >= THROTTLE_LEVEL_THROTTLED) {
1842 TAILQ_REMOVE(&info->throttle_uthlist[ut->uu_on_throttlelist], ut, uu_throttlelist);
1843 ut->uu_on_throttlelist = THROTTLE_LEVEL_NONE;
1844 }
1845 lck_mtx_unlock(&info->throttle_lock);
1846
1847 if (sleep_cnt) {
1848 KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_FSRW, 97)) | DBG_FUNC_END,
1849 throttle_windows_msecs[mylevel], info->throttle_io_periods[mylevel], info->throttle_io_count, 0, 0);
1850 /*
1851 * We update the stats for the last pid which opened a throttle window for the throttled thread.
1852 * This might not be completely accurate since the multiple throttles seen by the lower tier pid
1853 * might have been caused by various higher prio pids. However, updating these stats accurately
1854 * means doing a proc_find while holding the throttle lock which leads to deadlock.
1855 */
1856 throttle_update_proc_stats(info->throttle_last_IO_pid[throttling_level], sleep_cnt);
1857 }
1858
1859 ut->uu_throttle_info = NULL;
1860 ut->uu_throttle_bc = FALSE;
1861 ut->uu_lowpri_window = 0;
1862
1863 throttle_info_rel(info);
1864
1865 return (sleep_cnt);
1866 }
1867
1868 /*
1869 * KPI routine
1870 *
1871 * set a kernel thread's IO policy. policy can be:
1872 * IOPOL_NORMAL, IOPOL_THROTTLE, IOPOL_PASSIVE, IOPOL_UTILITY, IOPOL_STANDARD
1873 *
1874 * explanations about these policies are in the man page of setiopolicy_np
1875 */
1876 void throttle_set_thread_io_policy(int policy)
1877 {
1878 proc_set_thread_policy(current_thread(), TASK_POLICY_INTERNAL, TASK_POLICY_IOPOL, policy);
1879 }
1880
1881 void throttle_info_reset_window(uthread_t ut)
1882 {
1883 struct _throttle_io_info_t *info;
1884
1885 if (ut == NULL)
1886 ut = get_bsdthread_info(current_thread());
1887
1888 if ( (info = ut->uu_throttle_info) ) {
1889 throttle_info_rel(info);
1890
1891 ut->uu_throttle_info = NULL;
1892 ut->uu_lowpri_window = 0;
1893 ut->uu_throttle_bc = FALSE;
1894 }
1895 }
1896
1897 static
1898 void throttle_info_set_initial_window(uthread_t ut, struct _throttle_io_info_t *info, boolean_t BC_throttle, boolean_t isssd)
1899 {
1900 if (lowpri_throttle_enabled == 0 || info->throttle_disabled)
1901 return;
1902
1903 if (info->throttle_io_periods == 0) {
1904 throttle_init_throttle_period(info, isssd);
1905 }
1906 if (ut->uu_throttle_info == NULL) {
1907
1908 ut->uu_throttle_info = info;
1909 throttle_info_ref(info);
1910 DEBUG_ALLOC_THROTTLE_INFO("updating info = %p\n", info, info );
1911
1912 ut->uu_lowpri_window = 1;
1913 ut->uu_throttle_bc = BC_throttle;
1914 }
1915 }
1916
1917 /*
1918 * Update inflight IO count and throttling window
1919 * Should be called when an IO is done
1920 *
1921 * Only affects IO that was sent through spec_strategy
1922 */
1923 void throttle_info_end_io(buf_t bp) {
1924 mount_t mp;
1925 struct bufattr *bap;
1926 struct _throttle_io_info_t *info;
1927 int io_tier;
1928
1929 bap = &bp->b_attr;
1930 if (!ISSET(bap->ba_flags, BA_STRATEGY_TRACKED_IO)) {
1931 return;
1932 }
1933 CLR(bap->ba_flags, BA_STRATEGY_TRACKED_IO);
1934
1935 mp = buf_vnode(bp)->v_mount;
1936 if (mp != NULL) {
1937 info = &_throttle_io_info[mp->mnt_devbsdunit];
1938 } else {
1939 info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
1940 }
1941
1942 io_tier = GET_BUFATTR_IO_TIER(bap);
1943 if (ISSET(bap->ba_flags, BA_IO_TIER_UPGRADE)) {
1944 io_tier--;
1945 }
1946
1947 throttle_info_end_io_internal(info, io_tier);
1948 }
1949
1950 /*
1951 * Decrement inflight count initially incremented by throttle_info_update_internal
1952 */
1953 static
1954 void throttle_info_end_io_internal(struct _throttle_io_info_t *info, int throttle_level) {
1955 if (throttle_level == THROTTLE_LEVEL_NONE) {
1956 return;
1957 }
1958
1959 microuptime(&info->throttle_window_start_timestamp[throttle_level]);
1960 OSDecrementAtomic(&info->throttle_inflight_count[throttle_level]);
1961 assert(info->throttle_inflight_count[throttle_level] >= 0);
1962 }
1963
1964 /*
1965 * If inflight is TRUE and bap is NULL then the caller is responsible for calling
1966 * throttle_info_end_io_internal to avoid leaking in-flight I/O.
1967 */
1968 static
1969 int throttle_info_update_internal(struct _throttle_io_info_t *info, uthread_t ut, int flags, boolean_t isssd, boolean_t inflight, struct bufattr *bap)
1970 {
1971 int thread_throttle_level;
1972
1973 if (lowpri_throttle_enabled == 0 || info->throttle_disabled)
1974 return THROTTLE_LEVEL_NONE;
1975
1976 if (ut == NULL)
1977 ut = get_bsdthread_info(current_thread());
1978
1979 if (bap && inflight && !ut->uu_throttle_bc) {
1980 thread_throttle_level = GET_BUFATTR_IO_TIER(bap);
1981 if (ISSET(bap->ba_flags, BA_IO_TIER_UPGRADE)) {
1982 thread_throttle_level--;
1983 }
1984 } else {
1985 thread_throttle_level = throttle_get_thread_throttle_level(ut);
1986 }
1987
1988 if (thread_throttle_level != THROTTLE_LEVEL_NONE) {
1989 if(!ISSET(flags, B_PASSIVE)) {
1990 info->throttle_last_IO_pid[thread_throttle_level] = proc_selfpid();
1991 if (inflight && !ut->uu_throttle_bc) {
1992 if (NULL != bap) {
1993 SET(bap->ba_flags, BA_STRATEGY_TRACKED_IO);
1994 }
1995 OSIncrementAtomic(&info->throttle_inflight_count[thread_throttle_level]);
1996 } else {
1997 microuptime(&info->throttle_window_start_timestamp[thread_throttle_level]);
1998 }
1999 KERNEL_DEBUG_CONSTANT((FSDBG_CODE(DBG_THROTTLE, OPEN_THROTTLE_WINDOW)) | DBG_FUNC_NONE,
2000 current_proc()->p_pid, thread_throttle_level, 0, 0, 0);
2001 }
2002 microuptime(&info->throttle_last_IO_timestamp[thread_throttle_level]);
2003 }
2004
2005
2006 if (thread_throttle_level >= THROTTLE_LEVEL_THROTTLED) {
2007 /*
2008 * I'd really like to do the IOSleep here, but
2009 * we may be holding all kinds of filesystem related locks
2010 * and the pages for this I/O marked 'busy'...
2011 * we don't want to cause a normal task to block on
2012 * one of these locks while we're throttling a task marked
2013 * for low priority I/O... we'll mark the uthread and
2014 * do the delay just before we return from the system
2015 * call that triggered this I/O or from vnode_pagein
2016 */
2017 OSAddAtomic(1, &info->throttle_io_count);
2018
2019 throttle_info_set_initial_window(ut, info, FALSE, isssd);
2020 }
2021
2022 return thread_throttle_level;
2023 }
2024
2025 void *throttle_info_update_by_mount(mount_t mp)
2026 {
2027 struct _throttle_io_info_t *info;
2028 uthread_t ut;
2029 boolean_t isssd = FALSE;
2030
2031 ut = get_bsdthread_info(current_thread());
2032
2033 if (mp != NULL) {
2034 if ((mp->mnt_kern_flag & MNTK_SSD) && !ignore_is_ssd)
2035 isssd = TRUE;
2036 info = &_throttle_io_info[mp->mnt_devbsdunit];
2037 } else
2038 info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
2039
2040 if (!ut->uu_lowpri_window)
2041 throttle_info_set_initial_window(ut, info, FALSE, isssd);
2042
2043 return info;
2044 }
2045
2046
2047 /*
2048 * KPI routine
2049 *
2050 * this is usually called before every I/O, used for throttled I/O
2051 * book keeping. This routine has low overhead and does not sleep
2052 */
2053 void throttle_info_update(void *throttle_info, int flags)
2054 {
2055 if (throttle_info)
2056 throttle_info_update_internal(throttle_info, NULL, flags, FALSE, FALSE, NULL);
2057 }
2058
2059 /*
2060 * KPI routine
2061 *
2062 * this is usually called before every I/O, used for throttled I/O
2063 * book keeping. This routine has low overhead and does not sleep
2064 */
2065 void throttle_info_update_by_mask(void *throttle_info_handle, int flags)
2066 {
2067 void *throttle_info = throttle_info_handle;
2068
2069 /*
2070 * for now we only use the lowest bit of the throttle mask, so the
2071 * handle is the same as the throttle_info. Later if we store a
2072 * set of throttle infos in the handle, we will want to loop through
2073 * them and call throttle_info_update in a loop
2074 */
2075 throttle_info_update(throttle_info, flags);
2076 }
2077 /*
2078 * KPI routine
2079 *
2080 * This routine marks the throttle info as disabled. Used for mount points which
2081 * support I/O scheduling.
2082 */
2083
2084 void throttle_info_disable_throttle(int devno, boolean_t isfusion)
2085 {
2086 struct _throttle_io_info_t *info;
2087
2088 if (devno < 0 || devno >= LOWPRI_MAX_NUM_DEV)
2089 panic("Illegal devno (%d) passed into throttle_info_disable_throttle()", devno);
2090
2091 info = &_throttle_io_info[devno];
2092 // don't disable software throttling on devices that are part of a fusion device
2093 // and override the software throttle periods to use HDD periods
2094 if (isfusion) {
2095 info->throttle_is_fusion_with_priority = isfusion;
2096 throttle_init_throttle_period(info, FALSE);
2097 }
2098 info->throttle_disabled = !info->throttle_is_fusion_with_priority;
2099 return;
2100 }
2101
2102
2103 /*
2104 * KPI routine (private)
2105 * Called to determine if this IO is being throttled to this level so that it can be treated specially
2106 */
2107 int throttle_info_io_will_be_throttled(void * throttle_info, int policy)
2108 {
2109 struct _throttle_io_info_t *info = throttle_info;
2110 struct timeval elapsed;
2111 uint64_t elapsed_msecs;
2112 int throttle_level;
2113 int thread_throttle_level;
2114
2115 switch (policy) {
2116
2117 case IOPOL_THROTTLE:
2118 thread_throttle_level = THROTTLE_LEVEL_TIER3;
2119 break;
2120 case IOPOL_UTILITY:
2121 thread_throttle_level = THROTTLE_LEVEL_TIER2;
2122 break;
2123 case IOPOL_STANDARD:
2124 thread_throttle_level = THROTTLE_LEVEL_TIER1;
2125 break;
2126 default:
2127 thread_throttle_level = THROTTLE_LEVEL_TIER0;
2128 break;
2129 }
2130 for (throttle_level = THROTTLE_LEVEL_START; throttle_level < thread_throttle_level; throttle_level++) {
2131 if (info->throttle_inflight_count[throttle_level]) {
2132 break;
2133 }
2134
2135 microuptime(&elapsed);
2136 timevalsub(&elapsed, &info->throttle_window_start_timestamp[throttle_level]);
2137 elapsed_msecs = (uint64_t)elapsed.tv_sec * (uint64_t)1000 + (elapsed.tv_usec / 1000);
2138
2139 if (elapsed_msecs < (uint64_t)throttle_windows_msecs[thread_throttle_level])
2140 break;
2141 }
2142 if (throttle_level >= thread_throttle_level) {
2143 /*
2144 * we're beyond all of the throttle windows
2145 * so go ahead and treat as normal I/O
2146 */
2147 return (THROTTLE_DISENGAGED);
2148 }
2149 /*
2150 * we're in the throttle window
2151 */
2152 return (THROTTLE_ENGAGED);
2153 }
2154
2155 int throttle_lowpri_window(void)
2156 {
2157 struct uthread *ut = get_bsdthread_info(current_thread());
2158 return ut->uu_lowpri_window;
2159 }
2160
2161 int
2162 spec_strategy(struct vnop_strategy_args *ap)
2163 {
2164 buf_t bp;
2165 int bflags;
2166 int io_tier;
2167 int passive;
2168 dev_t bdev;
2169 uthread_t ut;
2170 mount_t mp;
2171 struct bufattr *bap;
2172 int strategy_ret;
2173 struct _throttle_io_info_t *throttle_info;
2174 boolean_t isssd = FALSE;
2175 boolean_t inflight = FALSE;
2176 boolean_t upgrade = FALSE;
2177 int code = 0;
2178
2179 proc_t curproc = current_proc();
2180
2181 bp = ap->a_bp;
2182 bdev = buf_device(bp);
2183 mp = buf_vnode(bp)->v_mount;
2184 bap = &bp->b_attr;
2185
2186 io_tier = throttle_get_io_policy(&ut);
2187 passive = throttle_get_passive_io_policy(&ut);
2188
2189 /*
2190 * Mark if the I/O was upgraded by throttle_get_thread_throttle_level
2191 * while preserving the original issued tier (throttle_get_io_policy
2192 * does not return upgraded tiers)
2193 */
2194 if (mp && io_tier > throttle_get_thread_throttle_level_internal(ut, io_tier)) {
2195 #if CONFIG_IOSCHED
2196 if (!(mp->mnt_ioflags & MNT_IOFLAGS_IOSCHED_SUPPORTED)) {
2197 upgrade = TRUE;
2198 }
2199 #else /* CONFIG_IOSCHED */
2200 upgrade = TRUE;
2201 #endif /* CONFIG_IOSCHED */
2202 }
2203
2204 if (bp->b_flags & B_META)
2205 bap->ba_flags |= BA_META;
2206
2207 #if CONFIG_IOSCHED
2208 /*
2209 * For I/O Scheduling, we currently do not have a way to track and expedite metadata I/Os.
2210 * To ensure we dont get into priority inversions due to metadata I/Os, we use the following rules:
2211 * For metadata reads, ceil all I/Os to IOSCHED_METADATA_TIER & mark them passive if the I/O tier was upgraded
2212 * For metadata writes, unconditionally mark them as IOSCHED_METADATA_TIER and passive
2213 */
2214 if (bap->ba_flags & BA_META) {
2215 if (mp && (mp->mnt_ioflags & MNT_IOFLAGS_IOSCHED_SUPPORTED)) {
2216 if (bp->b_flags & B_READ) {
2217 if (io_tier > IOSCHED_METADATA_TIER) {
2218 io_tier = IOSCHED_METADATA_TIER;
2219 passive = 1;
2220 }
2221 } else {
2222 io_tier = IOSCHED_METADATA_TIER;
2223 passive = 1;
2224 }
2225 }
2226 }
2227 #endif /* CONFIG_IOSCHED */
2228
2229 SET_BUFATTR_IO_TIER(bap, io_tier);
2230
2231 if (passive) {
2232 bp->b_flags |= B_PASSIVE;
2233 bap->ba_flags |= BA_PASSIVE;
2234 }
2235
2236 if ((curproc != NULL) && ((curproc->p_flag & P_DELAYIDLESLEEP) == P_DELAYIDLESLEEP))
2237 bap->ba_flags |= BA_DELAYIDLESLEEP;
2238
2239 bflags = bp->b_flags;
2240
2241 if (((bflags & B_READ) == 0) && ((bflags & B_ASYNC) == 0))
2242 bufattr_markquickcomplete(bap);
2243
2244 if (bflags & B_READ)
2245 code |= DKIO_READ;
2246 if (bflags & B_ASYNC)
2247 code |= DKIO_ASYNC;
2248
2249 if (bap->ba_flags & BA_META)
2250 code |= DKIO_META;
2251 else if (bflags & B_PAGEIO)
2252 code |= DKIO_PAGING;
2253
2254 if (io_tier != 0)
2255 code |= DKIO_THROTTLE;
2256
2257 code |= ((io_tier << DKIO_TIER_SHIFT) & DKIO_TIER_MASK);
2258
2259 if (bflags & B_PASSIVE)
2260 code |= DKIO_PASSIVE;
2261
2262 if (bap->ba_flags & BA_NOCACHE)
2263 code |= DKIO_NOCACHE;
2264
2265 if (upgrade) {
2266 code |= DKIO_TIER_UPGRADE;
2267 SET(bap->ba_flags, BA_IO_TIER_UPGRADE);
2268 }
2269
2270 if (kdebug_enable) {
2271 KERNEL_DEBUG_CONSTANT_IST(KDEBUG_COMMON, FSDBG_CODE(DBG_DKRW, code) | DBG_FUNC_NONE,
2272 buf_kernel_addrperm_addr(bp), bdev, (int)buf_blkno(bp), buf_count(bp), 0);
2273 }
2274
2275 thread_update_io_stats(current_thread(), buf_count(bp), code);
2276
2277 if (mp != NULL) {
2278 if ((mp->mnt_kern_flag & MNTK_SSD) && !ignore_is_ssd)
2279 isssd = TRUE;
2280 /*
2281 * Partially initialized mounts don't have a final devbsdunit and should not be tracked.
2282 * Verify that devbsdunit is initialized (non-zero) or that 0 is the correct initialized value
2283 * (mnt_throttle_mask is initialized and num_trailing_0 would be 0)
2284 */
2285 if (mp->mnt_devbsdunit || (mp->mnt_throttle_mask != LOWPRI_MAX_NUM_DEV - 1 && mp->mnt_throttle_mask & 0x1)) {
2286 inflight = TRUE;
2287 }
2288 throttle_info = &_throttle_io_info[mp->mnt_devbsdunit];
2289
2290 } else
2291 throttle_info = &_throttle_io_info[LOWPRI_MAX_NUM_DEV - 1];
2292
2293 throttle_info_update_internal(throttle_info, ut, bflags, isssd, inflight, bap);
2294
2295 if ((bflags & B_READ) == 0) {
2296 microuptime(&throttle_info->throttle_last_write_timestamp);
2297
2298 if (mp) {
2299 mp->mnt_last_write_issued_timestamp = throttle_info->throttle_last_write_timestamp;
2300 INCR_PENDING_IO(buf_count(bp), mp->mnt_pending_write_size);
2301 }
2302 } else if (mp) {
2303 INCR_PENDING_IO(buf_count(bp), mp->mnt_pending_read_size);
2304 }
2305 /*
2306 * The BootCache may give us special information about
2307 * the IO, so it returns special values that we check
2308 * for here.
2309 *
2310 * IO_SATISFIED_BY_CACHE
2311 * The read has been satisfied by the boot cache. Don't
2312 * throttle the thread unnecessarily.
2313 *
2314 * IO_SHOULD_BE_THROTTLED
2315 * The boot cache is playing back a playlist and this IO
2316 * cut through. Throttle it so we're not cutting through
2317 * the boot cache too often.
2318 *
2319 * Note that typical strategy routines are defined with
2320 * a void return so we'll get garbage here. In the
2321 * unlikely case the garbage matches our special return
2322 * value, it's not a big deal since we're only adjusting
2323 * the throttling delay.
2324 */
2325 #define IO_SATISFIED_BY_CACHE ((int)0xcafefeed)
2326 #define IO_SHOULD_BE_THROTTLED ((int)0xcafebeef)
2327 typedef int strategy_fcn_ret_t(struct buf *bp);
2328
2329 strategy_ret = (*(strategy_fcn_ret_t*)bdevsw[major(bdev)].d_strategy)(bp);
2330
2331 if (IO_SATISFIED_BY_CACHE == strategy_ret) {
2332 /*
2333 * If this was a throttled IO satisfied by the boot cache,
2334 * don't delay the thread.
2335 */
2336 throttle_info_reset_window(ut);
2337
2338 } else if (IO_SHOULD_BE_THROTTLED == strategy_ret) {
2339 /*
2340 * If the boot cache indicates this IO should be throttled,
2341 * delay the thread.
2342 */
2343 throttle_info_set_initial_window(ut, throttle_info, TRUE, isssd);
2344 }
2345 return (0);
2346 }
2347
2348
2349 /*
2350 * This is a noop, simply returning what one has been given.
2351 */
2352 int
2353 spec_blockmap(__unused struct vnop_blockmap_args *ap)
2354 {
2355 return (ENOTSUP);
2356 }
2357
2358
2359 /*
2360 * Device close routine
2361 */
2362 int
2363 spec_close(struct vnop_close_args *ap)
2364 {
2365 struct vnode *vp = ap->a_vp;
2366 dev_t dev = vp->v_rdev;
2367 int error = 0;
2368 int flags = ap->a_fflag;
2369 struct proc *p = vfs_context_proc(ap->a_context);
2370 struct session *sessp;
2371
2372 switch (vp->v_type) {
2373
2374 case VCHR:
2375 /*
2376 * Hack: a tty device that is a controlling terminal
2377 * has a reference from the session structure.
2378 * We cannot easily tell that a character device is
2379 * a controlling terminal, unless it is the closing
2380 * process' controlling terminal. In that case,
2381 * if the reference count is 1 (this is the very
2382 * last close)
2383 */
2384 sessp = proc_session(p);
2385 devsw_lock(dev, S_IFCHR);
2386 if (sessp != SESSION_NULL) {
2387 if (vp == sessp->s_ttyvp && vcount(vp) == 1) {
2388 struct tty *tp = TTY_NULL;
2389
2390 devsw_unlock(dev, S_IFCHR);
2391 session_lock(sessp);
2392 if (vp == sessp->s_ttyvp) {
2393 tp = SESSION_TP(sessp);
2394 sessp->s_ttyvp = NULL;
2395 sessp->s_ttyvid = 0;
2396 sessp->s_ttyp = TTY_NULL;
2397 sessp->s_ttypgrpid = NO_PID;
2398 }
2399 session_unlock(sessp);
2400
2401 if (tp != TTY_NULL) {
2402 /*
2403 * We may have won a race with a proc_exit
2404 * of the session leader, the winner
2405 * clears the flag (even if not set)
2406 */
2407 tty_lock(tp);
2408 ttyclrpgrphup(tp);
2409 tty_unlock(tp);
2410
2411 ttyfree(tp);
2412 }
2413 devsw_lock(dev, S_IFCHR);
2414 }
2415 session_rele(sessp);
2416 }
2417
2418 if (--vp->v_specinfo->si_opencount < 0)
2419 panic("negative open count (c, %u, %u)", major(dev), minor(dev));
2420
2421 /*
2422 * close on last reference or on vnode revoke call
2423 */
2424 if (vcount(vp) == 0 || (flags & IO_REVOKE) != 0)
2425 error = cdevsw[major(dev)].d_close(dev, flags, S_IFCHR, p);
2426
2427 devsw_unlock(dev, S_IFCHR);
2428 break;
2429
2430 case VBLK:
2431 /*
2432 * If there is more than one outstanding open, don't
2433 * send the close to the device.
2434 */
2435 devsw_lock(dev, S_IFBLK);
2436 if (vcount(vp) > 1) {
2437 vp->v_specinfo->si_opencount--;
2438 devsw_unlock(dev, S_IFBLK);
2439 return (0);
2440 }
2441 devsw_unlock(dev, S_IFBLK);
2442
2443 /*
2444 * On last close of a block device (that isn't mounted)
2445 * we must invalidate any in core blocks, so that
2446 * we can, for instance, change floppy disks.
2447 */
2448 if ((error = spec_fsync_internal(vp, MNT_WAIT, ap->a_context)))
2449 return (error);
2450
2451 error = buf_invalidateblks(vp, BUF_WRITE_DATA, 0, 0);
2452 if (error)
2453 return (error);
2454
2455 devsw_lock(dev, S_IFBLK);
2456
2457 if (--vp->v_specinfo->si_opencount < 0)
2458 panic("negative open count (b, %u, %u)", major(dev), minor(dev));
2459
2460 if (vcount(vp) == 0)
2461 error = bdevsw[major(dev)].d_close(dev, flags, S_IFBLK, p);
2462
2463 devsw_unlock(dev, S_IFBLK);
2464 break;
2465
2466 default:
2467 panic("spec_close: not special");
2468 return(EBADF);
2469 }
2470
2471 return error;
2472 }
2473
2474 /*
2475 * Return POSIX pathconf information applicable to special devices.
2476 */
2477 int
2478 spec_pathconf(struct vnop_pathconf_args *ap)
2479 {
2480
2481 switch (ap->a_name) {
2482 case _PC_LINK_MAX:
2483 *ap->a_retval = LINK_MAX;
2484 return (0);
2485 case _PC_MAX_CANON:
2486 *ap->a_retval = MAX_CANON;
2487 return (0);
2488 case _PC_MAX_INPUT:
2489 *ap->a_retval = MAX_INPUT;
2490 return (0);
2491 case _PC_PIPE_BUF:
2492 *ap->a_retval = PIPE_BUF;
2493 return (0);
2494 case _PC_CHOWN_RESTRICTED:
2495 *ap->a_retval = 200112; /* _POSIX_CHOWN_RESTRICTED */
2496 return (0);
2497 case _PC_VDISABLE:
2498 *ap->a_retval = _POSIX_VDISABLE;
2499 return (0);
2500 default:
2501 return (EINVAL);
2502 }
2503 /* NOTREACHED */
2504 }
2505
2506 /*
2507 * Special device failed operation
2508 */
2509 int
2510 spec_ebadf(__unused void *dummy)
2511 {
2512
2513 return (EBADF);
2514 }
2515
2516 /* Blktooff derives file offset from logical block number */
2517 int
2518 spec_blktooff(struct vnop_blktooff_args *ap)
2519 {
2520 struct vnode *vp = ap->a_vp;
2521
2522 switch (vp->v_type) {
2523 case VCHR:
2524 *ap->a_offset = (off_t)-1; /* failure */
2525 return (ENOTSUP);
2526
2527 case VBLK:
2528 printf("spec_blktooff: not implemented for VBLK\n");
2529 *ap->a_offset = (off_t)-1; /* failure */
2530 return (ENOTSUP);
2531
2532 default:
2533 panic("spec_blktooff type");
2534 }
2535 /* NOTREACHED */
2536
2537 return (0);
2538 }
2539
2540 /* Offtoblk derives logical block number from file offset */
2541 int
2542 spec_offtoblk(struct vnop_offtoblk_args *ap)
2543 {
2544 struct vnode *vp = ap->a_vp;
2545
2546 switch (vp->v_type) {
2547 case VCHR:
2548 *ap->a_lblkno = (daddr64_t)-1; /* failure */
2549 return (ENOTSUP);
2550
2551 case VBLK:
2552 printf("spec_offtoblk: not implemented for VBLK\n");
2553 *ap->a_lblkno = (daddr64_t)-1; /* failure */
2554 return (ENOTSUP);
2555
2556 default:
2557 panic("spec_offtoblk type");
2558 }
2559 /* NOTREACHED */
2560
2561 return (0);
2562 }
2563
2564 static void filt_specdetach(struct knote *kn);
2565 static int filt_spec(struct knote *kn, long hint);
2566 static int filt_spectouch(struct knote *kn, struct kevent_internal_s *kev);
2567 static int filt_specprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev);
2568 static unsigned filt_specpeek(struct knote *kn);
2569
2570 struct filterops spec_filtops = {
2571 .f_isfd = 1,
2572 .f_attach = filt_specattach,
2573 .f_detach = filt_specdetach,
2574 .f_event = filt_spec,
2575 .f_touch = filt_spectouch,
2576 .f_process = filt_specprocess,
2577 .f_peek = filt_specpeek
2578 };
2579
2580 static int
2581 filter_to_seltype(int16_t filter)
2582 {
2583 switch (filter) {
2584 case EVFILT_READ:
2585 return FREAD;
2586 case EVFILT_WRITE:
2587 return FWRITE;
2588 default:
2589 panic("filt_to_seltype(): invalid filter %d\n", filter);
2590 return 0;
2591 }
2592 }
2593
2594 static int
2595 filt_specattach(struct knote *kn)
2596 {
2597 vnode_t vp;
2598 dev_t dev;
2599
2600 vp = (vnode_t)kn->kn_fp->f_fglob->fg_data; /* Already have iocount, and vnode is alive */
2601
2602 assert(vnode_ischr(vp));
2603
2604 dev = vnode_specrdev(vp);
2605
2606 if (major(dev) > nchrdev) {
2607 kn->kn_flags |= EV_ERROR;
2608 kn->kn_data = ENXIO;
2609 return 0;
2610 }
2611
2612 /*
2613 * For a few special kinds of devices, we can attach knotes with
2614 * no restrictions because their "select" vectors return the amount
2615 * of data available. Others require an explicit NOTE_LOWAT with
2616 * data of 1, indicating that the caller doesn't care about actual
2617 * data counts, just an indication that the device has data.
2618 */
2619
2620 if ((cdevsw_flags[major(dev)] & CDEVSW_SELECT_KQUEUE) == 0 &&
2621 ((kn->kn_sfflags & NOTE_LOWAT) == 0 || kn->kn_sdata != 1)) {
2622 kn->kn_flags |= EV_ERROR;
2623 kn->kn_data = EINVAL;
2624 return 0;
2625 }
2626
2627 kn->kn_hook_data = 0;
2628
2629 kn->kn_filtid = EVFILTID_SPEC;
2630 kn->kn_hookid = vnode_vid(vp);
2631
2632 knote_markstayactive(kn);
2633
2634 return 0;
2635 }
2636
2637 static void
2638 filt_specdetach(struct knote *kn)
2639 {
2640 knote_clearstayactive(kn);
2641
2642 /*
2643 * This is potentially tricky: the device's selinfo waitq that was
2644 * tricked into being part of this knote's waitq set may not be a part
2645 * of any other set, and the device itself may have revoked the memory
2646 * in which the waitq was held. We use the knote's kn_hook_data field
2647 * to keep the ID of the waitq's prepost table object. This
2648 * object keeps a pointer back to the waitq, and gives us a safe way
2649 * to decouple the dereferencing of driver allocated memory: if the
2650 * driver goes away (taking the waitq with it) then the prepost table
2651 * object will be invalidated. The waitq details are handled in the
2652 * waitq API invoked here.
2653 */
2654 if (kn->kn_hook_data) {
2655 waitq_unlink_by_prepost_id(kn->kn_hook_data, &(knote_get_kq(kn)->kq_wqs));
2656 kn->kn_hook_data = 0;
2657 }
2658 }
2659
2660 static int
2661 filt_spec(__unused struct knote *kn, __unused long hint)
2662 {
2663 panic("filt_spec()");
2664 return 0;
2665 }
2666
2667
2668
2669 static int
2670 filt_spectouch(struct knote *kn, struct kevent_internal_s *kev)
2671 {
2672 kn->kn_sdata = kev->data;
2673 kn->kn_sfflags = kev->fflags;
2674 if ((kn->kn_status & KN_UDATA_SPECIFIC) == 0)
2675 kn->kn_udata = kev->udata;
2676
2677 /* stayqueued knotes don't need hints from touch */
2678 return 0;
2679 }
2680
2681 static int
2682 filt_specprocess(struct knote *kn, struct filt_process_s *data, struct kevent_internal_s *kev)
2683 {
2684 #pragma unused(data)
2685 vnode_t vp;
2686 uthread_t uth;
2687 struct waitq_set *old_wqs;
2688 vfs_context_t ctx;
2689 int res;
2690 int selres;
2691 int error;
2692 int use_offset;
2693 dev_t dev;
2694 uint64_t flags;
2695 uint64_t rsvd, rsvd_arg;
2696 uint64_t *rlptr = NULL;
2697
2698 uth = get_bsdthread_info(current_thread());
2699 ctx = vfs_context_current();
2700 vp = (vnode_t)kn->kn_fp->f_fglob->fg_data;
2701
2702 /* JMM - locking against touches? */
2703
2704 error = vnode_getwithvid(vp, kn->kn_hookid);
2705 if (error != 0) {
2706 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
2707 *kev = kn->kn_kevent;
2708 return 1;
2709 }
2710
2711 dev = vnode_specrdev(vp);
2712 flags = cdevsw_flags[major(dev)];
2713 use_offset = ((flags & CDEVSW_USE_OFFSET) != 0);
2714
2715 /*
2716 * This function may be called many times to link or re-link the
2717 * underlying vnode to the kqueue. If we've already linked the two,
2718 * we will have a valid kn_hook_data which ties us to the underlying
2719 * device's waitq via a the waitq's prepost table object. However,
2720 * devices can abort any select action by calling selthreadclear().
2721 * This is OK because the table object will be invalidated by the
2722 * driver (through a call to selthreadclear), so any attempt to access
2723 * the associated waitq will fail because the table object is invalid.
2724 *
2725 * Even if we've already registered, we need to pass a pointer
2726 * to a reserved link structure. Otherwise, selrecord() will
2727 * infer that we're in the second pass of select() and won't
2728 * actually do anything!
2729 */
2730 rsvd = rsvd_arg = waitq_link_reserve(NULL);
2731 rlptr = (void *)&rsvd_arg;
2732
2733 /*
2734 * Trick selrecord() into hooking kqueue's wait queue set
2735 * set into device's selinfo wait queue
2736 */
2737 old_wqs = uth->uu_wqset;
2738 uth->uu_wqset = &(knote_get_kq(kn)->kq_wqs);
2739 selres = VNOP_SELECT(vp, filter_to_seltype(kn->kn_filter),
2740 0, rlptr, ctx);
2741 uth->uu_wqset = old_wqs;
2742
2743 /*
2744 * make sure to cleanup the reserved link - this guards against
2745 * drivers that may not actually call selrecord().
2746 */
2747 waitq_link_release(rsvd);
2748 if (rsvd != rsvd_arg) {
2749 /* the driver / handler called selrecord() */
2750 struct waitq *wq;
2751 memcpy(&wq, rlptr, sizeof(void *));
2752
2753 /*
2754 * The waitq_get_prepost_id() function will (potentially)
2755 * allocate a prepost table object for the waitq and return
2756 * the table object's ID to us. It will also set the
2757 * waitq_prepost_id field within the waitq structure.
2758 *
2759 * We can just overwrite kn_hook_data because it's simply a
2760 * table ID used to grab a reference when needed.
2761 *
2762 * We have a reference on the vnode, so we know that the
2763 * device won't go away while we get this ID.
2764 */
2765 kn->kn_hook_data = waitq_get_prepost_id(wq);
2766 }
2767
2768 if (use_offset) {
2769 if (kn->kn_fp->f_fglob->fg_offset >= (uint32_t)selres) {
2770 kn->kn_data = 0;
2771 } else {
2772 kn->kn_data = ((uint32_t)selres) - kn->kn_fp->f_fglob->fg_offset;
2773 }
2774 } else {
2775 kn->kn_data = selres;
2776 }
2777
2778 vnode_put(vp);
2779
2780 res = ((kn->kn_sfflags & NOTE_LOWAT) != 0) ?
2781 (kn->kn_data >= kn->kn_sdata) : kn->kn_data;
2782
2783 if (res) {
2784 *kev = kn->kn_kevent;
2785 if (kn->kn_flags & EV_CLEAR) {
2786 kn->kn_fflags = 0;
2787 kn->kn_data = 0;
2788 }
2789 }
2790
2791 return res;
2792 }
2793
2794 static unsigned
2795 filt_specpeek(struct knote *kn)
2796 {
2797 vnode_t vp;
2798 uthread_t uth;
2799 struct waitq_set *old_wqs;
2800 vfs_context_t ctx;
2801 int error, selres;
2802 uint64_t rsvd, rsvd_arg;
2803 uint64_t *rlptr = NULL;
2804
2805 uth = get_bsdthread_info(current_thread());
2806 ctx = vfs_context_current();
2807 vp = (vnode_t)kn->kn_fp->f_fglob->fg_data;
2808
2809 error = vnode_getwithvid(vp, kn->kn_hookid);
2810 if (error != 0) {
2811 return 1; /* Just like VNOP_SELECT() on recycled vnode */
2812 }
2813
2814 /*
2815 * Even if we've already registered, we need to pass a pointer
2816 * to a reserved link structure. Otherwise, selrecord() will
2817 * infer that we're in the second pass of select() and won't
2818 * actually do anything!
2819 */
2820 rsvd = rsvd_arg = waitq_link_reserve(NULL);
2821 rlptr = (void *)&rsvd_arg;
2822
2823 old_wqs = uth->uu_wqset;
2824 uth->uu_wqset = &(knote_get_kq(kn)->kq_wqs);
2825 selres = VNOP_SELECT(vp, filter_to_seltype(kn->kn_filter),
2826 0, (void *)rlptr, ctx);
2827 uth->uu_wqset = old_wqs;
2828
2829 /*
2830 * make sure to cleanup the reserved link - this guards against
2831 * drivers that may not actually call selrecord()
2832 */
2833 waitq_link_release(rsvd);
2834 if (rsvd != rsvd_arg) {
2835 /* the driver / handler called selrecord() */
2836 struct waitq *wq;
2837 memcpy(&wq, rlptr, sizeof(void *));
2838
2839 /*
2840 * The waitq_get_prepost_id() function will (potentially)
2841 * allocate a prepost table object for the waitq and return
2842 * the table object's ID to us. It will also set the
2843 * waitq_prepost_id field within the waitq structure.
2844 *
2845 * We can just overwrite kn_hook_data because it's simply a
2846 * table ID used to grab a reference when needed.
2847 *
2848 * We have a reference on the vnode, so we know that the
2849 * device won't go away while we get this ID.
2850 */
2851 kn->kn_hook_data = waitq_get_prepost_id(wq);
2852 }
2853
2854 vnode_put(vp);
2855 return selres;
2856 }
2857
mirror of XNU