Commit | Line | Data |
---|---|---|
d7e09d03 PT |
1 | /* |
2 | * GPL HEADER START | |
3 | * | |
4 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License version 2 only, | |
8 | * as published by the Free Software Foundation. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, but | |
11 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
13 | * General Public License version 2 for more details (a copy is included | |
14 | * in the LICENSE file that accompanied this code). | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
17 | * version 2 along with this program; If not, see | |
18 | * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf | |
19 | * | |
20 | * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
21 | * CA 95054 USA or visit www.sun.com if you need additional information or | |
22 | * have any questions. | |
23 | * | |
24 | * GPL HEADER END | |
25 | */ | |
26 | /* | |
27 | * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved. | |
28 | * Use is subject to license terms. | |
29 | * | |
1dc563a6 | 30 | * Copyright (c) 2011, 2015, Intel Corporation. |
d7e09d03 PT |
31 | */ |
32 | /* | |
33 | * This file is part of Lustre, http://www.lustre.org/ | |
34 | * Lustre is a trademark of Sun Microsystems, Inc. | |
35 | * | |
36 | * Implementation of cl_page for OSC layer. | |
37 | * | |
38 | * Author: Nikita Danilov <nikita.danilov@sun.com> | |
39 | */ | |
40 | ||
41 | #define DEBUG_SUBSYSTEM S_OSC | |
42 | ||
43 | #include "osc_cl_internal.h" | |
44 | ||
45 | static void osc_lru_del(struct client_obd *cli, struct osc_page *opg, bool del); | |
46 | static void osc_lru_add(struct client_obd *cli, struct osc_page *opg); | |
47 | static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj, | |
48 | struct osc_page *opg); | |
49 | ||
50 | /** \addtogroup osc | |
51 | * @{ | |
52 | */ | |
53 | ||
d7e09d03 PT |
54 | static int osc_page_protected(const struct lu_env *env, |
55 | const struct osc_page *opg, | |
56 | enum cl_lock_mode mode, int unref) | |
57 | { | |
58 | return 1; | |
59 | } | |
d7e09d03 PT |
60 | |
61 | /***************************************************************************** | |
62 | * | |
63 | * Page operations. | |
64 | * | |
65 | */ | |
66 | static void osc_page_fini(const struct lu_env *env, | |
67 | struct cl_page_slice *slice) | |
68 | { | |
69 | struct osc_page *opg = cl2osc_page(slice); | |
50ffcb7e | 70 | |
d7e09d03 | 71 | CDEBUG(D_TRACE, "%p\n", opg); |
7f1ae4c0 | 72 | LASSERT(!opg->ops_lock); |
d7e09d03 PT |
73 | } |
74 | ||
75 | static void osc_page_transfer_get(struct osc_page *opg, const char *label) | |
76 | { | |
77 | struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page); | |
78 | ||
79 | LASSERT(!opg->ops_transfer_pinned); | |
80 | cl_page_get(page); | |
81 | lu_ref_add_atomic(&page->cp_reference, label, page); | |
82 | opg->ops_transfer_pinned = 1; | |
83 | } | |
84 | ||
85 | static void osc_page_transfer_put(const struct lu_env *env, | |
86 | struct osc_page *opg) | |
87 | { | |
88 | struct cl_page *page = cl_page_top(opg->ops_cl.cpl_page); | |
89 | ||
90 | if (opg->ops_transfer_pinned) { | |
91 | lu_ref_del(&page->cp_reference, "transfer", page); | |
92 | opg->ops_transfer_pinned = 0; | |
93 | cl_page_put(env, page); | |
94 | } | |
95 | } | |
96 | ||
97 | /** | |
98 | * This is called once for every page when it is submitted for a transfer | |
99 | * either opportunistic (osc_page_cache_add()), or immediate | |
100 | * (osc_page_submit()). | |
101 | */ | |
102 | static void osc_page_transfer_add(const struct lu_env *env, | |
103 | struct osc_page *opg, enum cl_req_type crt) | |
104 | { | |
105 | struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj); | |
106 | ||
107 | /* ops_lru and ops_inflight share the same field, so take it from LRU | |
30aa9c52 OD |
108 | * first and then use it as inflight. |
109 | */ | |
d7e09d03 PT |
110 | osc_lru_del(osc_cli(obj), opg, false); |
111 | ||
112 | spin_lock(&obj->oo_seatbelt); | |
113 | list_add(&opg->ops_inflight, &obj->oo_inflight[crt]); | |
114 | opg->ops_submitter = current; | |
115 | spin_unlock(&obj->oo_seatbelt); | |
116 | } | |
117 | ||
118 | static int osc_page_cache_add(const struct lu_env *env, | |
119 | const struct cl_page_slice *slice, | |
120 | struct cl_io *io) | |
121 | { | |
29ac6840 | 122 | struct osc_io *oio = osc_env_io(env); |
d7e09d03 PT |
123 | struct osc_page *opg = cl2osc_page(slice); |
124 | int result; | |
d7e09d03 PT |
125 | |
126 | LINVRNT(osc_page_protected(env, opg, CLM_WRITE, 0)); | |
127 | ||
128 | osc_page_transfer_get(opg, "transfer\0cache"); | |
129 | result = osc_queue_async_io(env, io, opg); | |
130 | if (result != 0) | |
131 | osc_page_transfer_put(env, opg); | |
132 | else | |
133 | osc_page_transfer_add(env, opg, CRT_WRITE); | |
134 | ||
135 | /* for sync write, kernel will wait for this page to be flushed before | |
136 | * osc_io_end() is called, so release it earlier. | |
30aa9c52 OD |
137 | * for mkwrite(), it's known there is no further pages. |
138 | */ | |
d7e09d03 | 139 | if (cl_io_is_sync_write(io) || cl_io_is_mkwrite(io)) { |
7f1ae4c0 | 140 | if (oio->oi_active) { |
d7e09d03 PT |
141 | osc_extent_release(env, oio->oi_active); |
142 | oio->oi_active = NULL; | |
143 | } | |
144 | } | |
145 | ||
0a3bdb00 | 146 | return result; |
d7e09d03 PT |
147 | } |
148 | ||
149 | void osc_index2policy(ldlm_policy_data_t *policy, const struct cl_object *obj, | |
150 | pgoff_t start, pgoff_t end) | |
151 | { | |
ec83e611 | 152 | memset(policy, 0, sizeof(*policy)); |
d7e09d03 | 153 | policy->l_extent.start = cl_offset(obj, start); |
29ac6840 | 154 | policy->l_extent.end = cl_offset(obj, end + 1) - 1; |
d7e09d03 PT |
155 | } |
156 | ||
157 | static int osc_page_addref_lock(const struct lu_env *env, | |
158 | struct osc_page *opg, | |
159 | struct cl_lock *lock) | |
160 | { | |
161 | struct osc_lock *olock; | |
29ac6840 | 162 | int rc; |
d7e09d03 | 163 | |
7f1ae4c0 | 164 | LASSERT(!opg->ops_lock); |
d7e09d03 PT |
165 | |
166 | olock = osc_lock_at(lock); | |
167 | if (atomic_inc_return(&olock->ols_pageref) <= 0) { | |
168 | atomic_dec(&olock->ols_pageref); | |
169 | rc = -ENODATA; | |
170 | } else { | |
171 | cl_lock_get(lock); | |
172 | opg->ops_lock = lock; | |
173 | rc = 0; | |
174 | } | |
175 | return rc; | |
176 | } | |
177 | ||
178 | static void osc_page_putref_lock(const struct lu_env *env, | |
179 | struct osc_page *opg) | |
180 | { | |
29ac6840 | 181 | struct cl_lock *lock = opg->ops_lock; |
d7e09d03 PT |
182 | struct osc_lock *olock; |
183 | ||
7f1ae4c0 | 184 | LASSERT(lock); |
d7e09d03 PT |
185 | olock = osc_lock_at(lock); |
186 | ||
187 | atomic_dec(&olock->ols_pageref); | |
188 | opg->ops_lock = NULL; | |
189 | ||
190 | cl_lock_put(env, lock); | |
191 | } | |
192 | ||
193 | static int osc_page_is_under_lock(const struct lu_env *env, | |
194 | const struct cl_page_slice *slice, | |
195 | struct cl_io *unused) | |
196 | { | |
197 | struct cl_lock *lock; | |
29ac6840 | 198 | int result = -ENODATA; |
d7e09d03 | 199 | |
d7e09d03 PT |
200 | lock = cl_lock_at_page(env, slice->cpl_obj, slice->cpl_page, |
201 | NULL, 1, 0); | |
7f1ae4c0 | 202 | if (lock) { |
d7e09d03 PT |
203 | if (osc_page_addref_lock(env, cl2osc_page(slice), lock) == 0) |
204 | result = -EBUSY; | |
205 | cl_lock_put(env, lock); | |
206 | } | |
0a3bdb00 | 207 | return result; |
d7e09d03 PT |
208 | } |
209 | ||
210 | static void osc_page_disown(const struct lu_env *env, | |
211 | const struct cl_page_slice *slice, | |
212 | struct cl_io *io) | |
213 | { | |
214 | struct osc_page *opg = cl2osc_page(slice); | |
215 | ||
216 | if (unlikely(opg->ops_lock)) | |
217 | osc_page_putref_lock(env, opg); | |
218 | } | |
219 | ||
220 | static void osc_page_completion_read(const struct lu_env *env, | |
221 | const struct cl_page_slice *slice, | |
222 | int ioret) | |
223 | { | |
29ac6840 | 224 | struct osc_page *opg = cl2osc_page(slice); |
d7e09d03 PT |
225 | struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj); |
226 | ||
227 | if (likely(opg->ops_lock)) | |
228 | osc_page_putref_lock(env, opg); | |
229 | osc_lru_add(osc_cli(obj), opg); | |
230 | } | |
231 | ||
232 | static void osc_page_completion_write(const struct lu_env *env, | |
233 | const struct cl_page_slice *slice, | |
234 | int ioret) | |
235 | { | |
29ac6840 | 236 | struct osc_page *opg = cl2osc_page(slice); |
d7e09d03 PT |
237 | struct osc_object *obj = cl2osc(slice->cpl_obj); |
238 | ||
239 | osc_lru_add(osc_cli(obj), opg); | |
240 | } | |
241 | ||
242 | static int osc_page_fail(const struct lu_env *env, | |
243 | const struct cl_page_slice *slice, | |
244 | struct cl_io *unused) | |
245 | { | |
246 | /* | |
247 | * Cached read? | |
248 | */ | |
249 | LBUG(); | |
250 | return 0; | |
251 | } | |
252 | ||
d7e09d03 PT |
253 | static const char *osc_list(struct list_head *head) |
254 | { | |
255 | return list_empty(head) ? "-" : "+"; | |
256 | } | |
257 | ||
a649ad1d | 258 | static inline unsigned long osc_submit_duration(struct osc_page *opg) |
d7e09d03 PT |
259 | { |
260 | if (opg->ops_submit_time == 0) | |
261 | return 0; | |
262 | ||
263 | return (cfs_time_current() - opg->ops_submit_time); | |
264 | } | |
265 | ||
266 | static int osc_page_print(const struct lu_env *env, | |
267 | const struct cl_page_slice *slice, | |
268 | void *cookie, lu_printer_t printer) | |
269 | { | |
29ac6840 | 270 | struct osc_page *opg = cl2osc_page(slice); |
d7e09d03 | 271 | struct osc_async_page *oap = &opg->ops_oap; |
29ac6840 CH |
272 | struct osc_object *obj = cl2osc(slice->cpl_obj); |
273 | struct client_obd *cli = &osc_export(obj)->exp_obd->u.cli; | |
d7e09d03 | 274 | |
2d00bd17 | 275 | return (*printer)(env, cookie, LUSTRE_OSC_NAME "-page@%p: 1< %#x %d %u %s %s > 2< %llu %u %u %#x %#x | %p %p %p > 3< %s %p %d %lu %d > 4< %d %d %d %lu %s | %s %s %s %s > 5< %s %s %s %s | %d %s | %d %s %s>\n", |
d7e09d03 PT |
276 | opg, |
277 | /* 1 */ | |
278 | oap->oap_magic, oap->oap_cmd, | |
279 | oap->oap_interrupted, | |
280 | osc_list(&oap->oap_pending_item), | |
281 | osc_list(&oap->oap_rpc_item), | |
282 | /* 2 */ | |
283 | oap->oap_obj_off, oap->oap_page_off, oap->oap_count, | |
284 | oap->oap_async_flags, oap->oap_brw_flags, | |
285 | oap->oap_request, oap->oap_cli, obj, | |
286 | /* 3 */ | |
287 | osc_list(&opg->ops_inflight), | |
288 | opg->ops_submitter, opg->ops_transfer_pinned, | |
289 | osc_submit_duration(opg), opg->ops_srvlock, | |
290 | /* 4 */ | |
291 | cli->cl_r_in_flight, cli->cl_w_in_flight, | |
292 | cli->cl_max_rpcs_in_flight, | |
293 | cli->cl_avail_grant, | |
294 | osc_list(&cli->cl_cache_waiters), | |
295 | osc_list(&cli->cl_loi_ready_list), | |
296 | osc_list(&cli->cl_loi_hp_ready_list), | |
297 | osc_list(&cli->cl_loi_write_list), | |
298 | osc_list(&cli->cl_loi_read_list), | |
299 | /* 5 */ | |
300 | osc_list(&obj->oo_ready_item), | |
301 | osc_list(&obj->oo_hp_ready_item), | |
302 | osc_list(&obj->oo_write_item), | |
303 | osc_list(&obj->oo_read_item), | |
304 | atomic_read(&obj->oo_nr_reads), | |
305 | osc_list(&obj->oo_reading_exts), | |
306 | atomic_read(&obj->oo_nr_writes), | |
307 | osc_list(&obj->oo_hp_exts), | |
308 | osc_list(&obj->oo_urgent_exts)); | |
309 | } | |
310 | ||
311 | static void osc_page_delete(const struct lu_env *env, | |
312 | const struct cl_page_slice *slice) | |
313 | { | |
29ac6840 | 314 | struct osc_page *opg = cl2osc_page(slice); |
d7e09d03 PT |
315 | struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj); |
316 | int rc; | |
317 | ||
318 | LINVRNT(opg->ops_temp || osc_page_protected(env, opg, CLM_READ, 1)); | |
319 | ||
d7e09d03 PT |
320 | CDEBUG(D_TRACE, "%p\n", opg); |
321 | osc_page_transfer_put(env, opg); | |
322 | rc = osc_teardown_async_page(env, obj, opg); | |
323 | if (rc) { | |
324 | CL_PAGE_DEBUG(D_ERROR, env, cl_page_top(slice->cpl_page), | |
325 | "Trying to teardown failed: %d\n", rc); | |
326 | LASSERT(0); | |
327 | } | |
328 | ||
329 | spin_lock(&obj->oo_seatbelt); | |
7f1ae4c0 | 330 | if (opg->ops_submitter) { |
d7e09d03 PT |
331 | LASSERT(!list_empty(&opg->ops_inflight)); |
332 | list_del_init(&opg->ops_inflight); | |
333 | opg->ops_submitter = NULL; | |
334 | } | |
335 | spin_unlock(&obj->oo_seatbelt); | |
336 | ||
337 | osc_lru_del(osc_cli(obj), opg, true); | |
d7e09d03 PT |
338 | } |
339 | ||
f6ab21b5 CFC |
340 | static void osc_page_clip(const struct lu_env *env, |
341 | const struct cl_page_slice *slice, int from, int to) | |
d7e09d03 | 342 | { |
29ac6840 | 343 | struct osc_page *opg = cl2osc_page(slice); |
d7e09d03 PT |
344 | struct osc_async_page *oap = &opg->ops_oap; |
345 | ||
346 | LINVRNT(osc_page_protected(env, opg, CLM_READ, 0)); | |
347 | ||
348 | opg->ops_from = from; | |
29ac6840 | 349 | opg->ops_to = to; |
d7e09d03 PT |
350 | spin_lock(&oap->oap_lock); |
351 | oap->oap_async_flags |= ASYNC_COUNT_STABLE; | |
352 | spin_unlock(&oap->oap_lock); | |
353 | } | |
354 | ||
355 | static int osc_page_cancel(const struct lu_env *env, | |
356 | const struct cl_page_slice *slice) | |
357 | { | |
358 | struct osc_page *opg = cl2osc_page(slice); | |
359 | int rc = 0; | |
360 | ||
361 | LINVRNT(osc_page_protected(env, opg, CLM_READ, 0)); | |
362 | ||
363 | /* Check if the transferring against this page | |
30aa9c52 OD |
364 | * is completed, or not even queued. |
365 | */ | |
d7e09d03 PT |
366 | if (opg->ops_transfer_pinned) |
367 | /* FIXME: may not be interrupted.. */ | |
368 | rc = osc_cancel_async_page(env, opg); | |
369 | LASSERT(ergo(rc == 0, opg->ops_transfer_pinned == 0)); | |
370 | return rc; | |
371 | } | |
372 | ||
373 | static int osc_page_flush(const struct lu_env *env, | |
374 | const struct cl_page_slice *slice, | |
375 | struct cl_io *io) | |
376 | { | |
377 | struct osc_page *opg = cl2osc_page(slice); | |
53a0d486 | 378 | int rc; |
29aaf496 | 379 | |
d7e09d03 | 380 | rc = osc_flush_async_page(env, io, opg); |
0a3bdb00 | 381 | return rc; |
d7e09d03 PT |
382 | } |
383 | ||
384 | static const struct cl_page_operations osc_page_ops = { | |
385 | .cpo_fini = osc_page_fini, | |
386 | .cpo_print = osc_page_print, | |
387 | .cpo_delete = osc_page_delete, | |
388 | .cpo_is_under_lock = osc_page_is_under_lock, | |
389 | .cpo_disown = osc_page_disown, | |
390 | .io = { | |
391 | [CRT_READ] = { | |
392 | .cpo_cache_add = osc_page_fail, | |
393 | .cpo_completion = osc_page_completion_read | |
394 | }, | |
395 | [CRT_WRITE] = { | |
396 | .cpo_cache_add = osc_page_cache_add, | |
397 | .cpo_completion = osc_page_completion_write | |
398 | } | |
399 | }, | |
400 | .cpo_clip = osc_page_clip, | |
401 | .cpo_cancel = osc_page_cancel, | |
402 | .cpo_flush = osc_page_flush | |
403 | }; | |
404 | ||
405 | int osc_page_init(const struct lu_env *env, struct cl_object *obj, | |
406 | struct cl_page *page, struct page *vmpage) | |
407 | { | |
408 | struct osc_object *osc = cl2osc(obj); | |
29ac6840 | 409 | struct osc_page *opg = cl_object_page_slice(obj, page); |
d7e09d03 PT |
410 | int result; |
411 | ||
412 | opg->ops_from = 0; | |
29ac6840 | 413 | opg->ops_to = PAGE_CACHE_SIZE; |
d7e09d03 PT |
414 | |
415 | result = osc_prep_async_page(osc, opg, vmpage, | |
416 | cl_offset(obj, page->cp_index)); | |
417 | if (result == 0) { | |
418 | struct osc_io *oio = osc_env_io(env); | |
50ffcb7e | 419 | |
d7e09d03 PT |
420 | opg->ops_srvlock = osc_io_srvlock(oio); |
421 | cl_page_slice_add(page, &opg->ops_cl, obj, | |
422 | &osc_page_ops); | |
423 | } | |
424 | /* | |
425 | * Cannot assert osc_page_protected() here as read-ahead | |
426 | * creates temporary pages outside of a lock. | |
427 | */ | |
428 | /* ops_inflight and ops_lru are the same field, but it doesn't | |
30aa9c52 OD |
429 | * hurt to initialize it twice :-) |
430 | */ | |
d7e09d03 PT |
431 | INIT_LIST_HEAD(&opg->ops_inflight); |
432 | INIT_LIST_HEAD(&opg->ops_lru); | |
433 | ||
434 | /* reserve an LRU space for this page */ | |
435 | if (page->cp_type == CPT_CACHEABLE && result == 0) | |
436 | result = osc_lru_reserve(env, osc, opg); | |
437 | ||
438 | return result; | |
439 | } | |
440 | ||
441 | /** | |
442 | * Helper function called by osc_io_submit() for every page in an immediate | |
443 | * transfer (i.e., transferred synchronously). | |
444 | */ | |
445 | void osc_page_submit(const struct lu_env *env, struct osc_page *opg, | |
446 | enum cl_req_type crt, int brw_flags) | |
447 | { | |
448 | struct osc_async_page *oap = &opg->ops_oap; | |
29ac6840 | 449 | struct osc_object *obj = oap->oap_obj; |
d7e09d03 PT |
450 | |
451 | LINVRNT(osc_page_protected(env, opg, | |
452 | crt == CRT_WRITE ? CLM_WRITE : CLM_READ, 1)); | |
453 | ||
2d00bd17 JP |
454 | LASSERTF(oap->oap_magic == OAP_MAGIC, "Bad oap magic: oap %p, magic 0x%x\n", |
455 | oap, oap->oap_magic); | |
d7e09d03 PT |
456 | LASSERT(oap->oap_async_flags & ASYNC_READY); |
457 | LASSERT(oap->oap_async_flags & ASYNC_COUNT_STABLE); | |
458 | ||
29ac6840 CH |
459 | oap->oap_cmd = crt == CRT_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ; |
460 | oap->oap_page_off = opg->ops_from; | |
461 | oap->oap_count = opg->ops_to - opg->ops_from; | |
40daf5ce | 462 | oap->oap_brw_flags = brw_flags | OBD_BRW_SYNC; |
d7e09d03 PT |
463 | |
464 | if (!client_is_remote(osc_export(obj)) && | |
2eb90a75 | 465 | capable(CFS_CAP_SYS_RESOURCE)) { |
d7e09d03 PT |
466 | oap->oap_brw_flags |= OBD_BRW_NOQUOTA; |
467 | oap->oap_cmd |= OBD_BRW_NOQUOTA; | |
468 | } | |
469 | ||
470 | opg->ops_submit_time = cfs_time_current(); | |
471 | osc_page_transfer_get(opg, "transfer\0imm"); | |
472 | osc_page_transfer_add(env, opg, crt); | |
473 | } | |
474 | ||
475 | /* --------------- LRU page management ------------------ */ | |
476 | ||
477 | /* OSC is a natural place to manage LRU pages as applications are specialized | |
478 | * to write OSC by OSC. Ideally, if one OSC is used more frequently it should | |
479 | * occupy more LRU slots. On the other hand, we should avoid using up all LRU | |
480 | * slots (client_obd::cl_lru_left) otherwise process has to be put into sleep | |
481 | * for free LRU slots - this will be very bad so the algorithm requires each | |
482 | * OSC to free slots voluntarily to maintain a reasonable number of free slots | |
483 | * at any time. | |
484 | */ | |
485 | ||
2f5723a1 | 486 | static DECLARE_WAIT_QUEUE_HEAD(osc_lru_waitq); |
d7e09d03 PT |
487 | static atomic_t osc_lru_waiters = ATOMIC_INIT(0); |
488 | /* LRU pages are freed in batch mode. OSC should at least free this | |
30aa9c52 OD |
489 | * number of pages to avoid running out of LRU budget, and.. |
490 | */ | |
d7e09d03 | 491 | static const int lru_shrink_min = 2 << (20 - PAGE_CACHE_SHIFT); /* 2M */ |
11d66e89 | 492 | /* free this number at most otherwise it will take too long time to finish. */ |
d7e09d03 PT |
493 | static const int lru_shrink_max = 32 << (20 - PAGE_CACHE_SHIFT); /* 32M */ |
494 | ||
495 | /* Check if we can free LRU slots from this OSC. If there exists LRU waiters, | |
496 | * we should free slots aggressively. In this way, slots are freed in a steady | |
497 | * step to maintain fairness among OSCs. | |
498 | * | |
30aa9c52 OD |
499 | * Return how many LRU pages should be freed. |
500 | */ | |
d7e09d03 PT |
501 | static int osc_cache_too_much(struct client_obd *cli) |
502 | { | |
503 | struct cl_client_cache *cache = cli->cl_cache; | |
504 | int pages = atomic_read(&cli->cl_lru_in_list) >> 1; | |
505 | ||
506 | if (atomic_read(&osc_lru_waiters) > 0 && | |
507 | atomic_read(cli->cl_lru_left) < lru_shrink_max) | |
508 | /* drop lru pages aggressively */ | |
509 | return min(pages, lru_shrink_max); | |
510 | ||
511 | /* if it's going to run out LRU slots, we should free some, but not | |
30aa9c52 OD |
512 | * too much to maintain fairness among OSCs. |
513 | */ | |
d7e09d03 PT |
514 | if (atomic_read(cli->cl_lru_left) < cache->ccc_lru_max >> 4) { |
515 | unsigned long tmp; | |
516 | ||
517 | tmp = cache->ccc_lru_max / atomic_read(&cache->ccc_users); | |
518 | if (pages > tmp) | |
519 | return min(pages, lru_shrink_max); | |
520 | ||
521 | return pages > lru_shrink_min ? lru_shrink_min : 0; | |
522 | } | |
523 | ||
524 | return 0; | |
525 | } | |
526 | ||
527 | /* Return how many pages are not discarded in @pvec. */ | |
528 | static int discard_pagevec(const struct lu_env *env, struct cl_io *io, | |
529 | struct cl_page **pvec, int max_index) | |
530 | { | |
531 | int count; | |
532 | int i; | |
533 | ||
534 | for (count = 0, i = 0; i < max_index; i++) { | |
535 | struct cl_page *page = pvec[i]; | |
50ffcb7e | 536 | |
d7e09d03 PT |
537 | if (cl_page_own_try(env, io, page) == 0) { |
538 | /* free LRU page only if nobody is using it. | |
539 | * This check is necessary to avoid freeing the pages | |
540 | * having already been removed from LRU and pinned | |
30aa9c52 OD |
541 | * for IO. |
542 | */ | |
d7e09d03 PT |
543 | if (!cl_page_in_use(page)) { |
544 | cl_page_unmap(env, io, page); | |
545 | cl_page_discard(env, io, page); | |
546 | ++count; | |
547 | } | |
548 | cl_page_disown(env, io, page); | |
549 | } | |
550 | cl_page_put(env, page); | |
551 | pvec[i] = NULL; | |
552 | } | |
553 | return max_index - count; | |
554 | } | |
555 | ||
556 | /** | |
557 | * Drop @target of pages from LRU at most. | |
558 | */ | |
559 | int osc_lru_shrink(struct client_obd *cli, int target) | |
560 | { | |
561 | struct cl_env_nest nest; | |
562 | struct lu_env *env; | |
563 | struct cl_io *io; | |
564 | struct cl_object *clobj = NULL; | |
565 | struct cl_page **pvec; | |
566 | struct osc_page *opg; | |
567 | int maxscan = 0; | |
568 | int count = 0; | |
569 | int index = 0; | |
570 | int rc = 0; | |
d7e09d03 PT |
571 | |
572 | LASSERT(atomic_read(&cli->cl_lru_in_list) >= 0); | |
573 | if (atomic_read(&cli->cl_lru_in_list) == 0 || target <= 0) | |
0a3bdb00 | 574 | return 0; |
d7e09d03 PT |
575 | |
576 | env = cl_env_nested_get(&nest); | |
577 | if (IS_ERR(env)) | |
0a3bdb00 | 578 | return PTR_ERR(env); |
d7e09d03 PT |
579 | |
580 | pvec = osc_env_info(env)->oti_pvec; | |
581 | io = &osc_env_info(env)->oti_io; | |
582 | ||
583 | client_obd_list_lock(&cli->cl_lru_list_lock); | |
584 | atomic_inc(&cli->cl_lru_shrinkers); | |
585 | maxscan = min(target << 1, atomic_read(&cli->cl_lru_in_list)); | |
586 | while (!list_empty(&cli->cl_lru_list)) { | |
587 | struct cl_page *page; | |
588 | ||
589 | if (--maxscan < 0) | |
590 | break; | |
591 | ||
592 | opg = list_entry(cli->cl_lru_list.next, struct osc_page, | |
593 | ops_lru); | |
594 | page = cl_page_top(opg->ops_cl.cpl_page); | |
595 | if (cl_page_in_use_noref(page)) { | |
596 | list_move_tail(&opg->ops_lru, &cli->cl_lru_list); | |
597 | continue; | |
598 | } | |
599 | ||
7f1ae4c0 | 600 | LASSERT(page->cp_obj); |
d7e09d03 PT |
601 | if (clobj != page->cp_obj) { |
602 | struct cl_object *tmp = page->cp_obj; | |
603 | ||
604 | cl_object_get(tmp); | |
605 | client_obd_list_unlock(&cli->cl_lru_list_lock); | |
606 | ||
7f1ae4c0 | 607 | if (clobj) { |
d7e09d03 PT |
608 | count -= discard_pagevec(env, io, pvec, index); |
609 | index = 0; | |
610 | ||
611 | cl_io_fini(env, io); | |
612 | cl_object_put(env, clobj); | |
613 | clobj = NULL; | |
614 | } | |
615 | ||
616 | clobj = tmp; | |
617 | io->ci_obj = clobj; | |
618 | io->ci_ignore_layout = 1; | |
619 | rc = cl_io_init(env, io, CIT_MISC, clobj); | |
620 | ||
621 | client_obd_list_lock(&cli->cl_lru_list_lock); | |
622 | ||
623 | if (rc != 0) | |
624 | break; | |
625 | ||
626 | ++maxscan; | |
627 | continue; | |
628 | } | |
629 | ||
630 | /* move this page to the end of list as it will be discarded | |
631 | * soon. The page will be finally removed from LRU list in | |
30aa9c52 OD |
632 | * osc_page_delete(). |
633 | */ | |
d7e09d03 PT |
634 | list_move_tail(&opg->ops_lru, &cli->cl_lru_list); |
635 | ||
636 | /* it's okay to grab a refcount here w/o holding lock because | |
30aa9c52 OD |
637 | * it has to grab cl_lru_list_lock to delete the page. |
638 | */ | |
d7e09d03 PT |
639 | cl_page_get(page); |
640 | pvec[index++] = page; | |
641 | if (++count >= target) | |
642 | break; | |
643 | ||
644 | if (unlikely(index == OTI_PVEC_SIZE)) { | |
645 | client_obd_list_unlock(&cli->cl_lru_list_lock); | |
646 | count -= discard_pagevec(env, io, pvec, index); | |
647 | index = 0; | |
648 | ||
649 | client_obd_list_lock(&cli->cl_lru_list_lock); | |
650 | } | |
651 | } | |
652 | client_obd_list_unlock(&cli->cl_lru_list_lock); | |
653 | ||
7f1ae4c0 | 654 | if (clobj) { |
d7e09d03 PT |
655 | count -= discard_pagevec(env, io, pvec, index); |
656 | ||
657 | cl_io_fini(env, io); | |
658 | cl_object_put(env, clobj); | |
659 | } | |
660 | cl_env_nested_put(&nest, env); | |
661 | ||
662 | atomic_dec(&cli->cl_lru_shrinkers); | |
0a3bdb00 | 663 | return count > 0 ? count : rc; |
d7e09d03 PT |
664 | } |
665 | ||
666 | static void osc_lru_add(struct client_obd *cli, struct osc_page *opg) | |
667 | { | |
668 | bool wakeup = false; | |
669 | ||
670 | if (!opg->ops_in_lru) | |
671 | return; | |
672 | ||
673 | atomic_dec(&cli->cl_lru_busy); | |
674 | client_obd_list_lock(&cli->cl_lru_list_lock); | |
675 | if (list_empty(&opg->ops_lru)) { | |
676 | list_move_tail(&opg->ops_lru, &cli->cl_lru_list); | |
677 | atomic_inc_return(&cli->cl_lru_in_list); | |
678 | wakeup = atomic_read(&osc_lru_waiters) > 0; | |
679 | } | |
680 | client_obd_list_unlock(&cli->cl_lru_list_lock); | |
681 | ||
682 | if (wakeup) { | |
683 | osc_lru_shrink(cli, osc_cache_too_much(cli)); | |
684 | wake_up_all(&osc_lru_waitq); | |
685 | } | |
686 | } | |
687 | ||
688 | /* delete page from LRUlist. The page can be deleted from LRUlist for two | |
30aa9c52 OD |
689 | * reasons: redirtied or deleted from page cache. |
690 | */ | |
d7e09d03 PT |
691 | static void osc_lru_del(struct client_obd *cli, struct osc_page *opg, bool del) |
692 | { | |
693 | if (opg->ops_in_lru) { | |
694 | client_obd_list_lock(&cli->cl_lru_list_lock); | |
695 | if (!list_empty(&opg->ops_lru)) { | |
696 | LASSERT(atomic_read(&cli->cl_lru_in_list) > 0); | |
697 | list_del_init(&opg->ops_lru); | |
698 | atomic_dec(&cli->cl_lru_in_list); | |
699 | if (!del) | |
700 | atomic_inc(&cli->cl_lru_busy); | |
701 | } else if (del) { | |
702 | LASSERT(atomic_read(&cli->cl_lru_busy) > 0); | |
703 | atomic_dec(&cli->cl_lru_busy); | |
704 | } | |
705 | client_obd_list_unlock(&cli->cl_lru_list_lock); | |
706 | if (del) { | |
707 | atomic_inc(cli->cl_lru_left); | |
708 | /* this is a great place to release more LRU pages if | |
709 | * this osc occupies too many LRU pages and kernel is | |
710 | * stealing one of them. | |
711 | * cl_lru_shrinkers is to avoid recursive call in case | |
30aa9c52 OD |
712 | * we're already in the context of osc_lru_shrink(). |
713 | */ | |
cad6fafa BJ |
714 | if (atomic_read(&cli->cl_lru_shrinkers) == 0 && |
715 | !memory_pressure_get()) | |
d7e09d03 PT |
716 | osc_lru_shrink(cli, osc_cache_too_much(cli)); |
717 | wake_up(&osc_lru_waitq); | |
718 | } | |
719 | } else { | |
720 | LASSERT(list_empty(&opg->ops_lru)); | |
721 | } | |
722 | } | |
723 | ||
724 | static inline int max_to_shrink(struct client_obd *cli) | |
725 | { | |
726 | return min(atomic_read(&cli->cl_lru_in_list) >> 1, lru_shrink_max); | |
727 | } | |
728 | ||
729 | static int osc_lru_reclaim(struct client_obd *cli) | |
730 | { | |
731 | struct cl_client_cache *cache = cli->cl_cache; | |
732 | int max_scans; | |
733 | int rc; | |
734 | ||
7f1ae4c0 | 735 | LASSERT(cache); |
d7e09d03 PT |
736 | |
737 | rc = osc_lru_shrink(cli, lru_shrink_min); | |
738 | if (rc != 0) { | |
739 | CDEBUG(D_CACHE, "%s: Free %d pages from own LRU: %p.\n", | |
740 | cli->cl_import->imp_obd->obd_name, rc, cli); | |
741 | return rc; | |
742 | } | |
743 | ||
744 | CDEBUG(D_CACHE, "%s: cli %p no free slots, pages: %d, busy: %d.\n", | |
745 | cli->cl_import->imp_obd->obd_name, cli, | |
746 | atomic_read(&cli->cl_lru_in_list), | |
747 | atomic_read(&cli->cl_lru_busy)); | |
748 | ||
749 | /* Reclaim LRU slots from other client_obd as it can't free enough | |
30aa9c52 OD |
750 | * from its own. This should rarely happen. |
751 | */ | |
d7e09d03 | 752 | spin_lock(&cache->ccc_lru_lock); |
0df83c18 HN |
753 | LASSERT(!list_empty(&cache->ccc_lru)); |
754 | ||
d7e09d03 PT |
755 | cache->ccc_lru_shrinkers++; |
756 | list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru); | |
757 | ||
758 | max_scans = atomic_read(&cache->ccc_users); | |
759 | while (--max_scans > 0 && !list_empty(&cache->ccc_lru)) { | |
760 | cli = list_entry(cache->ccc_lru.next, struct client_obd, | |
761 | cl_lru_osc); | |
762 | ||
763 | CDEBUG(D_CACHE, "%s: cli %p LRU pages: %d, busy: %d.\n", | |
764 | cli->cl_import->imp_obd->obd_name, cli, | |
765 | atomic_read(&cli->cl_lru_in_list), | |
766 | atomic_read(&cli->cl_lru_busy)); | |
767 | ||
768 | list_move_tail(&cli->cl_lru_osc, &cache->ccc_lru); | |
769 | if (atomic_read(&cli->cl_lru_in_list) > 0) { | |
770 | spin_unlock(&cache->ccc_lru_lock); | |
771 | ||
772 | rc = osc_lru_shrink(cli, max_to_shrink(cli)); | |
773 | spin_lock(&cache->ccc_lru_lock); | |
774 | if (rc != 0) | |
775 | break; | |
776 | } | |
777 | } | |
778 | spin_unlock(&cache->ccc_lru_lock); | |
779 | ||
780 | CDEBUG(D_CACHE, "%s: cli %p freed %d pages.\n", | |
781 | cli->cl_import->imp_obd->obd_name, cli, rc); | |
782 | return rc; | |
783 | } | |
784 | ||
785 | static int osc_lru_reserve(const struct lu_env *env, struct osc_object *obj, | |
786 | struct osc_page *opg) | |
787 | { | |
788 | struct l_wait_info lwi = LWI_INTR(LWI_ON_SIGNAL_NOOP, NULL); | |
789 | struct client_obd *cli = osc_cli(obj); | |
790 | int rc = 0; | |
d7e09d03 | 791 | |
7f1ae4c0 | 792 | if (!cli->cl_cache) /* shall not be in LRU */ |
0a3bdb00 | 793 | return 0; |
d7e09d03 PT |
794 | |
795 | LASSERT(atomic_read(cli->cl_lru_left) >= 0); | |
305ec768 | 796 | while (!atomic_add_unless(cli->cl_lru_left, -1, 0)) { |
d7e09d03 PT |
797 | int gen; |
798 | ||
799 | /* run out of LRU spaces, try to drop some by itself */ | |
800 | rc = osc_lru_reclaim(cli); | |
801 | if (rc < 0) | |
802 | break; | |
803 | if (rc > 0) | |
804 | continue; | |
805 | ||
806 | cond_resched(); | |
807 | ||
808 | /* slowest case, all of caching pages are busy, notifying | |
30aa9c52 OD |
809 | * other OSCs that we're lack of LRU slots. |
810 | */ | |
d7e09d03 PT |
811 | atomic_inc(&osc_lru_waiters); |
812 | ||
813 | gen = atomic_read(&cli->cl_lru_in_list); | |
814 | rc = l_wait_event(osc_lru_waitq, | |
815 | atomic_read(cli->cl_lru_left) > 0 || | |
816 | (atomic_read(&cli->cl_lru_in_list) > 0 && | |
817 | gen != atomic_read(&cli->cl_lru_in_list)), | |
818 | &lwi); | |
819 | ||
820 | atomic_dec(&osc_lru_waiters); | |
821 | if (rc < 0) | |
822 | break; | |
823 | } | |
824 | ||
825 | if (rc >= 0) { | |
826 | atomic_inc(&cli->cl_lru_busy); | |
827 | opg->ops_in_lru = 1; | |
828 | rc = 0; | |
829 | } | |
830 | ||
0a3bdb00 | 831 | return rc; |
d7e09d03 PT |
832 | } |
833 | ||
834 | /** @} osc */ |