Whamcloud - gitweb
LU-14711 osc: Do not attempt sending empty pages
[fs/lustre-release.git] / lustre / osc / osc_request.c
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.gnu.org/licenses/gpl-2.0.html
19  *
20  * GPL HEADER END
21  */
22 /*
23  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24  * Use is subject to license terms.
25  *
26  * Copyright (c) 2011, 2017, Intel Corporation.
27  */
28 /*
29  * This file is part of Lustre, http://www.lustre.org/
30  */
31
32 #define DEBUG_SUBSYSTEM S_OSC
33
34 #include <linux/workqueue.h>
35 #include <libcfs/libcfs.h>
36 #include <linux/falloc.h>
37 #include <lprocfs_status.h>
38 #include <lustre_dlm.h>
39 #include <lustre_fid.h>
40 #include <lustre_ha.h>
41 #include <uapi/linux/lustre/lustre_ioctl.h>
42 #include <lustre_net.h>
43 #include <lustre_obdo.h>
44 #include <obd.h>
45 #include <obd_cksum.h>
46 #include <obd_class.h>
47 #include <lustre_osc.h>
48 #include <linux/falloc.h>
49
50 #include "osc_internal.h"
51 #include <lnet/lnet_rdma.h>
52
53 atomic_t osc_pool_req_count;
54 unsigned int osc_reqpool_maxreqcount;
55 struct ptlrpc_request_pool *osc_rq_pool;
56
57 /* max memory used for request pool, unit is MB */
58 static unsigned int osc_reqpool_mem_max = 5;
59 module_param(osc_reqpool_mem_max, uint, 0444);
60
61 static int osc_idle_timeout = 20;
62 module_param(osc_idle_timeout, uint, 0644);
63
64 #define osc_grant_args osc_brw_async_args
65
66 struct osc_setattr_args {
67         struct obdo             *sa_oa;
68         obd_enqueue_update_f     sa_upcall;
69         void                    *sa_cookie;
70 };
71
72 struct osc_fsync_args {
73         struct osc_object       *fa_obj;
74         struct obdo             *fa_oa;
75         obd_enqueue_update_f    fa_upcall;
76         void                    *fa_cookie;
77 };
78
79 struct osc_ladvise_args {
80         struct obdo             *la_oa;
81         obd_enqueue_update_f     la_upcall;
82         void                    *la_cookie;
83 };
84
85 static void osc_release_ppga(struct brw_page **ppga, size_t count);
86 static int brw_interpret(const struct lu_env *env, struct ptlrpc_request *req,
87                          void *data, int rc);
88
89 void osc_pack_req_body(struct ptlrpc_request *req, struct obdo *oa)
90 {
91         struct ost_body *body;
92
93         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
94         LASSERT(body);
95
96         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
97 }
98
99 static int osc_getattr(const struct lu_env *env, struct obd_export *exp,
100                        struct obdo *oa)
101 {
102         struct ptlrpc_request   *req;
103         struct ost_body         *body;
104         int                      rc;
105
106         ENTRY;
107         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_GETATTR);
108         if (req == NULL)
109                 RETURN(-ENOMEM);
110
111         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GETATTR);
112         if (rc) {
113                 ptlrpc_request_free(req);
114                 RETURN(rc);
115         }
116
117         osc_pack_req_body(req, oa);
118
119         ptlrpc_request_set_replen(req);
120
121         rc = ptlrpc_queue_wait(req);
122         if (rc)
123                 GOTO(out, rc);
124
125         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
126         if (body == NULL)
127                 GOTO(out, rc = -EPROTO);
128
129         CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
130         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
131
132         oa->o_blksize = cli_brw_size(exp->exp_obd);
133         oa->o_valid |= OBD_MD_FLBLKSZ;
134
135         EXIT;
136 out:
137         ptlrpc_req_finished(req);
138
139         return rc;
140 }
141
142 static int osc_setattr(const struct lu_env *env, struct obd_export *exp,
143                        struct obdo *oa)
144 {
145         struct ptlrpc_request   *req;
146         struct ost_body         *body;
147         int                      rc;
148
149         ENTRY;
150         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
151
152         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
153         if (req == NULL)
154                 RETURN(-ENOMEM);
155
156         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
157         if (rc) {
158                 ptlrpc_request_free(req);
159                 RETURN(rc);
160         }
161
162         osc_pack_req_body(req, oa);
163
164         ptlrpc_request_set_replen(req);
165
166         rc = ptlrpc_queue_wait(req);
167         if (rc)
168                 GOTO(out, rc);
169
170         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
171         if (body == NULL)
172                 GOTO(out, rc = -EPROTO);
173
174         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
175
176         EXIT;
177 out:
178         ptlrpc_req_finished(req);
179
180         RETURN(rc);
181 }
182
183 static int osc_setattr_interpret(const struct lu_env *env,
184                                  struct ptlrpc_request *req, void *args, int rc)
185 {
186         struct osc_setattr_args *sa = args;
187         struct ost_body *body;
188
189         ENTRY;
190
191         if (rc != 0)
192                 GOTO(out, rc);
193
194         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
195         if (body == NULL)
196                 GOTO(out, rc = -EPROTO);
197
198         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, sa->sa_oa,
199                              &body->oa);
200 out:
201         rc = sa->sa_upcall(sa->sa_cookie, rc);
202         RETURN(rc);
203 }
204
205 int osc_setattr_async(struct obd_export *exp, struct obdo *oa,
206                       obd_enqueue_update_f upcall, void *cookie,
207                       struct ptlrpc_request_set *rqset)
208 {
209         struct ptlrpc_request   *req;
210         struct osc_setattr_args *sa;
211         int                      rc;
212
213         ENTRY;
214
215         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SETATTR);
216         if (req == NULL)
217                 RETURN(-ENOMEM);
218
219         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SETATTR);
220         if (rc) {
221                 ptlrpc_request_free(req);
222                 RETURN(rc);
223         }
224
225         osc_pack_req_body(req, oa);
226
227         ptlrpc_request_set_replen(req);
228
229         /* do mds to ost setattr asynchronously */
230         if (!rqset) {
231                 /* Do not wait for response. */
232                 ptlrpcd_add_req(req);
233         } else {
234                 req->rq_interpret_reply = osc_setattr_interpret;
235
236                 sa = ptlrpc_req_async_args(sa, req);
237                 sa->sa_oa = oa;
238                 sa->sa_upcall = upcall;
239                 sa->sa_cookie = cookie;
240
241                 ptlrpc_set_add_req(rqset, req);
242         }
243
244         RETURN(0);
245 }
246
247 static int osc_ladvise_interpret(const struct lu_env *env,
248                                  struct ptlrpc_request *req,
249                                  void *arg, int rc)
250 {
251         struct osc_ladvise_args *la = arg;
252         struct ost_body *body;
253         ENTRY;
254
255         if (rc != 0)
256                 GOTO(out, rc);
257
258         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
259         if (body == NULL)
260                 GOTO(out, rc = -EPROTO);
261
262         *la->la_oa = body->oa;
263 out:
264         rc = la->la_upcall(la->la_cookie, rc);
265         RETURN(rc);
266 }
267
268 /**
269  * If rqset is NULL, do not wait for response. Upcall and cookie could also
270  * be NULL in this case
271  */
272 int osc_ladvise_base(struct obd_export *exp, struct obdo *oa,
273                      struct ladvise_hdr *ladvise_hdr,
274                      obd_enqueue_update_f upcall, void *cookie,
275                      struct ptlrpc_request_set *rqset)
276 {
277         struct ptlrpc_request   *req;
278         struct ost_body         *body;
279         struct osc_ladvise_args *la;
280         int                      rc;
281         struct lu_ladvise       *req_ladvise;
282         struct lu_ladvise       *ladvise = ladvise_hdr->lah_advise;
283         int                      num_advise = ladvise_hdr->lah_count;
284         struct ladvise_hdr      *req_ladvise_hdr;
285         ENTRY;
286
287         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_LADVISE);
288         if (req == NULL)
289                 RETURN(-ENOMEM);
290
291         req_capsule_set_size(&req->rq_pill, &RMF_OST_LADVISE, RCL_CLIENT,
292                              num_advise * sizeof(*ladvise));
293         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_LADVISE);
294         if (rc != 0) {
295                 ptlrpc_request_free(req);
296                 RETURN(rc);
297         }
298         req->rq_request_portal = OST_IO_PORTAL;
299         ptlrpc_at_set_req_timeout(req);
300
301         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
302         LASSERT(body);
303         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa,
304                              oa);
305
306         req_ladvise_hdr = req_capsule_client_get(&req->rq_pill,
307                                                  &RMF_OST_LADVISE_HDR);
308         memcpy(req_ladvise_hdr, ladvise_hdr, sizeof(*ladvise_hdr));
309
310         req_ladvise = req_capsule_client_get(&req->rq_pill, &RMF_OST_LADVISE);
311         memcpy(req_ladvise, ladvise, sizeof(*ladvise) * num_advise);
312         ptlrpc_request_set_replen(req);
313
314         if (rqset == NULL) {
315                 /* Do not wait for response. */
316                 ptlrpcd_add_req(req);
317                 RETURN(0);
318         }
319
320         req->rq_interpret_reply = osc_ladvise_interpret;
321         la = ptlrpc_req_async_args(la, req);
322         la->la_oa = oa;
323         la->la_upcall = upcall;
324         la->la_cookie = cookie;
325
326         ptlrpc_set_add_req(rqset, req);
327
328         RETURN(0);
329 }
330
331 static int osc_create(const struct lu_env *env, struct obd_export *exp,
332                       struct obdo *oa)
333 {
334         struct ptlrpc_request *req;
335         struct ost_body       *body;
336         int                    rc;
337         ENTRY;
338
339         LASSERT(oa != NULL);
340         LASSERT(oa->o_valid & OBD_MD_FLGROUP);
341         LASSERT(fid_seq_is_echo(ostid_seq(&oa->o_oi)));
342
343         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_CREATE);
344         if (req == NULL)
345                 GOTO(out, rc = -ENOMEM);
346
347         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_CREATE);
348         if (rc) {
349                 ptlrpc_request_free(req);
350                 GOTO(out, rc);
351         }
352
353         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
354         LASSERT(body);
355
356         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
357
358         ptlrpc_request_set_replen(req);
359
360         rc = ptlrpc_queue_wait(req);
361         if (rc)
362                 GOTO(out_req, rc);
363
364         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
365         if (body == NULL)
366                 GOTO(out_req, rc = -EPROTO);
367
368         CDEBUG(D_INFO, "oa flags %x\n", oa->o_flags);
369         lustre_get_wire_obdo(&req->rq_import->imp_connect_data, oa, &body->oa);
370
371         oa->o_blksize = cli_brw_size(exp->exp_obd);
372         oa->o_valid |= OBD_MD_FLBLKSZ;
373
374         CDEBUG(D_HA, "transno: %lld\n",
375                lustre_msg_get_transno(req->rq_repmsg));
376 out_req:
377         ptlrpc_req_finished(req);
378 out:
379         RETURN(rc);
380 }
381
382 int osc_punch_send(struct obd_export *exp, struct obdo *oa,
383                    obd_enqueue_update_f upcall, void *cookie)
384 {
385         struct ptlrpc_request *req;
386         struct osc_setattr_args *sa;
387         struct obd_import *imp = class_exp2cliimp(exp);
388         struct ost_body *body;
389         int rc;
390
391         ENTRY;
392
393         req = ptlrpc_request_alloc(imp, &RQF_OST_PUNCH);
394         if (req == NULL)
395                 RETURN(-ENOMEM);
396
397         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_PUNCH);
398         if (rc < 0) {
399                 ptlrpc_request_free(req);
400                 RETURN(rc);
401         }
402
403         osc_set_io_portal(req);
404
405         ptlrpc_at_set_req_timeout(req);
406
407         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
408
409         lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
410
411         ptlrpc_request_set_replen(req);
412
413         req->rq_interpret_reply = osc_setattr_interpret;
414         sa = ptlrpc_req_async_args(sa, req);
415         sa->sa_oa = oa;
416         sa->sa_upcall = upcall;
417         sa->sa_cookie = cookie;
418
419         ptlrpcd_add_req(req);
420
421         RETURN(0);
422 }
423 EXPORT_SYMBOL(osc_punch_send);
424
425 /**
426  * osc_fallocate_base() - Handles fallocate request.
427  *
428  * @exp:        Export structure
429  * @oa:         Attributes passed to OSS from client (obdo structure)
430  * @upcall:     Primary & supplementary group information
431  * @cookie:     Exclusive identifier
432  * @rqset:      Request list.
433  * @mode:       Operation done on given range.
434  *
435  * osc_fallocate_base() - Handles fallocate requests only. Only block
436  * allocation or standard preallocate operation is supported currently.
437  * Other mode flags is not supported yet. ftruncate(2) or truncate(2)
438  * is supported via SETATTR request.
439  *
440  * Return: Non-zero on failure and O on success.
441  */
442 int osc_fallocate_base(struct obd_export *exp, struct obdo *oa,
443                        obd_enqueue_update_f upcall, void *cookie, int mode)
444 {
445         struct ptlrpc_request *req;
446         struct osc_setattr_args *sa;
447         struct ost_body *body;
448         struct obd_import *imp = class_exp2cliimp(exp);
449         int rc;
450         ENTRY;
451
452         oa->o_falloc_mode = mode;
453         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
454                                    &RQF_OST_FALLOCATE);
455         if (req == NULL)
456                 RETURN(-ENOMEM);
457
458         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_FALLOCATE);
459         if (rc != 0) {
460                 ptlrpc_request_free(req);
461                 RETURN(rc);
462         }
463
464         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
465         LASSERT(body);
466
467         lustre_set_wire_obdo(&imp->imp_connect_data, &body->oa, oa);
468
469         ptlrpc_request_set_replen(req);
470
471         req->rq_interpret_reply = osc_setattr_interpret;
472         BUILD_BUG_ON(sizeof(*sa) > sizeof(req->rq_async_args));
473         sa = ptlrpc_req_async_args(sa, req);
474         sa->sa_oa = oa;
475         sa->sa_upcall = upcall;
476         sa->sa_cookie = cookie;
477
478         ptlrpcd_add_req(req);
479
480         RETURN(0);
481 }
482 EXPORT_SYMBOL(osc_fallocate_base);
483
484 static int osc_sync_interpret(const struct lu_env *env,
485                               struct ptlrpc_request *req, void *args, int rc)
486 {
487         struct osc_fsync_args *fa = args;
488         struct ost_body *body;
489         struct cl_attr *attr = &osc_env_info(env)->oti_attr;
490         unsigned long valid = 0;
491         struct cl_object *obj;
492         ENTRY;
493
494         if (rc != 0)
495                 GOTO(out, rc);
496
497         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
498         if (body == NULL) {
499                 CERROR("can't unpack ost_body\n");
500                 GOTO(out, rc = -EPROTO);
501         }
502
503         *fa->fa_oa = body->oa;
504         obj = osc2cl(fa->fa_obj);
505
506         /* Update osc object's blocks attribute */
507         cl_object_attr_lock(obj);
508         if (body->oa.o_valid & OBD_MD_FLBLOCKS) {
509                 attr->cat_blocks = body->oa.o_blocks;
510                 valid |= CAT_BLOCKS;
511         }
512
513         if (valid != 0)
514                 cl_object_attr_update(env, obj, attr, valid);
515         cl_object_attr_unlock(obj);
516
517 out:
518         rc = fa->fa_upcall(fa->fa_cookie, rc);
519         RETURN(rc);
520 }
521
522 int osc_sync_base(struct osc_object *obj, struct obdo *oa,
523                   obd_enqueue_update_f upcall, void *cookie,
524                   struct ptlrpc_request_set *rqset)
525 {
526         struct obd_export     *exp = osc_export(obj);
527         struct ptlrpc_request *req;
528         struct ost_body       *body;
529         struct osc_fsync_args *fa;
530         int                    rc;
531         ENTRY;
532
533         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_SYNC);
534         if (req == NULL)
535                 RETURN(-ENOMEM);
536
537         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_SYNC);
538         if (rc) {
539                 ptlrpc_request_free(req);
540                 RETURN(rc);
541         }
542
543         /* overload the size and blocks fields in the oa with start/end */
544         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
545         LASSERT(body);
546         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
547
548         ptlrpc_request_set_replen(req);
549         req->rq_interpret_reply = osc_sync_interpret;
550
551         fa = ptlrpc_req_async_args(fa, req);
552         fa->fa_obj = obj;
553         fa->fa_oa = oa;
554         fa->fa_upcall = upcall;
555         fa->fa_cookie = cookie;
556
557         ptlrpc_set_add_req(rqset, req);
558
559         RETURN (0);
560 }
561
562 /* Find and cancel locally locks matched by @mode in the resource found by
563  * @objid. Found locks are added into @cancel list. Returns the amount of
564  * locks added to @cancels list. */
565 static int osc_resource_get_unused(struct obd_export *exp, struct obdo *oa,
566                                    struct list_head *cancels,
567                                    enum ldlm_mode mode, __u64 lock_flags)
568 {
569         struct ldlm_namespace *ns = exp->exp_obd->obd_namespace;
570         struct ldlm_res_id res_id;
571         struct ldlm_resource *res;
572         int count;
573         ENTRY;
574
575         /* Return, i.e. cancel nothing, only if ELC is supported (flag in
576          * export) but disabled through procfs (flag in NS).
577          *
578          * This distinguishes from a case when ELC is not supported originally,
579          * when we still want to cancel locks in advance and just cancel them
580          * locally, without sending any RPC. */
581         if (exp_connect_cancelset(exp) && !ns_connect_cancelset(ns))
582                 RETURN(0);
583
584         ostid_build_res_name(&oa->o_oi, &res_id);
585         res = ldlm_resource_get(ns, NULL, &res_id, 0, 0);
586         if (IS_ERR(res))
587                 RETURN(0);
588
589         LDLM_RESOURCE_ADDREF(res);
590         count = ldlm_cancel_resource_local(res, cancels, NULL, mode,
591                                            lock_flags, 0, NULL);
592         LDLM_RESOURCE_DELREF(res);
593         ldlm_resource_putref(res);
594         RETURN(count);
595 }
596
597 static int osc_destroy_interpret(const struct lu_env *env,
598                                  struct ptlrpc_request *req, void *args, int rc)
599 {
600         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
601
602         atomic_dec(&cli->cl_destroy_in_flight);
603         wake_up(&cli->cl_destroy_waitq);
604
605         return 0;
606 }
607
608 static int osc_can_send_destroy(struct client_obd *cli)
609 {
610         if (atomic_inc_return(&cli->cl_destroy_in_flight) <=
611             cli->cl_max_rpcs_in_flight) {
612                 /* The destroy request can be sent */
613                 return 1;
614         }
615         if (atomic_dec_return(&cli->cl_destroy_in_flight) <
616             cli->cl_max_rpcs_in_flight) {
617                 /*
618                  * The counter has been modified between the two atomic
619                  * operations.
620                  */
621                 wake_up(&cli->cl_destroy_waitq);
622         }
623         return 0;
624 }
625
626 static int osc_destroy(const struct lu_env *env, struct obd_export *exp,
627                        struct obdo *oa)
628 {
629         struct client_obd     *cli = &exp->exp_obd->u.cli;
630         struct ptlrpc_request *req;
631         struct ost_body       *body;
632         LIST_HEAD(cancels);
633         int rc, count;
634         ENTRY;
635
636         if (!oa) {
637                 CDEBUG(D_INFO, "oa NULL\n");
638                 RETURN(-EINVAL);
639         }
640
641         count = osc_resource_get_unused(exp, oa, &cancels, LCK_PW,
642                                         LDLM_FL_DISCARD_DATA);
643
644         req = ptlrpc_request_alloc(class_exp2cliimp(exp), &RQF_OST_DESTROY);
645         if (req == NULL) {
646                 ldlm_lock_list_put(&cancels, l_bl_ast, count);
647                 RETURN(-ENOMEM);
648         }
649
650         rc = ldlm_prep_elc_req(exp, req, LUSTRE_OST_VERSION, OST_DESTROY,
651                                0, &cancels, count);
652         if (rc) {
653                 ptlrpc_request_free(req);
654                 RETURN(rc);
655         }
656
657         req->rq_request_portal = OST_IO_PORTAL; /* bug 7198 */
658         ptlrpc_at_set_req_timeout(req);
659
660         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
661         LASSERT(body);
662         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
663
664         ptlrpc_request_set_replen(req);
665
666         req->rq_interpret_reply = osc_destroy_interpret;
667         if (!osc_can_send_destroy(cli)) {
668                 /*
669                  * Wait until the number of on-going destroy RPCs drops
670                  * under max_rpc_in_flight
671                  */
672                 rc = l_wait_event_abortable_exclusive(
673                         cli->cl_destroy_waitq,
674                         osc_can_send_destroy(cli));
675                 if (rc) {
676                         ptlrpc_req_finished(req);
677                         RETURN(-EINTR);
678                 }
679         }
680
681         /* Do not wait for response */
682         ptlrpcd_add_req(req);
683         RETURN(0);
684 }
685
686 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
687                                 long writing_bytes)
688 {
689         u64 bits = OBD_MD_FLBLOCKS | OBD_MD_FLGRANT;
690
691         LASSERT(!(oa->o_valid & bits));
692
693         oa->o_valid |= bits;
694         spin_lock(&cli->cl_loi_list_lock);
695         if (cli->cl_ocd_grant_param)
696                 oa->o_dirty = cli->cl_dirty_grant;
697         else
698                 oa->o_dirty = cli->cl_dirty_pages << PAGE_SHIFT;
699         if (unlikely(cli->cl_dirty_pages > cli->cl_dirty_max_pages)) {
700                 CERROR("dirty %lu > dirty_max %lu\n",
701                        cli->cl_dirty_pages,
702                        cli->cl_dirty_max_pages);
703                 oa->o_undirty = 0;
704         } else if (unlikely(atomic_long_read(&obd_dirty_pages) >
705                             (long)(obd_max_dirty_pages + 1))) {
706                 /* The atomic_read() allowing the atomic_inc() are
707                  * not covered by a lock thus they may safely race and trip
708                  * this CERROR() unless we add in a small fudge factor (+1). */
709                 CERROR("%s: dirty %ld > system dirty_max %ld\n",
710                        cli_name(cli), atomic_long_read(&obd_dirty_pages),
711                        obd_max_dirty_pages);
712                 oa->o_undirty = 0;
713         } else if (unlikely(cli->cl_dirty_max_pages - cli->cl_dirty_pages >
714                             0x7fffffff)) {
715                 CERROR("dirty %lu - dirty_max %lu too big???\n",
716                        cli->cl_dirty_pages, cli->cl_dirty_max_pages);
717                 oa->o_undirty = 0;
718         } else {
719                 unsigned long nrpages;
720                 unsigned long undirty;
721
722                 nrpages = cli->cl_max_pages_per_rpc;
723                 nrpages *= cli->cl_max_rpcs_in_flight + 1;
724                 nrpages = max(nrpages, cli->cl_dirty_max_pages);
725                 undirty = nrpages << PAGE_SHIFT;
726                 if (cli->cl_ocd_grant_param) {
727                         int nrextents;
728
729                         /* take extent tax into account when asking for more
730                          * grant space */
731                         nrextents = (nrpages + cli->cl_max_extent_pages - 1) /
732                                      cli->cl_max_extent_pages;
733                         undirty += nrextents * cli->cl_grant_extent_tax;
734                 }
735                 /* Do not ask for more than OBD_MAX_GRANT - a margin for server
736                  * to add extent tax, etc.
737                  */
738                 oa->o_undirty = min(undirty, OBD_MAX_GRANT &
739                                     ~(PTLRPC_MAX_BRW_SIZE * 4UL));
740         }
741         oa->o_grant = cli->cl_avail_grant + cli->cl_reserved_grant;
742         /* o_dropped AKA o_misc is 32 bits, but cl_lost_grant is 64 bits */
743         if (cli->cl_lost_grant > INT_MAX) {
744                 CDEBUG(D_CACHE,
745                       "%s: avoided o_dropped overflow: cl_lost_grant %lu\n",
746                       cli_name(cli), cli->cl_lost_grant);
747                 oa->o_dropped = INT_MAX;
748         } else {
749                 oa->o_dropped = cli->cl_lost_grant;
750         }
751         cli->cl_lost_grant -= oa->o_dropped;
752         spin_unlock(&cli->cl_loi_list_lock);
753         CDEBUG(D_CACHE, "%s: dirty: %llu undirty: %u dropped %u grant: %llu"
754                " cl_lost_grant %lu\n", cli_name(cli), oa->o_dirty,
755                oa->o_undirty, oa->o_dropped, oa->o_grant, cli->cl_lost_grant);
756 }
757
758 void osc_update_next_shrink(struct client_obd *cli)
759 {
760         cli->cl_next_shrink_grant = ktime_get_seconds() +
761                                     cli->cl_grant_shrink_interval;
762
763         CDEBUG(D_CACHE, "next time %lld to shrink grant\n",
764                cli->cl_next_shrink_grant);
765 }
766
767 static void __osc_update_grant(struct client_obd *cli, u64 grant)
768 {
769         spin_lock(&cli->cl_loi_list_lock);
770         cli->cl_avail_grant += grant;
771         spin_unlock(&cli->cl_loi_list_lock);
772 }
773
774 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
775 {
776         if (body->oa.o_valid & OBD_MD_FLGRANT) {
777                 CDEBUG(D_CACHE, "got %llu extra grant\n", body->oa.o_grant);
778                 __osc_update_grant(cli, body->oa.o_grant);
779         }
780 }
781
782 /**
783  * grant thread data for shrinking space.
784  */
785 struct grant_thread_data {
786         struct list_head        gtd_clients;
787         struct mutex            gtd_mutex;
788         unsigned long           gtd_stopped:1;
789 };
790 static struct grant_thread_data client_gtd;
791
792 static int osc_shrink_grant_interpret(const struct lu_env *env,
793                                       struct ptlrpc_request *req,
794                                       void *args, int rc)
795 {
796         struct osc_grant_args *aa = args;
797         struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
798         struct ost_body *body;
799
800         if (rc != 0) {
801                 __osc_update_grant(cli, aa->aa_oa->o_grant);
802                 GOTO(out, rc);
803         }
804
805         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
806         LASSERT(body);
807         osc_update_grant(cli, body);
808 out:
809         OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
810         aa->aa_oa = NULL;
811
812         return rc;
813 }
814
815 static void osc_shrink_grant_local(struct client_obd *cli, struct obdo *oa)
816 {
817         spin_lock(&cli->cl_loi_list_lock);
818         oa->o_grant = cli->cl_avail_grant / 4;
819         cli->cl_avail_grant -= oa->o_grant;
820         spin_unlock(&cli->cl_loi_list_lock);
821         if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
822                 oa->o_valid |= OBD_MD_FLFLAGS;
823                 oa->o_flags = 0;
824         }
825         oa->o_flags |= OBD_FL_SHRINK_GRANT;
826         osc_update_next_shrink(cli);
827 }
828
829 /* Shrink the current grant, either from some large amount to enough for a
830  * full set of in-flight RPCs, or if we have already shrunk to that limit
831  * then to enough for a single RPC.  This avoids keeping more grant than
832  * needed, and avoids shrinking the grant piecemeal. */
833 static int osc_shrink_grant(struct client_obd *cli)
834 {
835         __u64 target_bytes = (cli->cl_max_rpcs_in_flight + 1) *
836                              (cli->cl_max_pages_per_rpc << PAGE_SHIFT);
837
838         spin_lock(&cli->cl_loi_list_lock);
839         if (cli->cl_avail_grant <= target_bytes)
840                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
841         spin_unlock(&cli->cl_loi_list_lock);
842
843         return osc_shrink_grant_to_target(cli, target_bytes);
844 }
845
846 int osc_shrink_grant_to_target(struct client_obd *cli, __u64 target_bytes)
847 {
848         int                     rc = 0;
849         struct ost_body        *body;
850         ENTRY;
851
852         spin_lock(&cli->cl_loi_list_lock);
853         /* Don't shrink if we are already above or below the desired limit
854          * We don't want to shrink below a single RPC, as that will negatively
855          * impact block allocation and long-term performance. */
856         if (target_bytes < cli->cl_max_pages_per_rpc << PAGE_SHIFT)
857                 target_bytes = cli->cl_max_pages_per_rpc << PAGE_SHIFT;
858
859         if (target_bytes >= cli->cl_avail_grant) {
860                 spin_unlock(&cli->cl_loi_list_lock);
861                 RETURN(0);
862         }
863         spin_unlock(&cli->cl_loi_list_lock);
864
865         OBD_ALLOC_PTR(body);
866         if (!body)
867                 RETURN(-ENOMEM);
868
869         osc_announce_cached(cli, &body->oa, 0);
870
871         spin_lock(&cli->cl_loi_list_lock);
872         if (target_bytes >= cli->cl_avail_grant) {
873                 /* available grant has changed since target calculation */
874                 spin_unlock(&cli->cl_loi_list_lock);
875                 GOTO(out_free, rc = 0);
876         }
877         body->oa.o_grant = cli->cl_avail_grant - target_bytes;
878         cli->cl_avail_grant = target_bytes;
879         spin_unlock(&cli->cl_loi_list_lock);
880         if (!(body->oa.o_valid & OBD_MD_FLFLAGS)) {
881                 body->oa.o_valid |= OBD_MD_FLFLAGS;
882                 body->oa.o_flags = 0;
883         }
884         body->oa.o_flags |= OBD_FL_SHRINK_GRANT;
885         osc_update_next_shrink(cli);
886
887         rc = osc_set_info_async(NULL, cli->cl_import->imp_obd->obd_self_export,
888                                 sizeof(KEY_GRANT_SHRINK), KEY_GRANT_SHRINK,
889                                 sizeof(*body), body, NULL);
890         if (rc != 0)
891                 __osc_update_grant(cli, body->oa.o_grant);
892 out_free:
893         OBD_FREE_PTR(body);
894         RETURN(rc);
895 }
896
897 static int osc_should_shrink_grant(struct client_obd *client)
898 {
899         time64_t next_shrink = client->cl_next_shrink_grant;
900
901         if (client->cl_import == NULL)
902                 return 0;
903
904         if (!OCD_HAS_FLAG(&client->cl_import->imp_connect_data, GRANT_SHRINK) ||
905             client->cl_import->imp_grant_shrink_disabled) {
906                 osc_update_next_shrink(client);
907                 return 0;
908         }
909
910         if (ktime_get_seconds() >= next_shrink - 5) {
911                 /* Get the current RPC size directly, instead of going via:
912                  * cli_brw_size(obd->u.cli.cl_import->imp_obd->obd_self_export)
913                  * Keep comment here so that it can be found by searching. */
914                 int brw_size = client->cl_max_pages_per_rpc << PAGE_SHIFT;
915
916                 if (client->cl_import->imp_state == LUSTRE_IMP_FULL &&
917                     client->cl_avail_grant > brw_size)
918                         return 1;
919                 else
920                         osc_update_next_shrink(client);
921         }
922         return 0;
923 }
924
925 #define GRANT_SHRINK_RPC_BATCH  100
926
927 static struct delayed_work work;
928
929 static void osc_grant_work_handler(struct work_struct *data)
930 {
931         struct client_obd *cli;
932         int rpc_sent;
933         bool init_next_shrink = true;
934         time64_t next_shrink = ktime_get_seconds() + GRANT_SHRINK_INTERVAL;
935
936         rpc_sent = 0;
937         mutex_lock(&client_gtd.gtd_mutex);
938         list_for_each_entry(cli, &client_gtd.gtd_clients,
939                             cl_grant_chain) {
940                 if (rpc_sent < GRANT_SHRINK_RPC_BATCH &&
941                     osc_should_shrink_grant(cli)) {
942                         osc_shrink_grant(cli);
943                         rpc_sent++;
944                 }
945
946                 if (!init_next_shrink) {
947                         if (cli->cl_next_shrink_grant < next_shrink &&
948                             cli->cl_next_shrink_grant > ktime_get_seconds())
949                                 next_shrink = cli->cl_next_shrink_grant;
950                 } else {
951                         init_next_shrink = false;
952                         next_shrink = cli->cl_next_shrink_grant;
953                 }
954         }
955         mutex_unlock(&client_gtd.gtd_mutex);
956
957         if (client_gtd.gtd_stopped == 1)
958                 return;
959
960         if (next_shrink > ktime_get_seconds()) {
961                 time64_t delay = next_shrink - ktime_get_seconds();
962
963                 schedule_delayed_work(&work, cfs_time_seconds(delay));
964         } else {
965                 schedule_work(&work.work);
966         }
967 }
968
969 void osc_schedule_grant_work(void)
970 {
971         cancel_delayed_work_sync(&work);
972         schedule_work(&work.work);
973 }
974
975 /**
976  * Start grant thread for returing grant to server for idle clients.
977  */
978 static int osc_start_grant_work(void)
979 {
980         client_gtd.gtd_stopped = 0;
981         mutex_init(&client_gtd.gtd_mutex);
982         INIT_LIST_HEAD(&client_gtd.gtd_clients);
983
984         INIT_DELAYED_WORK(&work, osc_grant_work_handler);
985         schedule_work(&work.work);
986
987         return 0;
988 }
989
990 static void osc_stop_grant_work(void)
991 {
992         client_gtd.gtd_stopped = 1;
993         cancel_delayed_work_sync(&work);
994 }
995
996 static void osc_add_grant_list(struct client_obd *client)
997 {
998         mutex_lock(&client_gtd.gtd_mutex);
999         list_add(&client->cl_grant_chain, &client_gtd.gtd_clients);
1000         mutex_unlock(&client_gtd.gtd_mutex);
1001 }
1002
1003 static void osc_del_grant_list(struct client_obd *client)
1004 {
1005         if (list_empty(&client->cl_grant_chain))
1006                 return;
1007
1008         mutex_lock(&client_gtd.gtd_mutex);
1009         list_del_init(&client->cl_grant_chain);
1010         mutex_unlock(&client_gtd.gtd_mutex);
1011 }
1012
1013 void osc_init_grant(struct client_obd *cli, struct obd_connect_data *ocd)
1014 {
1015         /*
1016          * ocd_grant is the total grant amount we're expect to hold: if we've
1017          * been evicted, it's the new avail_grant amount, cl_dirty_pages will
1018          * drop to 0 as inflight RPCs fail out; otherwise, it's avail_grant +
1019          * dirty.
1020          *
1021          * race is tolerable here: if we're evicted, but imp_state already
1022          * left EVICTED state, then cl_dirty_pages must be 0 already.
1023          */
1024         spin_lock(&cli->cl_loi_list_lock);
1025         cli->cl_avail_grant = ocd->ocd_grant;
1026         if (cli->cl_import->imp_state != LUSTRE_IMP_EVICTED) {
1027                 unsigned long consumed = cli->cl_reserved_grant;
1028
1029                 if (OCD_HAS_FLAG(ocd, GRANT_PARAM))
1030                         consumed += cli->cl_dirty_grant;
1031                 else
1032                         consumed += cli->cl_dirty_pages << PAGE_SHIFT;
1033                 if (cli->cl_avail_grant < consumed) {
1034                         CERROR("%s: granted %ld but already consumed %ld\n",
1035                                cli_name(cli), cli->cl_avail_grant, consumed);
1036                         cli->cl_avail_grant = 0;
1037                 } else {
1038                         cli->cl_avail_grant -= consumed;
1039                 }
1040         }
1041
1042         if (OCD_HAS_FLAG(ocd, GRANT_PARAM)) {
1043                 u64 size;
1044                 int chunk_mask;
1045
1046                 /* overhead for each extent insertion */
1047                 cli->cl_grant_extent_tax = ocd->ocd_grant_tax_kb << 10;
1048                 /* determine the appropriate chunk size used by osc_extent. */
1049                 cli->cl_chunkbits = max_t(int, PAGE_SHIFT,
1050                                           ocd->ocd_grant_blkbits);
1051                 /* max_pages_per_rpc must be chunk aligned */
1052                 chunk_mask = ~((1 << (cli->cl_chunkbits - PAGE_SHIFT)) - 1);
1053                 cli->cl_max_pages_per_rpc = (cli->cl_max_pages_per_rpc +
1054                                              ~chunk_mask) & chunk_mask;
1055                 /* determine maximum extent size, in #pages */
1056                 size = (u64)ocd->ocd_grant_max_blks << ocd->ocd_grant_blkbits;
1057                 cli->cl_max_extent_pages = (size >> PAGE_SHIFT) ?: 1;
1058                 cli->cl_ocd_grant_param = 1;
1059         } else {
1060                 cli->cl_ocd_grant_param = 0;
1061                 cli->cl_grant_extent_tax = 0;
1062                 cli->cl_chunkbits = PAGE_SHIFT;
1063                 cli->cl_max_extent_pages = DT_MAX_BRW_PAGES;
1064         }
1065         spin_unlock(&cli->cl_loi_list_lock);
1066
1067         CDEBUG(D_CACHE,
1068                "%s, setting cl_avail_grant: %ld cl_lost_grant: %ld. chunk bits: %d cl_max_extent_pages: %d\n",
1069                cli_name(cli),
1070                cli->cl_avail_grant, cli->cl_lost_grant, cli->cl_chunkbits,
1071                cli->cl_max_extent_pages);
1072
1073         if (OCD_HAS_FLAG(ocd, GRANT_SHRINK) && list_empty(&cli->cl_grant_chain))
1074                 osc_add_grant_list(cli);
1075 }
1076 EXPORT_SYMBOL(osc_init_grant);
1077
1078 /* We assume that the reason this OSC got a short read is because it read
1079  * beyond the end of a stripe file; i.e. lustre is reading a sparse file
1080  * via the LOV, and it _knows_ it's reading inside the file, it's just that
1081  * this stripe never got written at or beyond this stripe offset yet. */
1082 static void handle_short_read(int nob_read, size_t page_count,
1083                               struct brw_page **pga)
1084 {
1085         char *ptr;
1086         int i = 0;
1087
1088         /* skip bytes read OK */
1089         while (nob_read > 0) {
1090                 LASSERT (page_count > 0);
1091
1092                 if (pga[i]->count > nob_read) {
1093                         /* EOF inside this page */
1094                         ptr = kmap(pga[i]->pg) +
1095                                 (pga[i]->off & ~PAGE_MASK);
1096                         memset(ptr + nob_read, 0, pga[i]->count - nob_read);
1097                         kunmap(pga[i]->pg);
1098                         page_count--;
1099                         i++;
1100                         break;
1101                 }
1102
1103                 nob_read -= pga[i]->count;
1104                 page_count--;
1105                 i++;
1106         }
1107
1108         /* zero remaining pages */
1109         while (page_count-- > 0) {
1110                 ptr = kmap(pga[i]->pg) + (pga[i]->off & ~PAGE_MASK);
1111                 memset(ptr, 0, pga[i]->count);
1112                 kunmap(pga[i]->pg);
1113                 i++;
1114         }
1115 }
1116
1117 static int check_write_rcs(struct ptlrpc_request *req,
1118                            int requested_nob, int niocount,
1119                            size_t page_count, struct brw_page **pga)
1120 {
1121         int     i;
1122         __u32   *remote_rcs;
1123
1124         remote_rcs = req_capsule_server_sized_get(&req->rq_pill, &RMF_RCS,
1125                                                   sizeof(*remote_rcs) *
1126                                                   niocount);
1127         if (remote_rcs == NULL) {
1128                 CDEBUG(D_INFO, "Missing/short RC vector on BRW_WRITE reply\n");
1129                 return(-EPROTO);
1130         }
1131
1132         /* return error if any niobuf was in error */
1133         for (i = 0; i < niocount; i++) {
1134                 if ((int)remote_rcs[i] < 0) {
1135                         CDEBUG(D_INFO, "rc[%d]: %d req %p\n",
1136                                i, remote_rcs[i], req);
1137                         return remote_rcs[i];
1138                 }
1139
1140                 if (remote_rcs[i] != 0) {
1141                         CDEBUG(D_INFO, "rc[%d] invalid (%d) req %p\n",
1142                                 i, remote_rcs[i], req);
1143                         return(-EPROTO);
1144                 }
1145         }
1146         if (req->rq_bulk != NULL &&
1147             req->rq_bulk->bd_nob_transferred != requested_nob) {
1148                 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
1149                        req->rq_bulk->bd_nob_transferred, requested_nob);
1150                 return(-EPROTO);
1151         }
1152
1153         return (0);
1154 }
1155
1156 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
1157 {
1158         if (p1->flag != p2->flag) {
1159                 unsigned mask = ~(OBD_BRW_FROM_GRANT | OBD_BRW_NOCACHE |
1160                                   OBD_BRW_SYNC       | OBD_BRW_ASYNC   |
1161                                   OBD_BRW_NOQUOTA    | OBD_BRW_SOFT_SYNC);
1162
1163                 /* warn if we try to combine flags that we don't know to be
1164                  * safe to combine */
1165                 if (unlikely((p1->flag & mask) != (p2->flag & mask))) {
1166                         CWARN("Saw flags 0x%x and 0x%x in the same brw, please "
1167                               "report this at https://jira.whamcloud.com/\n",
1168                               p1->flag, p2->flag);
1169                 }
1170                 return 0;
1171         }
1172
1173         return (p1->off + p1->count == p2->off);
1174 }
1175
1176 #if IS_ENABLED(CONFIG_CRC_T10DIF)
1177 static int osc_checksum_bulk_t10pi(const char *obd_name, int nob,
1178                                    size_t pg_count, struct brw_page **pga,
1179                                    int opc, obd_dif_csum_fn *fn,
1180                                    int sector_size,
1181                                    u32 *check_sum, bool resend)
1182 {
1183         struct ahash_request *req;
1184         /* Used Adler as the default checksum type on top of DIF tags */
1185         unsigned char cfs_alg = cksum_obd2cfs(OBD_CKSUM_T10_TOP);
1186         struct page *__page;
1187         unsigned char *buffer;
1188         __u16 *guard_start;
1189         unsigned int bufsize;
1190         int guard_number;
1191         int used_number = 0;
1192         int used;
1193         u32 cksum;
1194         int rc = 0;
1195         int i = 0;
1196
1197         LASSERT(pg_count > 0);
1198
1199         __page = alloc_page(GFP_KERNEL);
1200         if (__page == NULL)
1201                 return -ENOMEM;
1202
1203         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1204         if (IS_ERR(req)) {
1205                 rc = PTR_ERR(req);
1206                 CERROR("%s: unable to initialize checksum hash %s: rc = %d\n",
1207                        obd_name, cfs_crypto_hash_name(cfs_alg), rc);
1208                 GOTO(out, rc);
1209         }
1210
1211         buffer = kmap(__page);
1212         guard_start = (__u16 *)buffer;
1213         guard_number = PAGE_SIZE / sizeof(*guard_start);
1214         CDEBUG(D_PAGE | (resend ? D_HA : 0),
1215                "GRD tags per page=%u, resend=%u, bytes=%u, pages=%zu\n",
1216                guard_number, resend, nob, pg_count);
1217
1218         while (nob > 0 && pg_count > 0) {
1219                 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1220
1221                 /* corrupt the data before we compute the checksum, to
1222                  * simulate an OST->client data error */
1223                 if (unlikely(i == 0 && opc == OST_READ &&
1224                              OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE))) {
1225                         unsigned char *ptr = kmap(pga[i]->pg);
1226                         int off = pga[i]->off & ~PAGE_MASK;
1227
1228                         memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1229                         kunmap(pga[i]->pg);
1230                 }
1231
1232                 /*
1233                  * The left guard number should be able to hold checksums of a
1234                  * whole page
1235                  */
1236                 rc = obd_page_dif_generate_buffer(obd_name, pga[i]->pg,
1237                                                   pga[i]->off & ~PAGE_MASK,
1238                                                   count,
1239                                                   guard_start + used_number,
1240                                                   guard_number - used_number,
1241                                                   &used, sector_size,
1242                                                   fn);
1243                 if (unlikely(resend))
1244                         CDEBUG(D_PAGE | D_HA,
1245                                "pga[%u]: used %u off %llu+%u gen checksum: %*phN\n",
1246                                i, used, pga[i]->off & ~PAGE_MASK, count,
1247                                (int)(used * sizeof(*guard_start)),
1248                                guard_start + used_number);
1249                 if (rc)
1250                         break;
1251
1252                 used_number += used;
1253                 if (used_number == guard_number) {
1254                         cfs_crypto_hash_update_page(req, __page, 0,
1255                                 used_number * sizeof(*guard_start));
1256                         used_number = 0;
1257                 }
1258
1259                 nob -= pga[i]->count;
1260                 pg_count--;
1261                 i++;
1262         }
1263         kunmap(__page);
1264         if (rc)
1265                 GOTO(out, rc);
1266
1267         if (used_number != 0)
1268                 cfs_crypto_hash_update_page(req, __page, 0,
1269                         used_number * sizeof(*guard_start));
1270
1271         bufsize = sizeof(cksum);
1272         cfs_crypto_hash_final(req, (unsigned char *)&cksum, &bufsize);
1273
1274         /* For sending we only compute the wrong checksum instead
1275          * of corrupting the data so it is still correct on a redo */
1276         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1277                 cksum++;
1278
1279         *check_sum = cksum;
1280 out:
1281         __free_page(__page);
1282         return rc;
1283 }
1284 #else /* !CONFIG_CRC_T10DIF */
1285 #define obd_dif_ip_fn NULL
1286 #define obd_dif_crc_fn NULL
1287 #define osc_checksum_bulk_t10pi(name, nob, pgc, pga, opc, fn, ssize, csum, re) \
1288         -EOPNOTSUPP
1289 #endif /* CONFIG_CRC_T10DIF */
1290
1291 static int osc_checksum_bulk(int nob, size_t pg_count,
1292                              struct brw_page **pga, int opc,
1293                              enum cksum_types cksum_type,
1294                              u32 *cksum)
1295 {
1296         int                             i = 0;
1297         struct ahash_request           *req;
1298         unsigned int                    bufsize;
1299         unsigned char                   cfs_alg = cksum_obd2cfs(cksum_type);
1300
1301         LASSERT(pg_count > 0);
1302
1303         req = cfs_crypto_hash_init(cfs_alg, NULL, 0);
1304         if (IS_ERR(req)) {
1305                 CERROR("Unable to initialize checksum hash %s\n",
1306                        cfs_crypto_hash_name(cfs_alg));
1307                 return PTR_ERR(req);
1308         }
1309
1310         while (nob > 0 && pg_count > 0) {
1311                 unsigned int count = pga[i]->count > nob ? nob : pga[i]->count;
1312
1313                 /* corrupt the data before we compute the checksum, to
1314                  * simulate an OST->client data error */
1315                 if (i == 0 && opc == OST_READ &&
1316                     OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_RECEIVE)) {
1317                         unsigned char *ptr = kmap(pga[i]->pg);
1318                         int off = pga[i]->off & ~PAGE_MASK;
1319
1320                         memcpy(ptr + off, "bad1", min_t(typeof(nob), 4, nob));
1321                         kunmap(pga[i]->pg);
1322                 }
1323                 cfs_crypto_hash_update_page(req, pga[i]->pg,
1324                                             pga[i]->off & ~PAGE_MASK,
1325                                             count);
1326                 LL_CDEBUG_PAGE(D_PAGE, pga[i]->pg, "off %d\n",
1327                                (int)(pga[i]->off & ~PAGE_MASK));
1328
1329                 nob -= pga[i]->count;
1330                 pg_count--;
1331                 i++;
1332         }
1333
1334         bufsize = sizeof(*cksum);
1335         cfs_crypto_hash_final(req, (unsigned char *)cksum, &bufsize);
1336
1337         /* For sending we only compute the wrong checksum instead
1338          * of corrupting the data so it is still correct on a redo */
1339         if (opc == OST_WRITE && OBD_FAIL_CHECK(OBD_FAIL_OSC_CHECKSUM_SEND))
1340                 (*cksum)++;
1341
1342         return 0;
1343 }
1344
1345 static int osc_checksum_bulk_rw(const char *obd_name,
1346                                 enum cksum_types cksum_type,
1347                                 int nob, size_t pg_count,
1348                                 struct brw_page **pga, int opc,
1349                                 u32 *check_sum, bool resend)
1350 {
1351         obd_dif_csum_fn *fn = NULL;
1352         int sector_size = 0;
1353         int rc;
1354
1355         ENTRY;
1356         obd_t10_cksum2dif(cksum_type, &fn, &sector_size);
1357
1358         if (fn)
1359                 rc = osc_checksum_bulk_t10pi(obd_name, nob, pg_count, pga,
1360                                              opc, fn, sector_size, check_sum,
1361                                              resend);
1362         else
1363                 rc = osc_checksum_bulk(nob, pg_count, pga, opc, cksum_type,
1364                                        check_sum);
1365
1366         RETURN(rc);
1367 }
1368
1369 static inline void osc_release_bounce_pages(struct brw_page **pga,
1370                                             u32 page_count)
1371 {
1372 #ifdef HAVE_LUSTRE_CRYPTO
1373         int i;
1374
1375         for (i = 0; i < page_count; i++) {
1376                 /* Bounce pages allocated by a call to
1377                  * llcrypt_encrypt_pagecache_blocks() in osc_brw_prep_request()
1378                  * are identified thanks to the PageChecked flag.
1379                  */
1380                 if (PageChecked(pga[i]->pg))
1381                         llcrypt_finalize_bounce_page(&pga[i]->pg);
1382                 pga[i]->count -= pga[i]->bp_count_diff;
1383                 pga[i]->off += pga[i]->bp_off_diff;
1384         }
1385 #endif
1386 }
1387
1388 static int
1389 osc_brw_prep_request(int cmd, struct client_obd *cli, struct obdo *oa,
1390                      u32 page_count, struct brw_page **pga,
1391                      struct ptlrpc_request **reqp, int resend)
1392 {
1393         struct ptlrpc_request *req;
1394         struct ptlrpc_bulk_desc *desc;
1395         struct ost_body *body;
1396         struct obd_ioobj *ioobj;
1397         struct niobuf_remote *niobuf;
1398         int niocount, i, requested_nob, opc, rc, short_io_size = 0;
1399         struct osc_brw_async_args *aa;
1400         struct req_capsule *pill;
1401         struct brw_page *pg_prev;
1402         void *short_io_buf;
1403         const char *obd_name = cli->cl_import->imp_obd->obd_name;
1404         struct inode *inode = NULL;
1405         bool directio = false;
1406         bool enable_checksum = true;
1407
1408         ENTRY;
1409         if (pga[0]->pg) {
1410                 inode = page2inode(pga[0]->pg);
1411                 if (inode == NULL) {
1412                         /* Try to get reference to inode from cl_page if we are
1413                          * dealing with direct IO, as handled pages are not
1414                          * actual page cache pages.
1415                          */
1416                         struct osc_async_page *oap = brw_page2oap(pga[0]);
1417                         struct cl_page *clpage = oap2cl_page(oap);
1418
1419                         inode = clpage->cp_inode;
1420                         if (inode)
1421                                 directio = true;
1422                 }
1423         }
1424         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ))
1425                 RETURN(-ENOMEM); /* Recoverable */
1426         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_BRW_PREP_REQ2))
1427                 RETURN(-EINVAL); /* Fatal */
1428
1429         if ((cmd & OBD_BRW_WRITE) != 0) {
1430                 opc = OST_WRITE;
1431                 req = ptlrpc_request_alloc_pool(cli->cl_import,
1432                                                 osc_rq_pool,
1433                                                 &RQF_OST_BRW_WRITE);
1434         } else {
1435                 opc = OST_READ;
1436                 req = ptlrpc_request_alloc(cli->cl_import, &RQF_OST_BRW_READ);
1437         }
1438         if (req == NULL)
1439                 RETURN(-ENOMEM);
1440
1441         if (opc == OST_WRITE && inode && IS_ENCRYPTED(inode)) {
1442                 for (i = 0; i < page_count; i++) {
1443                         struct brw_page *pg = pga[i];
1444                         struct page *data_page = NULL;
1445                         bool retried = false;
1446                         bool lockedbymyself;
1447                         u32 nunits = (pg->off & ~PAGE_MASK) + pg->count;
1448                         struct address_space *map_orig = NULL;
1449                         pgoff_t index_orig;
1450
1451 retry_encrypt:
1452                         if (nunits & ~LUSTRE_ENCRYPTION_MASK)
1453                                 nunits = (nunits & LUSTRE_ENCRYPTION_MASK) +
1454                                         LUSTRE_ENCRYPTION_UNIT_SIZE;
1455                         /* The page can already be locked when we arrive here.
1456                          * This is possible when cl_page_assume/vvp_page_assume
1457                          * is stuck on wait_on_page_writeback with page lock
1458                          * held. In this case there is no risk for the lock to
1459                          * be released while we are doing our encryption
1460                          * processing, because writeback against that page will
1461                          * end in vvp_page_completion_write/cl_page_completion,
1462                          * which means only once the page is fully processed.
1463                          */
1464                         lockedbymyself = trylock_page(pg->pg);
1465                         if (directio) {
1466                                 map_orig = pg->pg->mapping;
1467                                 pg->pg->mapping = inode->i_mapping;
1468                                 index_orig = pg->pg->index;
1469                                 pg->pg->index = pg->off >> PAGE_SHIFT;
1470                         }
1471                         data_page =
1472                                 llcrypt_encrypt_pagecache_blocks(pg->pg,
1473                                                                  nunits, 0,
1474                                                                  GFP_NOFS);
1475                         if (directio) {
1476                                 pg->pg->mapping = map_orig;
1477                                 pg->pg->index = index_orig;
1478                         }
1479                         if (lockedbymyself)
1480                                 unlock_page(pg->pg);
1481                         if (IS_ERR(data_page)) {
1482                                 rc = PTR_ERR(data_page);
1483                                 if (rc == -ENOMEM && !retried) {
1484                                         retried = true;
1485                                         rc = 0;
1486                                         goto retry_encrypt;
1487                                 }
1488                                 ptlrpc_request_free(req);
1489                                 RETURN(rc);
1490                         }
1491                         /* Set PageChecked flag on bounce page for
1492                          * disambiguation in osc_release_bounce_pages().
1493                          */
1494                         SetPageChecked(data_page);
1495                         pg->pg = data_page;
1496                         /* there should be no gap in the middle of page array */
1497                         if (i == page_count - 1) {
1498                                 struct osc_async_page *oap = brw_page2oap(pg);
1499
1500                                 oa->o_size = oap->oap_count +
1501                                         oap->oap_obj_off + oap->oap_page_off;
1502                         }
1503                         /* len is forced to nunits, and relative offset to 0
1504                          * so store the old, clear text info
1505                          */
1506                         pg->bp_count_diff = nunits - pg->count;
1507                         pg->count = nunits;
1508                         pg->bp_off_diff = pg->off & ~PAGE_MASK;
1509                         pg->off = pg->off & PAGE_MASK;
1510                 }
1511         } else if (opc == OST_READ && inode && IS_ENCRYPTED(inode)) {
1512                 for (i = 0; i < page_count; i++) {
1513                         struct brw_page *pg = pga[i];
1514                         u32 nunits = (pg->off & ~PAGE_MASK) + pg->count;
1515
1516                         if (nunits & ~LUSTRE_ENCRYPTION_MASK)
1517                                 nunits = (nunits & LUSTRE_ENCRYPTION_MASK) +
1518                                         LUSTRE_ENCRYPTION_UNIT_SIZE;
1519                         /* count/off are forced to cover the whole encryption
1520                          * unit size so that all encrypted data is stored on the
1521                          * OST, so adjust bp_{count,off}_diff for the size of
1522                          * the clear text.
1523                          */
1524                         pg->bp_count_diff = nunits - pg->count;
1525                         pg->count = nunits;
1526                         pg->bp_off_diff = pg->off & ~PAGE_MASK;
1527                         pg->off = pg->off & PAGE_MASK;
1528                 }
1529         }
1530
1531         for (niocount = i = 1; i < page_count; i++) {
1532                 if (!can_merge_pages(pga[i - 1], pga[i]))
1533                         niocount++;
1534         }
1535
1536         pill = &req->rq_pill;
1537         req_capsule_set_size(pill, &RMF_OBD_IOOBJ, RCL_CLIENT,
1538                              sizeof(*ioobj));
1539         req_capsule_set_size(pill, &RMF_NIOBUF_REMOTE, RCL_CLIENT,
1540                              niocount * sizeof(*niobuf));
1541
1542         for (i = 0; i < page_count; i++) {
1543                 short_io_size += pga[i]->count;
1544                 if (!inode || !IS_ENCRYPTED(inode)) {
1545                         pga[i]->bp_count_diff = 0;
1546                         pga[i]->bp_off_diff = 0;
1547                 }
1548         }
1549
1550         if (lnet_is_rdma_only_page(pga[0]->pg)) {
1551                 enable_checksum = false;
1552                 short_io_size = 0;
1553         }
1554
1555         /* Check if read/write is small enough to be a short io. */
1556         if (short_io_size > cli->cl_max_short_io_bytes || niocount > 1 ||
1557             !imp_connect_shortio(cli->cl_import))
1558                 short_io_size = 0;
1559
1560         /* If this is an empty RPC to old server, just ignore it */
1561         if (!short_io_size && !pga[0]->pg) {
1562                 ptlrpc_request_free(req);
1563                 RETURN(-ENODATA);
1564         }
1565
1566         req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_CLIENT,
1567                              opc == OST_READ ? 0 : short_io_size);
1568         if (opc == OST_READ)
1569                 req_capsule_set_size(pill, &RMF_SHORT_IO, RCL_SERVER,
1570                                      short_io_size);
1571
1572         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, opc);
1573         if (rc) {
1574                 ptlrpc_request_free(req);
1575                 RETURN(rc);
1576         }
1577         osc_set_io_portal(req);
1578
1579         ptlrpc_at_set_req_timeout(req);
1580         /* ask ptlrpc not to resend on EINPROGRESS since BRWs have their own
1581          * retry logic */
1582         req->rq_no_retry_einprogress = 1;
1583
1584         if (short_io_size != 0) {
1585                 desc = NULL;
1586                 short_io_buf = NULL;
1587                 goto no_bulk;
1588         }
1589
1590         desc = ptlrpc_prep_bulk_imp(req, page_count,
1591                 cli->cl_import->imp_connect_data.ocd_brw_size >> LNET_MTU_BITS,
1592                 (opc == OST_WRITE ? PTLRPC_BULK_GET_SOURCE :
1593                         PTLRPC_BULK_PUT_SINK),
1594                 OST_BULK_PORTAL,
1595                 &ptlrpc_bulk_kiov_pin_ops);
1596
1597         if (desc == NULL)
1598                 GOTO(out, rc = -ENOMEM);
1599         /* NB request now owns desc and will free it when it gets freed */
1600 no_bulk:
1601         body = req_capsule_client_get(pill, &RMF_OST_BODY);
1602         ioobj = req_capsule_client_get(pill, &RMF_OBD_IOOBJ);
1603         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1604         LASSERT(body != NULL && ioobj != NULL && niobuf != NULL);
1605
1606         lustre_set_wire_obdo(&req->rq_import->imp_connect_data, &body->oa, oa);
1607
1608         /* For READ and WRITE, we can't fill o_uid and o_gid using from_kuid()
1609          * and from_kgid(), because they are asynchronous. Fortunately, variable
1610          * oa contains valid o_uid and o_gid in these two operations.
1611          * Besides, filling o_uid and o_gid is enough for nrs-tbf, see LU-9658.
1612          * OBD_MD_FLUID and OBD_MD_FLUID is not set in order to avoid breaking
1613          * other process logic */
1614         body->oa.o_uid = oa->o_uid;
1615         body->oa.o_gid = oa->o_gid;
1616
1617         obdo_to_ioobj(oa, ioobj);
1618         ioobj->ioo_bufcnt = niocount;
1619         /* The high bits of ioo_max_brw tells server _maximum_ number of bulks
1620          * that might be send for this request.  The actual number is decided
1621          * when the RPC is finally sent in ptlrpc_register_bulk(). It sends
1622          * "max - 1" for old client compatibility sending "0", and also so the
1623          * the actual maximum is a power-of-two number, not one less. LU-1431 */
1624         if (desc != NULL)
1625                 ioobj_max_brw_set(ioobj, desc->bd_md_max_brw);
1626         else /* short io */
1627                 ioobj_max_brw_set(ioobj, 0);
1628
1629         if (short_io_size != 0) {
1630                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1631                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1632                         body->oa.o_flags = 0;
1633                 }
1634                 body->oa.o_flags |= OBD_FL_SHORT_IO;
1635                 CDEBUG(D_CACHE, "Using short io for data transfer, size = %d\n",
1636                        short_io_size);
1637                 if (opc == OST_WRITE) {
1638                         short_io_buf = req_capsule_client_get(pill,
1639                                                               &RMF_SHORT_IO);
1640                         LASSERT(short_io_buf != NULL);
1641                 }
1642         }
1643
1644         LASSERT(page_count > 0);
1645         pg_prev = pga[0];
1646         for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
1647                 struct brw_page *pg = pga[i];
1648                 int poff = pg->off & ~PAGE_MASK;
1649
1650                 LASSERT(pg->count > 0);
1651                 /* make sure there is no gap in the middle of page array */
1652                 LASSERTF(page_count == 1 ||
1653                          (ergo(i == 0, poff + pg->count == PAGE_SIZE) &&
1654                           ergo(i > 0 && i < page_count - 1,
1655                                poff == 0 && pg->count == PAGE_SIZE)   &&
1656                           ergo(i == page_count - 1, poff == 0)),
1657                          "i: %d/%d pg: %p off: %llu, count: %u\n",
1658                          i, page_count, pg, pg->off, pg->count);
1659                 LASSERTF(i == 0 || pg->off > pg_prev->off,
1660                          "i %d p_c %u pg %p [pri %lu ind %lu] off %llu"
1661                          " prev_pg %p [pri %lu ind %lu] off %llu\n",
1662                          i, page_count,
1663                          pg->pg, page_private(pg->pg), pg->pg->index, pg->off,
1664                          pg_prev->pg, page_private(pg_prev->pg),
1665                          pg_prev->pg->index, pg_prev->off);
1666                 LASSERT((pga[0]->flag & OBD_BRW_SRVLOCK) ==
1667                         (pg->flag & OBD_BRW_SRVLOCK));
1668                 if (short_io_size != 0 && opc == OST_WRITE) {
1669                         unsigned char *ptr = kmap_atomic(pg->pg);
1670
1671                         LASSERT(short_io_size >= requested_nob + pg->count);
1672                         memcpy(short_io_buf + requested_nob,
1673                                ptr + poff,
1674                                pg->count);
1675                         kunmap_atomic(ptr);
1676                 } else if (short_io_size == 0) {
1677                         desc->bd_frag_ops->add_kiov_frag(desc, pg->pg, poff,
1678                                                          pg->count);
1679                 }
1680                 requested_nob += pg->count;
1681
1682                 if (i > 0 && can_merge_pages(pg_prev, pg)) {
1683                         niobuf--;
1684                         niobuf->rnb_len += pg->count;
1685                 } else {
1686                         niobuf->rnb_offset = pg->off;
1687                         niobuf->rnb_len    = pg->count;
1688                         niobuf->rnb_flags  = pg->flag;
1689                 }
1690                 pg_prev = pg;
1691         }
1692
1693         LASSERTF((void *)(niobuf - niocount) ==
1694                 req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE),
1695                 "want %p - real %p\n", req_capsule_client_get(&req->rq_pill,
1696                 &RMF_NIOBUF_REMOTE), (void *)(niobuf - niocount));
1697
1698         osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
1699         if (resend) {
1700                 if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0) {
1701                         body->oa.o_valid |= OBD_MD_FLFLAGS;
1702                         body->oa.o_flags = 0;
1703                 }
1704                 body->oa.o_flags |= OBD_FL_RECOV_RESEND;
1705         }
1706
1707         if (osc_should_shrink_grant(cli))
1708                 osc_shrink_grant_local(cli, &body->oa);
1709
1710         if (!cli->cl_checksum || sptlrpc_flavor_has_bulk(&req->rq_flvr))
1711                 enable_checksum = false;
1712
1713         /* size[REQ_REC_OFF] still sizeof (*body) */
1714         if (opc == OST_WRITE) {
1715                 if (enable_checksum) {
1716                         /* store cl_cksum_type in a local variable since
1717                          * it can be changed via lprocfs */
1718                         enum cksum_types cksum_type = cli->cl_cksum_type;
1719
1720                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1721                                 body->oa.o_flags = 0;
1722
1723                         body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1724                                                                 cksum_type);
1725                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1726
1727                         rc = osc_checksum_bulk_rw(obd_name, cksum_type,
1728                                                   requested_nob, page_count,
1729                                                   pga, OST_WRITE,
1730                                                   &body->oa.o_cksum, resend);
1731                         if (rc < 0) {
1732                                 CDEBUG(D_PAGE, "failed to checksum: rc = %d\n",
1733                                        rc);
1734                                 GOTO(out, rc);
1735                         }
1736                         CDEBUG(D_PAGE | (resend ? D_HA : 0),
1737                                "checksum at write origin: %x (%x)\n",
1738                                body->oa.o_cksum, cksum_type);
1739
1740                         /* save this in 'oa', too, for later checking */
1741                         oa->o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1742                         oa->o_flags |= obd_cksum_type_pack(obd_name,
1743                                                            cksum_type);
1744                 } else {
1745                         /* clear out the checksum flag, in case this is a
1746                          * resend but cl_checksum is no longer set. b=11238 */
1747                         oa->o_valid &= ~OBD_MD_FLCKSUM;
1748                 }
1749                 oa->o_cksum = body->oa.o_cksum;
1750                 /* 1 RC per niobuf */
1751                 req_capsule_set_size(pill, &RMF_RCS, RCL_SERVER,
1752                                      sizeof(__u32) * niocount);
1753         } else {
1754                 if (enable_checksum) {
1755                         if ((body->oa.o_valid & OBD_MD_FLFLAGS) == 0)
1756                                 body->oa.o_flags = 0;
1757                         body->oa.o_flags |= obd_cksum_type_pack(obd_name,
1758                                 cli->cl_cksum_type);
1759                         body->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1760                 }
1761
1762                 /* Client cksum has been already copied to wire obdo in previous
1763                  * lustre_set_wire_obdo(), and in the case a bulk-read is being
1764                  * resent due to cksum error, this will allow Server to
1765                  * check+dump pages on its side */
1766         }
1767         ptlrpc_request_set_replen(req);
1768
1769         aa = ptlrpc_req_async_args(aa, req);
1770         aa->aa_oa = oa;
1771         aa->aa_requested_nob = requested_nob;
1772         aa->aa_nio_count = niocount;
1773         aa->aa_page_count = page_count;
1774         aa->aa_resends = 0;
1775         aa->aa_ppga = pga;
1776         aa->aa_cli = cli;
1777         INIT_LIST_HEAD(&aa->aa_oaps);
1778
1779         *reqp = req;
1780         niobuf = req_capsule_client_get(pill, &RMF_NIOBUF_REMOTE);
1781         CDEBUG(D_RPCTRACE, "brw rpc %p - object "DOSTID" offset %lld<>%lld\n",
1782                 req, POSTID(&oa->o_oi), niobuf[0].rnb_offset,
1783                 niobuf[niocount - 1].rnb_offset + niobuf[niocount - 1].rnb_len);
1784         RETURN(0);
1785
1786  out:
1787         ptlrpc_req_finished(req);
1788         RETURN(rc);
1789 }
1790
1791 char dbgcksum_file_name[PATH_MAX];
1792
1793 static void dump_all_bulk_pages(struct obdo *oa, __u32 page_count,
1794                                 struct brw_page **pga, __u32 server_cksum,
1795                                 __u32 client_cksum)
1796 {
1797         struct file *filp;
1798         int rc, i;
1799         unsigned int len;
1800         char *buf;
1801
1802         /* will only keep dump of pages on first error for the same range in
1803          * file/fid, not during the resends/retries. */
1804         snprintf(dbgcksum_file_name, sizeof(dbgcksum_file_name),
1805                  "%s-checksum_dump-osc-"DFID":[%llu-%llu]-%x-%x",
1806                  (strncmp(libcfs_debug_file_path, "NONE", 4) != 0 ?
1807                   libcfs_debug_file_path : LIBCFS_DEBUG_FILE_PATH_DEFAULT),
1808                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : 0ULL,
1809                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1810                  oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1811                  pga[0]->off,
1812                  pga[page_count-1]->off + pga[page_count-1]->count - 1,
1813                  client_cksum, server_cksum);
1814         CWARN("dumping checksum data to %s\n", dbgcksum_file_name);
1815         filp = filp_open(dbgcksum_file_name,
1816                          O_CREAT | O_EXCL | O_WRONLY | O_LARGEFILE, 0600);
1817         if (IS_ERR(filp)) {
1818                 rc = PTR_ERR(filp);
1819                 if (rc == -EEXIST)
1820                         CDEBUG(D_INFO, "%s: can't open to dump pages with "
1821                                "checksum error: rc = %d\n", dbgcksum_file_name,
1822                                rc);
1823                 else
1824                         CERROR("%s: can't open to dump pages with checksum "
1825                                "error: rc = %d\n", dbgcksum_file_name, rc);
1826                 return;
1827         }
1828
1829         for (i = 0; i < page_count; i++) {
1830                 len = pga[i]->count;
1831                 buf = kmap(pga[i]->pg);
1832                 while (len != 0) {
1833                         rc = cfs_kernel_write(filp, buf, len, &filp->f_pos);
1834                         if (rc < 0) {
1835                                 CERROR("%s: wanted to write %u but got %d "
1836                                        "error\n", dbgcksum_file_name, len, rc);
1837                                 break;
1838                         }
1839                         len -= rc;
1840                         buf += rc;
1841                 }
1842                 kunmap(pga[i]->pg);
1843         }
1844
1845         rc = vfs_fsync_range(filp, 0, LLONG_MAX, 1);
1846         if (rc)
1847                 CERROR("%s: sync returns %d\n", dbgcksum_file_name, rc);
1848         filp_close(filp, NULL);
1849
1850         libcfs_debug_dumplog();
1851 }
1852
1853 static int
1854 check_write_checksum(struct obdo *oa, const struct lnet_process_id *peer,
1855                      __u32 client_cksum, __u32 server_cksum,
1856                      struct osc_brw_async_args *aa)
1857 {
1858         const char *obd_name = aa->aa_cli->cl_import->imp_obd->obd_name;
1859         enum cksum_types cksum_type;
1860         obd_dif_csum_fn *fn = NULL;
1861         int sector_size = 0;
1862         __u32 new_cksum;
1863         char *msg;
1864         int rc;
1865
1866         if (server_cksum == client_cksum) {
1867                 CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
1868                 return 0;
1869         }
1870
1871         if (aa->aa_cli->cl_checksum_dump)
1872                 dump_all_bulk_pages(oa, aa->aa_page_count, aa->aa_ppga,
1873                                     server_cksum, client_cksum);
1874
1875         cksum_type = obd_cksum_type_unpack(oa->o_valid & OBD_MD_FLFLAGS ?
1876                                            oa->o_flags : 0);
1877
1878         switch (cksum_type) {
1879         case OBD_CKSUM_T10IP512:
1880                 fn = obd_dif_ip_fn;
1881                 sector_size = 512;
1882                 break;
1883         case OBD_CKSUM_T10IP4K:
1884                 fn = obd_dif_ip_fn;
1885                 sector_size = 4096;
1886                 break;
1887         case OBD_CKSUM_T10CRC512:
1888                 fn = obd_dif_crc_fn;
1889                 sector_size = 512;
1890                 break;
1891         case OBD_CKSUM_T10CRC4K:
1892                 fn = obd_dif_crc_fn;
1893                 sector_size = 4096;
1894                 break;
1895         default:
1896                 break;
1897         }
1898
1899         if (fn)
1900                 rc = osc_checksum_bulk_t10pi(obd_name, aa->aa_requested_nob,
1901                                              aa->aa_page_count, aa->aa_ppga,
1902                                              OST_WRITE, fn, sector_size,
1903                                              &new_cksum, true);
1904         else
1905                 rc = osc_checksum_bulk(aa->aa_requested_nob, aa->aa_page_count,
1906                                        aa->aa_ppga, OST_WRITE, cksum_type,
1907                                        &new_cksum);
1908
1909         if (rc < 0)
1910                 msg = "failed to calculate the client write checksum";
1911         else if (cksum_type != obd_cksum_type_unpack(aa->aa_oa->o_flags))
1912                 msg = "the server did not use the checksum type specified in "
1913                       "the original request - likely a protocol problem";
1914         else if (new_cksum == server_cksum)
1915                 msg = "changed on the client after we checksummed it - "
1916                       "likely false positive due to mmap IO (bug 11742)";
1917         else if (new_cksum == client_cksum)
1918                 msg = "changed in transit before arrival at OST";
1919         else
1920                 msg = "changed in transit AND doesn't match the original - "
1921                       "likely false positive due to mmap IO (bug 11742)";
1922
1923         LCONSOLE_ERROR_MSG(0x132, "%s: BAD WRITE CHECKSUM: %s: from %s inode "
1924                            DFID " object "DOSTID" extent [%llu-%llu], original "
1925                            "client csum %x (type %x), server csum %x (type %x),"
1926                            " client csum now %x\n",
1927                            obd_name, msg, libcfs_nid2str(peer->nid),
1928                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_seq : (__u64)0,
1929                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_oid : 0,
1930                            oa->o_valid & OBD_MD_FLFID ? oa->o_parent_ver : 0,
1931                            POSTID(&oa->o_oi), aa->aa_ppga[0]->off,
1932                            aa->aa_ppga[aa->aa_page_count - 1]->off +
1933                                 aa->aa_ppga[aa->aa_page_count-1]->count - 1,
1934                            client_cksum,
1935                            obd_cksum_type_unpack(aa->aa_oa->o_flags),
1936                            server_cksum, cksum_type, new_cksum);
1937         return 1;
1938 }
1939
1940 /* Note rc enters this function as number of bytes transferred */
1941 static int osc_brw_fini_request(struct ptlrpc_request *req, int rc)
1942 {
1943         struct osc_brw_async_args *aa = (void *)&req->rq_async_args;
1944         struct client_obd *cli = aa->aa_cli;
1945         const char *obd_name = cli->cl_import->imp_obd->obd_name;
1946         const struct lnet_process_id *peer =
1947                 &req->rq_import->imp_connection->c_peer;
1948         struct ost_body *body;
1949         u32 client_cksum = 0;
1950         struct inode *inode;
1951         unsigned int blockbits = 0, blocksize = 0;
1952
1953         ENTRY;
1954
1955         if (rc < 0 && rc != -EDQUOT) {
1956                 DEBUG_REQ(D_INFO, req, "Failed request: rc = %d", rc);
1957                 RETURN(rc);
1958         }
1959
1960         LASSERTF(req->rq_repmsg != NULL, "rc = %d\n", rc);
1961         body = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1962         if (body == NULL) {
1963                 DEBUG_REQ(D_INFO, req, "cannot unpack body");
1964                 RETURN(-EPROTO);
1965         }
1966
1967         /* set/clear over quota flag for a uid/gid/projid */
1968         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE &&
1969             body->oa.o_valid & (OBD_MD_FLALLQUOTA)) {
1970                 unsigned qid[LL_MAXQUOTAS] = {
1971                                          body->oa.o_uid, body->oa.o_gid,
1972                                          body->oa.o_projid };
1973                 CDEBUG(D_QUOTA,
1974                        "setdq for [%u %u %u] with valid %#llx, flags %x\n",
1975                        body->oa.o_uid, body->oa.o_gid, body->oa.o_projid,
1976                        body->oa.o_valid, body->oa.o_flags);
1977                        osc_quota_setdq(cli, req->rq_xid, qid, body->oa.o_valid,
1978                                        body->oa.o_flags);
1979         }
1980
1981         osc_update_grant(cli, body);
1982
1983         if (rc < 0)
1984                 RETURN(rc);
1985
1986         if (aa->aa_oa->o_valid & OBD_MD_FLCKSUM)
1987                 client_cksum = aa->aa_oa->o_cksum; /* save for later */
1988
1989         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
1990                 if (rc > 0) {
1991                         CERROR("%s: unexpected positive size %d\n",
1992                                obd_name, rc);
1993                         RETURN(-EPROTO);
1994                 }
1995
1996                 if (req->rq_bulk != NULL &&
1997                     sptlrpc_cli_unwrap_bulk_write(req, req->rq_bulk))
1998                         RETURN(-EAGAIN);
1999
2000                 if ((aa->aa_oa->o_valid & OBD_MD_FLCKSUM) && client_cksum &&
2001                     check_write_checksum(&body->oa, peer, client_cksum,
2002                                          body->oa.o_cksum, aa))
2003                         RETURN(-EAGAIN);
2004
2005                 rc = check_write_rcs(req, aa->aa_requested_nob,
2006                                      aa->aa_nio_count, aa->aa_page_count,
2007                                      aa->aa_ppga);
2008                 GOTO(out, rc);
2009         }
2010
2011         /* The rest of this function executes only for OST_READs */
2012
2013         if (req->rq_bulk == NULL) {
2014                 rc = req_capsule_get_size(&req->rq_pill, &RMF_SHORT_IO,
2015                                           RCL_SERVER);
2016                 LASSERT(rc == req->rq_status);
2017         } else {
2018                 /* if unwrap_bulk failed, return -EAGAIN to retry */
2019                 rc = sptlrpc_cli_unwrap_bulk_read(req, req->rq_bulk, rc);
2020         }
2021         if (rc < 0)
2022                 GOTO(out, rc = -EAGAIN);
2023
2024         if (rc > aa->aa_requested_nob) {
2025                 CERROR("%s: unexpected size %d, requested %d\n", obd_name,
2026                        rc, aa->aa_requested_nob);
2027                 RETURN(-EPROTO);
2028         }
2029
2030         if (req->rq_bulk != NULL && rc != req->rq_bulk->bd_nob_transferred) {
2031                 CERROR("%s: unexpected size %d, transferred %d\n", obd_name,
2032                        rc, req->rq_bulk->bd_nob_transferred);
2033                 RETURN(-EPROTO);
2034         }
2035
2036         if (req->rq_bulk == NULL) {
2037                 /* short io */
2038                 int nob, pg_count, i = 0;
2039                 unsigned char *buf;
2040
2041                 CDEBUG(D_CACHE, "Using short io read, size %d\n", rc);
2042                 pg_count = aa->aa_page_count;
2043                 buf = req_capsule_server_sized_get(&req->rq_pill, &RMF_SHORT_IO,
2044                                                    rc);
2045                 nob = rc;
2046                 while (nob > 0 && pg_count > 0) {
2047                         unsigned char *ptr;
2048                         int count = aa->aa_ppga[i]->count > nob ?
2049                                     nob : aa->aa_ppga[i]->count;
2050
2051                         CDEBUG(D_CACHE, "page %p count %d\n",
2052                                aa->aa_ppga[i]->pg, count);
2053                         ptr = kmap_atomic(aa->aa_ppga[i]->pg);
2054                         memcpy(ptr + (aa->aa_ppga[i]->off & ~PAGE_MASK), buf,
2055                                count);
2056                         kunmap_atomic((void *) ptr);
2057
2058                         buf += count;
2059                         nob -= count;
2060                         i++;
2061                         pg_count--;
2062                 }
2063         }
2064
2065         if (rc < aa->aa_requested_nob)
2066                 handle_short_read(rc, aa->aa_page_count, aa->aa_ppga);
2067
2068         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
2069                 static int cksum_counter;
2070                 u32 server_cksum = body->oa.o_cksum;
2071                 int nob = rc;
2072                 char *via = "";
2073                 char *router = "";
2074                 enum cksum_types cksum_type;
2075                 u32 o_flags = body->oa.o_valid & OBD_MD_FLFLAGS ?
2076                         body->oa.o_flags : 0;
2077
2078                 cksum_type = obd_cksum_type_unpack(o_flags);
2079                 rc = osc_checksum_bulk_rw(obd_name, cksum_type, nob,
2080                                           aa->aa_page_count, aa->aa_ppga,
2081                                           OST_READ, &client_cksum, false);
2082                 if (rc < 0)
2083                         GOTO(out, rc);
2084
2085                 if (req->rq_bulk != NULL &&
2086                     peer->nid != req->rq_bulk->bd_sender) {
2087                         via = " via ";
2088                         router = libcfs_nid2str(req->rq_bulk->bd_sender);
2089                 }
2090
2091                 if (server_cksum != client_cksum) {
2092                         struct ost_body *clbody;
2093                         __u32 client_cksum2;
2094                         u32 page_count = aa->aa_page_count;
2095
2096                         osc_checksum_bulk_rw(obd_name, cksum_type, nob,
2097                                              page_count, aa->aa_ppga,
2098                                              OST_READ, &client_cksum2, true);
2099                         clbody = req_capsule_client_get(&req->rq_pill,
2100                                                         &RMF_OST_BODY);
2101                         if (cli->cl_checksum_dump)
2102                                 dump_all_bulk_pages(&clbody->oa, page_count,
2103                                                     aa->aa_ppga, server_cksum,
2104                                                     client_cksum);
2105
2106                         LCONSOLE_ERROR_MSG(0x133, "%s: BAD READ CHECKSUM: from "
2107                                            "%s%s%s inode "DFID" object "DOSTID
2108                                            " extent [%llu-%llu], client %x/%x, "
2109                                            "server %x, cksum_type %x\n",
2110                                            obd_name,
2111                                            libcfs_nid2str(peer->nid),
2112                                            via, router,
2113                                            clbody->oa.o_valid & OBD_MD_FLFID ?
2114                                                 clbody->oa.o_parent_seq : 0ULL,
2115                                            clbody->oa.o_valid & OBD_MD_FLFID ?
2116                                                 clbody->oa.o_parent_oid : 0,
2117                                            clbody->oa.o_valid & OBD_MD_FLFID ?
2118                                                 clbody->oa.o_parent_ver : 0,
2119                                            POSTID(&body->oa.o_oi),
2120                                            aa->aa_ppga[0]->off,
2121                                            aa->aa_ppga[page_count-1]->off +
2122                                            aa->aa_ppga[page_count-1]->count - 1,
2123                                            client_cksum, client_cksum2,
2124                                            server_cksum, cksum_type);
2125                         cksum_counter = 0;
2126                         aa->aa_oa->o_cksum = client_cksum;
2127                         rc = -EAGAIN;
2128                 } else {
2129                         cksum_counter++;
2130                         CDEBUG(D_PAGE, "checksum %x confirmed\n", client_cksum);
2131                         rc = 0;
2132                 }
2133         } else if (unlikely(client_cksum)) {
2134                 static int cksum_missed;
2135
2136                 cksum_missed++;
2137                 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
2138                         CERROR("%s: checksum %u requested from %s but not sent\n",
2139                                obd_name, cksum_missed,
2140                                libcfs_nid2str(peer->nid));
2141         } else {
2142                 rc = 0;
2143         }
2144
2145         inode = page2inode(aa->aa_ppga[0]->pg);
2146         if (inode == NULL) {
2147                 /* Try to get reference to inode from cl_page if we are
2148                  * dealing with direct IO, as handled pages are not
2149                  * actual page cache pages.
2150                  */
2151                 struct osc_async_page *oap = brw_page2oap(aa->aa_ppga[0]);
2152
2153                 inode = oap2cl_page(oap)->cp_inode;
2154                 if (inode) {
2155                         blockbits = inode->i_blkbits;
2156                         blocksize = 1 << blockbits;
2157                 }
2158         }
2159         if (inode && IS_ENCRYPTED(inode)) {
2160                 int idx;
2161
2162                 if (!llcrypt_has_encryption_key(inode)) {
2163                         CDEBUG(D_SEC, "no enc key for ino %lu\n", inode->i_ino);
2164                         GOTO(out, rc);
2165                 }
2166                 for (idx = 0; idx < aa->aa_page_count; idx++) {
2167                         struct brw_page *pg = aa->aa_ppga[idx];
2168                         unsigned int offs = 0;
2169
2170                         while (offs < PAGE_SIZE) {
2171                                 /* do not decrypt if page is all 0s */
2172                                 if (memchr_inv(page_address(pg->pg) + offs, 0,
2173                                          LUSTRE_ENCRYPTION_UNIT_SIZE) == NULL) {
2174                                         /* if page is empty forward info to
2175                                          * upper layers (ll_io_zero_page) by
2176                                          * clearing PagePrivate2
2177                                          */
2178                                         if (!offs)
2179                                                 ClearPagePrivate2(pg->pg);
2180                                         break;
2181                                 }
2182
2183                                 if (blockbits) {
2184                                         /* This is direct IO case. Directly call
2185                                          * decrypt function that takes inode as
2186                                          * input parameter. Page does not need
2187                                          * to be locked.
2188                                          */
2189                                         u64 lblk_num =
2190                                                 ((u64)(pg->off >> PAGE_SHIFT) <<
2191                                                      (PAGE_SHIFT - blockbits)) +
2192                                                        (offs >> blockbits);
2193                                         unsigned int i;
2194
2195                                         for (i = offs;
2196                                              i < offs +
2197                                                     LUSTRE_ENCRYPTION_UNIT_SIZE;
2198                                              i += blocksize, lblk_num++) {
2199                                                 rc =
2200                                                   llcrypt_decrypt_block_inplace(
2201                                                           inode, pg->pg,
2202                                                           blocksize, i,
2203                                                           lblk_num);
2204                                                 if (rc)
2205                                                         break;
2206                                         }
2207                                 } else {
2208                                         rc = llcrypt_decrypt_pagecache_blocks(
2209                                                 pg->pg,
2210                                                 LUSTRE_ENCRYPTION_UNIT_SIZE,
2211                                                 offs);
2212                                 }
2213                                 if (rc)
2214                                         GOTO(out, rc);
2215
2216                                 offs += LUSTRE_ENCRYPTION_UNIT_SIZE;
2217                         }
2218                 }
2219         }
2220
2221 out:
2222         if (rc >= 0)
2223                 lustre_get_wire_obdo(&req->rq_import->imp_connect_data,
2224                                      aa->aa_oa, &body->oa);
2225
2226         RETURN(rc);
2227 }
2228
2229 static int osc_brw_redo_request(struct ptlrpc_request *request,
2230                                 struct osc_brw_async_args *aa, int rc)
2231 {
2232         struct ptlrpc_request *new_req;
2233         struct osc_brw_async_args *new_aa;
2234         struct osc_async_page *oap;
2235         ENTRY;
2236
2237         /* The below message is checked in replay-ost-single.sh test_8ae*/
2238         DEBUG_REQ(rc == -EINPROGRESS ? D_RPCTRACE : D_ERROR, request,
2239                   "redo for recoverable error %d", rc);
2240
2241         rc = osc_brw_prep_request(lustre_msg_get_opc(request->rq_reqmsg) ==
2242                                 OST_WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
2243                                   aa->aa_cli, aa->aa_oa, aa->aa_page_count,
2244                                   aa->aa_ppga, &new_req, 1);
2245         if (rc)
2246                 RETURN(rc);
2247
2248         list_for_each_entry(oap, &aa->aa_oaps, oap_rpc_item) {
2249                 if (oap->oap_request != NULL) {
2250                         LASSERTF(request == oap->oap_request,
2251                                  "request %p != oap_request %p\n",
2252                                  request, oap->oap_request);
2253                 }
2254         }
2255         /*
2256          * New request takes over pga and oaps from old request.
2257          * Note that copying a list_head doesn't work, need to move it...
2258          */
2259         aa->aa_resends++;
2260         new_req->rq_interpret_reply = request->rq_interpret_reply;
2261         new_req->rq_async_args = request->rq_async_args;
2262         new_req->rq_commit_cb = request->rq_commit_cb;
2263         /* cap resend delay to the current request timeout, this is similar to
2264          * what ptlrpc does (see after_reply()) */
2265         if (aa->aa_resends > new_req->rq_timeout)
2266                 new_req->rq_sent = ktime_get_real_seconds() + new_req->rq_timeout;
2267         else
2268                 new_req->rq_sent = ktime_get_real_seconds() + aa->aa_resends;
2269         new_req->rq_generation_set = 1;
2270         new_req->rq_import_generation = request->rq_import_generation;
2271
2272         new_aa = ptlrpc_req_async_args(new_aa, new_req);
2273
2274         INIT_LIST_HEAD(&new_aa->aa_oaps);
2275         list_splice_init(&aa->aa_oaps, &new_aa->aa_oaps);
2276         INIT_LIST_HEAD(&new_aa->aa_exts);
2277         list_splice_init(&aa->aa_exts, &new_aa->aa_exts);
2278         new_aa->aa_resends = aa->aa_resends;
2279
2280         list_for_each_entry(oap, &new_aa->aa_oaps, oap_rpc_item) {
2281                 if (oap->oap_request) {
2282                         ptlrpc_req_finished(oap->oap_request);
2283                         oap->oap_request = ptlrpc_request_addref(new_req);
2284                 }
2285         }
2286
2287         /* XXX: This code will run into problem if we're going to support
2288          * to add a series of BRW RPCs into a self-defined ptlrpc_request_set
2289          * and wait for all of them to be finished. We should inherit request
2290          * set from old request. */
2291         ptlrpcd_add_req(new_req);
2292
2293         DEBUG_REQ(D_INFO, new_req, "new request");
2294         RETURN(0);
2295 }
2296
2297 /*
2298  * ugh, we want disk allocation on the target to happen in offset order.  we'll
2299  * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
2300  * fine for our small page arrays and doesn't require allocation.  its an
2301  * insertion sort that swaps elements that are strides apart, shrinking the
2302  * stride down until its '1' and the array is sorted.
2303  */
2304 static void sort_brw_pages(struct brw_page **array, int num)
2305 {
2306         int stride, i, j;
2307         struct brw_page *tmp;
2308
2309         if (num == 1)
2310                 return;
2311         for (stride = 1; stride < num ; stride = (stride * 3) + 1)
2312                 ;
2313
2314         do {
2315                 stride /= 3;
2316                 for (i = stride ; i < num ; i++) {
2317                         tmp = array[i];
2318                         j = i;
2319                         while (j >= stride && array[j - stride]->off > tmp->off) {
2320                                 array[j] = array[j - stride];
2321                                 j -= stride;
2322                         }
2323                         array[j] = tmp;
2324                 }
2325         } while (stride > 1);
2326 }
2327
2328 static void osc_release_ppga(struct brw_page **ppga, size_t count)
2329 {
2330         LASSERT(ppga != NULL);
2331         OBD_FREE_PTR_ARRAY_LARGE(ppga, count);
2332 }
2333
2334 static int brw_interpret(const struct lu_env *env,
2335                          struct ptlrpc_request *req, void *args, int rc)
2336 {
2337         struct osc_brw_async_args *aa = args;
2338         struct osc_extent *ext;
2339         struct osc_extent *tmp;
2340         struct client_obd *cli = aa->aa_cli;
2341         unsigned long transferred = 0;
2342
2343         ENTRY;
2344
2345         rc = osc_brw_fini_request(req, rc);
2346         CDEBUG(D_INODE, "request %p aa %p rc %d\n", req, aa, rc);
2347
2348         /* restore clear text pages */
2349         osc_release_bounce_pages(aa->aa_ppga, aa->aa_page_count);
2350
2351         /*
2352          * When server returns -EINPROGRESS, client should always retry
2353          * regardless of the number of times the bulk was resent already.
2354          */
2355         if (osc_recoverable_error(rc) && !req->rq_no_delay) {
2356                 if (req->rq_import_generation !=
2357                     req->rq_import->imp_generation) {
2358                         CDEBUG(D_HA, "%s: resend cross eviction for object: "
2359                                ""DOSTID", rc = %d.\n",
2360                                req->rq_import->imp_obd->obd_name,
2361                                POSTID(&aa->aa_oa->o_oi), rc);
2362                 } else if (rc == -EINPROGRESS ||
2363                            client_should_resend(aa->aa_resends, aa->aa_cli)) {
2364                         rc = osc_brw_redo_request(req, aa, rc);
2365                 } else {
2366                         CERROR("%s: too many resent retries for object: "
2367                                "%llu:%llu, rc = %d.\n",
2368                                req->rq_import->imp_obd->obd_name,
2369                                POSTID(&aa->aa_oa->o_oi), rc);
2370                 }
2371
2372                 if (rc == 0)
2373                         RETURN(0);
2374                 else if (rc == -EAGAIN || rc == -EINPROGRESS)
2375                         rc = -EIO;
2376         }
2377
2378         if (rc == 0) {
2379                 struct obdo *oa = aa->aa_oa;
2380                 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
2381                 unsigned long valid = 0;
2382                 struct cl_object *obj;
2383                 struct osc_async_page *last;
2384
2385                 last = brw_page2oap(aa->aa_ppga[aa->aa_page_count - 1]);
2386                 obj = osc2cl(last->oap_obj);
2387
2388                 cl_object_attr_lock(obj);
2389                 if (oa->o_valid & OBD_MD_FLBLOCKS) {
2390                         attr->cat_blocks = oa->o_blocks;
2391                         valid |= CAT_BLOCKS;
2392                 }
2393                 if (oa->o_valid & OBD_MD_FLMTIME) {
2394                         attr->cat_mtime = oa->o_mtime;
2395                         valid |= CAT_MTIME;
2396                 }
2397                 if (oa->o_valid & OBD_MD_FLATIME) {
2398                         attr->cat_atime = oa->o_atime;
2399                         valid |= CAT_ATIME;
2400                 }
2401                 if (oa->o_valid & OBD_MD_FLCTIME) {
2402                         attr->cat_ctime = oa->o_ctime;
2403                         valid |= CAT_CTIME;
2404                 }
2405
2406                 if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE) {
2407                         struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
2408                         loff_t last_off = last->oap_count + last->oap_obj_off +
2409                                 last->oap_page_off;
2410
2411                         /* Change file size if this is an out of quota or
2412                          * direct IO write and it extends the file size */
2413                         if (loi->loi_lvb.lvb_size < last_off) {
2414                                 attr->cat_size = last_off;
2415                                 valid |= CAT_SIZE;
2416                         }
2417                         /* Extend KMS if it's not a lockless write */
2418                         if (loi->loi_kms < last_off &&
2419                             oap2osc_page(last)->ops_srvlock == 0) {
2420                                 attr->cat_kms = last_off;
2421                                 valid |= CAT_KMS;
2422                         }
2423                 }
2424
2425                 if (valid != 0)
2426                         cl_object_attr_update(env, obj, attr, valid);
2427                 cl_object_attr_unlock(obj);
2428         }
2429         OBD_SLAB_FREE_PTR(aa->aa_oa, osc_obdo_kmem);
2430         aa->aa_oa = NULL;
2431
2432         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE && rc == 0)
2433                 osc_inc_unstable_pages(req);
2434
2435         list_for_each_entry_safe(ext, tmp, &aa->aa_exts, oe_link) {
2436                 list_del_init(&ext->oe_link);
2437                 osc_extent_finish(env, ext, 1,
2438                                   rc && req->rq_no_delay ? -EAGAIN : rc);
2439         }
2440         LASSERT(list_empty(&aa->aa_exts));
2441         LASSERT(list_empty(&aa->aa_oaps));
2442
2443         transferred = (req->rq_bulk == NULL ? /* short io */
2444                        aa->aa_requested_nob :
2445                        req->rq_bulk->bd_nob_transferred);
2446
2447         osc_release_ppga(aa->aa_ppga, aa->aa_page_count);
2448         ptlrpc_lprocfs_brw(req, transferred);
2449
2450         spin_lock(&cli->cl_loi_list_lock);
2451         /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
2452          * is called so we know whether to go to sync BRWs or wait for more
2453          * RPCs to complete */
2454         if (lustre_msg_get_opc(req->rq_reqmsg) == OST_WRITE)
2455                 cli->cl_w_in_flight--;
2456         else
2457                 cli->cl_r_in_flight--;
2458         osc_wake_cache_waiters(cli);
2459         spin_unlock(&cli->cl_loi_list_lock);
2460
2461         osc_io_unplug(env, cli, NULL);
2462         RETURN(rc);
2463 }
2464
2465 static void brw_commit(struct ptlrpc_request *req)
2466 {
2467         /* If osc_inc_unstable_pages (via osc_extent_finish) races with
2468          * this called via the rq_commit_cb, I need to ensure
2469          * osc_dec_unstable_pages is still called. Otherwise unstable
2470          * pages may be leaked. */
2471         spin_lock(&req->rq_lock);
2472         if (likely(req->rq_unstable)) {
2473                 req->rq_unstable = 0;
2474                 spin_unlock(&req->rq_lock);
2475
2476                 osc_dec_unstable_pages(req);
2477         } else {
2478                 req->rq_committed = 1;
2479                 spin_unlock(&req->rq_lock);
2480         }
2481 }
2482
2483 /**
2484  * Build an RPC by the list of extent @ext_list. The caller must ensure
2485  * that the total pages in this list are NOT over max pages per RPC.
2486  * Extents in the list must be in OES_RPC state.
2487  */
2488 int osc_build_rpc(const struct lu_env *env, struct client_obd *cli,
2489                   struct list_head *ext_list, int cmd)
2490 {
2491         struct ptlrpc_request           *req = NULL;
2492         struct osc_extent               *ext;
2493         struct brw_page                 **pga = NULL;
2494         struct osc_brw_async_args       *aa = NULL;
2495         struct obdo                     *oa = NULL;
2496         struct osc_async_page           *oap;
2497         struct osc_object               *obj = NULL;
2498         struct cl_req_attr              *crattr = NULL;
2499         loff_t                          starting_offset = OBD_OBJECT_EOF;
2500         loff_t                          ending_offset = 0;
2501         /* '1' for consistency with code that checks !mpflag to restore */
2502         int mpflag = 1;
2503         int                             mem_tight = 0;
2504         int                             page_count = 0;
2505         bool                            soft_sync = false;
2506         bool                            ndelay = false;
2507         int                             i;
2508         int                             grant = 0;
2509         int                             rc;
2510         __u32                           layout_version = 0;
2511         LIST_HEAD(rpc_list);
2512         struct ost_body                 *body;
2513         ENTRY;
2514         LASSERT(!list_empty(ext_list));
2515
2516         /* add pages into rpc_list to build BRW rpc */
2517         list_for_each_entry(ext, ext_list, oe_link) {
2518                 LASSERT(ext->oe_state == OES_RPC);
2519                 mem_tight |= ext->oe_memalloc;
2520                 grant += ext->oe_grants;
2521                 page_count += ext->oe_nr_pages;
2522                 layout_version = max(layout_version, ext->oe_layout_version);
2523                 if (obj == NULL)
2524                         obj = ext->oe_obj;
2525         }
2526
2527         soft_sync = osc_over_unstable_soft_limit(cli);
2528         if (mem_tight)
2529                 mpflag = memalloc_noreclaim_save();
2530
2531         OBD_ALLOC_PTR_ARRAY_LARGE(pga, page_count);
2532         if (pga == NULL)
2533                 GOTO(out, rc = -ENOMEM);
2534
2535         OBD_SLAB_ALLOC_PTR_GFP(oa, osc_obdo_kmem, GFP_NOFS);
2536         if (oa == NULL)
2537                 GOTO(out, rc = -ENOMEM);
2538
2539         i = 0;
2540         list_for_each_entry(ext, ext_list, oe_link) {
2541                 list_for_each_entry(oap, &ext->oe_pages, oap_pending_item) {
2542                         if (mem_tight)
2543                                 oap->oap_brw_flags |= OBD_BRW_MEMALLOC;
2544                         if (soft_sync)
2545                                 oap->oap_brw_flags |= OBD_BRW_SOFT_SYNC;
2546                         pga[i] = &oap->oap_brw_page;
2547                         pga[i]->off = oap->oap_obj_off + oap->oap_page_off;
2548                         i++;
2549
2550                         list_add_tail(&oap->oap_rpc_item, &rpc_list);
2551                         if (starting_offset == OBD_OBJECT_EOF ||
2552                             starting_offset > oap->oap_obj_off)
2553                                 starting_offset = oap->oap_obj_off;
2554                         else
2555                                 LASSERT(oap->oap_page_off == 0);
2556                         if (ending_offset < oap->oap_obj_off + oap->oap_count)
2557                                 ending_offset = oap->oap_obj_off +
2558                                                 oap->oap_count;
2559                         else
2560                                 LASSERT(oap->oap_page_off + oap->oap_count ==
2561                                         PAGE_SIZE);
2562                 }
2563                 if (ext->oe_ndelay)
2564                         ndelay = true;
2565         }
2566
2567         /* first page in the list */
2568         oap = list_first_entry(&rpc_list, typeof(*oap), oap_rpc_item);
2569
2570         crattr = &osc_env_info(env)->oti_req_attr;
2571         memset(crattr, 0, sizeof(*crattr));
2572         crattr->cra_type = (cmd & OBD_BRW_WRITE) ? CRT_WRITE : CRT_READ;
2573         crattr->cra_flags = ~0ULL;
2574         crattr->cra_page = oap2cl_page(oap);
2575         crattr->cra_oa = oa;
2576         cl_req_attr_set(env, osc2cl(obj), crattr);
2577
2578         if (cmd == OBD_BRW_WRITE) {
2579                 oa->o_grant_used = grant;
2580                 if (layout_version > 0) {
2581                         CDEBUG(D_LAYOUT, DFID": write with layout version %u\n",
2582                                PFID(&oa->o_oi.oi_fid), layout_version);
2583
2584                         oa->o_layout_version = layout_version;
2585                         oa->o_valid |= OBD_MD_LAYOUT_VERSION;
2586                 }
2587         }
2588
2589         sort_brw_pages(pga, page_count);
2590         rc = osc_brw_prep_request(cmd, cli, oa, page_count, pga, &req, 0);
2591         if (rc != 0) {
2592                 CERROR("prep_req failed: %d\n", rc);
2593                 GOTO(out, rc);
2594         }
2595
2596         req->rq_commit_cb = brw_commit;
2597         req->rq_interpret_reply = brw_interpret;
2598         req->rq_memalloc = mem_tight != 0;
2599         oap->oap_request = ptlrpc_request_addref(req);
2600         if (ndelay) {
2601                 req->rq_no_resend = req->rq_no_delay = 1;
2602                 /* probably set a shorter timeout value.
2603                  * to handle ETIMEDOUT in brw_interpret() correctly. */
2604                 /* lustre_msg_set_timeout(req, req->rq_timeout / 2); */
2605         }
2606
2607         /* Need to update the timestamps after the request is built in case
2608          * we race with setattr (locally or in queue at OST).  If OST gets
2609          * later setattr before earlier BRW (as determined by the request xid),
2610          * the OST will not use BRW timestamps.  Sadly, there is no obvious
2611          * way to do this in a single call.  bug 10150 */
2612         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2613         crattr->cra_oa = &body->oa;
2614         crattr->cra_flags = OBD_MD_FLMTIME | OBD_MD_FLCTIME | OBD_MD_FLATIME;
2615         cl_req_attr_set(env, osc2cl(obj), crattr);
2616         lustre_msg_set_jobid(req->rq_reqmsg, crattr->cra_jobid);
2617
2618         aa = ptlrpc_req_async_args(aa, req);
2619         INIT_LIST_HEAD(&aa->aa_oaps);
2620         list_splice_init(&rpc_list, &aa->aa_oaps);
2621         INIT_LIST_HEAD(&aa->aa_exts);
2622         list_splice_init(ext_list, &aa->aa_exts);
2623
2624         spin_lock(&cli->cl_loi_list_lock);
2625         starting_offset >>= PAGE_SHIFT;
2626         if (cmd == OBD_BRW_READ) {
2627                 cli->cl_r_in_flight++;
2628                 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
2629                 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
2630                 lprocfs_oh_tally_log2(&cli->cl_read_offset_hist,
2631                                       starting_offset + 1);
2632         } else {
2633                 cli->cl_w_in_flight++;
2634                 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
2635                 lprocfs_oh_tally(&cli->cl_write_rpc_hist, cli->cl_w_in_flight);
2636                 lprocfs_oh_tally_log2(&cli->cl_write_offset_hist,
2637                                       starting_offset + 1);
2638         }
2639         spin_unlock(&cli->cl_loi_list_lock);
2640
2641         DEBUG_REQ(D_INODE, req, "%d pages, aa %p, now %ur/%uw in flight",
2642                   page_count, aa, cli->cl_r_in_flight,
2643                   cli->cl_w_in_flight);
2644         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_DELAY_IO, cfs_fail_val);
2645
2646         ptlrpcd_add_req(req);
2647         rc = 0;
2648         EXIT;
2649
2650 out:
2651         if (mem_tight)
2652                 memalloc_noreclaim_restore(mpflag);
2653
2654         if (rc != 0) {
2655                 LASSERT(req == NULL);
2656
2657                 if (oa)
2658                         OBD_SLAB_FREE_PTR(oa, osc_obdo_kmem);
2659                 if (pga) {
2660                         osc_release_bounce_pages(pga, page_count);
2661                         osc_release_ppga(pga, page_count);
2662                 }
2663                 /* this should happen rarely and is pretty bad, it makes the
2664                  * pending list not follow the dirty order
2665                  */
2666                 while ((ext = list_first_entry_or_null(ext_list,
2667                                                        struct osc_extent,
2668                                                        oe_link)) != NULL) {
2669                         list_del_init(&ext->oe_link);
2670                         osc_extent_finish(env, ext, 0, rc);
2671                 }
2672         }
2673         RETURN(rc);
2674 }
2675
2676 /* This is to refresh our lock in face of no RPCs. */
2677 void osc_send_empty_rpc(struct osc_object *osc, pgoff_t start)
2678 {
2679         struct ptlrpc_request *req;
2680         struct obdo oa;
2681         struct brw_page bpg = { .off = start, .count = 1};
2682         struct brw_page *pga = &bpg;
2683         int rc;
2684
2685         memset(&oa, 0, sizeof(oa));
2686         oa.o_oi = osc->oo_oinfo->loi_oi;
2687         oa.o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLFLAGS;
2688         /* For updated servers - don't do a read */
2689         oa.o_flags = OBD_FL_NORPC;
2690
2691         rc = osc_brw_prep_request(OBD_BRW_READ, osc_cli(osc), &oa, 1, &pga,
2692                                   &req, 0);
2693
2694         /* If we succeeded we ship it off, if not there's no point in doing
2695          * anything. Also no resends.
2696          * No interpret callback, no commit callback.
2697          */
2698         if (!rc) {
2699                 req->rq_no_resend = 1;
2700                 ptlrpcd_add_req(req);
2701         }
2702 }
2703
2704 static int osc_set_lock_data(struct ldlm_lock *lock, void *data)
2705 {
2706         int set = 0;
2707
2708         LASSERT(lock != NULL);
2709
2710         lock_res_and_lock(lock);
2711
2712         if (lock->l_ast_data == NULL)
2713                 lock->l_ast_data = data;
2714         if (lock->l_ast_data == data)
2715                 set = 1;
2716
2717         unlock_res_and_lock(lock);
2718
2719         return set;
2720 }
2721
2722 int osc_enqueue_fini(struct ptlrpc_request *req, osc_enqueue_upcall_f upcall,
2723                      void *cookie, struct lustre_handle *lockh,
2724                      enum ldlm_mode mode, __u64 *flags, bool speculative,
2725                      int errcode)
2726 {
2727         bool intent = *flags & LDLM_FL_HAS_INTENT;
2728         int rc;
2729         ENTRY;
2730
2731         /* The request was created before ldlm_cli_enqueue call. */
2732         if (intent && errcode == ELDLM_LOCK_ABORTED) {
2733                 struct ldlm_reply *rep;
2734
2735                 rep = req_capsule_server_get(&req->rq_pill, &RMF_DLM_REP);
2736                 LASSERT(rep != NULL);
2737
2738                 rep->lock_policy_res1 =
2739                         ptlrpc_status_ntoh(rep->lock_policy_res1);
2740                 if (rep->lock_policy_res1)
2741                         errcode = rep->lock_policy_res1;
2742                 if (!speculative)
2743                         *flags |= LDLM_FL_LVB_READY;
2744         } else if (errcode == ELDLM_OK) {
2745                 *flags |= LDLM_FL_LVB_READY;
2746         }
2747
2748         /* Call the update callback. */
2749         rc = (*upcall)(cookie, lockh, errcode);
2750
2751         /* release the reference taken in ldlm_cli_enqueue() */
2752         if (errcode == ELDLM_LOCK_MATCHED)
2753                 errcode = ELDLM_OK;
2754         if (errcode == ELDLM_OK && lustre_handle_is_used(lockh))
2755                 ldlm_lock_decref(lockh, mode);
2756
2757         RETURN(rc);
2758 }
2759
2760 int osc_enqueue_interpret(const struct lu_env *env, struct ptlrpc_request *req,
2761                           void *args, int rc)
2762 {
2763         struct osc_enqueue_args *aa = args;
2764         struct ldlm_lock *lock;
2765         struct lustre_handle *lockh = &aa->oa_lockh;
2766         enum ldlm_mode mode = aa->oa_mode;
2767         struct ost_lvb *lvb = aa->oa_lvb;
2768         __u32 lvb_len = sizeof(*lvb);
2769         __u64 flags = 0;
2770         struct ldlm_enqueue_info einfo = {
2771                 .ei_type = aa->oa_type,
2772                 .ei_mode = mode,
2773         };
2774
2775         ENTRY;
2776
2777         /* ldlm_cli_enqueue is holding a reference on the lock, so it must
2778          * be valid. */
2779         lock = ldlm_handle2lock(lockh);
2780         LASSERTF(lock != NULL,
2781                  "lockh %#llx, req %p, aa %p - client evicted?\n",
2782                  lockh->cookie, req, aa);
2783
2784         /* Take an additional reference so that a blocking AST that
2785          * ldlm_cli_enqueue_fini() might post for a failed lock, is guaranteed
2786          * to arrive after an upcall has been executed by
2787          * osc_enqueue_fini(). */
2788         ldlm_lock_addref(lockh, mode);
2789
2790         /* Let cl_lock_state_wait fail with -ERESTARTSYS to unuse sublocks. */
2791         OBD_FAIL_TIMEOUT(OBD_FAIL_LDLM_ENQUEUE_HANG, 2);
2792
2793         /* Let CP AST to grant the lock first. */
2794         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_ENQ_RACE, 1);
2795
2796         if (aa->oa_speculative) {
2797                 LASSERT(aa->oa_lvb == NULL);
2798                 LASSERT(aa->oa_flags == NULL);
2799                 aa->oa_flags = &flags;
2800         }
2801
2802         /* Complete obtaining the lock procedure. */
2803         rc = ldlm_cli_enqueue_fini(aa->oa_exp, req, &einfo, 1, aa->oa_flags,
2804                                    lvb, lvb_len, lockh, rc);
2805         /* Complete osc stuff. */
2806         rc = osc_enqueue_fini(req, aa->oa_upcall, aa->oa_cookie, lockh, mode,
2807                               aa->oa_flags, aa->oa_speculative, rc);
2808
2809         OBD_FAIL_TIMEOUT(OBD_FAIL_OSC_CP_CANCEL_RACE, 10);
2810
2811         ldlm_lock_decref(lockh, mode);
2812         LDLM_LOCK_PUT(lock);
2813         RETURN(rc);
2814 }
2815
2816 /* When enqueuing asynchronously, locks are not ordered, we can obtain a lock
2817  * from the 2nd OSC before a lock from the 1st one. This does not deadlock with
2818  * other synchronous requests, however keeping some locks and trying to obtain
2819  * others may take a considerable amount of time in a case of ost failure; and
2820  * when other sync requests do not get released lock from a client, the client
2821  * is evicted from the cluster -- such scenarious make the life difficult, so
2822  * release locks just after they are obtained. */
2823 int osc_enqueue_base(struct obd_export *exp, struct ldlm_res_id *res_id,
2824                      __u64 *flags, union ldlm_policy_data *policy,
2825                      struct ost_lvb *lvb, osc_enqueue_upcall_f upcall,
2826                      void *cookie, struct ldlm_enqueue_info *einfo,
2827                      struct ptlrpc_request_set *rqset, int async,
2828                      bool speculative)
2829 {
2830         struct obd_device *obd = exp->exp_obd;
2831         struct lustre_handle lockh = { 0 };
2832         struct ptlrpc_request *req = NULL;
2833         int intent = *flags & LDLM_FL_HAS_INTENT;
2834         __u64 match_flags = *flags;
2835         enum ldlm_mode mode;
2836         int rc;
2837         ENTRY;
2838
2839         /* Filesystem lock extents are extended to page boundaries so that
2840          * dealing with the page cache is a little smoother.  */
2841         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2842         policy->l_extent.end |= ~PAGE_MASK;
2843
2844         /* Next, search for already existing extent locks that will cover us */
2845         /* If we're trying to read, we also search for an existing PW lock.  The
2846          * VFS and page cache already protect us locally, so lots of readers/
2847          * writers can share a single PW lock.
2848          *
2849          * There are problems with conversion deadlocks, so instead of
2850          * converting a read lock to a write lock, we'll just enqueue a new
2851          * one.
2852          *
2853          * At some point we should cancel the read lock instead of making them
2854          * send us a blocking callback, but there are problems with canceling
2855          * locks out from other users right now, too. */
2856         mode = einfo->ei_mode;
2857         if (einfo->ei_mode == LCK_PR)
2858                 mode |= LCK_PW;
2859         /* Normal lock requests must wait for the LVB to be ready before
2860          * matching a lock; speculative lock requests do not need to,
2861          * because they will not actually use the lock. */
2862         if (!speculative)
2863                 match_flags |= LDLM_FL_LVB_READY;
2864         if (intent != 0)
2865                 match_flags |= LDLM_FL_BLOCK_GRANTED;
2866         mode = ldlm_lock_match(obd->obd_namespace, match_flags, res_id,
2867                                einfo->ei_type, policy, mode, &lockh);
2868         if (mode) {
2869                 struct ldlm_lock *matched;
2870
2871                 if (*flags & LDLM_FL_TEST_LOCK)
2872                         RETURN(ELDLM_OK);
2873
2874                 matched = ldlm_handle2lock(&lockh);
2875                 if (speculative) {
2876                         /* This DLM lock request is speculative, and does not
2877                          * have an associated IO request. Therefore if there
2878                          * is already a DLM lock, it wll just inform the
2879                          * caller to cancel the request for this stripe.*/
2880                         lock_res_and_lock(matched);
2881                         if (ldlm_extent_equal(&policy->l_extent,
2882                             &matched->l_policy_data.l_extent))
2883                                 rc = -EEXIST;
2884                         else
2885                                 rc = -ECANCELED;
2886                         unlock_res_and_lock(matched);
2887
2888                         ldlm_lock_decref(&lockh, mode);
2889                         LDLM_LOCK_PUT(matched);
2890                         RETURN(rc);
2891                 } else if (osc_set_lock_data(matched, einfo->ei_cbdata)) {
2892                         *flags |= LDLM_FL_LVB_READY;
2893
2894                         /* We already have a lock, and it's referenced. */
2895                         (*upcall)(cookie, &lockh, ELDLM_LOCK_MATCHED);
2896
2897                         ldlm_lock_decref(&lockh, mode);
2898                         LDLM_LOCK_PUT(matched);
2899                         RETURN(ELDLM_OK);
2900                 } else {
2901                         ldlm_lock_decref(&lockh, mode);
2902                         LDLM_LOCK_PUT(matched);
2903                 }
2904         }
2905
2906         if (*flags & (LDLM_FL_TEST_LOCK | LDLM_FL_MATCH_LOCK))
2907                 RETURN(-ENOLCK);
2908
2909         /* users of osc_enqueue() can pass this flag for ldlm_lock_match() */
2910         *flags &= ~LDLM_FL_BLOCK_GRANTED;
2911
2912         rc = ldlm_cli_enqueue(exp, &req, einfo, res_id, policy, flags, lvb,
2913                               sizeof(*lvb), LVB_T_OST, &lockh, async);
2914         if (async) {
2915                 if (!rc) {
2916                         struct osc_enqueue_args *aa;
2917                         aa = ptlrpc_req_async_args(aa, req);
2918                         aa->oa_exp         = exp;
2919                         aa->oa_mode        = einfo->ei_mode;
2920                         aa->oa_type        = einfo->ei_type;
2921                         lustre_handle_copy(&aa->oa_lockh, &lockh);
2922                         aa->oa_upcall      = upcall;
2923                         aa->oa_cookie      = cookie;
2924                         aa->oa_speculative = speculative;
2925                         if (!speculative) {
2926                                 aa->oa_flags  = flags;
2927                                 aa->oa_lvb    = lvb;
2928                         } else {
2929                                 /* speculative locks are essentially to enqueue
2930                                  * a DLM lock  in advance, so we don't care
2931                                  * about the result of the enqueue. */
2932                                 aa->oa_lvb    = NULL;
2933                                 aa->oa_flags  = NULL;
2934                         }
2935
2936                         req->rq_interpret_reply = osc_enqueue_interpret;
2937                         ptlrpc_set_add_req(rqset, req);
2938                 }
2939                 RETURN(rc);
2940         }
2941
2942         rc = osc_enqueue_fini(req, upcall, cookie, &lockh, einfo->ei_mode,
2943                               flags, speculative, rc);
2944
2945         RETURN(rc);
2946 }
2947
2948 int osc_match_base(const struct lu_env *env, struct obd_export *exp,
2949                    struct ldlm_res_id *res_id, enum ldlm_type type,
2950                    union ldlm_policy_data *policy, enum ldlm_mode mode,
2951                    __u64 *flags, struct osc_object *obj,
2952                    struct lustre_handle *lockh, enum ldlm_match_flags match_flags)
2953 {
2954         struct obd_device *obd = exp->exp_obd;
2955         __u64 lflags = *flags;
2956         enum ldlm_mode rc;
2957         ENTRY;
2958
2959         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_MATCH))
2960                 RETURN(-EIO);
2961
2962         /* Filesystem lock extents are extended to page boundaries so that
2963          * dealing with the page cache is a little smoother */
2964         policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2965         policy->l_extent.end |= ~PAGE_MASK;
2966
2967         /* Next, search for already existing extent locks that will cover us */
2968         rc = ldlm_lock_match_with_skip(obd->obd_namespace, lflags, 0,
2969                                         res_id, type, policy, mode, lockh,
2970                                         match_flags);
2971         if (rc == 0 || lflags & LDLM_FL_TEST_LOCK)
2972                 RETURN(rc);
2973
2974         if (obj != NULL) {
2975                 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2976
2977                 LASSERT(lock != NULL);
2978                 if (osc_set_lock_data(lock, obj)) {
2979                         lock_res_and_lock(lock);
2980                         if (!ldlm_is_lvb_cached(lock)) {
2981                                 LASSERT(lock->l_ast_data == obj);
2982                                 osc_lock_lvb_update(env, obj, lock, NULL);
2983                                 ldlm_set_lvb_cached(lock);
2984                         }
2985                         unlock_res_and_lock(lock);
2986                 } else {
2987                         ldlm_lock_decref(lockh, rc);
2988                         rc = 0;
2989                 }
2990                 LDLM_LOCK_PUT(lock);
2991         }
2992         RETURN(rc);
2993 }
2994
2995 static int osc_statfs_interpret(const struct lu_env *env,
2996                                 struct ptlrpc_request *req, void *args, int rc)
2997 {
2998         struct osc_async_args *aa = args;
2999         struct obd_statfs *msfs;
3000
3001         ENTRY;
3002         if (rc == -EBADR)
3003                 /*
3004                  * The request has in fact never been sent due to issues at
3005                  * a higher level (LOV).  Exit immediately since the caller
3006                  * is aware of the problem and takes care of the clean up.
3007                  */
3008                 RETURN(rc);
3009
3010         if ((rc == -ENOTCONN || rc == -EAGAIN) &&
3011             (aa->aa_oi->oi_flags & OBD_STATFS_NODELAY))
3012                 GOTO(out, rc = 0);
3013
3014         if (rc != 0)
3015                 GOTO(out, rc);
3016
3017         msfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
3018         if (msfs == NULL)
3019                 GOTO(out, rc = -EPROTO);
3020
3021         *aa->aa_oi->oi_osfs = *msfs;
3022 out:
3023         rc = aa->aa_oi->oi_cb_up(aa->aa_oi, rc);
3024
3025         RETURN(rc);
3026 }
3027
3028 static int osc_statfs_async(struct obd_export *exp,
3029                             struct obd_info *oinfo, time64_t max_age,
3030                             struct ptlrpc_request_set *rqset)
3031 {
3032         struct obd_device     *obd = class_exp2obd(exp);
3033         struct ptlrpc_request *req;
3034         struct osc_async_args *aa;
3035         int rc;
3036         ENTRY;
3037
3038         if (obd->obd_osfs_age >= max_age) {
3039                 CDEBUG(D_SUPER,
3040                        "%s: use %p cache blocks %llu/%llu objects %llu/%llu\n",
3041                        obd->obd_name, &obd->obd_osfs,
3042                        obd->obd_osfs.os_bavail, obd->obd_osfs.os_blocks,
3043                        obd->obd_osfs.os_ffree, obd->obd_osfs.os_files);
3044                 spin_lock(&obd->obd_osfs_lock);
3045                 memcpy(oinfo->oi_osfs, &obd->obd_osfs, sizeof(*oinfo->oi_osfs));
3046                 spin_unlock(&obd->obd_osfs_lock);
3047                 oinfo->oi_flags |= OBD_STATFS_FROM_CACHE;
3048                 if (oinfo->oi_cb_up)
3049                         oinfo->oi_cb_up(oinfo, 0);
3050
3051                 RETURN(0);
3052         }
3053
3054         /* We could possibly pass max_age in the request (as an absolute
3055          * timestamp or a "seconds.usec ago") so the target can avoid doing
3056          * extra calls into the filesystem if that isn't necessary (e.g.
3057          * during mount that would help a bit).  Having relative timestamps
3058          * is not so great if request processing is slow, while absolute
3059          * timestamps are not ideal because they need time synchronization. */
3060         req = ptlrpc_request_alloc(obd->u.cli.cl_import, &RQF_OST_STATFS);
3061         if (req == NULL)
3062                 RETURN(-ENOMEM);
3063
3064         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_STATFS);
3065         if (rc) {
3066                 ptlrpc_request_free(req);
3067                 RETURN(rc);
3068         }
3069         ptlrpc_request_set_replen(req);
3070         req->rq_request_portal = OST_CREATE_PORTAL;
3071         ptlrpc_at_set_req_timeout(req);
3072
3073         if (oinfo->oi_flags & OBD_STATFS_NODELAY) {
3074                 /* procfs requests not want stat in wait for avoid deadlock */
3075                 req->rq_no_resend = 1;
3076                 req->rq_no_delay = 1;
3077         }
3078
3079         req->rq_interpret_reply = osc_statfs_interpret;
3080         aa = ptlrpc_req_async_args(aa, req);
3081         aa->aa_oi = oinfo;
3082
3083         ptlrpc_set_add_req(rqset, req);
3084         RETURN(0);
3085 }
3086
3087 static int osc_statfs(const struct lu_env *env, struct obd_export *exp,
3088                       struct obd_statfs *osfs, time64_t max_age, __u32 flags)
3089 {
3090         struct obd_device     *obd = class_exp2obd(exp);
3091         struct obd_statfs     *msfs;
3092         struct ptlrpc_request *req;
3093         struct obd_import     *imp, *imp0;
3094         int rc;
3095         ENTRY;
3096
3097         /*Since the request might also come from lprocfs, so we need
3098          *sync this with client_disconnect_export Bug15684
3099          */
3100         with_imp_locked(obd, imp0, rc)
3101                 imp = class_import_get(imp0);
3102         if (rc)
3103                 RETURN(rc);
3104
3105         /* We could possibly pass max_age in the request (as an absolute
3106          * timestamp or a "seconds.usec ago") so the target can avoid doing
3107          * extra calls into the filesystem if that isn't necessary (e.g.
3108          * during mount that would help a bit).  Having relative timestamps