1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (C) 2001-2003 Cluster File Systems, Inc.
5 * Author Peter Braam <braam@clusterfs.com>
7 * This file is part of Lustre, http://www.lustre.org.
9 * Lustre is free software; you can redistribute it and/or
10 * modify it under the terms of version 2 of the GNU General Public
11 * License as published by the Free Software Foundation.
13 * Lustre is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with Lustre; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
22 * For testing and management it is treated as an obd_device,
23 * although * it does not export a full OBD method table (the
24 * requests are coming * in over the wire, so object target modules
25 * do not have a full * method table.)
30 # define EXPORT_SYMTAB
32 #define DEBUG_SUBSYSTEM S_OSC
35 # include <linux/version.h>
36 # include <linux/module.h>
37 # include <linux/mm.h>
38 # include <linux/highmem.h>
39 # include <linux/ctype.h>
40 # include <linux/init.h>
41 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
42 # include <linux/workqueue.h>
43 # include <linux/smp_lock.h>
45 # include <linux/locks.h>
47 #else /* __KERNEL__ */
48 # include <liblustre.h>
51 #include <linux/lustre_dlm.h>
52 #include <libcfs/kp30.h>
53 #include <linux/lustre_net.h>
54 #include <linux/lustre_sec.h>
55 #include <lustre/lustre_user.h>
56 #include <linux/obd_ost.h>
57 #include <linux/obd_lov.h>
63 #include <linux/lustre_ha.h>
64 #include <linux/lprocfs_status.h>
65 #include <linux/lustre_log.h>
66 #include <linux/lustre_lite.h>
67 #include <linux/lustre_audit.h>
68 #include <linux/lustre_gs.h>
70 #include "osc_internal.h"
72 /* Pack OSC object metadata for disk storage (LE byte order). */
73 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
74 struct lov_stripe_md *lsm)
79 lmm_size = sizeof(**lmmp);
84 OBD_FREE(*lmmp, lmm_size);
90 OBD_ALLOC(*lmmp, lmm_size);
96 LASSERT(lsm->lsm_object_id);
97 LASSERT(lsm->lsm_object_gr);
98 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
99 (*lmmp)->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
105 /* Unpack OSC object metadata from disk storage (LE byte order). */
106 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
107 struct lov_mds_md *lmm, int lmm_bytes)
113 if (lmm_bytes < sizeof (*lmm)) {
114 CERROR("lov_mds_md too small: %d, need %d\n",
115 lmm_bytes, (int)sizeof(*lmm));
118 /* XXX LOV_MAGIC etc check? */
120 if (lmm->lmm_object_id == 0) {
121 CERROR("lov_mds_md: zero lmm_object_id\n");
126 lsm_size = lov_stripe_md_size(1);
130 if (*lsmp != NULL && lmm == NULL) {
131 OBD_FREE(*lsmp, lsm_size);
137 OBD_ALLOC(*lsmp, lsm_size);
140 loi_init((*lsmp)->lsm_oinfo);
144 /* XXX zero *lsmp? */
145 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
146 (*lsmp)->lsm_object_gr = le64_to_cpu (lmm->lmm_object_gr);
147 LASSERT((*lsmp)->lsm_object_id);
148 LASSERT((*lsmp)->lsm_object_gr);
151 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
156 static int osc_getattr_interpret(struct ptlrpc_request *req,
157 struct osc_getattr_async_args *aa, int rc)
159 struct ost_body *body;
165 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
167 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
168 memcpy(aa->aa_oa, &body->oa, sizeof(*aa->aa_oa));
170 /* This should really be sent by the OST */
171 aa->aa_oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
172 aa->aa_oa->o_valid |= OBD_MD_FLBLKSZ;
174 CERROR("can't unpack ost_body\n");
176 aa->aa_oa->o_valid = 0;
182 static int osc_getattr_async(struct obd_export *exp, struct obdo *oa,
183 struct lov_stripe_md *md,
184 struct ptlrpc_request_set *set)
186 struct ptlrpc_request *request;
187 struct ost_body *body;
188 int size = sizeof(*body);
189 struct osc_getattr_async_args *aa;
192 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
193 OST_GETATTR, 1, &size, NULL);
197 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
198 memcpy(&body->oa, oa, sizeof(*oa));
200 request->rq_replen = lustre_msg_size(1, &size);
201 request->rq_interpret_reply = osc_getattr_interpret;
203 LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
204 aa = (struct osc_getattr_async_args *)&request->rq_async_args;
207 ptlrpc_set_add_req (set, request);
211 static int osc_getattr(struct obd_export *exp, struct obdo *oa,
212 struct lov_stripe_md *md)
214 struct ptlrpc_request *request;
215 struct ost_body *body;
216 int rc, size = sizeof(*body);
219 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
220 OST_GETATTR, 1, &size, NULL);
224 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
225 memcpy(&body->oa, oa, sizeof(*oa));
227 request->rq_replen = lustre_msg_size(1, &size);
229 rc = ptlrpc_queue_wait(request);
231 CERROR("%s failed: rc = %d\n", __FUNCTION__, rc);
235 body = lustre_swab_repbuf(request, 0, sizeof (*body),
236 lustre_swab_ost_body);
238 CERROR ("can't unpack ost_body\n");
239 GOTO (out, rc = -EPROTO);
242 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
243 memcpy(oa, &body->oa, sizeof(*oa));
245 /* This should really be sent by the OST */
246 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
247 oa->o_valid |= OBD_MD_FLBLKSZ;
251 ptlrpc_req_finished(request);
255 static int osc_setattr(struct obd_export *exp, struct obdo *oa,
256 struct lov_stripe_md *md, struct obd_trans_info *oti,
257 struct lustre_capa *capa)
259 struct ptlrpc_request *request;
260 struct ost_body *body;
261 int rc, size = sizeof(*body);
264 LASSERT(!(oa->o_valid & OBD_MD_FLGROUP) || oa->o_gr > 0);
266 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
267 OST_SETATTR, 1, &size, NULL);
271 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
272 memcpy(&body->oa, oa, sizeof(*oa));
274 request->rq_replen = lustre_msg_size(1, &size);
276 if (oti != NULL && (oti->oti_flags & OBD_MODE_ASYNC)) {
277 ptlrpcd_add_req(request);
280 rc = ptlrpc_queue_wait(request);
284 body = lustre_swab_repbuf(request, 0, sizeof(*body),
285 lustre_swab_ost_body);
287 GOTO(out, rc = -EPROTO);
289 memcpy(oa, &body->oa, sizeof(*oa));
293 ptlrpc_req_finished(request);
297 int osc_real_create(struct obd_export *exp, struct obdo *oa,
298 struct lov_stripe_md **ea, struct obd_trans_info *oti)
300 struct osc_creator *oscc = &exp->exp_obd->u.cli.cl_oscc;
301 struct ptlrpc_request *request;
302 struct ost_body *body;
303 struct lov_stripe_md *lsm;
304 int rc, size = sizeof(*body);
312 rc = obd_alloc_memmd(exp, &lsm);
317 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
318 OST_CREATE, 1, &size, NULL);
320 GOTO(out, rc = -ENOMEM);
322 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
323 memcpy(&body->oa, oa, sizeof(body->oa));
325 request->rq_replen = lustre_msg_size(1, &size);
326 if (oa->o_valid & OBD_MD_FLINLINE) {
327 LASSERT((oa->o_valid & OBD_MD_FLFLAGS) &&
328 oa->o_flags == OBD_FL_DELORPHAN);
329 DEBUG_REQ(D_HA, request,
330 "delorphan from OST integration");
331 /* Don't resend the delorphan request */
332 request->rq_no_resend = request->rq_no_delay = 1;
335 rc = ptlrpc_queue_wait(request);
339 body = lustre_swab_repbuf(request, 0, sizeof(*body),
340 lustre_swab_ost_body);
342 CERROR ("can't unpack ost_body\n");
343 GOTO (out_req, rc = -EPROTO);
346 if ((oa->o_valid & OBD_MD_FLFLAGS) && oa->o_flags == OBD_FL_DELORPHAN) {
347 struct obd_import *imp = class_exp2cliimp(exp);
348 /* MDS declares last known object, OSS responses
349 * with next possible object -bzzz */
350 spin_lock(&oscc->oscc_lock);
351 oscc->oscc_next_id = body->oa.o_id;
352 spin_unlock(&oscc->oscc_lock);
353 CDEBUG(D_HA, "%s: set nextid "LPD64" after recovery\n",
354 imp->imp_target_uuid.uuid, oa->o_id);
356 memcpy(oa, &body->oa, sizeof(*oa));
358 /* This should really be sent by the OST */
359 oa->o_blksize = PTLRPC_MAX_BRW_SIZE;
360 oa->o_valid |= OBD_MD_FLBLKSZ;
362 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
363 * have valid lsm_oinfo data structs, so don't go touching that.
364 * This needs to be fixed in a big way.
366 lsm->lsm_object_id = oa->o_id;
367 lsm->lsm_object_gr = oa->o_gr;
371 oti->oti_transno = request->rq_repmsg->transno;
373 if (oa->o_valid & OBD_MD_FLCOOKIE) {
374 if (!oti->oti_logcookies)
375 oti_alloc_cookies(oti, 1);
376 memcpy(oti->oti_logcookies, obdo_logcookie(oa),
377 sizeof(oti->oti_onecookie));
381 CDEBUG(D_HA, "transno: "LPD64"\n", request->rq_repmsg->transno);
384 ptlrpc_req_finished(request);
387 obd_free_memmd(exp, &lsm);
391 static int osc_punch(struct obd_export *exp, struct obdo *oa,
392 struct lov_stripe_md *md, obd_size start,
393 obd_size end, struct obd_trans_info *oti,
394 struct lustre_capa *capa)
396 struct ptlrpc_request *request;
397 struct ost_body *body;
398 struct lustre_capa *req_capa;
400 int rc, size[2] = { sizeof(*body), sizeof(*capa) };
408 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
409 OST_PUNCH, capa ? 2 : 1, size, NULL);
413 body = lustre_msg_buf(request->rq_reqmsg, bufcnt++, sizeof (*body));
415 memcpy(&body->oa, oa, sizeof(*oa));
417 /* overload the size and blocks fields in the oa with start/end */
418 body->oa.o_size = start;
419 body->oa.o_blocks = end;
420 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
423 req_capa = lustre_msg_buf(request->rq_reqmsg, bufcnt++,
425 capa_dup2(req_capa, capa);
426 body->oa.o_valid |= OBD_MD_CAPA;
429 request->rq_replen = lustre_msg_size(1, size);
431 rc = ptlrpc_queue_wait(request);
435 body = lustre_swab_repbuf (request, 0, sizeof (*body),
436 lustre_swab_ost_body);
438 CERROR ("can't unpack ost_body\n");
439 GOTO (out, rc = -EPROTO);
442 memcpy(oa, &body->oa, sizeof(*oa));
446 ptlrpc_req_finished(request);
450 static int osc_sync(struct obd_export *exp, struct obdo *oa,
451 struct lov_stripe_md *md, obd_size start,
454 struct ptlrpc_request *request;
455 struct ost_body *body;
456 int rc, size = sizeof(*body);
464 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
465 OST_SYNC, 1, &size, NULL);
469 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
470 memcpy(&body->oa, oa, sizeof(*oa));
472 /* overload the size and blocks fields in the oa with start/end */
473 body->oa.o_size = start;
474 body->oa.o_blocks = end;
475 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
477 request->rq_replen = lustre_msg_size(1, &size);
479 rc = ptlrpc_queue_wait(request);
483 body = lustre_swab_repbuf(request, 0, sizeof(*body),
484 lustre_swab_ost_body);
486 CERROR ("can't unpack ost_body\n");
487 GOTO (out, rc = -EPROTO);
490 memcpy(oa, &body->oa, sizeof(*oa));
494 ptlrpc_req_finished(request);
498 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
499 struct lov_stripe_md *ea, struct obd_trans_info *oti)
501 struct ptlrpc_request *request;
502 struct ost_body *body;
503 int rc, size = sizeof(*body);
511 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
512 OST_DESTROY, 1, &size, NULL);
515 request->rq_request_portal = OST_DESTROY_PORTAL;
517 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
519 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) {
520 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
521 sizeof(*oti->oti_logcookies));
522 oti->oti_logcookies++;
525 memcpy(&body->oa, oa, sizeof(*oa));
526 request->rq_replen = lustre_msg_size(1, &size);
528 if (oti != NULL && (oti->oti_flags & OBD_MODE_ASYNC)) {
529 ptlrpcd_add_req(request);
532 rc = ptlrpc_queue_wait(request);
538 ptlrpc_req_finished(request);
542 body = lustre_swab_repbuf(request, 0, sizeof(*body),
543 lustre_swab_ost_body);
545 CERROR ("Can't unpack body\n");
546 ptlrpc_req_finished(request);
550 memcpy(oa, &body->oa, sizeof(*oa));
551 ptlrpc_req_finished(request);
556 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
559 obd_valid bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
561 LASSERT(!(oa->o_valid & bits));
564 spin_lock(&cli->cl_loi_list_lock);
565 oa->o_dirty = cli->cl_dirty;
566 oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
567 oa->o_grant = cli->cl_avail_grant;
568 oa->o_dropped = cli->cl_lost_grant;
569 cli->cl_lost_grant = 0;
570 spin_unlock(&cli->cl_loi_list_lock);
571 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
572 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
575 /* caller must hold loi_list_lock */
576 static void osc_consume_write_grant(struct client_obd *cli,
577 struct osc_async_page *oap)
579 cli->cl_dirty += PAGE_SIZE;
580 cli->cl_avail_grant -= PAGE_SIZE;
581 oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
582 CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
583 LASSERT(cli->cl_avail_grant >= 0);
586 static unsigned long rpcs_in_flight(struct client_obd *cli)
588 return cli->cl_r_in_flight + cli->cl_w_in_flight;
591 /* caller must hold loi_list_lock */
592 void osc_wake_cache_waiters(struct client_obd *cli)
594 struct list_head *l, *tmp;
595 struct osc_cache_waiter *ocw;
597 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
598 /* if we can't dirty more, we must wait until some is written */
599 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
600 CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
601 cli->cl_dirty, cli->cl_dirty_max);
605 /* if still dirty cache but no grant wait for pending RPCs that
606 * may yet return us some grant before doing sync writes */
607 if (cli->cl_w_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
608 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
609 cli->cl_w_in_flight);
611 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
612 list_del_init(&ocw->ocw_entry);
613 if (cli->cl_avail_grant < PAGE_SIZE) {
614 /* no more RPCs in flight to return grant, do sync IO */
615 ocw->ocw_rc = -EDQUOT;
616 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
618 osc_consume_write_grant(cli, ocw->ocw_oap);
621 wake_up(&ocw->ocw_waitq);
627 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
629 spin_lock(&cli->cl_loi_list_lock);
630 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
631 cli->cl_avail_grant += body->oa.o_grant;
632 /* waiters are woken in brw_interpret_oap */
633 spin_unlock(&cli->cl_loi_list_lock);
636 /* We assume that the reason this OSC got a short read is because it read
637 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
638 * via the LOV, and it _knows_ it's reading inside the file, it's just that
639 * this stripe never got written at or beyond this stripe offset yet. */
640 static void handle_short_read(int nob_read, obd_count page_count,
641 struct brw_page *pga)
645 /* skip bytes read OK */
646 while (nob_read > 0) {
647 LASSERT (page_count > 0);
649 if (pga->count > nob_read) {
650 /* EOF inside this page */
651 ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
652 memset(ptr + nob_read, 0, pga->count - nob_read);
659 nob_read -= pga->count;
664 /* zero remaining pages */
665 while (page_count-- > 0) {
666 ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
667 memset(ptr, 0, pga->count);
673 static int check_write_rcs(struct ptlrpc_request *request,
674 int requested_nob, int niocount,
675 obd_count page_count, struct brw_page *pga)
679 /* return error if any niobuf was in error */
680 remote_rcs = lustre_swab_repbuf(request, 1,
681 sizeof(*remote_rcs) * niocount, NULL);
682 if (remote_rcs == NULL) {
683 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
686 if (lustre_msg_swabbed(request->rq_repmsg))
687 for (i = 0; i < niocount; i++)
688 __swab32s((__u32 *)&remote_rcs[i]);
690 for (i = 0; i < niocount; i++) {
691 if (remote_rcs[i] < 0)
692 return(remote_rcs[i]);
694 if (remote_rcs[i] != 0) {
695 CERROR("rc[%d] invalid (%d) req %p\n",
696 i, remote_rcs[i], request);
701 if (request->rq_bulk->bd_nob_transferred != requested_nob) {
702 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
703 requested_nob, request->rq_bulk->bd_nob_transferred);
710 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
712 if (p1->flag != p2->flag) {
713 unsigned mask = ~OBD_BRW_FROM_GRANT;
715 /* warn if we try to combine flags that we don't know to be
717 if ((p1->flag & mask) != (p2->flag & mask))
718 CERROR("is it ok to have flags 0x%x and 0x%x in the "
719 "same brw?\n", p1->flag, p2->flag);
723 return (p1->disk_offset + p1->count == p2->disk_offset);
727 static obd_count cksum_pages(int nob, obd_count page_count,
728 struct brw_page *pga)
734 LASSERT (page_count > 0);
737 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
738 pga->count > nob ? nob : pga->count);
750 #define osc_encrypt_page(page, off, count) \
751 osc_crypt_page(page, off, count, ENCRYPT_DATA)
752 #define osc_decrypt_page(page, off, count) \
753 osc_crypt_page(page, off, count, DECRYPT_DATA)
754 /*Put a global call back var here is Ugly, but put it to client_obd
755 *also seems not a good idea, WangDi*/
756 crypt_cb_t osc_crypt_cb = NULL;
758 static int osc_crypt_page(struct page *page, obd_off page_off, obd_off count,
764 if (osc_crypt_cb != NULL)
765 rc = osc_crypt_cb(page, page_off, count, flags);
767 CERROR("crypt page error %d \n", rc);
771 static int osc_decrypt_pages(struct brw_page *pga, int page_count)
776 for (i = 0; i < page_count; i++) {
777 struct brw_page *pg = &pga[i];
778 osc_decrypt_page(pg->pg, pg->page_offset, pg->count);
783 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
784 struct lov_stripe_md *lsm, obd_count page_count,
785 struct brw_page *pga, int *requested_nobp,
786 int *niocountp, struct ptlrpc_request **reqp)
788 struct ptlrpc_request *req;
789 struct ptlrpc_bulk_desc *desc;
790 struct client_obd *cli = &imp->imp_obd->u.cli;
791 struct ost_body *body;
792 struct lustre_id *raw_id = obdo_id(oa);
793 struct obd_capa *ocapa = NULL;
794 struct lustre_capa *capa = NULL;
795 struct obd_ioobj *ioobj;
796 struct niobuf_remote *niobuf;
805 opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
807 for (niocount = i = 1; i < page_count; i++)
808 if (!can_merge_pages(&pga[i - 1], &pga[i]))
811 capa_op = (opc == OST_WRITE) ? MAY_WRITE : MAY_READ;
813 ocapa = capa_get(oa->o_fsuid, capa_op, raw_id->li_fid.lf_group,
814 raw_id->li_stc.u.e3s.l3s_ino, CLIENT_CAPA,
817 if (opc == OST_READ && capa_op == MAY_READ) {
818 /* partial write might cause read, MAY_WRITE capability
819 * should be used here */
825 size[bufcnt++] = sizeof(*body);
826 size[bufcnt++] = sizeof(*ioobj);
828 size[bufcnt++] = sizeof(*capa);
829 size[bufcnt++] = niocount * sizeof(*niobuf);
831 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, opc, bufcnt, size, NULL);
835 if (opc == OST_WRITE)
836 desc = ptlrpc_prep_bulk_imp (req, page_count,
837 BULK_GET_SOURCE, OST_BULK_PORTAL);
839 desc = ptlrpc_prep_bulk_imp (req, page_count,
840 BULK_PUT_SINK, OST_BULK_PORTAL);
842 GOTO(out, rc = -ENOMEM);
843 /* NB request now owns desc and will free it when it gets freed */
846 body = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*body));
847 ioobj = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*ioobj));
849 capa = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*capa));
850 niobuf = lustre_msg_buf(req->rq_reqmsg, bufcnt++,
851 niocount * sizeof(*niobuf));
853 memcpy(&body->oa, oa, sizeof(*oa));
855 obdo_to_ioobj(oa, ioobj);
856 ioobj->ioo_bufcnt = niocount;
859 capa_dup(capa, ocapa);
860 body->oa.o_valid |= OBD_MD_CAPA;
861 capa_put(ocapa, CLIENT_CAPA);
864 LASSERT (page_count > 0);
866 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
867 struct brw_page *pg = &pga[i];
868 struct brw_page *pg_prev = pg - 1;
870 LASSERT(pg->count > 0);
871 LASSERTF((pg->page_offset & ~PAGE_MASK)+ pg->count <= PAGE_SIZE,
872 "i: %d pg: %p pg_off: "LPU64", count: %u\n", i, pg,
873 pg->page_offset, pg->count);
874 LASSERTF(i == 0 || pg->disk_offset > pg_prev->disk_offset,
875 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
876 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
878 pg->pg, pg->pg->private, pg->pg->index, pg->disk_offset,
879 pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
880 pg_prev->disk_offset);
882 if (opc == OST_WRITE) {
883 rc = osc_encrypt_page(pg->pg, pg->page_offset, pg->count);
888 ptlrpc_prep_bulk_page(desc, pg->pg,
889 pg->page_offset & ~PAGE_MASK, pg->count);
890 requested_nob += pg->count;
892 if (i > 0 && can_merge_pages(pg_prev, pg)) {
894 niobuf->len += pg->count;
896 niobuf->offset = pg->disk_offset;
897 niobuf->len = pg->count;
898 niobuf->flags = pg->flag;
902 LASSERT((void *)(niobuf - niocount) ==
903 lustre_msg_buf(req->rq_reqmsg, bufcnt - 1,
904 niocount * sizeof(*niobuf)));
905 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
907 /* size[0] still sizeof (*body) */
908 if (opc == OST_WRITE) {
910 body->oa.o_valid |= OBD_MD_FLCKSUM;
911 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
913 /* 1 RC per niobuf */
914 size[1] = sizeof(__u32) * niocount;
915 req->rq_replen = lustre_msg_size(2, size);
917 /* 1 RC for the whole I/O */
918 req->rq_replen = lustre_msg_size(1, size);
921 *niocountp = niocount;
922 *requested_nobp = requested_nob;
927 ptlrpc_req_finished (req);
931 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
932 int requested_nob, int niocount,
933 obd_count page_count, struct brw_page *pga,
936 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
937 struct ost_body *body;
943 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
945 CERROR ("Can't unpack body\n");
949 osc_update_grant(cli, body);
950 memcpy(oa, &body->oa, sizeof(*oa));
952 if (req->rq_reqmsg->opc == OST_WRITE) {
954 CERROR ("Unexpected +ve rc %d\n", rc);
957 LASSERT (req->rq_bulk->bd_nob == requested_nob);
958 osc_decrypt_pages(pga, page_count);
959 RETURN(check_write_rcs(req, requested_nob, niocount,
963 if (rc > requested_nob) {
964 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
968 if (rc != req->rq_bulk->bd_nob_transferred) {
969 CERROR ("Unexpected rc %d (%d transferred)\n",
970 rc, req->rq_bulk->bd_nob_transferred);
974 if (rc < requested_nob)
975 handle_short_read(rc, page_count, pga);
978 if (oa->o_valid & OBD_MD_FLCKSUM) {
979 const struct ptlrpc_peer *peer =
980 &req->rq_import->imp_connection->c_peer;
981 static int cksum_counter;
982 obd_count server_cksum = oa->o_cksum;
983 obd_count cksum = cksum_pages(rc, page_count, pga);
984 char str[PTL_NALFMT_SIZE];
986 ptlrpc_peernid2str(peer, str);
989 if (server_cksum != cksum) {
990 CERROR("Bad checksum: server %x, client %x, server NID "
991 LPX64" (%s)\n", server_cksum, cksum,
992 peer->peer_id.nid, str);
995 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
996 CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
997 cksum_counter, peer->peer_id.nid, str, cksum);
1000 static int cksum_missed;
1003 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1004 CERROR("Request checksum %u from "LPX64", no reply\n",
1006 req->rq_import->imp_connection->c_peer.peer_id.nid);
1009 osc_decrypt_pages(pga, page_count);
1013 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
1014 struct lov_stripe_md *lsm,
1015 obd_count page_count, struct brw_page *pga)
1019 struct ptlrpc_request *request;
1024 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
1025 page_count, pga, &requested_nob, &niocount,
1030 rc = ptlrpc_queue_wait(request);
1032 if (rc == -ETIMEDOUT && request->rq_resend) {
1033 DEBUG_REQ(D_HA, request, "BULK TIMEOUT");
1034 ptlrpc_req_finished(request);
1038 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
1039 page_count, pga, rc);
1041 ptlrpc_req_finished(request);
1045 static int brw_interpret(struct ptlrpc_request *request,
1046 struct osc_brw_async_args *aa, int rc)
1048 struct obdo *oa = aa->aa_oa;
1049 int requested_nob = aa->aa_requested_nob;
1050 int niocount = aa->aa_nio_count;
1051 obd_count page_count = aa->aa_page_count;
1052 struct brw_page *pga = aa->aa_pga;
1055 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
1056 page_count, pga, rc);
1060 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1061 struct lov_stripe_md *lsm, obd_count page_count,
1062 struct brw_page *pga, struct ptlrpc_request_set *set)
1064 struct ptlrpc_request *request;
1067 struct osc_brw_async_args *aa;
1071 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
1072 page_count, pga, &requested_nob, &nio_count,
1075 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1076 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1078 aa->aa_requested_nob = requested_nob;
1079 aa->aa_nio_count = nio_count;
1080 aa->aa_page_count = page_count;
1083 request->rq_interpret_reply = brw_interpret;
1084 ptlrpc_set_add_req(set, request);
1090 #define min_t(type,x,y) \
1091 ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
1095 * ugh, we want disk allocation on the target to happen in offset order. we'll
1096 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1097 * fine for our small page arrays and doesn't require allocation. its an
1098 * insertion sort that swaps elements that are strides apart, shrinking the
1099 * stride down until its '1' and the array is sorted.
1101 static void sort_brw_pages(struct brw_page *array, int num)
1104 struct brw_page tmp;
1108 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1113 for (i = stride ; i < num ; i++) {
1116 while (j >= stride && array[j - stride].disk_offset >
1118 array[j] = array[j - stride];
1123 } while (stride > 1);
1126 /* make sure we the regions we're passing to elan don't violate its '4
1127 * fragments' constraint. portal headers are a fragment, all full
1128 * PAGE_SIZE long pages count as 1 fragment, and each partial page
1129 * counts as a fragment. I think. see bug 934. */
1130 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1133 int saw_whole_frag = 0;
1136 for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1137 if (pg->count == PAGE_SIZE) {
1138 if (!saw_whole_frag) {
1149 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1150 struct lov_stripe_md *lsm, obd_count page_count,
1151 struct brw_page *pga, struct obd_trans_info *oti)
1155 if (cmd == OBD_BRW_CHECK) {
1156 /* The caller just wants to know if there's a chance that this
1157 * I/O can succeed */
1158 struct obd_import *imp = class_exp2cliimp(exp);
1160 if (imp == NULL || imp->imp_invalid)
1165 while (page_count) {
1166 obd_count pages_per_brw;
1169 if (page_count > PTLRPC_MAX_BRW_PAGES)
1170 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1172 pages_per_brw = page_count;
1174 sort_brw_pages(pga, pages_per_brw);
1175 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1177 rc = osc_brw_internal(cmd, exp, oa, lsm, pages_per_brw, pga);
1182 page_count -= pages_per_brw;
1183 pga += pages_per_brw;
1188 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1189 struct lov_stripe_md *lsm, obd_count page_count,
1190 struct brw_page *pga, struct ptlrpc_request_set *set,
1191 struct obd_trans_info *oti)
1195 if (cmd == OBD_BRW_CHECK) {
1196 /* The caller just wants to know if there's a chance that this
1197 * I/O can succeed */
1198 struct obd_import *imp = class_exp2cliimp(exp);
1200 if (imp == NULL || imp->imp_invalid)
1205 while (page_count) {
1206 obd_count pages_per_brw;
1209 if (page_count > PTLRPC_MAX_BRW_PAGES)
1210 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1212 pages_per_brw = page_count;
1214 sort_brw_pages(pga, pages_per_brw);
1215 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1217 rc = async_internal(cmd, exp, oa, lsm, pages_per_brw, pga, set);
1222 page_count -= pages_per_brw;
1223 pga += pages_per_brw;
1228 static void osc_check_rpcs(struct client_obd *cli);
1229 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1231 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1232 static void lop_update_pending(struct client_obd *cli,
1233 struct loi_oap_pages *lop, int cmd, int delta);
1235 /* this is called when a sync waiter receives an interruption. Its job is to
1236 * get the caller woken as soon as possible. If its page hasn't been put in an
1237 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1238 * desiring interruption which will forcefully complete the rpc once the rpc
1240 static void osc_occ_interrupted(struct oig_callback_context *occ)
1242 struct osc_async_page *oap;
1243 struct loi_oap_pages *lop;
1244 struct lov_oinfo *loi;
1247 /* XXX member_of() */
1248 oap = list_entry(occ, struct osc_async_page, oap_occ);
1250 spin_lock(&oap->oap_cli->cl_loi_list_lock);
1252 oap->oap_interrupted = 1;
1254 /* ok, it's been put in an rpc. */
1255 if (oap->oap_request != NULL) {
1256 ptlrpc_mark_interrupted(oap->oap_request);
1257 ptlrpcd_wake(oap->oap_request);
1261 /* we don't get interruption callbacks until osc_trigger_sync_io()
1262 * has been called and put the sync oaps in the pending/urgent lists.*/
1263 if (!list_empty(&oap->oap_pending_item)) {
1264 list_del_init(&oap->oap_pending_item);
1265 if (oap->oap_async_flags & ASYNC_URGENT)
1266 list_del_init(&oap->oap_urgent_item);
1269 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1270 &loi->loi_write_lop : &loi->loi_read_lop;
1271 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1272 loi_list_maint(oap->oap_cli, oap->oap_loi);
1274 oig_complete_one(oap->oap_oig, &oap->oap_occ, 0);
1275 oap->oap_oig = NULL;
1279 spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1282 /* this must be called holding the loi list lock to give coverage to exit_cache,
1283 * async_flag maintenance, and oap_request */
1284 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1285 struct osc_async_page *oap, int sent, int rc)
1287 osc_exit_cache(cli, oap, sent);
1288 oap->oap_async_flags = 0;
1289 oap->oap_interrupted = 0;
1291 if (oap->oap_request != NULL) {
1292 ptlrpc_req_finished(oap->oap_request);
1293 oap->oap_request = NULL;
1296 if (rc == 0 && oa != NULL)
1297 oap->oap_loi->loi_blocks = oa->o_blocks;
1300 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1301 oap->oap_oig = NULL;
1306 oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1310 static int brw_interpret_oap(struct ptlrpc_request *request,
1311 struct osc_brw_async_args *aa, int rc)
1313 struct osc_async_page *oap;
1314 struct client_obd *cli;
1315 struct list_head *pos, *n;
1319 do_gettimeofday(&now);
1320 rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1321 aa->aa_nio_count, aa->aa_page_count,
1324 CDEBUG(D_INODE, "request %p aa %p rc %d\n", request, aa, rc);
1328 /* in failout recovery we ignore writeback failure and want
1329 * to just tell llite to unlock the page and continue */
1330 if (request->rq_reqmsg->opc == OST_WRITE &&
1331 (cli->cl_import == NULL || cli->cl_import->imp_invalid)) {
1332 CDEBUG(D_INODE, "flipping to rc 0 imp %p inv %d\n",
1334 cli->cl_import ? cli->cl_import->imp_invalid : -1);
1338 spin_lock(&cli->cl_loi_list_lock);
1340 if (request->rq_reqmsg->opc == OST_WRITE)
1341 lprocfs_stime_record(&cli->cl_write_stime, &now,
1342 &request->rq_rpcd_start);
1344 lprocfs_stime_record(&cli->cl_read_stime, &now,
1345 &request->rq_rpcd_start);
1347 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1348 * is called so we know whether to go to sync BRWs or wait for more
1349 * RPCs to complete */
1350 if (request->rq_reqmsg->opc == OST_WRITE)
1351 cli->cl_w_in_flight--;
1353 cli->cl_r_in_flight--;
1355 /* the caller may re-use the oap after the completion call so
1356 * we need to clean it up a little */
1357 list_for_each_safe(pos, n, &aa->aa_oaps) {
1358 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1360 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1361 //oap->oap_page, oap->oap_page->index, oap);
1363 list_del_init(&oap->oap_rpc_item);
1364 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
1367 /* no write RPCs in flight, reset the time */
1368 if (request->rq_reqmsg->opc == OST_WRITE && cli->cl_w_in_flight == 0)
1369 do_gettimeofday(&cli->cl_last_write_time);
1371 osc_wake_cache_waiters(cli);
1372 osc_check_rpcs(cli);
1373 spin_unlock(&cli->cl_loi_list_lock);
1375 obdo_free(aa->aa_oa);
1376 OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1381 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1382 struct list_head *rpc_list,
1383 int page_count, int cmd)
1385 struct ptlrpc_request *req;
1386 struct brw_page *pga = NULL;
1387 int requested_nob, nio_count;
1388 struct osc_brw_async_args *aa;
1389 struct obdo *oa = NULL;
1390 struct obd_async_page_ops *ops = NULL;
1391 void *caller_data = NULL;
1392 struct list_head *pos;
1395 LASSERT(!list_empty(rpc_list));
1397 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1399 RETURN(ERR_PTR(-ENOMEM));
1403 GOTO(out, req = ERR_PTR(-ENOMEM));
1406 list_for_each(pos, rpc_list) {
1407 struct osc_async_page *oap;
1409 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1411 ops = oap->oap_caller_ops;
1412 caller_data = oap->oap_caller_data;
1414 pga[i].disk_offset = oap->oap_obj_off + oap->oap_page_off;
1415 pga[i].page_offset = pga[i].disk_offset;
1416 pga[i].pg = oap->oap_page;
1417 pga[i].count = oap->oap_count;
1418 pga[i].flag = oap->oap_brw_flags;
1419 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1420 pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1424 /* always get the data for the obdo for the rpc */
1425 LASSERT(ops != NULL);
1426 ops->ap_fill_obdo(caller_data, cmd, oa);
1428 sort_brw_pages(pga, page_count);
1429 rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1430 pga, &requested_nob, &nio_count, &req);
1432 CERROR("prep_req failed: %d\n", rc);
1433 GOTO(out, req = ERR_PTR(rc));
1436 LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1437 aa = (struct osc_brw_async_args *)&req->rq_async_args;
1439 aa->aa_requested_nob = requested_nob;
1440 aa->aa_nio_count = nio_count;
1441 aa->aa_page_count = page_count;
1450 OBD_FREE(pga, sizeof(*pga) * page_count);
1455 /* strange write gap too long (15s) */
1456 #define CLI_ODD_WRITE_GAP 15000000
1458 static void lop_update_pending(struct client_obd *cli,
1459 struct loi_oap_pages *lop, int cmd, int delta)
1461 lop->lop_num_pending += delta;
1462 if (cmd == OBD_BRW_WRITE)
1463 cli->cl_pending_w_pages += delta;
1465 cli->cl_pending_r_pages += delta;
1468 static long timeval_sub(struct timeval *large, struct timeval *small)
1470 return (large->tv_sec - small->tv_sec) * 1000000 +
1471 (large->tv_usec - small->tv_usec);
1474 /* the loi lock is held across this function but it's allowed to release
1475 * and reacquire it during its work */
1476 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1477 int cmd, struct loi_oap_pages *lop)
1479 struct ptlrpc_request *request;
1480 obd_count page_count = 0;
1481 struct list_head *tmp, *pos;
1482 struct osc_async_page *oap = NULL;
1483 struct osc_brw_async_args *aa;
1484 struct obd_async_page_ops *ops;
1485 LIST_HEAD(rpc_list);
1488 LASSERT(lop != LP_POISON);
1489 LASSERT(lop->lop_pending.next != LP_POISON);
1491 /* first we find the pages we're allowed to work with */
1492 list_for_each_safe(pos, tmp, &lop->lop_pending) {
1493 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1494 ops = oap->oap_caller_ops;
1496 LASSERT(oap->oap_magic == OAP_MAGIC);
1498 /* in llite being 'ready' equates to the page being locked
1499 * until completion unlocks it. commit_write submits a page
1500 * as not ready because its unlock will happen unconditionally
1501 * as the call returns. if we race with commit_write giving
1502 * us that page we dont' want to create a hole in the page
1503 * stream, so we stop and leave the rpc to be fired by
1504 * another dirtier or kupdated interval (the not ready page
1505 * will still be on the dirty list). we could call in
1506 * at the end of ll_file_write to process the queue again. */
1507 if (!(oap->oap_async_flags & ASYNC_READY)) {
1508 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1510 CDEBUG(D_INODE, "oap %p page %p returned %d "
1511 "instead of ready\n", oap,
1515 /* llite is telling us that the page is still
1516 * in commit_write and that we should try
1517 * and put it in an rpc again later. we
1518 * break out of the loop so we don't create
1519 * a hole in the sequence of pages in the rpc
1524 /* the io isn't needed.. tell the checks
1525 * below to complete the rpc with EINTR */
1526 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1527 oap->oap_count = -EINTR;
1530 oap->oap_async_flags |= ASYNC_READY;
1533 LASSERTF(0, "oap %p page %p returned %d "
1534 "from make_ready\n", oap,
1542 * Page submitted for IO has to be locked. Either by
1543 * ->ap_make_ready() or by higher layers.
1545 * XXX nikita: this assertion should be adjusted when lustre
1546 * starts using PG_writeback for pages being written out.
1548 LASSERT(PageLocked(oap->oap_page));
1550 /* take the page out of our book-keeping */
1551 list_del_init(&oap->oap_pending_item);
1552 lop_update_pending(cli, lop, cmd, -1);
1553 list_del_init(&oap->oap_urgent_item);
1555 /* ask the caller for the size of the io as the rpc leaves. */
1556 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1558 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1559 if (oap->oap_count <= 0) {
1560 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1562 osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1566 /* now put the page back in our accounting */
1567 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1568 if (++page_count >= cli->cl_max_pages_per_rpc)
1572 osc_wake_cache_waiters(cli);
1574 if (page_count == 0)
1577 loi_list_maint(cli, loi);
1578 spin_unlock(&cli->cl_loi_list_lock);
1580 request = osc_build_req(cli, &rpc_list, page_count, cmd);
1581 if (IS_ERR(request)) {
1582 /* this should happen rarely and is pretty bad, it makes the
1583 * pending list not follow the dirty order */
1584 spin_lock(&cli->cl_loi_list_lock);
1585 list_for_each_safe(pos, tmp, &rpc_list) {
1586 oap = list_entry(pos, struct osc_async_page,
1588 list_del_init(&oap->oap_rpc_item);
1590 /* queued sync pages can be torn down while the pages
1591 * were between the pending list and the rpc */
1592 if (oap->oap_interrupted) {
1593 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1594 osc_ap_completion(cli, NULL, oap, 0,
1599 /* put the page back in the loi/lop lists */
1600 list_add_tail(&oap->oap_pending_item,
1602 lop_update_pending(cli, lop, cmd, 1);
1603 if (oap->oap_async_flags & ASYNC_URGENT)
1604 list_add(&oap->oap_urgent_item,
1607 loi_list_maint(cli, loi);
1608 RETURN(PTR_ERR(request));
1611 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1612 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1613 INIT_LIST_HEAD(&aa->aa_oaps);
1614 list_splice(&rpc_list, &aa->aa_oaps);
1615 INIT_LIST_HEAD(&rpc_list);
1618 if (cmd == OBD_BRW_READ) {
1619 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1620 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1622 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1623 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1624 cli->cl_w_in_flight);
1628 spin_lock(&cli->cl_loi_list_lock);
1630 /* collect write gaps and sum of them */
1631 if (cmd == OBD_BRW_WRITE && cli->cl_w_in_flight == 0) {
1635 do_gettimeofday(&now);
1637 if (cli->cl_last_write_time.tv_sec) {
1638 diff = timeval_sub(&now, &cli->cl_last_write_time);
1639 if (diff < CLI_ODD_WRITE_GAP) {
1640 cli->cl_write_gap_sum += diff;
1641 cli->cl_write_gaps++;
1646 if (cmd == OBD_BRW_READ) {
1647 cli->cl_r_in_flight++;
1650 cli->cl_w_in_flight++;
1651 cli->cl_write_num++;
1654 /* queued sync pages can be torn down while the pages
1655 * were between the pending list and the rpc */
1656 list_for_each(pos, &aa->aa_oaps) {
1657 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1658 if (oap->oap_interrupted) {
1659 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1661 ptlrpc_mark_interrupted(request);
1666 CDEBUG(D_INODE, "req %p: %d pages, aa %p. now %dr/%dw in flight\n",
1667 request, page_count, aa, cli->cl_r_in_flight,
1668 cli->cl_w_in_flight);
1670 oap->oap_request = ptlrpc_request_addref(request);
1671 request->rq_interpret_reply = brw_interpret_oap;
1673 ptlrpcd_add_req(request);
1677 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1683 if (lop->lop_num_pending == 0)
1686 /* if we have an invalid import we want to drain the queued pages
1687 * by forcing them through rpcs that immediately fail and complete
1688 * the pages. recovery relies on this to empty the queued pages
1689 * before canceling the locks and evicting down the llite pages */
1690 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1693 /* stream rpcs in queue order as long as as there is an urgent page
1694 * queued. this is our cheap solution for good batching in the case
1695 * where writepage marks some random page in the middle of the file as
1696 * urgent because of, say, memory pressure */
1697 if (!list_empty(&lop->lop_urgent))
1700 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1701 optimal = cli->cl_max_pages_per_rpc;
1702 if (cmd == OBD_BRW_WRITE) {
1703 /* trigger a write rpc stream as long as there are dirtiers
1704 * waiting for space. as they're waiting, they're not going to
1705 * create more pages to coallesce with what's waiting.. */
1706 if (!list_empty(&cli->cl_cache_waiters))
1709 /* *2 to avoid triggering rpcs that would want to include pages
1710 * that are being queued but which can't be made ready until
1711 * the queuer finishes with the page. this is a wart for
1712 * llite::commit_write() */
1715 if (lop->lop_num_pending >= optimal)
1721 static void on_list(struct list_head *item, struct list_head *list,
1724 if (list_empty(item) && should_be_on)
1725 list_add_tail(item, list);
1726 else if (!list_empty(item) && !should_be_on)
1727 list_del_init(item);
1730 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1731 * can find pages to build into rpcs quickly */
1732 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1734 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1735 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1736 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1738 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1739 loi->loi_write_lop.lop_num_pending);
1741 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1742 loi->loi_read_lop.lop_num_pending);
1745 #define LOI_DEBUG(LOI, STR, args...) \
1746 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
1747 !list_empty(&(LOI)->loi_cli_item), \
1748 (LOI)->loi_write_lop.lop_num_pending, \
1749 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
1750 (LOI)->loi_read_lop.lop_num_pending, \
1751 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
1754 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1757 /* first return all objects which we already know to have
1758 * pages ready to be stuffed into rpcs */
1759 if (!list_empty(&cli->cl_loi_ready_list))
1760 RETURN(list_entry(cli->cl_loi_ready_list.next,
1761 struct lov_oinfo, loi_cli_item));
1763 /* then if we have cache waiters, return all objects with queued
1764 * writes. This is especially important when many small files
1765 * have filled up the cache and not been fired into rpcs because
1766 * they don't pass the nr_pending/object threshhold */
1767 if (!list_empty(&cli->cl_cache_waiters) &&
1768 !list_empty(&cli->cl_loi_write_list))
1769 RETURN(list_entry(cli->cl_loi_write_list.next,
1770 struct lov_oinfo, loi_write_item));
1772 /* then return all queued objects when we have an invalid import
1773 * so that they get flushed */
1774 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1775 if (!list_empty(&cli->cl_loi_write_list))
1776 RETURN(list_entry(cli->cl_loi_write_list.next,
1777 struct lov_oinfo, loi_write_item));
1778 if (!list_empty(&cli->cl_loi_read_list))
1779 RETURN(list_entry(cli->cl_loi_read_list.next,
1780 struct lov_oinfo, loi_read_item));
1785 /* called with the loi list lock held */
1786 static void osc_check_rpcs(struct client_obd *cli)
1788 struct lov_oinfo *loi;
1789 int rc = 0, race_counter = 0;
1792 while ((loi = osc_next_loi(cli)) != NULL) {
1793 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
1794 LASSERT(loi->loi_ost_idx != LL_POISON);
1796 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
1799 /* attempt some read/write balancing by alternating between
1800 * reads and writes in an object. The makes_rpc checks here
1801 * would be redundant if we were getting read/write work items
1802 * instead of objects. we don't want send_oap_rpc to drain a
1803 * partial read pending queue when we're given this object to
1804 * do io on writes while there are cache waiters */
1805 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1806 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1807 &loi->loi_write_lop);
1815 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1816 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1817 &loi->loi_read_lop);
1826 /* attempt some inter-object balancing by issueing rpcs
1827 * for each object in turn */
1828 if (!list_empty(&loi->loi_cli_item))
1829 list_del_init(&loi->loi_cli_item);
1830 if (!list_empty(&loi->loi_write_item))
1831 list_del_init(&loi->loi_write_item);
1832 if (!list_empty(&loi->loi_read_item))
1833 list_del_init(&loi->loi_read_item);
1835 loi_list_maint(cli, loi);
1837 /* send_oap_rpc fails with 0 when make_ready tells it to
1838 * back off. llite's make_ready does this when it tries
1839 * to lock a page queued for write that is already locked.
1840 * we want to try sending rpcs from many objects, but we
1841 * don't want to spin failing with 0. */
1842 if (race_counter == 10)
1848 /* we're trying to queue a page in the osc so we're subject to the
1849 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1850 * If the osc's queued pages are already at that limit, then we want to sleep
1851 * until there is space in the osc's queue for us. We also may be waiting for
1852 * write credits from the OST if there are RPCs in flight that may return some
1853 * before we fall back to sync writes.
1855 * We need this know our allocation was granted in the presence of signals */
1856 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1860 spin_lock(&cli->cl_loi_list_lock);
1861 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
1862 spin_unlock(&cli->cl_loi_list_lock);
1866 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1867 * grant or cache space. */
1868 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1869 struct osc_async_page *oap)
1871 struct osc_cache_waiter ocw;
1872 struct l_wait_info lwi = { 0 };
1873 struct timeval start, stop;
1876 CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1877 cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1878 cli->cl_avail_grant);
1880 if (cli->cl_dirty_max < PAGE_SIZE)
1883 if (~0ul - cli->cl_dirty_sum <= cli->cl_dirty) {
1884 cli->cl_dirty_av = (cli->cl_dirty_av +
1885 (cli->cl_dirty_sum / cli->cl_dirty_num)) / 2;
1886 cli->cl_dirty_num = 0;
1887 cli->cl_dirty_sum = 0;
1889 if (cli->cl_dirty_num)
1890 cli->cl_dirty_av = (cli->cl_dirty_sum / cli->cl_dirty_num);
1893 cli->cl_dirty_num++;
1894 cli->cl_dirty_sum += cli->cl_dirty;
1896 if (cli->cl_dirty > cli->cl_dirty_dmax)
1897 cli->cl_dirty_dmax = cli->cl_dirty;
1898 if (cli->cl_dirty < cli->cl_dirty_dmin)
1899 cli->cl_dirty_dmin = cli->cl_dirty;
1901 /* Hopefully normal case - cache space and write credits available */
1902 if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1903 cli->cl_avail_grant >= PAGE_SIZE) {
1904 /* account for ourselves */
1905 osc_consume_write_grant(cli, oap);
1909 /* Make sure that there are write rpcs in flight to wait for. This
1910 * is a little silly as this object may not have any pending but
1911 * other objects sure might. */
1912 if (cli->cl_w_in_flight) {
1913 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1914 init_waitqueue_head(&ocw.ocw_waitq);
1918 loi_list_maint(cli, loi);
1919 osc_check_rpcs(cli);
1920 spin_unlock(&cli->cl_loi_list_lock);
1922 CDEBUG(0, "sleeping for cache space\n");
1923 do_gettimeofday(&start);
1924 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1925 do_gettimeofday(&stop);
1927 cli->cl_cache_wait_num++;
1928 cli->cl_cache_wait_sum += timeval_sub(&stop, &start);
1930 spin_lock(&cli->cl_loi_list_lock);
1931 lprocfs_stime_record(&cli->cl_enter_stime, &stop, &start);
1932 if (!list_empty(&ocw.ocw_entry)) {
1933 list_del(&ocw.ocw_entry);
1941 cli->cl_sync_rpcs++;
1945 /* the companion to enter_cache, called when an oap is no longer part of the
1946 * dirty accounting.. so writeback completes or truncate happens before writing
1947 * starts. must be called with the loi lock held. */
1948 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1951 int blocksize = cli->cl_import->imp_obd->obd_osfs.os_bsize ? : 4096;
1954 if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1959 if (~0ul - cli->cl_dirty_sum <= cli->cl_dirty) {
1960 cli->cl_dirty_av = (cli->cl_dirty_av +
1961 (cli->cl_dirty_sum / cli->cl_dirty_num)) / 2;
1962 cli->cl_dirty_num = 0;
1963 cli->cl_dirty_sum = 0;
1965 if (cli->cl_dirty_num)
1966 cli->cl_dirty_av = (cli->cl_dirty_sum / cli->cl_dirty_num);
1969 cli->cl_dirty_num++;
1970 cli->cl_dirty_sum += cli->cl_dirty;
1972 if (cli->cl_dirty > cli->cl_dirty_dmax)
1973 cli->cl_dirty_dmax = cli->cl_dirty;
1974 if (cli->cl_dirty < cli->cl_dirty_dmin)
1975 cli->cl_dirty_dmin = cli->cl_dirty;
1977 oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1978 cli->cl_dirty -= PAGE_SIZE;
1980 cli->cl_lost_grant += PAGE_SIZE;
1981 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1982 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1983 } else if (PAGE_SIZE != blocksize && oap->oap_count != PAGE_SIZE) {
1984 /* For short writes we shouldn't count parts of pages that span
1985 * a whole block on the OST side, or our accounting goes wrong.
1986 * Should match the code in filter_grant_check. */
1987 int offset = (oap->oap_obj_off + oap->oap_page_off) & ~PAGE_MASK;
1988 int count = oap->oap_count + (offset & (blocksize - 1));
1989 int end = (offset + oap->oap_count) & (blocksize - 1);
1991 count += blocksize - end;
1993 cli->cl_lost_grant += PAGE_SIZE - count;
1994 CDEBUG(D_CACHE, "lost %lu grant: %lu avail: %lu dirty: %lu\n",
1995 PAGE_SIZE - count, cli->cl_lost_grant,
1996 cli->cl_avail_grant, cli->cl_dirty);
2002 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2003 struct lov_oinfo *loi, struct page *page,
2004 obd_off offset, struct obd_async_page_ops *ops,
2005 void *data, void **res)
2007 struct osc_async_page *oap;
2010 OBD_ALLOC(oap, sizeof(*oap));
2014 oap->oap_magic = OAP_MAGIC;
2015 oap->oap_cli = &exp->exp_obd->u.cli;
2018 oap->oap_caller_ops = ops;
2019 oap->oap_caller_data = data;
2021 oap->oap_page = page;
2022 oap->oap_obj_off = offset;
2024 INIT_LIST_HEAD(&oap->oap_pending_item);
2025 INIT_LIST_HEAD(&oap->oap_urgent_item);
2026 INIT_LIST_HEAD(&oap->oap_rpc_item);
2028 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
2030 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2035 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2036 struct lov_oinfo *loi, void *cookie,
2037 int cmd, obd_off off, int count,
2038 obd_flags brw_flags, enum async_flags async_flags)
2040 struct client_obd *cli = &exp->exp_obd->u.cli;
2041 struct osc_async_page *oap;
2042 struct loi_oap_pages *lop;
2046 oap = OAP_FROM_COOKIE(cookie);
2048 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2051 if (!list_empty(&oap->oap_pending_item) ||
2052 !list_empty(&oap->oap_urgent_item) ||
2053 !list_empty(&oap->oap_rpc_item))
2057 loi = &lsm->lsm_oinfo[0];
2059 spin_lock(&cli->cl_loi_list_lock);
2062 oap->oap_async_flags = async_flags;
2063 oap->oap_page_off = off;
2064 oap->oap_count = count;
2065 oap->oap_brw_flags = brw_flags;
2067 if (cmd == OBD_BRW_WRITE) {
2068 rc = osc_enter_cache(cli, loi, oap);
2070 spin_unlock(&cli->cl_loi_list_lock);
2073 lop = &loi->loi_write_lop;
2075 lop = &loi->loi_read_lop;
2078 if (oap->oap_async_flags & ASYNC_URGENT)
2079 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2080 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2081 lop_update_pending(cli, lop, cmd, 1);
2083 loi_list_maint(cli, loi);
2085 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2088 osc_check_rpcs(cli);
2089 spin_unlock(&cli->cl_loi_list_lock);
2094 /* aka (~was & now & flag), but this is more clear :) */
2095 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2097 static int osc_set_async_flags(struct obd_export *exp,
2098 struct lov_stripe_md *lsm,
2099 struct lov_oinfo *loi, void *cookie,
2100 obd_flags async_flags)
2102 struct client_obd *cli = &exp->exp_obd->u.cli;
2103 struct loi_oap_pages *lop;
2104 struct osc_async_page *oap;
2108 oap = OAP_FROM_COOKIE(cookie);
2110 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2114 loi = &lsm->lsm_oinfo[0];
2116 if (oap->oap_cmd == OBD_BRW_WRITE) {
2117 lop = &loi->loi_write_lop;
2119 lop = &loi->loi_read_lop;
2122 spin_lock(&cli->cl_loi_list_lock);
2124 if (list_empty(&oap->oap_pending_item))
2125 GOTO(out, rc = -EINVAL);
2127 if ((oap->oap_async_flags & async_flags) == async_flags)
2130 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2131 oap->oap_async_flags |= ASYNC_READY;
2133 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2134 if (list_empty(&oap->oap_rpc_item)) {
2135 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2136 loi_list_maint(cli, loi);
2140 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2141 oap->oap_async_flags);
2143 osc_check_rpcs(cli);
2144 spin_unlock(&cli->cl_loi_list_lock);
2148 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2149 struct lov_oinfo *loi,
2150 struct obd_io_group *oig, void *cookie,
2151 int cmd, obd_off off, int count,
2152 obd_flags brw_flags,
2153 obd_flags async_flags)
2155 struct client_obd *cli = &exp->exp_obd->u.cli;
2156 struct osc_async_page *oap;
2157 struct loi_oap_pages *lop;
2160 oap = OAP_FROM_COOKIE(cookie);
2162 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2165 if (!list_empty(&oap->oap_pending_item) ||
2166 !list_empty(&oap->oap_urgent_item) ||
2167 !list_empty(&oap->oap_rpc_item))
2171 loi = &lsm->lsm_oinfo[0];
2173 spin_lock(&cli->cl_loi_list_lock);
2176 oap->oap_page_off = off;
2177 oap->oap_count = count;
2178 oap->oap_brw_flags = brw_flags;
2179 oap->oap_async_flags = async_flags;
2181 if (cmd == OBD_BRW_WRITE)
2182 lop = &loi->loi_write_lop;
2184 lop = &loi->loi_read_lop;
2186 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2187 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2189 oig_add_one(oig, &oap->oap_occ);
2192 LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
2194 spin_unlock(&cli->cl_loi_list_lock);
2199 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2200 struct loi_oap_pages *lop, int cmd)
2202 struct list_head *pos, *tmp;
2203 struct osc_async_page *oap;
2205 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2206 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2207 list_del(&oap->oap_pending_item);
2208 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2209 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2210 lop_update_pending(cli, lop, cmd, 1);
2212 loi_list_maint(cli, loi);
2215 static int osc_trigger_group_io(struct obd_export *exp,
2216 struct lov_stripe_md *lsm,
2217 struct lov_oinfo *loi,
2218 struct obd_io_group *oig)
2220 struct client_obd *cli = &exp->exp_obd->u.cli;
2224 loi = &lsm->lsm_oinfo[0];
2226 spin_lock(&cli->cl_loi_list_lock);
2228 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2229 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2231 osc_check_rpcs(cli);
2232 spin_unlock(&cli->cl_loi_list_lock);
2237 static int osc_teardown_async_page(struct obd_export *exp,
2238 struct lov_stripe_md *lsm,
2239 struct lov_oinfo *loi, void *cookie)
2241 struct client_obd *cli = &exp->exp_obd->u.cli;
2242 struct loi_oap_pages *lop;
2243 struct osc_async_page *oap;
2247 oap = OAP_FROM_COOKIE(cookie);
2250 loi = &lsm->lsm_oinfo[0];
2252 if (oap->oap_cmd == OBD_BRW_WRITE) {
2253 lop = &loi->loi_write_lop;
2255 lop = &loi->loi_read_lop;
2258 spin_lock(&cli->cl_loi_list_lock);
2260 if (!list_empty(&oap->oap_rpc_item))
2261 GOTO(out, rc = -EBUSY);
2263 osc_exit_cache(cli, oap, 0);
2264 osc_wake_cache_waiters(cli);
2266 if (!list_empty(&oap->oap_urgent_item)) {
2267 list_del_init(&oap->oap_urgent_item);
2268 oap->oap_async_flags &= ~ASYNC_URGENT;
2270 if (!list_empty(&oap->oap_pending_item)) {
2271 list_del_init(&oap->oap_pending_item);
2272 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2274 loi_list_maint(cli, loi);
2276 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2278 spin_unlock(&cli->cl_loi_list_lock);
2280 OBD_FREE(oap, sizeof(*oap));
2285 /* Note: caller will lock/unlock, and set uptodate on the pages */
2286 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2287 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2288 struct lov_stripe_md *lsm, obd_count page_count,
2289 struct brw_page *pga)
2291 struct ptlrpc_request *request = NULL;
2292 struct ost_body *body;
2293 struct niobuf_remote *nioptr;
2294 struct obd_ioobj *iooptr;
2295 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2299 /* XXX does not handle 'new' brw protocol */
2301 size[1] = sizeof(struct obd_ioobj);
2302 size[2] = page_count * sizeof(*nioptr);
2304 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2305 OST_SAN_READ, 3, size, NULL);
2309 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2310 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2311 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2312 sizeof(*nioptr) * page_count);
2314 memcpy(&body->oa, oa, sizeof(body->oa));
2316 obdo_to_ioobj(oa, iooptr);
2317 iooptr->ioo_bufcnt = page_count;
2319 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2320 LASSERT(PageLocked(pga[mapped].pg));
2321 LASSERT(mapped == 0 ||
2322 pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2324 nioptr->offset = pga[mapped].disk_offset;
2325 nioptr->len = pga[mapped].count;
2326 nioptr->flags = pga[mapped].flag;
2329 size[1] = page_count * sizeof(*nioptr);
2330 request->rq_replen = lustre_msg_size(2, size);
2332 rc = ptlrpc_queue_wait(request);
2336 body = lustre_swab_repbuf(request, 0, sizeof(*body),
2337 lustre_swab_ost_body);
2339 CERROR("Can't unpack body\n");
2340 GOTO(out_req, rc = -EPROTO);
2343 memcpy(oa, &body->oa, sizeof(*oa));
2345 swab = lustre_msg_swabbed(request->rq_repmsg);
2346 LASSERT_REPSWAB(request, 1);
2347 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2349 /* nioptr missing or short */
2350 GOTO(out_req, rc = -EPROTO);
2354 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2355 struct page *page = pga[mapped].pg;
2356 struct buffer_head *bh;
2360 lustre_swab_niobuf_remote (nioptr);
2362 /* got san device associated */
2363 LASSERT(exp->exp_obd != NULL);
2364 dev = exp->exp_obd->u.cli.cl_sandev;
2367 if (!nioptr->offset) {
2368 CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2369 page->mapping->host->i_ino,
2371 memset(page_address(page), 0, PAGE_SIZE);
2375 if (!page->buffers) {
2376 create_empty_buffers(page, dev, PAGE_SIZE);
2379 clear_bit(BH_New, &bh->b_state);
2380 set_bit(BH_Mapped, &bh->b_state);
2381 bh->b_blocknr = (unsigned long)nioptr->offset;
2383 clear_bit(BH_Uptodate, &bh->b_state);
2385 ll_rw_block(READ, 1, &bh);
2389 /* if buffer already existed, it must be the
2390 * one we mapped before, check it */
2391 LASSERT(!test_bit(BH_New, &bh->b_state));
2392 LASSERT(test_bit(BH_Mapped, &bh->b_state));
2393 LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2395 /* wait it's io completion */
2396 if (test_bit(BH_Lock, &bh->b_state))
2399 if (!test_bit(BH_Uptodate, &bh->b_state))
2400 ll_rw_block(READ, 1, &bh);
2404 /* must do syncronous write here */
2406 if (!buffer_uptodate(bh)) {
2414 ptlrpc_req_finished(request);
2418 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2419 struct lov_stripe_md *lsm, obd_count page_count,
2420 struct brw_page *pga)
2422 struct ptlrpc_request *request = NULL;
2423 struct ost_body *body;
2424 struct niobuf_remote *nioptr;
2425 struct obd_ioobj *iooptr;
2426 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2430 size[1] = sizeof(struct obd_ioobj);
2431 size[2] = page_count * sizeof(*nioptr);
2433 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2434 OST_SAN_WRITE, 3, size, NULL);
2438 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2439 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2440 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2441 sizeof (*nioptr) * page_count);
2443 memcpy(&body->oa, oa, sizeof(body->oa));
2445 obdo_to_ioobj(oa, iooptr);
2446 iooptr->ioo_bufcnt = page_count;
2449 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2450 LASSERT(PageLocked(pga[mapped].pg));
2451 LASSERT(mapped == 0 ||
2452 pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2454 nioptr->offset = pga[mapped].disk_offset;
2455 nioptr->len = pga[mapped].count;
2456 nioptr->flags = pga[mapped].flag;
2459 size[1] = page_count * sizeof(*nioptr);
2460 request->rq_replen = lustre_msg_size(2, size);
2462 rc = ptlrpc_queue_wait(request);
2466 swab = lustre_msg_swabbed (request->rq_repmsg);
2467 LASSERT_REPSWAB (request, 1);
2468 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2470 CERROR("absent/short niobuf array\n");
2471 GOTO(out_req, rc = -EPROTO);
2475 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2476 struct page *page = pga[mapped].pg;
2477 struct buffer_head *bh;
2481 lustre_swab_niobuf_remote (nioptr);
2483 /* got san device associated */
2484 LASSERT(exp->exp_obd != NULL);
2485 dev = exp->exp_obd->u.cli.cl_sandev;
2487 if (!page->buffers) {
2488 create_empty_buffers(page, dev, PAGE_SIZE);
2491 LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2492 LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2493 LASSERT(page->buffers->b_blocknr ==
2494 (unsigned long)nioptr->offset);
2500 /* if buffer locked, wait it's io completion */
2501 if (test_bit(BH_Lock, &bh->b_state))
2504 clear_bit(BH_New, &bh->b_state);
2505 set_bit(BH_Mapped, &bh->b_state);
2507 /* override the block nr */
2508 bh->b_blocknr = (unsigned long)nioptr->offset;
2510 /* we are about to write it, so set it
2512 * page lock should garentee no race condition here */
2513 set_bit(BH_Uptodate, &bh->b_state);
2514 set_bit(BH_Dirty, &bh->b_state);
2516 ll_rw_block(WRITE, 1, &bh);
2518 /* must do syncronous write here */
2520 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2528 ptlrpc_req_finished(request);
2532 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2533 struct lov_stripe_md *lsm, obd_count page_count,
2534 struct brw_page *pga, struct obd_trans_info *oti)
2538 while (page_count) {
2539 obd_count pages_per_brw;
2542 if (page_count > PTLRPC_MAX_BRW_PAGES)
2543 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2545 pages_per_brw = page_count;
2547 if (cmd & OBD_BRW_WRITE)
2548 rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2550 rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2555 page_count -= pages_per_brw;
2556 pga += pages_per_brw;
2563 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2565 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2568 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
2572 lock_res_and_lock(lock);
2574 if (lock->l_ast_data && lock->l_ast_data != data) {
2575 struct inode *new_inode = data;
2576 struct inode *old_inode = lock->l_ast_data;
2577 if (!(old_inode->i_state & I_FREEING))
2578 LDLM_ERROR(lock, "inconsistent l_ast_data found");
2579 LASSERTF(old_inode->i_state & I_FREEING,
2580 "Found existing inode %p/%lu/%u state %lu in lock: "
2581 "setting data to %p/%lu/%u\n", old_inode,
2582 old_inode->i_ino, old_inode->i_generation,
2584 new_inode, new_inode->i_ino, new_inode->i_generation);
2587 lock->l_ast_data = data;
2588 unlock_res_and_lock(lock);
2589 LDLM_LOCK_PUT(lock);
2592 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2593 ldlm_iterator_t replace, void *data)
2595 struct ldlm_res_id res_id = { .name = {0} };
2596 struct obd_device *obd = class_exp2obd(exp);
2598 res_id.name[0] = lsm->lsm_object_id;
2599 res_id.name[2] = lsm->lsm_object_gr;
2600 ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2604 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2605 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2606 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2607 void *data, __u32 lvb_len, void *lvb_swabber,
2608 struct lustre_handle *lockh)
2610 struct obd_device *obd = exp->exp_obd;
2611 struct ldlm_res_id res_id = { .name = {0} };
2613 struct ldlm_reply *rep;
2614 struct ptlrpc_request *req = NULL;
2618 res_id.name[0] = lsm->lsm_object_id;
2619 res_id.name[2] = lsm->lsm_object_gr;
2621 /* Filesystem lock extents are extended to page boundaries so that
2622 * dealing with the page cache is a little smoother. */
2623 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2624 policy->l_extent.end |= ~PAGE_MASK;
2626 if (lsm->lsm_oinfo->loi_kms_valid == 0)
2629 /* Next, search for already existing extent locks that will cover us */
2630 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2633 if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2634 /* return immediately if no credential held */
2635 ldlm_lock_decref(lockh, mode);
2639 osc_set_data_with_check(lockh, data);
2640 if (*flags & LDLM_FL_HAS_INTENT) {
2641 /* I would like to be able to ASSERT here that rss <=
2642 * kms, but I can't, for reasons which are explained in
2645 /* We already have a lock, and it's referenced */
2649 /* If we're trying to read, we also search for an existing PW lock. The
2650 * VFS and page cache already protect us locally, so lots of readers/
2651 * writers can share a single PW lock.
2653 * There are problems with conversion deadlocks, so instead of
2654 * converting a read lock to a write lock, we'll just enqueue a new
2657 * At some point we should cancel the read lock instead of making them
2658 * send us a blocking callback, but there are problems with canceling
2659 * locks out from other users right now, too. */
2661 if (mode == LCK_PR) {
2662 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2663 policy, LCK_PW, lockh);
2665 if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2666 /* return immediately if no credential held */
2667 ldlm_lock_decref(lockh, LCK_PW);
2671 /* FIXME: This is not incredibly elegant, but it might
2672 * be more elegant than adding another parameter to
2673 * lock_match. I want a second opinion. */
2674 ldlm_lock_addref(lockh, LCK_PR);
2675 ldlm_lock_decref(lockh, LCK_PW);
2676 osc_set_data_with_check(lockh, data);
2680 if (mode == LCK_PW) {
2681 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2682 policy, LCK_PR, lockh);
2684 rc = ldlm_cli_convert(lockh, mode, flags);
2686 /* Update readers/writers accounting */
2687 ldlm_lock_addref(lockh, LCK_PW);
2688 ldlm_lock_decref(lockh, LCK_PR);
2689 osc_set_data_with_check(lockh, data);
2692 /* If the conversion failed, we need to drop refcount
2693 on matched lock before we get new one */
2694 /* XXX Won't it save us some efforts if we cancel PR
2695 lock here? We are going to take PW lock anyway and it
2696 will invalidate PR lock */
2697 ldlm_lock_decref(lockh, LCK_PR);
2698 if (rc != EDEADLOCK) {
2705 if (*flags & LDLM_FL_HAS_INTENT) {
2706 int size[2] = {0, sizeof(struct ldlm_request)};
2708 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
2709 LDLM_ENQUEUE, 2, size, NULL);
2713 size[0] = sizeof(*rep);
2714 size[1] = sizeof(lvb);
2715 req->rq_replen = lustre_msg_size(2, size);
2717 rc = ldlm_cli_enqueue(exp, req, obd->obd_namespace, res_id, type,
2718 policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2719 &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2721 if (rc == ELDLM_LOCK_ABORTED) {
2722 /* swabbed by ldlm_cli_enqueue() */
2723 LASSERT_REPSWABBED(req, 0);
2724 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
2725 LASSERT(rep != NULL);
2726 if (rep->lock_policy_res1)
2727 rc = rep->lock_policy_res1;
2729 ptlrpc_req_finished(req);
2732 if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2733 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2734 lvb.lvb_size, lvb.lvb_blocks);
2735 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2736 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2742 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2743 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2744 int *flags, void *data, struct lustre_handle *lockh)
2746 struct ldlm_res_id res_id = { .name = {0} };
2747 struct obd_device *obd = exp->exp_obd;
2751 res_id.name[0] = lsm->lsm_object_id;
2752 res_id.name[2] = lsm->lsm_object_gr;
2754 OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2756 /* Filesystem lock extents are extended to page boundaries so that
2757 * dealing with the page cache is a little smoother */
2758 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2759 policy->l_extent.end |= ~PAGE_MASK;
2761 /* Next, search for already existing extent locks that will cover us */
2762 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2763 policy, mode, lockh);
2765 // if (!(*flags & LDLM_FL_TEST_LOCK))
2766 osc_set_data_with_check(lockh, data);
2769 /* If we're trying to read, we also search for an existing PW lock. The
2770 * VFS and page cache already protect us locally, so lots of readers/
2771 * writers can share a single PW lock. */
2772 if (mode == LCK_PR) {
2773 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2774 policy, LCK_PW, lockh);
2775 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2776 /* FIXME: This is not incredibly elegant, but it might
2777 * be more elegant than adding another parameter to
2778 * lock_match. I want a second opinion. */
2779 osc_set_data_with_check(lockh, data);
2780 ldlm_lock_addref(lockh, LCK_PR);
2781 ldlm_lock_decref(lockh, LCK_PW);
2787 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2788 __u32 mode, struct lustre_handle *lockh)
2792 if (mode == LCK_GROUP)
2793 ldlm_lock_decref_and_cancel(lockh, mode);
2795 ldlm_lock_decref(lockh, mode);
2800 static int osc_cancel_unused(struct obd_export *exp,
2801 struct lov_stripe_md *lsm,
2802 int flags, void *opaque)
2804 struct obd_device *obd = class_exp2obd(exp);
2805 struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
2808 res_id.name[0] = lsm->lsm_object_id;
2809 res_id.name[2] = lsm->lsm_object_gr;
2813 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2816 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2817 unsigned long max_age)
2819 struct obd_statfs *msfs;
2820 struct ptlrpc_request *request;
2821 int rc, size = sizeof(*osfs);
2824 /* We could possibly pass max_age in the request (as an absolute
2825 * timestamp or a "seconds.usec ago") so the target can avoid doing
2826 * extra calls into the filesystem if that isn't necessary (e.g.
2827 * during mount that would help a bit). Having relative timestamps
2828 * is not so great if request processing is slow, while absolute
2829 * timestamps are not ideal because they need time synchronization. */
2830 request = ptlrpc_prep_req(obd->u.cli.cl_import, LUSTRE_OBD_VERSION,
2831 OST_STATFS, 0, NULL, NULL);
2835 request->rq_replen = lustre_msg_size(1, &size);
2836 request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2838 rc = ptlrpc_queue_wait(request);
2842 msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2843 lustre_swab_obd_statfs);
2845 CERROR("Can't unpack obd_statfs\n");
2846 GOTO(out, rc = -EPROTO);
2849 memcpy(osfs, msfs, sizeof(*osfs));
2853 ptlrpc_req_finished(request);
2857 /* Retrieve object striping information.
2859 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2860 * the maximum number of OST indices which will fit in the user buffer.
2861 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2863 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2865 struct lov_user_md lum, *lumk;
2872 rc = copy_from_user(&lum, lump, sizeof(lum));
2876 if (lum.lmm_magic != LOV_USER_MAGIC)
2879 if (lum.lmm_stripe_count > 0) {
2880 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2881 OBD_ALLOC(lumk, lum_size);
2885 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2886 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
2888 lum_size = sizeof(lum);
2892 lumk->lmm_object_id = lsm->lsm_object_id;
2893 lumk->lmm_object_gr = lsm->lsm_object_gr;
2894 lumk->lmm_stripe_count = 1;
2896 if (copy_to_user(lump, lumk, lum_size))
2900 OBD_FREE(lumk, lum_size);
2905 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2906 void *karg, void *uarg)
2908 struct obd_device *obd = exp->exp_obd;
2909 struct obd_ioctl_data *data = karg;
2913 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2916 if (!try_module_get(THIS_MODULE)) {
2917 CERROR("Can't get module. Is it alive?");
2922 case OBD_IOC_LOV_GET_CONFIG: {
2924 struct lov_desc *desc;
2925 struct obd_uuid uuid;
2929 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2930 GOTO(out, err = -EINVAL);
2932 data = (struct obd_ioctl_data *)buf;
2934 if (sizeof(*desc) > data->ioc_inllen1) {
2936 GOTO(out, err = -EINVAL);
2939 if (data->ioc_inllen2 < sizeof(uuid)) {
2941 GOTO(out, err = -EINVAL);
2944 if (data->ioc_inllen3 < sizeof(__u32)) {
2946 GOTO(out, err = -EINVAL);
2949 desc = (struct lov_desc *)data->ioc_inlbuf1;
2950 desc->ld_tgt_count = 1;
2951 desc->ld_active_tgt_count = 1;
2952 desc->ld_default_stripe_count = 1;
2953 desc->ld_default_stripe_size = 0;
2954 desc->ld_default_stripe_offset = 0;
2955 desc->ld_pattern = 0;
2956 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2957 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2958 *((__u32 *)data->ioc_inlbuf3) = 1;
2960 err = copy_to_user((void *)uarg, buf, len);
2963 obd_ioctl_freedata(buf, len);
2966 case LL_IOC_LOV_SETSTRIPE:
2967 err = obd_alloc_memmd(exp, karg);
2971 case LL_IOC_LOV_GETSTRIPE:
2972 err = osc_getstripe(karg, uarg);
2974 case OBD_IOC_CLIENT_RECOVER:
2975 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2980 case IOC_OSC_SET_ACTIVE:
2981 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2984 case IOC_OSC_CTL_RECOVERY:
2985 err = ptlrpc_import_control_recovery(obd->u.cli.cl_import,
2989 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2990 GOTO(out, err = -ENOTTY);
2993 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2996 module_put(THIS_MODULE);
3001 static int osc_get_info(struct obd_export *exp, __u32 keylen,
3002 void *key, __u32 *vallen, void *val)
3005 if (!vallen || !val)
3008 if (keylen > strlen("lock_to_stripe") &&
3009 strcmp(key, "lock_to_stripe") == 0) {
3010 __u32 *stripe = val;
3011 *vallen = sizeof(*stripe);
3014 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
3015 struct ptlrpc_request *req;
3017 char *bufs[1] = {key};
3019 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3020 OST_GET_INFO, 1, (int *)&keylen, bufs);
3024 req->rq_replen = lustre_msg_size(1, (int *)vallen);
3025 rc = ptlrpc_queue_wait(req);
3029 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
3030 lustre_swab_ost_last_id);
3031 if (reply == NULL) {
3032 CERROR("Can't unpack OST last ID\n");
3033 GOTO(out, rc = -EPROTO);
3035 *((obd_id *)val) = *reply;
3037 ptlrpc_req_finished(req);
3039 } else if (keylen == 10 && strcmp(key, "client_nid") == 0) {
3040 struct ptlrpc_connection * conn;
3041 ptl_nid_t * nid = val;
3042 ptl_process_id_t id;
3045 *vallen = sizeof(*nid);
3046 conn = class_exp2cliimp(exp)->imp_connection;
3048 if (!conn || !conn->c_peer.peer_ni)
3051 rc = PtlGetId(conn->c_peer.peer_ni->pni_ni_h, &id);
3060 static int osc_set_info(struct obd_export *exp, obd_count keylen,
3061 void *key, obd_count vallen, void *val)
3063 struct obd_device *obd = exp->exp_obd;
3064 struct obd_import *imp = class_exp2cliimp(exp);
3065 struct llog_ctxt *ctxt;
3069 if (keylen == strlen("unlinked") &&
3070 memcmp(key, "unlinked", keylen) == 0) {
3071 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3072 spin_lock(&oscc->oscc_lock);
3073 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3074 spin_unlock(&oscc->oscc_lock);
3078 if (keylen == strlen("unrecovery") &&
3079 memcmp(key, "unrecovery", keylen) == 0) {
3080 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3081 spin_lock(&oscc->oscc_lock);
3082 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3083 spin_unlock(&oscc->oscc_lock);
3087 if (keylen == strlen("initial_recov") &&
3088 memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
3089 struct obd_import *imp = class_exp2cliimp(exp);
3090 if (vallen != sizeof(int))
3092 imp->imp_initial_recov = *(int *)val;
3093 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
3094 exp->exp_obd->obd_name,
3095 imp->imp_initial_recov);
3099 if (keylen == strlen("async") &&
3100 memcmp(key, "async", keylen) == 0) {
3101 struct client_obd *cl = &obd->u.cli;
3102 if (vallen != sizeof(int))
3104 cl->cl_async = *(int *)val;
3105 CDEBUG(D_HA, "%s: set async = %d\n",
3106 obd->obd_name, cl->cl_async);
3110 if (keylen == 5 && strcmp(key, "audit") == 0) {
3111 struct ptlrpc_request *req;
3112 char *bufs[2] = {key, val};
3113 int size[2] = {keylen, vallen};
3115 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3116 OST_SET_INFO, 2, size, bufs);
3120 req->rq_replen = lustre_msg_size(0, size);
3121 lustre_swab_reqbuf(req, 1, sizeof(struct audit_attr_msg),
3122 lustre_swab_audit_attr);
3123 rc = ptlrpc_queue_wait(req);
3125 ptlrpc_req_finished(req);
3129 if (keylen == 9 && strcmp(key, "audit_obj") == 0) {
3130 struct ptlrpc_request *req;
3131 char *bufs[2] = {key, val};
3132 int size[2] = {keylen, vallen};
3134 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3135 OST_SET_INFO, 2, size, bufs);
3139 req->rq_replen = lustre_msg_size(0, size);
3140 lustre_swab_reqbuf(req, 1, sizeof(struct obdo),
3142 rc = ptlrpc_queue_wait(req);
3144 ptlrpc_req_finished(req);
3148 if (keylen == 8 && memcmp(key, "auditlog", 8) == 0) {
3149 struct ptlrpc_request *req;
3150 char *bufs[2] = {key, val};
3151 int size[2] = {keylen, vallen};
3153 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3154 OST_SET_INFO, 2, size, bufs);
3158 req->rq_replen = lustre_msg_size(0, size);
3159 lustre_swab_reqbuf(req, 1, sizeof(struct audit_msg),
3160 lustre_swab_audit_msg);
3161 rc = ptlrpc_queue_wait(req);
3163 ptlrpc_req_finished(req);
3167 if (keylen == strlen("sec") && memcmp(key, "sec", keylen) == 0) {
3168 struct client_obd *cli = &exp->exp_obd->u.cli;
3170 cli->cl_sec_flavor = ptlrpcs_name2flavor(val);
3171 if (cli->cl_sec_flavor == PTLRPCS_FLVR_INVALID) {
3172 CERROR("unrecognized security flavor %s\n", (char*) val);
3179 if (keylen == strlen("sec_flags") &&
3180 memcmp(key, "sec_flags", keylen) == 0) {
3181 struct client_obd *cli = &exp->exp_obd->u.cli;
3183 cli->cl_sec_flags = *((unsigned long *) val);
3187 if (keylen == strlen("flush_cred") &&
3188 memcmp(key, "flush_cred", keylen) == 0) {
3189 struct client_obd *cli = &exp->exp_obd->u.cli;
3192 ptlrpcs_import_flush_current_creds(cli->cl_import);
3195 if (keylen == strlen("crypto_cb") &&
3196 memcmp(key, "crypto_cb", keylen) == 0) {
3197 LASSERT(vallen == sizeof(crypt_cb_t));
3198 osc_crypt_cb = (crypt_cb_t)val;
3202 if (keylen == 8 && memcmp(key, "capa_key", 8) == 0) {
3203 struct ptlrpc_request *req;
3204 char *bufs[2] = {key, val};
3205 unsigned long irqflags;
3206 int rc, size[2] = {keylen, vallen};
3208 LASSERT(vallen == sizeof(struct lustre_capa_key));
3210 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3211 OST_SET_INFO, 2, size, bufs);
3215 spin_lock_irqsave (&req->rq_lock, irqflags);
3217 spin_unlock_irqrestore (&req->rq_lock, irqflags);
3219 req->rq_replen = lustre_msg_size(0, NULL);
3220 rc = ptlrpc_queue_wait(req);
3221 ptlrpc_req_finished(req);
3225 if (keylen == strlen("setext") &&
3226 memcmp(key, "setext", keylen) == 0) {
3227 struct client_obd *cli = &exp->exp_obd->u.cli;
3228 struct osc_creator *oscc = &cli->cl_oscc;
3229 struct fid_extent *ext = val;
3231 oscc->oscc_next_id = (obd_id)ext->fe_start;
3235 if (keylen < strlen("mds_conn") ||
3236 memcmp(key, "mds_conn", keylen) != 0)
3239 ctxt = llog_get_context(&exp->exp_obd->obd_llogs,
3240 LLOG_UNLINK_ORIG_CTXT);
3243 rc = llog_initiator_connect(ctxt);
3245 CERROR("cannot establish the connect for "
3246 "ctxt %p: %d\n", ctxt, rc);
3249 imp->imp_server_timeout = 1;
3250 CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
3251 imp->imp_pingable = 1;
3257 static struct llog_operations osc_size_repl_logops = {
3258 lop_cancel: llog_obd_repl_cancel
3261 static struct llog_operations osc_unlink_orig_logops;
3263 static int osc_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
3264 struct obd_device *tgt, int count,
3265 struct llog_catid *catid)
3270 osc_unlink_orig_logops = llog_lvfs_ops;
3271 osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
3272 osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
3273 osc_unlink_orig_logops.lop_add = llog_catalog_add;
3274 osc_unlink_orig_logops.lop_connect = llog_origin_connect;
3276 rc = obd_llog_setup(obd, llogs, LLOG_UNLINK_ORIG_CTXT, tgt, count,
3277 &catid->lci_logid, &osc_unlink_orig_logops);
3281 rc = obd_llog_setup(obd, llogs, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
3282 &osc_size_repl_logops);
3286 static int osc_llog_finish(struct obd_device *obd,
3287 struct obd_llogs *llogs, int count)
3292 rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_ORIG_CTXT));
3296 rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_SIZE_REPL_CTXT));
3300 static int osc_connect(struct lustre_handle *exph,
3301 struct obd_device *obd, struct obd_uuid *cluuid,
3302 struct obd_connect_data *data,
3303 unsigned long connect_flags)
3307 rc = client_connect_import(exph, obd, cluuid, data, connect_flags);
3311 static int osc_disconnect(struct obd_export *exp, unsigned long flags)
3313 struct obd_device *obd = class_exp2obd(exp);
3314 struct llog_ctxt *ctxt;
3318 ctxt = llog_get_context(&obd->obd_llogs, LLOG_SIZE_REPL_CTXT);
3319 if (obd->u.cli.cl_conn_count == 1)
3320 /* flush any remaining cancel messages out to the target */
3321 llog_sync(ctxt, exp);
3323 rc = client_disconnect_export(exp, flags);
3327 static int osc_import_event(struct obd_device *obd,
3328 struct obd_import *imp,
3329 enum obd_import_event event)
3331 struct client_obd *cli;
3335 LASSERT(imp->imp_obd == obd);
3338 case IMP_EVENT_DISCON: {
3339 /* Only do this on the MDS OSC's */
3340 if (imp->imp_server_timeout) {
3341 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3343 spin_lock(&oscc->oscc_lock);
3344 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3345 spin_unlock(&oscc->oscc_lock);
3349 case IMP_EVENT_INACTIVE: {
3350 if (obd->obd_observer)
3351 rc = obd_notify(obd->obd_observer, obd, 0, 0);
3354 case IMP_EVENT_INVALIDATE: {
3355 struct ldlm_namespace *ns = obd->obd_namespace;
3359 spin_lock(&cli->cl_loi_list_lock);
3360 cli->cl_avail_grant = 0;
3361 cli->cl_lost_grant = 0;
3362 /* all pages go to failing rpcs due to the invalid import */
3363 osc_check_rpcs(cli);
3364 spin_unlock(&cli->cl_loi_list_lock);
3366 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3370 case IMP_EVENT_ACTIVE: {
3371 /* Only do this on the MDS OSC's */
3372 if (imp->imp_server_timeout) {
3373 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3375 spin_lock(&oscc->oscc_lock);
3376 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3377 spin_unlock(&oscc->oscc_lock);
3380 if (obd->obd_observer)
3381 rc = obd_notify(obd->obd_observer, obd, 1, 0);
3385 CERROR("Unknown import event %d\n", event);
3391 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
3393 struct lprocfs_static_vars lvars;
3397 lprocfs_init_vars(osc,&lvars);
3398 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3402 rc = lproc_osc_attach_seqstat(dev);
3404 lprocfs_obd_detach(dev);
3408 ptlrpc_lprocfs_register_obd(dev);
3412 static int osc_detach(struct obd_device *dev)
3414 ptlrpc_lprocfs_unregister_obd(dev);
3415 return lprocfs_obd_detach(dev);
3418 static int osc_setup(struct obd_device *obd, obd_count len, void *buf)
3422 rc = ptlrpcd_addref();
3426 rc = client_obd_setup(obd, len, buf);
3435 static int osc_cleanup(struct obd_device *obd, int flags)
3437 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3440 rc = ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
3441 LDLM_FL_CONFIG_CHANGE, NULL);
3445 spin_lock(&oscc->oscc_lock);
3446 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3447 oscc->oscc_flags |= OSCC_FLAG_EXITING;
3448 spin_unlock(&oscc->oscc_lock);
3450 rc = client_obd_cleanup(obd, flags);
3456 struct obd_ops osc_obd_ops = {
3457 .o_owner = THIS_MODULE,
3458 .o_attach = osc_attach,
3459 .o_detach = osc_detach,
3460 .o_setup = osc_setup,
3461 .o_cleanup = osc_cleanup,
3462 .o_add_conn = client_import_add_conn,
3463 .o_del_conn = client_import_del_conn,
3464 .o_connect = osc_connect,
3465 .o_disconnect = osc_disconnect,
3466 .o_statfs = osc_statfs,
3467 .o_packmd = osc_packmd,
3468 .o_unpackmd = osc_unpackmd,
3469 .o_create = osc_create,
3470 .o_destroy = osc_destroy,
3471 .o_getattr = osc_getattr,
3472 .o_getattr_async = osc_getattr_async,
3473 .o_setattr = osc_setattr,
3475 .o_brw_async = osc_brw_async,
3476 .o_prep_async_page = osc_prep_async_page,
3477 .o_queue_async_io = osc_queue_async_io,
3478 .o_set_async_flags = osc_set_async_flags,
3479 .o_queue_group_io = osc_queue_group_io,
3480 .o_trigger_group_io = osc_trigger_group_io,
3481 .o_teardown_async_page = osc_teardown_async_page,
3482 .o_punch = osc_punch,
3484 .o_enqueue = osc_enqueue,
3485 .o_match = osc_match,
3486 .o_change_cbdata = osc_change_cbdata,
3487 .o_cancel = osc_cancel,
3488 .o_cancel_unused = osc_cancel_unused,
3489 .o_iocontrol = osc_iocontrol,
3490 .o_get_info = osc_get_info,
3491 .o_set_info = osc_set_info,
3492 .o_import_event = osc_import_event,
3493 .o_llog_init = osc_llog_init,
3494 .o_llog_finish = osc_llog_finish,
3497 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3498 struct obd_ops sanosc_obd_ops = {
3499 .o_owner = THIS_MODULE,
3500 .o_attach = osc_attach,
3501 .o_detach = osc_detach,
3502 .o_cleanup = client_obd_cleanup,
3503 .o_add_conn = client_import_add_conn,
3504 .o_del_conn = client_import_del_conn,
3505 .o_connect = osc_connect,
3506 .o_disconnect = client_disconnect_export,
3507 .o_statfs = osc_statfs,
3508 .o_packmd = osc_packmd,
3509 .o_unpackmd = osc_unpackmd,
3510 .o_create = osc_real_create,
3511 .o_destroy = osc_destroy,
3512 .o_getattr = osc_getattr,
3513 .o_getattr_async = osc_getattr_async,
3514 .o_setattr = osc_setattr,
3515 .o_setup = client_sanobd_setup,
3516 .o_brw = sanosc_brw,
3517 .o_punch = osc_punch,
3519 .o_enqueue = osc_enqueue,
3520 .o_match = osc_match,
3521 .o_change_cbdata = osc_change_cbdata,
3522 .o_cancel = osc_cancel,
3523 .o_cancel_unused = osc_cancel_unused,
3524 .o_iocontrol = osc_iocontrol,
3525 .o_import_event = osc_import_event,
3526 .o_llog_init = osc_llog_init,
3527 .o_llog_finish = osc_llog_finish,
3531 int __init osc_init(void)
3533 struct lprocfs_static_vars lvars;
3534 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3535 struct lprocfs_static_vars sanlvars;
3540 lprocfs_init_vars(osc, &lvars);
3541 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3542 lprocfs_init_vars(osc, &sanlvars);
3545 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3546 OBD_OSC_DEVICENAME);
3550 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3551 rc = class_register_type(&sanosc_obd_ops, NULL, sanlvars.module_vars,
3552 OBD_SANOSC_DEVICENAME);
3554 class_unregister_type(OBD_OSC_DEVICENAME);
3561 static void /*__exit*/ osc_exit(void)
3563 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3564 class_unregister_type(OBD_SANOSC_DEVICENAME);
3566 class_unregister_type(OBD_OSC_DEVICENAME);
3569 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3570 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3571 MODULE_LICENSE("GPL");
3573 module_init(osc_init);
3574 module_exit(osc_exit);