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/lustre_dlm.h>
40 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
41 # include <linux/workqueue.h>
42 # include <linux/smp_lock.h>
44 # include <linux/locks.h>
46 #else /* __KERNEL__ */
47 # include <liblustre.h>
50 #include <linux/kp30.h>
51 #include <linux/lustre_net.h>
52 #include <linux/lustre_user.h>
53 #include <linux/obd_ost.h>
54 #include <linux/obd_lov.h>
57 # include <linux/ctype.h>
58 # include <linux/init.h>
63 #include <linux/lustre_ha.h>
64 #include <linux/lprocfs_status.h>
65 #include <linux/lustre_log.h>
66 #include "osc_internal.h"
69 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
71 struct lprocfs_static_vars lvars;
75 lprocfs_init_vars(osc,&lvars);
76 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
80 rc = lproc_osc_attach_seqstat(dev);
82 lprocfs_obd_detach(dev);
86 ptlrpc_lprocfs_register_obd(dev);
90 static int osc_detach(struct obd_device *dev)
92 ptlrpc_lprocfs_unregister_obd(dev);
93 return lprocfs_obd_detach(dev);
97 /* Pack OSC object metadata for disk storage (LE byte order). */
98 static int osc_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
99 struct lov_stripe_md *lsm)
104 lmm_size = sizeof(**lmmp);
109 OBD_FREE(*lmmp, lmm_size);
115 OBD_ALLOC(*lmmp, lmm_size);
121 LASSERT(lsm->lsm_object_id);
122 (*lmmp)->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
128 /* Unpack OSC object metadata from disk storage (LE byte order). */
129 static int osc_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
130 struct lov_mds_md *lmm, int lmm_bytes)
136 if (lmm_bytes < sizeof (*lmm)) {
137 CERROR("lov_mds_md too small: %d, need %d\n",
138 lmm_bytes, (int)sizeof(*lmm));
141 /* XXX LOV_MAGIC etc check? */
143 if (lmm->lmm_object_id == 0) {
144 CERROR("lov_mds_md: zero lmm_object_id\n");
149 lsm_size = lov_stripe_md_size(1);
153 if (*lsmp != NULL && lmm == NULL) {
154 OBD_FREE(*lsmp, lsm_size);
160 OBD_ALLOC(*lsmp, lsm_size);
163 loi_init((*lsmp)->lsm_oinfo);
167 /* XXX zero *lsmp? */
168 (*lsmp)->lsm_object_id = le64_to_cpu (lmm->lmm_object_id);
169 LASSERT((*lsmp)->lsm_object_id);
172 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES;
177 static int osc_getattr_interpret(struct ptlrpc_request *req,
178 struct osc_getattr_async_args *aa, int rc)
180 struct ost_body *body;
186 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
188 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
189 memcpy(aa->aa_oa, &body->oa, sizeof(*aa->aa_oa));
191 /* This should really be sent by the OST */
192 aa->aa_oa->o_blksize = OSC_BRW_MAX_SIZE;
193 aa->aa_oa->o_valid |= OBD_MD_FLBLKSZ;
195 CERROR("can't unpack ost_body\n");
197 aa->aa_oa->o_valid = 0;
203 static int osc_getattr_async(struct obd_export *exp, struct obdo *oa,
204 struct lov_stripe_md *md,
205 struct ptlrpc_request_set *set)
207 struct ptlrpc_request *request;
208 struct ost_body *body;
209 int size = sizeof(*body);
210 struct osc_getattr_async_args *aa;
213 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GETATTR, 1,
218 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
219 memcpy(&body->oa, oa, sizeof(*oa));
221 request->rq_replen = lustre_msg_size(1, &size);
222 request->rq_interpret_reply = osc_getattr_interpret;
224 LASSERT (sizeof (*aa) <= sizeof (request->rq_async_args));
225 aa = (struct osc_getattr_async_args *)&request->rq_async_args;
228 ptlrpc_set_add_req (set, request);
232 static int osc_getattr(struct obd_export *exp, struct obdo *oa,
233 struct lov_stripe_md *md)
235 struct ptlrpc_request *request;
236 struct ost_body *body;
237 int rc, size = sizeof(*body);
240 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GETATTR, 1,
245 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
246 memcpy(&body->oa, oa, sizeof(*oa));
248 request->rq_replen = lustre_msg_size(1, &size);
250 rc = ptlrpc_queue_wait(request);
252 CERROR("%s failed: rc = %d\n", __FUNCTION__, rc);
256 body = lustre_swab_repbuf(request, 0, sizeof (*body),
257 lustre_swab_ost_body);
259 CERROR ("can't unpack ost_body\n");
260 GOTO (out, rc = -EPROTO);
263 CDEBUG(D_INODE, "mode: %o\n", body->oa.o_mode);
264 memcpy(oa, &body->oa, sizeof(*oa));
266 /* This should really be sent by the OST */
267 oa->o_blksize = OSC_BRW_MAX_SIZE;
268 oa->o_valid |= OBD_MD_FLBLKSZ;
272 ptlrpc_req_finished(request);
276 static int osc_setattr(struct obd_export *exp, struct obdo *oa,
277 struct lov_stripe_md *md, struct obd_trans_info *oti)
279 struct ptlrpc_request *request;
280 struct ost_body *body;
281 int rc, size = sizeof(*body);
284 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SETATTR, 1, &size,
289 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
290 memcpy(&body->oa, oa, sizeof(*oa));
292 request->rq_replen = lustre_msg_size(1, &size);
294 rc = ptlrpc_queue_wait(request);
298 body = lustre_swab_repbuf(request, 0, sizeof(*body),
299 lustre_swab_ost_body);
301 GOTO(out, rc = -EPROTO);
303 memcpy(oa, &body->oa, sizeof(*oa));
307 ptlrpc_req_finished(request);
311 int osc_real_create(struct obd_export *exp, struct obdo *oa,
312 struct lov_stripe_md **ea, struct obd_trans_info *oti)
314 struct ptlrpc_request *request;
315 struct ost_body *body;
316 struct lov_stripe_md *lsm;
317 int rc, size = sizeof(*body);
325 rc = obd_alloc_memmd(exp, &lsm);
330 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_CREATE, 1, &size,
333 GOTO(out, rc = -ENOMEM);
335 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
336 memcpy(&body->oa, oa, sizeof(body->oa));
338 request->rq_replen = lustre_msg_size(1, &size);
339 if (oa->o_valid & OBD_MD_FLINLINE) {
340 LASSERT((oa->o_valid & OBD_MD_FLFLAGS) &&
341 oa->o_flags == OBD_FL_DELORPHAN);
342 DEBUG_REQ(D_HA, request,
343 "delorphan from OST integration");
346 rc = ptlrpc_queue_wait(request);
350 body = lustre_swab_repbuf(request, 0, sizeof(*body),
351 lustre_swab_ost_body);
353 CERROR ("can't unpack ost_body\n");
354 GOTO (out_req, rc = -EPROTO);
357 memcpy(oa, &body->oa, sizeof(*oa));
359 /* This should really be sent by the OST */
360 oa->o_blksize = OSC_BRW_MAX_SIZE;
361 oa->o_valid |= OBD_MD_FLBLKSZ;
363 /* XXX LOV STACKING: the lsm that is passed to us from LOV does not
364 * have valid lsm_oinfo data structs, so don't go touching that.
365 * This needs to be fixed in a big way.
367 lsm->lsm_object_id = oa->o_id;
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)
395 struct ptlrpc_request *request;
396 struct ost_body *body;
397 int rc, size = sizeof(*body);
405 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_PUNCH, 1, &size,
410 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
411 memcpy(&body->oa, oa, sizeof(*oa));
413 /* overload the size and blocks fields in the oa with start/end */
414 body->oa.o_size = start;
415 body->oa.o_blocks = end;
416 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
418 request->rq_replen = lustre_msg_size(1, &size);
420 rc = ptlrpc_queue_wait(request);
424 body = lustre_swab_repbuf (request, 0, sizeof (*body),
425 lustre_swab_ost_body);
427 CERROR ("can't unpack ost_body\n");
428 GOTO (out, rc = -EPROTO);
431 memcpy(oa, &body->oa, sizeof(*oa));
435 ptlrpc_req_finished(request);
439 static int osc_sync(struct obd_export *exp, struct obdo *oa,
440 struct lov_stripe_md *md, obd_size start, obd_size end)
442 struct ptlrpc_request *request;
443 struct ost_body *body;
444 int rc, size = sizeof(*body);
452 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SYNC, 1, &size,
457 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
458 memcpy(&body->oa, oa, sizeof(*oa));
460 /* overload the size and blocks fields in the oa with start/end */
461 body->oa.o_size = start;
462 body->oa.o_blocks = end;
463 body->oa.o_valid |= (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS);
465 request->rq_replen = lustre_msg_size(1, &size);
467 rc = ptlrpc_queue_wait(request);
471 body = lustre_swab_repbuf(request, 0, sizeof(*body),
472 lustre_swab_ost_body);
474 CERROR ("can't unpack ost_body\n");
475 GOTO (out, rc = -EPROTO);
478 memcpy(oa, &body->oa, sizeof(*oa));
482 ptlrpc_req_finished(request);
486 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
487 struct lov_stripe_md *ea, struct obd_trans_info *oti)
489 struct ptlrpc_request *request;
490 struct ost_body *body;
491 int rc, size = sizeof(*body);
499 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_DESTROY, 1,
504 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
506 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) {
507 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
508 sizeof(*oti->oti_logcookies));
509 oti->oti_logcookies++;
512 memcpy(&body->oa, oa, sizeof(*oa));
513 request->rq_replen = lustre_msg_size(1, &size);
515 rc = ptlrpc_queue_wait(request);
519 body = lustre_swab_repbuf(request, 0, sizeof(*body),
520 lustre_swab_ost_body);
522 CERROR ("Can't unpack body\n");
523 GOTO (out, rc = -EPROTO);
526 memcpy(oa, &body->oa, sizeof(*oa));
530 ptlrpc_req_finished(request);
534 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
537 obd_flag bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
539 LASSERT(!(oa->o_valid & bits));
542 spin_lock(&cli->cl_loi_list_lock);
543 oa->o_dirty = cli->cl_dirty;
544 oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
545 oa->o_grant = cli->cl_avail_grant;
546 oa->o_dropped = cli->cl_lost_grant;
547 cli->cl_lost_grant = 0;
548 spin_unlock(&cli->cl_loi_list_lock);
549 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
550 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
553 /* caller must hold loi_list_lock */
554 static void osc_consume_write_grant(struct client_obd *cli,
555 struct osc_async_page *oap)
557 cli->cl_dirty += PAGE_SIZE;
558 cli->cl_avail_grant -= PAGE_SIZE;
559 oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
560 CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
561 LASSERT(cli->cl_avail_grant >= 0);
564 /* caller must hold loi_list_lock */
565 void osc_wake_cache_waiters(struct client_obd *cli)
567 struct list_head *l, *tmp;
568 struct osc_cache_waiter *ocw;
570 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
571 /* if we can't dirty more, we must wait until some is written */
572 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
573 CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
574 cli->cl_dirty, cli->cl_dirty_max);
578 /* if still dirty cache but no grant wait for pending RPCs that
579 * may yet return us some grant before doing sync writes */
580 if (cli->cl_brw_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
581 CDEBUG(D_CACHE, "%d BRWs in flight, no grant\n",
582 cli->cl_brw_in_flight);
586 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
587 list_del_init(&ocw->ocw_entry);
588 if (cli->cl_avail_grant < PAGE_SIZE) {
589 /* no more RPCs in flight to return grant, do sync IO */
590 ocw->ocw_rc = -EDQUOT;
591 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
593 osc_consume_write_grant(cli, ocw->ocw_oap);
596 wake_up(&ocw->ocw_waitq);
602 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
604 spin_lock(&cli->cl_loi_list_lock);
605 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
606 cli->cl_avail_grant += body->oa.o_grant;
607 /* waiters are woken in brw_interpret_oap */
608 spin_unlock(&cli->cl_loi_list_lock);
611 /* We assume that the reason this OSC got a short read is because it read
612 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
613 * via the LOV, and it _knows_ it's reading inside the file, it's just that
614 * this stripe never got written at or beyond this stripe offset yet. */
615 static void handle_short_read(int nob_read, obd_count page_count,
616 struct brw_page *pga)
620 /* skip bytes read OK */
621 while (nob_read > 0) {
622 LASSERT (page_count > 0);
624 if (pga->count > nob_read) {
625 /* EOF inside this page */
626 ptr = kmap(pga->pg) + (pga->off & ~PAGE_MASK);
627 memset(ptr + nob_read, 0, pga->count - nob_read);
634 nob_read -= pga->count;
639 /* zero remaining pages */
640 while (page_count-- > 0) {
641 ptr = kmap(pga->pg) + (pga->off & ~PAGE_MASK);
642 memset(ptr, 0, pga->count);
648 static int check_write_rcs(struct ptlrpc_request *request,
649 int requested_nob, int niocount,
650 obd_count page_count, struct brw_page *pga)
654 /* return error if any niobuf was in error */
655 remote_rcs = lustre_swab_repbuf(request, 1,
656 sizeof(*remote_rcs) * niocount, NULL);
657 if (remote_rcs == NULL) {
658 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
661 if (lustre_msg_swabbed(request->rq_repmsg))
662 for (i = 0; i < niocount; i++)
663 __swab32s(&remote_rcs[i]);
665 for (i = 0; i < niocount; i++) {
666 if (remote_rcs[i] < 0)
667 return(remote_rcs[i]);
669 if (remote_rcs[i] != 0) {
670 CERROR("rc[%d] invalid (%d) req %p\n",
671 i, remote_rcs[i], request);
676 if (request->rq_bulk->bd_nob_transferred != requested_nob) {
677 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
678 requested_nob, request->rq_bulk->bd_nob_transferred);
685 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
687 if (p1->flag != p2->flag) {
688 unsigned mask = ~OBD_BRW_FROM_GRANT;
690 /* warn if we try to combine flags that we don't know to be
692 if ((p1->flag & mask) != (p2->flag & mask))
693 CERROR("is it ok to have flags 0x%x and 0x%x in the "
694 "same brw?\n", p1->flag, p2->flag);
698 return (p1->off + p1->count == p2->off);
702 static obd_count cksum_pages(int nob, obd_count page_count,
703 struct brw_page *pga)
709 LASSERT (page_count > 0);
712 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
713 pga->count > nob ? nob : pga->count);
725 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
726 struct lov_stripe_md *lsm, obd_count page_count,
727 struct brw_page *pga, int *requested_nobp,
728 int *niocountp, struct ptlrpc_request **reqp)
730 struct ptlrpc_request *req;
731 struct ptlrpc_bulk_desc *desc;
732 struct client_obd *cli = &imp->imp_obd->u.cli;
733 struct ost_body *body;
734 struct obd_ioobj *ioobj;
735 struct niobuf_remote *niobuf;
744 opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
746 for (niocount = i = 1; i < page_count; i++)
747 if (!can_merge_pages(&pga[i - 1], &pga[i]))
750 size[0] = sizeof(*body);
751 size[1] = sizeof(*ioobj);
752 size[2] = niocount * sizeof(*niobuf);
754 req = ptlrpc_prep_req(imp, opc, 3, size, NULL);
758 if (opc == OST_WRITE)
759 desc = ptlrpc_prep_bulk_imp (req, page_count,
760 BULK_GET_SOURCE, OST_BULK_PORTAL);
762 desc = ptlrpc_prep_bulk_imp (req, page_count,
763 BULK_PUT_SINK, OST_BULK_PORTAL);
765 GOTO(out, rc = -ENOMEM);
766 /* NB request now owns desc and will free it when it gets freed */
768 body = lustre_msg_buf(req->rq_reqmsg, 0, sizeof(*body));
769 ioobj = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*ioobj));
770 niobuf = lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf));
772 memcpy(&body->oa, oa, sizeof(*oa));
774 obdo_to_ioobj(oa, ioobj);
775 ioobj->ioo_bufcnt = niocount;
777 LASSERT (page_count > 0);
778 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
779 struct brw_page *pg = &pga[i];
780 struct brw_page *pg_prev = pg - 1;
782 LASSERT(pg->count > 0);
783 LASSERT((pg->off & ~PAGE_MASK) + pg->count <= PAGE_SIZE);
784 LASSERTF(i == 0 || pg->off > pg_prev->off,
785 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
786 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
788 pg->pg, pg->pg->private, pg->pg->index, pg->off,
789 pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
792 ptlrpc_prep_bulk_page(desc, pg->pg, pg->off & ~PAGE_MASK,
794 requested_nob += pg->count;
796 if (i > 0 && can_merge_pages(pg_prev, pg)) {
798 niobuf->len += pg->count;
800 niobuf->offset = pg->off;
801 niobuf->len = pg->count;
802 niobuf->flags = pg->flag;
806 LASSERT((void *)(niobuf - niocount) ==
807 lustre_msg_buf(req->rq_reqmsg, 2, niocount * sizeof(*niobuf)));
808 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
809 spin_lock_irqsave(&req->rq_lock, flags);
810 req->rq_no_resend = 1;
811 spin_unlock_irqrestore(&req->rq_lock, flags);
813 /* size[0] still sizeof (*body) */
814 if (opc == OST_WRITE) {
816 body->oa.o_valid |= OBD_MD_FLCKSUM;
817 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
819 /* 1 RC per niobuf */
820 size[1] = sizeof(__u32) * niocount;
821 req->rq_replen = lustre_msg_size(2, size);
823 /* 1 RC for the whole I/O */
824 req->rq_replen = lustre_msg_size(1, size);
827 *niocountp = niocount;
828 *requested_nobp = requested_nob;
833 ptlrpc_req_finished (req);
837 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
838 int requested_nob, int niocount,
839 obd_count page_count, struct brw_page *pga,
842 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
843 struct ost_body *body;
849 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
851 CERROR ("Can't unpack body\n");
855 osc_update_grant(cli, body);
857 if (req->rq_reqmsg->opc == OST_WRITE) {
859 CERROR ("Unexpected +ve rc %d\n", rc);
862 LASSERT (req->rq_bulk->bd_nob == requested_nob);
864 RETURN(check_write_rcs(req, requested_nob, niocount,
868 if (rc > requested_nob) {
869 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
873 if (rc != req->rq_bulk->bd_nob_transferred) {
874 CERROR ("Unexpected rc %d (%d transferred)\n",
875 rc, req->rq_bulk->bd_nob_transferred);
879 if (rc < requested_nob)
880 handle_short_read(rc, page_count, pga);
882 memcpy(oa, &body->oa, sizeof(*oa));
885 if (oa->o_valid & OBD_MD_FLCKSUM) {
886 const struct ptlrpc_peer *peer =
887 &req->rq_import->imp_connection->c_peer;
888 static int cksum_counter;
889 obd_count server_cksum = oa->o_cksum;
890 obd_count cksum = cksum_pages(rc, page_count, pga);
891 char str[PTL_NALFMT_SIZE];
893 portals_nid2str(peer->peer_ni->pni_number, peer->peer_nid, str);
896 if (server_cksum != cksum) {
897 CERROR("Bad checksum: server %x, client %x, server NID "
898 LPX64" (%s)\n", server_cksum, cksum,
899 peer->peer_nid, str);
902 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
903 CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
904 cksum_counter, peer->peer_nid, str, cksum);
907 static int cksum_missed;
910 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
911 CERROR("Request checksum %u from "LPX64", no reply\n",
913 req->rq_import->imp_connection->c_peer.peer_nid);
919 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
920 struct lov_stripe_md *lsm,
921 obd_count page_count, struct brw_page *pga)
925 struct ptlrpc_request *request;
930 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
931 page_count, pga, &requested_nob, &niocount,
933 /* NB ^ sets rq_no_resend */
938 rc = ptlrpc_queue_wait(request);
940 if (rc == -ETIMEDOUT && request->rq_resend) {
941 DEBUG_REQ(D_HA, request, "BULK TIMEOUT");
942 ptlrpc_req_finished(request);
946 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
947 page_count, pga, rc);
949 ptlrpc_req_finished(request);
953 static int brw_interpret(struct ptlrpc_request *request,
954 struct osc_brw_async_args *aa, int rc)
956 struct obdo *oa = aa->aa_oa;
957 int requested_nob = aa->aa_requested_nob;
958 int niocount = aa->aa_nio_count;
959 obd_count page_count = aa->aa_page_count;
960 struct brw_page *pga = aa->aa_pga;
963 /* XXX bug 937 here */
964 if (rc == -ETIMEDOUT && request->rq_resend) {
965 DEBUG_REQ(D_HA, request, "BULK TIMEOUT");
966 LBUG(); /* re-send. later. */
970 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
971 page_count, pga, rc);
975 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
976 struct lov_stripe_md *lsm, obd_count page_count,
977 struct brw_page *pga, struct ptlrpc_request_set *set)
979 struct ptlrpc_request *request;
982 struct osc_brw_async_args *aa;
986 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
987 page_count, pga, &requested_nob, &nio_count,
989 /* NB ^ sets rq_no_resend */
992 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
993 aa = (struct osc_brw_async_args *)&request->rq_async_args;
995 aa->aa_requested_nob = requested_nob;
996 aa->aa_nio_count = nio_count;
997 aa->aa_page_count = page_count;
1000 request->rq_interpret_reply = brw_interpret;
1001 ptlrpc_set_add_req(set, request);
1007 #define min_t(type,x,y) \
1008 ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
1012 * ugh, we want disk allocation on the target to happen in offset order. we'll
1013 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1014 * fine for our small page arrays and doesn't require allocation. its an
1015 * insertion sort that swaps elements that are strides apart, shrinking the
1016 * stride down until its '1' and the array is sorted.
1018 static void sort_brw_pages(struct brw_page *array, int num)
1021 struct brw_page tmp;
1025 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1030 for (i = stride ; i < num ; i++) {
1033 while (j >= stride && array[j - stride].off > tmp.off) {
1034 array[j] = array[j - stride];
1039 } while (stride > 1);
1042 /* make sure we the regions we're passing to elan don't violate its '4
1043 * fragments' constraint. portal headers are a fragment, all full
1044 * PAGE_SIZE long pages count as 1 fragment, and each partial page
1045 * counts as a fragment. I think. see bug 934. */
1046 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1049 int saw_whole_frag = 0;
1052 for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1053 if (pg->count == PAGE_SIZE) {
1054 if (!saw_whole_frag) {
1065 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1066 struct lov_stripe_md *md, obd_count page_count,
1067 struct brw_page *pga, struct obd_trans_info *oti)
1071 if (cmd == OBD_BRW_CHECK) {
1072 /* The caller just wants to know if there's a chance that this
1073 * I/O can succeed */
1074 struct obd_import *imp = class_exp2cliimp(exp);
1076 if (imp == NULL || imp->imp_invalid)
1081 while (page_count) {
1082 obd_count pages_per_brw;
1085 if (page_count > OSC_BRW_MAX_IOV)
1086 pages_per_brw = OSC_BRW_MAX_IOV;
1088 pages_per_brw = page_count;
1090 sort_brw_pages(pga, pages_per_brw);
1091 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1093 rc = osc_brw_internal(cmd, exp, oa, md, pages_per_brw, pga);
1098 page_count -= pages_per_brw;
1099 pga += pages_per_brw;
1104 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1105 struct lov_stripe_md *md, obd_count page_count,
1106 struct brw_page *pga, struct ptlrpc_request_set *set,
1107 struct obd_trans_info *oti)
1111 if (cmd == OBD_BRW_CHECK) {
1112 /* The caller just wants to know if there's a chance that this
1113 * I/O can succeed */
1114 struct obd_import *imp = class_exp2cliimp(exp);
1116 if (imp == NULL || imp->imp_invalid)
1121 while (page_count) {
1122 obd_count pages_per_brw;
1125 if (page_count > OSC_BRW_MAX_IOV)
1126 pages_per_brw = OSC_BRW_MAX_IOV;
1128 pages_per_brw = page_count;
1130 sort_brw_pages(pga, pages_per_brw);
1131 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1133 rc = async_internal(cmd, exp, oa, md, pages_per_brw, pga, set);
1138 page_count -= pages_per_brw;
1139 pga += pages_per_brw;
1144 static void osc_check_rpcs(struct client_obd *cli);
1145 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1147 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1148 static void lop_update_pending(struct client_obd *cli,
1149 struct loi_oap_pages *lop, int cmd, int delta);
1151 /* this is called when a sync waiter receives an interruption. Its job is to
1152 * get the caller woken as soon as possible. If its page hasn't been put in an
1153 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1154 * desiring interruption which will forcefully complete the rpc once the rpc
1156 static void osc_occ_interrupted(struct oig_callback_context *occ)
1158 struct osc_async_page *oap;
1159 struct loi_oap_pages *lop;
1160 struct lov_oinfo *loi;
1163 /* XXX member_of() */
1164 oap = list_entry(occ, struct osc_async_page, oap_occ);
1166 spin_lock(&oap->oap_cli->cl_loi_list_lock);
1168 oap->oap_interrupted = 1;
1170 /* ok, it's been put in an rpc. */
1171 if (oap->oap_request != NULL) {
1172 ptlrpc_mark_interrupted(oap->oap_request);
1177 /* we don't get interruption callbacks until osc_trigger_sync_io()
1178 * has been called and put the sync oaps in the pending/urgent lists.*/
1179 if (!list_empty(&oap->oap_pending_item)) {
1180 list_del_init(&oap->oap_pending_item);
1181 if (oap->oap_async_flags & ASYNC_URGENT)
1182 list_del_init(&oap->oap_urgent_item);
1185 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1186 &loi->loi_write_lop : &loi->loi_read_lop;
1187 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1188 loi_list_maint(oap->oap_cli, oap->oap_loi);
1190 oig_complete_one(oap->oap_oig, &oap->oap_occ, 0);
1191 oap->oap_oig = NULL;
1195 spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1198 /* this must be called holding the loi list lock to give coverage to exit_cache,
1199 * async_flag maintenance, and oap_request */
1200 static void osc_complete_oap(struct client_obd *cli,
1201 struct osc_async_page *oap, int sent, int rc)
1203 osc_exit_cache(cli, oap, sent);
1204 oap->oap_async_flags = 0;
1205 oap->oap_interrupted = 0;
1207 if (oap->oap_request != NULL) {
1208 ptlrpc_req_finished(oap->oap_request);
1209 oap->oap_request = NULL;
1213 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1214 oap->oap_oig = NULL;
1219 oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1223 static int brw_interpret_oap(struct ptlrpc_request *request,
1224 struct osc_brw_async_args *aa, int rc)
1226 struct osc_async_page *oap;
1227 struct client_obd *cli;
1228 struct list_head *pos, *n;
1231 CDEBUG(D_INODE, "request %p aa %p\n", request, aa);
1233 rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1234 aa->aa_nio_count, aa->aa_page_count,
1238 /* in failout recovery we ignore writeback failure and want
1239 * to just tell llite to unlock the page and continue */
1240 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1243 spin_lock(&cli->cl_loi_list_lock);
1245 /* We need to decrement before osc_complete_oap->osc_wake_cache_waiters
1246 * is called so we know whether to go to sync BRWs or wait for more
1247 * RPCs to complete */
1248 cli->cl_brw_in_flight--;
1250 /* the caller may re-use the oap after the completion call so
1251 * we need to clean it up a little */
1252 list_for_each_safe(pos, n, &aa->aa_oaps) {
1253 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1255 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1256 //oap->oap_page, oap->oap_page->index, oap);
1258 list_del_init(&oap->oap_rpc_item);
1259 osc_complete_oap(cli, oap, 1, rc);
1262 osc_wake_cache_waiters(cli);
1263 osc_check_rpcs(cli);
1265 spin_unlock(&cli->cl_loi_list_lock);
1267 obdo_free(aa->aa_oa);
1268 OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1273 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1274 struct list_head *rpc_list,
1275 int page_count, int cmd)
1277 struct ptlrpc_request *req;
1278 struct brw_page *pga = NULL;
1279 int requested_nob, nio_count;
1280 struct osc_brw_async_args *aa;
1281 struct obdo *oa = NULL;
1282 struct obd_async_page_ops *ops = NULL;
1283 void *caller_data = NULL;
1284 struct list_head *pos;
1287 LASSERT(!list_empty(rpc_list));
1289 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1291 RETURN(ERR_PTR(-ENOMEM));
1295 GOTO(out, req = ERR_PTR(-ENOMEM));
1298 list_for_each(pos, rpc_list) {
1299 struct osc_async_page *oap;
1301 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1303 ops = oap->oap_caller_ops;
1304 caller_data = oap->oap_caller_data;
1306 pga[i].off = oap->oap_obj_off + oap->oap_page_off;
1307 pga[i].pg = oap->oap_page;
1308 pga[i].count = oap->oap_count;
1309 pga[i].flag = oap->oap_brw_flags;
1310 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1311 pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1315 /* always get the data for the obdo for the rpc */
1316 LASSERT(ops != NULL);
1317 ops->ap_fill_obdo(caller_data, cmd, oa);
1319 sort_brw_pages(pga, page_count);
1320 rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1321 pga, &requested_nob, &nio_count, &req);
1323 CERROR("prep_req failed: %d\n", rc);
1324 GOTO(out, req = ERR_PTR(rc));
1327 LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1328 aa = (struct osc_brw_async_args *)&req->rq_async_args;
1330 aa->aa_requested_nob = requested_nob;
1331 aa->aa_nio_count = nio_count;
1332 aa->aa_page_count = page_count;
1341 OBD_FREE(pga, sizeof(*pga) * page_count);
1346 static void lop_update_pending(struct client_obd *cli,
1347 struct loi_oap_pages *lop, int cmd, int delta)
1349 lop->lop_num_pending += delta;
1350 if (cmd == OBD_BRW_WRITE)
1351 cli->cl_pending_w_pages += delta;
1353 cli->cl_pending_r_pages += delta;
1356 /* the loi lock is held across this function but it's allowed to release
1357 * and reacquire it during its work */
1358 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1359 int cmd, struct loi_oap_pages *lop)
1361 struct ptlrpc_request *request;
1362 obd_count page_count = 0;
1363 struct list_head *tmp, *pos;
1364 struct osc_async_page *oap = NULL;
1365 struct osc_brw_async_args *aa;
1366 struct obd_async_page_ops *ops;
1367 LIST_HEAD(rpc_list);
1370 /* first we find the pages we're allowed to work with */
1371 list_for_each_safe(pos, tmp, &lop->lop_pending) {
1372 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1373 ops = oap->oap_caller_ops;
1375 LASSERT(oap->oap_magic == OAP_MAGIC);
1377 /* in llite being 'ready' equates to the page being locked
1378 * until completion unlocks it. commit_write submits a page
1379 * as not ready because its unlock will happen unconditionally
1380 * as the call returns. if we race with commit_write giving
1381 * us that page we dont' want to create a hole in the page
1382 * stream, so we stop and leave the rpc to be fired by
1383 * another dirtier or kupdated interval (the not ready page
1384 * will still be on the dirty list). we could call in
1385 * at the end of ll_file_write to process the queue again. */
1386 if (!(oap->oap_async_flags & ASYNC_READY)) {
1387 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1389 CDEBUG(D_INODE, "oap %p page %p returned %d "
1390 "instead of ready\n", oap,
1394 /* llite is telling us that the page is still
1395 * in commit_write and that we should try
1396 * and put it in an rpc again later. we
1397 * break out of the loop so we don't create
1398 * a hole in the sequence of pages in the rpc
1403 /* the io isn't needed.. tell the checks
1404 * below to complete the rpc with EINTR */
1405 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1406 oap->oap_count = -EINTR;
1409 oap->oap_async_flags |= ASYNC_READY;
1412 LASSERTF(0, "oap %p page %p returned %d "
1413 "from make_ready\n", oap,
1421 /* take the page out of our book-keeping */
1422 list_del_init(&oap->oap_pending_item);
1423 lop_update_pending(cli, lop, cmd, -1);
1424 if (!list_empty(&oap->oap_urgent_item))
1425 list_del_init(&oap->oap_urgent_item);
1427 /* ask the caller for the size of the io as the rpc leaves. */
1428 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1430 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1431 if (oap->oap_count <= 0) {
1432 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1434 osc_complete_oap(cli, oap, 0, oap->oap_count);
1438 /* now put the page back in our accounting */
1439 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1440 if (++page_count >= cli->cl_max_pages_per_rpc)
1444 osc_wake_cache_waiters(cli);
1446 if (page_count == 0)
1449 loi_list_maint(cli, loi);
1450 spin_unlock(&cli->cl_loi_list_lock);
1452 request = osc_build_req(cli, &rpc_list, page_count, cmd);
1453 if (IS_ERR(request)) {
1454 /* this should happen rarely and is pretty bad, it makes the
1455 * pending list not follow the dirty order */
1456 spin_lock(&cli->cl_loi_list_lock);
1457 list_for_each_safe(pos, tmp, &rpc_list) {
1458 oap = list_entry(pos, struct osc_async_page,
1460 list_del_init(&oap->oap_rpc_item);
1462 /* queued sync pages can be torn down while the pages
1463 * were between the pending list and the rpc */
1464 if (oap->oap_interrupted) {
1465 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1466 osc_complete_oap(cli, oap, 0, oap->oap_count);
1470 /* put the page back in the loi/lop lists */
1471 list_add_tail(&oap->oap_pending_item,
1473 lop_update_pending(cli, lop, cmd, 1);
1474 if (oap->oap_async_flags & ASYNC_URGENT)
1475 list_add(&oap->oap_urgent_item,
1478 loi_list_maint(cli, loi);
1479 RETURN(PTR_ERR(request));
1482 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1483 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1484 INIT_LIST_HEAD(&aa->aa_oaps);
1485 list_splice(&rpc_list, &aa->aa_oaps);
1486 INIT_LIST_HEAD(&rpc_list);
1489 if (cmd == OBD_BRW_READ) {
1490 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1491 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_brw_in_flight);
1493 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1494 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1495 cli->cl_brw_in_flight);
1499 spin_lock(&cli->cl_loi_list_lock);
1501 cli->cl_brw_in_flight++;
1502 /* queued sync pages can be torn down while the pages
1503 * were between the pending list and the rpc */
1504 list_for_each(pos, &aa->aa_oaps) {
1505 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1506 if (oap->oap_interrupted) {
1507 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1509 ptlrpc_mark_interrupted(request);
1514 CDEBUG(D_INODE, "req %p: %d pages, aa %p. now %d in flight\n", request,
1515 page_count, aa, cli->cl_brw_in_flight);
1517 oap->oap_request = ptlrpc_request_addref(request);
1518 request->rq_interpret_reply = brw_interpret_oap;
1519 ptlrpcd_add_req(request);
1523 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1529 if (lop->lop_num_pending == 0)
1532 /* stream rpcs in queue order as long as as there is an urgent page
1533 * queued. this is our cheap solution for good batching in the case
1534 * where writepage marks some random page in the middle of the file as
1535 * urgent because of, say, memory pressure */
1536 if (!list_empty(&lop->lop_urgent))
1539 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1540 optimal = cli->cl_max_pages_per_rpc;
1541 if (cmd == OBD_BRW_WRITE) {
1542 /* trigger a write rpc stream as long as there are dirtiers
1543 * waiting for space. as they're waiting, they're not going to
1544 * create more pages to coallesce with what's waiting.. */
1545 if (!list_empty(&cli->cl_cache_waiters))
1548 /* *2 to avoid triggering rpcs that would want to include pages
1549 * that are being queued but which can't be made ready until
1550 * the queuer finishes with the page. this is a wart for
1551 * llite::commit_write() */
1554 if (lop->lop_num_pending >= optimal)
1560 static void on_list(struct list_head *item, struct list_head *list,
1563 if (list_empty(item) && should_be_on)
1564 list_add_tail(item, list);
1565 else if (!list_empty(item) && !should_be_on)
1566 list_del_init(item);
1569 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1570 * can find pages to build into rpcs quickly */
1571 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1573 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1574 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1575 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1577 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1578 loi->loi_write_lop.lop_num_pending);
1581 #define LOI_DEBUG(LOI, STR, args...) \
1582 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
1583 !list_empty(&(LOI)->loi_cli_item), \
1584 (LOI)->loi_write_lop.lop_num_pending, \
1585 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
1586 (LOI)->loi_read_lop.lop_num_pending, \
1587 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
1590 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1593 /* first return all objects which we already know to have
1594 * pages ready to be stuffed into rpcs */
1595 if (!list_empty(&cli->cl_loi_ready_list))
1596 RETURN(list_entry(cli->cl_loi_ready_list.next,
1597 struct lov_oinfo, loi_cli_item));
1599 /* then if we have cache waiters, return all objects with queued
1600 * writes. This is especially important when many small files
1601 * have filled up the cache and not been fired into rpcs because
1602 * they don't pass the nr_pending/object threshhold */
1603 if (!list_empty(&cli->cl_cache_waiters) &&
1604 !list_empty(&cli->cl_loi_write_list))
1605 RETURN(list_entry(cli->cl_loi_write_list.next,
1606 struct lov_oinfo, loi_write_item));
1610 /* called with the loi list lock held */
1611 static void osc_check_rpcs(struct client_obd *cli)
1613 struct lov_oinfo *loi;
1614 int rc = 0, race_counter = 0;
1617 while ((loi = osc_next_loi(cli)) != NULL) {
1618 LOI_DEBUG(loi, "%d in flight\n", cli->cl_brw_in_flight);
1620 if (cli->cl_brw_in_flight >= cli->cl_max_rpcs_in_flight)
1623 /* attempt some read/write balancing by alternating between
1624 * reads and writes in an object. The makes_rpc checks here
1625 * would be redundant if we were getting read/write work items
1626 * instead of objects. we don't want send_oap_rpc to drain a
1627 * partial read pending queue when we're given this object to
1628 * do io on writes while there are cache waiters */
1629 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1630 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1631 &loi->loi_write_lop);
1639 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1640 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1641 &loi->loi_read_lop);
1650 /* attempt some inter-object balancing by issueing rpcs
1651 * for each object in turn */
1652 if (!list_empty(&loi->loi_cli_item))
1653 list_del_init(&loi->loi_cli_item);
1654 if (!list_empty(&loi->loi_write_item))
1655 list_del_init(&loi->loi_write_item);
1657 loi_list_maint(cli, loi);
1659 /* send_oap_rpc fails with 0 when make_ready tells it to
1660 * back off. llite's make_ready does this when it tries
1661 * to lock a page queued for write that is already locked.
1662 * we want to try sending rpcs from many objects, but we
1663 * don't want to spin failing with 0. */
1664 if (race_counter == 10)
1670 /* we're trying to queue a page in the osc so we're subject to the
1671 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1672 * If the osc's queued pages are already at that limit, then we want to sleep
1673 * until there is space in the osc's queue for us. We also may be waiting for
1674 * write credits from the OST if there are RPCs in flight that may return some
1675 * before we fall back to sync writes.
1677 * We need this know our allocation was granted in the presence of signals */
1678 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1682 spin_lock(&cli->cl_loi_list_lock);
1683 rc = list_empty(&ocw->ocw_entry) || cli->cl_brw_in_flight == 0;
1684 spin_unlock(&cli->cl_loi_list_lock);
1688 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1689 * grant or cache space. */
1690 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1691 struct osc_async_page *oap)
1693 struct osc_cache_waiter ocw;
1694 struct l_wait_info lwi = { 0 };
1696 CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1697 cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1698 cli->cl_avail_grant);
1700 if (cli->cl_dirty_max < PAGE_SIZE)
1703 /* Hopefully normal case - cache space and write credits available */
1704 if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1705 cli->cl_avail_grant >= PAGE_SIZE) {
1706 /* account for ourselves */
1707 osc_consume_write_grant(cli, oap);
1711 /* Make sure that there are write rpcs in flight to wait for. This
1712 * is a little silly as this object may not have any pending but
1713 * other objects sure might. */
1714 if (cli->cl_brw_in_flight) {
1715 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1716 init_waitqueue_head(&ocw.ocw_waitq);
1720 loi_list_maint(cli, loi);
1721 osc_check_rpcs(cli);
1722 spin_unlock(&cli->cl_loi_list_lock);
1724 CDEBUG(0, "sleeping for cache space\n");
1725 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1727 spin_lock(&cli->cl_loi_list_lock);
1728 if (!list_empty(&ocw.ocw_entry)) {
1729 list_del(&ocw.ocw_entry);
1738 /* the companion to enter_cache, called when an oap is no longer part of the
1739 * dirty accounting.. so writeback completes or truncate happens before writing
1740 * starts. must be called with the loi lock held. */
1741 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1746 if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1751 oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
1752 cli->cl_dirty -= PAGE_SIZE;
1754 cli->cl_lost_grant += PAGE_SIZE;
1755 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
1756 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
1762 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
1763 struct lov_oinfo *loi, struct page *page,
1764 obd_off offset, struct obd_async_page_ops *ops,
1765 void *data, void **res)
1767 struct osc_async_page *oap;
1770 OBD_ALLOC(oap, sizeof(*oap));
1774 oap->oap_magic = OAP_MAGIC;
1775 oap->oap_cli = &exp->exp_obd->u.cli;
1778 oap->oap_caller_ops = ops;
1779 oap->oap_caller_data = data;
1781 oap->oap_page = page;
1782 oap->oap_obj_off = offset;
1784 INIT_LIST_HEAD(&oap->oap_pending_item);
1785 INIT_LIST_HEAD(&oap->oap_urgent_item);
1786 INIT_LIST_HEAD(&oap->oap_rpc_item);
1788 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
1790 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
1795 struct osc_async_page *oap_from_cookie(void *cookie)
1797 struct osc_async_page *oap = cookie;
1798 if (oap->oap_magic != OAP_MAGIC)
1799 return ERR_PTR(-EINVAL);
1803 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1804 struct lov_oinfo *loi, void *cookie,
1805 int cmd, obd_off off, int count,
1806 obd_flag brw_flags, enum async_flags async_flags)
1808 struct client_obd *cli = &exp->exp_obd->u.cli;
1809 struct osc_async_page *oap;
1810 struct loi_oap_pages *lop;
1814 oap = oap_from_cookie(cookie);
1816 RETURN(PTR_ERR(oap));
1818 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1821 if (!list_empty(&oap->oap_pending_item) ||
1822 !list_empty(&oap->oap_urgent_item) ||
1823 !list_empty(&oap->oap_rpc_item))
1827 loi = &lsm->lsm_oinfo[0];
1829 spin_lock(&cli->cl_loi_list_lock);
1832 oap->oap_async_flags = async_flags;
1833 oap->oap_page_off = off;
1834 oap->oap_count = count;
1835 oap->oap_brw_flags = brw_flags;
1837 if (cmd == OBD_BRW_WRITE) {
1838 rc = osc_enter_cache(cli, loi, oap);
1840 spin_unlock(&cli->cl_loi_list_lock);
1843 lop = &loi->loi_write_lop;
1845 lop = &loi->loi_read_lop;
1848 if (oap->oap_async_flags & ASYNC_URGENT)
1849 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1850 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1851 lop_update_pending(cli, lop, cmd, 1);
1853 loi_list_maint(cli, loi);
1855 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
1858 osc_check_rpcs(cli);
1859 spin_unlock(&cli->cl_loi_list_lock);
1864 /* aka (~was & now & flag), but this is more clear :) */
1865 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
1867 static int osc_set_async_flags(struct obd_export *exp,
1868 struct lov_stripe_md *lsm,
1869 struct lov_oinfo *loi, void *cookie,
1870 obd_flag async_flags)
1872 struct client_obd *cli = &exp->exp_obd->u.cli;
1873 struct loi_oap_pages *lop;
1874 struct osc_async_page *oap;
1878 oap = oap_from_cookie(cookie);
1880 RETURN(PTR_ERR(oap));
1882 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1886 loi = &lsm->lsm_oinfo[0];
1888 if (oap->oap_cmd == OBD_BRW_WRITE) {
1889 lop = &loi->loi_write_lop;
1891 lop = &loi->loi_read_lop;
1894 spin_lock(&cli->cl_loi_list_lock);
1896 if (list_empty(&oap->oap_pending_item))
1897 GOTO(out, rc = -EINVAL);
1899 if ((oap->oap_async_flags & async_flags) == async_flags)
1902 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
1903 oap->oap_async_flags |= ASYNC_READY;
1905 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
1906 if (list_empty(&oap->oap_rpc_item)) {
1907 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1908 loi_list_maint(cli, loi);
1912 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
1913 oap->oap_async_flags);
1915 osc_check_rpcs(cli);
1916 spin_unlock(&cli->cl_loi_list_lock);
1920 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
1921 struct lov_oinfo *loi,
1922 struct obd_io_group *oig, void *cookie,
1923 int cmd, obd_off off, int count,
1925 obd_flag async_flags)
1927 struct client_obd *cli = &exp->exp_obd->u.cli;
1928 struct osc_async_page *oap;
1929 struct loi_oap_pages *lop;
1932 oap = oap_from_cookie(cookie);
1934 RETURN(PTR_ERR(oap));
1936 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1939 if (!list_empty(&oap->oap_pending_item) ||
1940 !list_empty(&oap->oap_urgent_item) ||
1941 !list_empty(&oap->oap_rpc_item))
1945 loi = &lsm->lsm_oinfo[0];
1947 spin_lock(&cli->cl_loi_list_lock);
1950 oap->oap_page_off = off;
1951 oap->oap_count = count;
1952 oap->oap_brw_flags = brw_flags;
1953 oap->oap_async_flags = async_flags;
1955 if (cmd == OBD_BRW_WRITE)
1956 lop = &loi->loi_write_lop;
1958 lop = &loi->loi_read_lop;
1960 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
1961 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
1963 oig_add_one(oig, &oap->oap_occ);
1966 LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
1968 spin_unlock(&cli->cl_loi_list_lock);
1973 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
1974 struct loi_oap_pages *lop, int cmd)
1976 struct list_head *pos, *tmp;
1977 struct osc_async_page *oap;
1979 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
1980 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1981 list_del(&oap->oap_pending_item);
1982 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
1983 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
1984 lop_update_pending(cli, lop, cmd, 1);
1986 loi_list_maint(cli, loi);
1989 static int osc_trigger_group_io(struct obd_export *exp,
1990 struct lov_stripe_md *lsm,
1991 struct lov_oinfo *loi,
1992 struct obd_io_group *oig)
1994 struct client_obd *cli = &exp->exp_obd->u.cli;
1998 loi = &lsm->lsm_oinfo[0];
2000 spin_lock(&cli->cl_loi_list_lock);
2002 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2003 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2005 osc_check_rpcs(cli);
2006 spin_unlock(&cli->cl_loi_list_lock);
2011 static int osc_teardown_async_page(struct obd_export *exp,
2012 struct lov_stripe_md *lsm,
2013 struct lov_oinfo *loi, void *cookie)
2015 struct client_obd *cli = &exp->exp_obd->u.cli;
2016 struct loi_oap_pages *lop;
2017 struct osc_async_page *oap;
2021 oap = oap_from_cookie(cookie);
2023 RETURN(PTR_ERR(oap));
2026 loi = &lsm->lsm_oinfo[0];
2028 if (oap->oap_cmd == OBD_BRW_WRITE) {
2029 lop = &loi->loi_write_lop;
2031 lop = &loi->loi_read_lop;
2034 spin_lock(&cli->cl_loi_list_lock);
2036 if (!list_empty(&oap->oap_rpc_item))
2037 GOTO(out, rc = -EBUSY);
2039 osc_exit_cache(cli, oap, 0);
2040 osc_wake_cache_waiters(cli);
2042 if (!list_empty(&oap->oap_urgent_item)) {
2043 list_del_init(&oap->oap_urgent_item);
2044 oap->oap_async_flags &= ~ASYNC_URGENT;
2046 if (!list_empty(&oap->oap_pending_item)) {
2047 list_del_init(&oap->oap_pending_item);
2048 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2050 loi_list_maint(cli, loi);
2052 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2054 spin_unlock(&cli->cl_loi_list_lock);
2056 OBD_FREE(oap, sizeof(*oap));
2061 /* Note: caller will lock/unlock, and set uptodate on the pages */
2062 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2063 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2064 struct lov_stripe_md *lsm, obd_count page_count,
2065 struct brw_page *pga)
2067 struct ptlrpc_request *request = NULL;
2068 struct ost_body *body;
2069 struct niobuf_remote *nioptr;
2070 struct obd_ioobj *iooptr;
2071 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2075 /* XXX does not handle 'new' brw protocol */
2077 size[1] = sizeof(struct obd_ioobj);
2078 size[2] = page_count * sizeof(*nioptr);
2080 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_READ, 3,
2085 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2086 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2087 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2088 sizeof(*nioptr) * page_count);
2090 memcpy(&body->oa, oa, sizeof(body->oa));
2092 obdo_to_ioobj(oa, iooptr);
2093 iooptr->ioo_bufcnt = page_count;
2095 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2096 LASSERT(PageLocked(pga[mapped].pg));
2097 LASSERT(mapped == 0 || pga[mapped].off > pga[mapped - 1].off);
2099 nioptr->offset = pga[mapped].off;
2100 nioptr->len = pga[mapped].count;
2101 nioptr->flags = pga[mapped].flag;
2104 size[1] = page_count * sizeof(*nioptr);
2105 request->rq_replen = lustre_msg_size(2, size);
2107 rc = ptlrpc_queue_wait(request);
2111 body = lustre_swab_repbuf(request, 0, sizeof(*body),
2112 lustre_swab_ost_body);
2114 CERROR("Can't unpack body\n");
2115 GOTO(out_req, rc = -EPROTO);
2118 memcpy(oa, &body->oa, sizeof(*oa));
2120 swab = lustre_msg_swabbed(request->rq_repmsg);
2121 LASSERT_REPSWAB(request, 1);
2122 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2124 /* nioptr missing or short */
2125 GOTO(out_req, rc = -EPROTO);
2129 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2130 struct page *page = pga[mapped].pg;
2131 struct buffer_head *bh;
2135 lustre_swab_niobuf_remote (nioptr);
2137 /* got san device associated */
2138 LASSERT(exp->exp_obd != NULL);
2139 dev = exp->exp_obd->u.cli.cl_sandev;
2142 if (!nioptr->offset) {
2143 CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2144 page->mapping->host->i_ino,
2146 memset(page_address(page), 0, PAGE_SIZE);
2150 if (!page->buffers) {
2151 create_empty_buffers(page, dev, PAGE_SIZE);
2154 clear_bit(BH_New, &bh->b_state);
2155 set_bit(BH_Mapped, &bh->b_state);
2156 bh->b_blocknr = (unsigned long)nioptr->offset;
2158 clear_bit(BH_Uptodate, &bh->b_state);
2160 ll_rw_block(READ, 1, &bh);
2164 /* if buffer already existed, it must be the
2165 * one we mapped before, check it */
2166 LASSERT(!test_bit(BH_New, &bh->b_state));
2167 LASSERT(test_bit(BH_Mapped, &bh->b_state));
2168 LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2170 /* wait it's io completion */
2171 if (test_bit(BH_Lock, &bh->b_state))
2174 if (!test_bit(BH_Uptodate, &bh->b_state))
2175 ll_rw_block(READ, 1, &bh);
2179 /* must do syncronous write here */
2181 if (!buffer_uptodate(bh)) {
2189 ptlrpc_req_finished(request);
2193 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2194 struct lov_stripe_md *lsm, obd_count page_count,
2195 struct brw_page *pga)
2197 struct ptlrpc_request *request = NULL;
2198 struct ost_body *body;
2199 struct niobuf_remote *nioptr;
2200 struct obd_ioobj *iooptr;
2201 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2205 size[1] = sizeof(struct obd_ioobj);
2206 size[2] = page_count * sizeof(*nioptr);
2208 request = ptlrpc_prep_req(class_exp2cliimp(exp), OST_SAN_WRITE,
2213 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2214 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2215 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2216 sizeof (*nioptr) * page_count);
2218 memcpy(&body->oa, oa, sizeof(body->oa));
2220 obdo_to_ioobj(oa, iooptr);
2221 iooptr->ioo_bufcnt = page_count;
2224 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2225 LASSERT(PageLocked(pga[mapped].pg));
2226 LASSERT(mapped == 0 || pga[mapped].off > pga[mapped - 1].off);
2228 nioptr->offset = pga[mapped].off;
2229 nioptr->len = pga[mapped].count;
2230 nioptr->flags = pga[mapped].flag;
2233 size[1] = page_count * sizeof(*nioptr);
2234 request->rq_replen = lustre_msg_size(2, size);
2236 rc = ptlrpc_queue_wait(request);
2240 swab = lustre_msg_swabbed (request->rq_repmsg);
2241 LASSERT_REPSWAB (request, 1);
2242 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2244 CERROR("absent/short niobuf array\n");
2245 GOTO(out_req, rc = -EPROTO);
2249 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2250 struct page *page = pga[mapped].pg;
2251 struct buffer_head *bh;
2255 lustre_swab_niobuf_remote (nioptr);
2257 /* got san device associated */
2258 LASSERT(exp->exp_obd != NULL);
2259 dev = exp->exp_obd->u.cli.cl_sandev;
2261 if (!page->buffers) {
2262 create_empty_buffers(page, dev, PAGE_SIZE);
2265 LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2266 LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2267 LASSERT(page->buffers->b_blocknr ==
2268 (unsigned long)nioptr->offset);
2274 /* if buffer locked, wait it's io completion */
2275 if (test_bit(BH_Lock, &bh->b_state))
2278 clear_bit(BH_New, &bh->b_state);
2279 set_bit(BH_Mapped, &bh->b_state);
2281 /* override the block nr */
2282 bh->b_blocknr = (unsigned long)nioptr->offset;
2284 /* we are about to write it, so set it
2286 * page lock should garentee no race condition here */
2287 set_bit(BH_Uptodate, &bh->b_state);
2288 set_bit(BH_Dirty, &bh->b_state);
2290 ll_rw_block(WRITE, 1, &bh);
2292 /* must do syncronous write here */
2294 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2302 ptlrpc_req_finished(request);
2306 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2307 struct lov_stripe_md *lsm, obd_count page_count,
2308 struct brw_page *pga, struct obd_trans_info *oti)
2312 while (page_count) {
2313 obd_count pages_per_brw;
2316 if (page_count > OSC_BRW_MAX_IOV)
2317 pages_per_brw = OSC_BRW_MAX_IOV;
2319 pages_per_brw = page_count;
2321 if (cmd & OBD_BRW_WRITE)
2322 rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2324 rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2329 page_count -= pages_per_brw;
2330 pga += pages_per_brw;
2337 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2339 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2341 LASSERT(lock != NULL);
2342 l_lock(&lock->l_resource->lr_namespace->ns_lock);
2344 if (lock->l_ast_data && lock->l_ast_data != data) {
2345 struct inode *new_inode = data;
2346 struct inode *old_inode = lock->l_ast_data;
2347 LASSERTF(old_inode->i_state & I_FREEING,
2348 "Found existing inode %p/%lu/%u state %lu in lock: "
2349 "setting data to %p/%lu/%u\n", old_inode,
2350 old_inode->i_ino, old_inode->i_generation,
2352 new_inode, new_inode->i_ino, new_inode->i_generation);
2355 lock->l_ast_data = data;
2356 l_unlock(&lock->l_resource->lr_namespace->ns_lock);
2357 LDLM_LOCK_PUT(lock);
2360 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2361 ldlm_iterator_t replace, void *data)
2363 struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2364 struct obd_device *obd = class_exp2obd(exp);
2366 ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2370 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2371 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2372 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2373 void *data, __u32 lvb_len, void *lvb_swabber,
2374 struct lustre_handle *lockh)
2376 struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2377 struct obd_device *obd = exp->exp_obd;
2382 /* Filesystem lock extents are extended to page boundaries so that
2383 * dealing with the page cache is a little smoother. */
2384 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2385 policy->l_extent.end |= ~PAGE_MASK;
2387 /* Next, search for already existing extent locks that will cover us */
2388 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2391 osc_set_data_with_check(lockh, data);
2392 if (*flags & LDLM_FL_HAS_INTENT) {
2393 /* I would like to be able to ASSERT here that rss <=
2394 * kms, but I can't, for reasons which are explained in
2397 /* We already have a lock, and it's referenced */
2401 /* If we're trying to read, we also search for an existing PW lock. The
2402 * VFS and page cache already protect us locally, so lots of readers/
2403 * writers can share a single PW lock.
2405 * There are problems with conversion deadlocks, so instead of
2406 * converting a read lock to a write lock, we'll just enqueue a new
2409 * At some point we should cancel the read lock instead of making them
2410 * send us a blocking callback, but there are problems with canceling
2411 * locks out from other users right now, too. */
2413 if (mode == LCK_PR) {
2414 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2415 policy, LCK_PW, lockh);
2417 /* FIXME: This is not incredibly elegant, but it might
2418 * be more elegant than adding another parameter to
2419 * lock_match. I want a second opinion. */
2420 ldlm_lock_addref(lockh, LCK_PR);
2421 ldlm_lock_decref(lockh, LCK_PW);
2422 osc_set_data_with_check(lockh, data);
2427 rc = ldlm_cli_enqueue(exp, NULL, obd->obd_namespace, res_id, type,
2428 policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2429 &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2431 if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2432 CDEBUG(D_INODE, "received kms == "LPU64"\n", lvb.lvb_size);
2433 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2439 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2440 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2441 int *flags, void *data, struct lustre_handle *lockh)
2443 struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2444 struct obd_device *obd = exp->exp_obd;
2448 OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2450 /* Filesystem lock extents are extended to page boundaries so that
2451 * dealing with the page cache is a little smoother */
2452 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2453 policy->l_extent.end |= ~PAGE_MASK;
2455 /* Next, search for already existing extent locks that will cover us */
2456 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2457 policy, mode, lockh);
2459 osc_set_data_with_check(lockh, data);
2462 /* If we're trying to read, we also search for an existing PW lock. The
2463 * VFS and page cache already protect us locally, so lots of readers/
2464 * writers can share a single PW lock. */
2465 if (mode == LCK_PR) {
2466 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2467 policy, LCK_PW, lockh);
2468 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2469 /* FIXME: This is not incredibly elegant, but it might
2470 * be more elegant than adding another parameter to
2471 * lock_match. I want a second opinion. */
2472 osc_set_data_with_check(lockh, data);
2473 ldlm_lock_addref(lockh, LCK_PR);
2474 ldlm_lock_decref(lockh, LCK_PW);
2480 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2481 __u32 mode, struct lustre_handle *lockh)
2485 ldlm_lock_decref(lockh, mode);
2490 static int osc_cancel_unused(struct obd_export *exp,
2491 struct lov_stripe_md *lsm, int flags, void *opaque)
2493 struct obd_device *obd = class_exp2obd(exp);
2494 struct ldlm_res_id res_id = { .name = {lsm->lsm_object_id} };
2496 return ldlm_cli_cancel_unused(obd->obd_namespace, &res_id, flags,
2500 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2501 unsigned long max_age)
2503 struct obd_statfs *msfs;
2504 struct ptlrpc_request *request;
2505 int rc, size = sizeof(*osfs);
2508 /* We could possibly pass max_age in the request (as an absolute
2509 * timestamp or a "seconds.usec ago") so the target can avoid doing
2510 * extra calls into the filesystem if that isn't necessary (e.g.
2511 * during mount that would help a bit). Having relative timestamps
2512 * is not so great if request processing is slow, while absolute
2513 * timestamps are not ideal because they need time synchronization. */
2514 request = ptlrpc_prep_req(obd->u.cli.cl_import, OST_STATFS,0,NULL,NULL);
2518 request->rq_replen = lustre_msg_size(1, &size);
2519 request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2521 rc = ptlrpc_queue_wait(request);
2525 msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2526 lustre_swab_obd_statfs);
2528 CERROR("Can't unpack obd_statfs\n");
2529 GOTO(out, rc = -EPROTO);
2532 memcpy(osfs, msfs, sizeof(*osfs));
2536 ptlrpc_req_finished(request);
2540 /* Retrieve object striping information.
2542 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2543 * the maximum number of OST indices which will fit in the user buffer.
2544 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2546 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2548 struct lov_user_md lum, *lumk;
2555 rc = copy_from_user(&lum, lump, sizeof(lum));
2559 if (lum.lmm_magic != LOV_USER_MAGIC)
2562 if (lum.lmm_stripe_count > 0) {
2563 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2564 OBD_ALLOC(lumk, lum_size);
2568 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2570 lum_size = sizeof(lum);
2574 lumk->lmm_object_id = lsm->lsm_object_id;
2575 lumk->lmm_stripe_count = 1;
2577 if (copy_to_user(lump, lumk, lum_size))
2581 OBD_FREE(lumk, lum_size);
2586 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2587 void *karg, void *uarg)
2589 struct obd_device *obd = exp->exp_obd;
2590 struct obd_ioctl_data *data = karg;
2597 case OBD_IOC_LOV_GET_CONFIG: {
2599 struct lov_desc *desc;
2600 struct obd_uuid uuid;
2604 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2605 GOTO(out, err = -EINVAL);
2607 data = (struct obd_ioctl_data *)buf;
2609 if (sizeof(*desc) > data->ioc_inllen1) {
2611 GOTO(out, err = -EINVAL);
2614 if (data->ioc_inllen2 < sizeof(uuid)) {
2616 GOTO(out, err = -EINVAL);
2619 desc = (struct lov_desc *)data->ioc_inlbuf1;
2620 desc->ld_tgt_count = 1;
2621 desc->ld_active_tgt_count = 1;
2622 desc->ld_default_stripe_count = 1;
2623 desc->ld_default_stripe_size = 0;
2624 desc->ld_default_stripe_offset = 0;
2625 desc->ld_pattern = 0;
2626 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2628 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2630 err = copy_to_user((void *)uarg, buf, len);
2633 obd_ioctl_freedata(buf, len);
2636 case LL_IOC_LOV_SETSTRIPE:
2637 err = obd_alloc_memmd(exp, karg);
2641 case LL_IOC_LOV_GETSTRIPE:
2642 err = osc_getstripe(karg, uarg);
2644 case OBD_IOC_CLIENT_RECOVER:
2645 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2650 case IOC_OSC_SET_ACTIVE:
2651 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2655 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
2656 GOTO(out, err = -ENOTTY);
2663 static int osc_get_info(struct obd_export *exp, obd_count keylen,
2664 void *key, __u32 *vallen, void *val)
2667 if (!vallen || !val)
2670 if (keylen > strlen("lock_to_stripe") &&
2671 strcmp(key, "lock_to_stripe") == 0) {
2672 __u32 *stripe = val;
2673 *vallen = sizeof(*stripe);
2676 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
2677 struct ptlrpc_request *req;
2679 char *bufs[1] = {key};
2681 req = ptlrpc_prep_req(class_exp2cliimp(exp), OST_GET_INFO, 1,
2686 req->rq_replen = lustre_msg_size(1, vallen);
2687 rc = ptlrpc_queue_wait(req);
2691 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
2692 lustre_swab_ost_last_id);
2693 if (reply == NULL) {
2694 CERROR("Can't unpack OST last ID\n");
2695 GOTO(out, rc = -EPROTO);
2697 *((obd_id *)val) = *reply;
2699 ptlrpc_req_finished(req);
2705 static int osc_set_info(struct obd_export *exp, obd_count keylen,
2706 void *key, obd_count vallen, void *val)
2708 struct ptlrpc_request *req;
2709 struct obd_device *obd = exp->exp_obd;
2710 struct obd_import *imp = class_exp2cliimp(exp);
2711 struct llog_ctxt *ctxt;
2712 int rc, size = keylen;
2713 char *bufs[1] = {key};
2716 if (keylen == strlen("next_id") &&
2717 memcmp(key, "next_id", strlen("next_id")) == 0) {
2718 if (vallen != sizeof(obd_id))
2720 obd->u.cli.cl_oscc.oscc_next_id = *((obd_id*)val) + 1;
2721 CDEBUG(D_INODE, "%s: set oscc_next_id = "LPU64"\n",
2722 exp->exp_obd->obd_name,
2723 obd->u.cli.cl_oscc.oscc_next_id);
2728 if (keylen == strlen("growth_count") &&
2729 memcmp(key, "growth_count", strlen("growth_count")) == 0) {
2730 if (vallen != sizeof(int))
2732 obd->u.cli.cl_oscc.oscc_grow_count = *((int*)val);
2736 if (keylen == strlen("unlinked") &&
2737 memcmp(key, "unlinked", keylen) == 0) {
2738 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2739 spin_lock(&oscc->oscc_lock);
2740 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
2741 spin_unlock(&oscc->oscc_lock);
2746 if (keylen == strlen("initial_recov") &&
2747 memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
2748 struct obd_import *imp = exp->exp_obd->u.cli.cl_import;
2749 if (vallen != sizeof(int))
2751 imp->imp_initial_recov = *(int *)val;
2752 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
2753 exp->exp_obd->obd_name,
2754 imp->imp_initial_recov);
2758 if (keylen < strlen("mds_conn") ||
2759 memcmp(key, "mds_conn", strlen("mds_conn")) != 0)
2763 req = ptlrpc_prep_req(imp, OST_SET_INFO, 1, &size, bufs);
2767 req->rq_replen = lustre_msg_size(0, NULL);
2768 rc = ptlrpc_queue_wait(req);
2769 ptlrpc_req_finished(req);
2771 ctxt = llog_get_context(exp->exp_obd, LLOG_UNLINK_ORIG_CTXT);
2773 rc = llog_initiator_connect(ctxt);
2778 imp->imp_server_timeout = 1;
2779 CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
2780 imp->imp_pingable = 1;
2786 static struct llog_operations osc_size_repl_logops = {
2787 lop_cancel: llog_obd_repl_cancel
2790 static struct llog_operations osc_unlink_orig_logops;
2791 static int osc_llog_init(struct obd_device *obd, struct obd_device *tgt,
2792 int count, struct llog_catid *catid)
2797 osc_unlink_orig_logops = llog_lvfs_ops;
2798 osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
2799 osc_unlink_orig_logops.lop_cleanup = llog_obd_origin_cleanup;
2800 osc_unlink_orig_logops.lop_add = llog_obd_origin_add;
2801 osc_unlink_orig_logops.lop_connect = llog_origin_connect;
2803 rc = llog_setup(obd, LLOG_UNLINK_ORIG_CTXT, tgt, count,
2804 &catid->lci_logid, &osc_unlink_orig_logops);
2808 rc = llog_setup(obd, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
2809 &osc_size_repl_logops);
2813 static int osc_llog_finish(struct obd_device *obd, int count)
2818 rc = llog_cleanup(llog_get_context(obd, LLOG_UNLINK_ORIG_CTXT));
2822 rc = llog_cleanup(llog_get_context(obd, LLOG_SIZE_REPL_CTXT));
2827 static int osc_connect(struct lustre_handle *exph,
2828 struct obd_device *obd, struct obd_uuid *cluuid)
2832 rc = client_connect_import(exph, obd, cluuid);
2837 static int osc_disconnect(struct obd_export *exp, int flags)
2839 struct obd_device *obd = class_exp2obd(exp);
2840 struct llog_ctxt *ctxt = llog_get_context(obd, LLOG_SIZE_REPL_CTXT);
2843 if (obd->u.cli.cl_conn_count == 1)
2844 /* flush any remaining cancel messages out to the target */
2845 llog_sync(ctxt, exp);
2847 rc = client_disconnect_export(exp, flags);
2851 static int osc_import_event(struct obd_device *obd,
2852 struct obd_import *imp,
2853 enum obd_import_event event)
2855 struct client_obd *cli;
2858 LASSERT(imp->imp_obd == obd);
2861 case IMP_EVENT_DISCON: {
2862 /* Only do this on the MDS OSC's */
2863 if (imp->imp_server_timeout) {
2864 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
2866 spin_lock(&oscc->oscc_lock);
2867 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
2868 spin_unlock(&oscc->oscc_lock);
2872 case IMP_EVENT_INVALIDATE: {
2873 struct ldlm_namespace *ns = obd->obd_namespace;
2875 /* this used to try and tear down queued pages, but it was
2876 * not correctly implemented. We'll have to do it again once
2877 * we call obd_invalidate_import() agian */
2878 /* XXX And we still need to do this */
2880 /* Reset grants, too */
2882 spin_lock(&cli->cl_loi_list_lock);
2883 cli->cl_avail_grant = 0;
2884 cli->cl_lost_grant = 0;
2885 spin_unlock(&cli->cl_loi_list_lock);
2887 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
2889 if (obd->obd_observer)
2890 rc = obd_notify(obd->obd_observer, obd, 0);
2893 case IMP_EVENT_ACTIVE: {
2894 if (obd->obd_observer)
2895 rc = obd_notify(obd->obd_observer, obd, 1);
2899 CERROR("Unknown import event %d\n", event);
2905 int osc_setup(struct obd_device *obd, obd_count len, void *buf)
2909 rc = ptlrpcd_addref();
2913 rc = client_obd_setup(obd, len, buf);
2922 int osc_cleanup(struct obd_device *obd, int flags)
2926 rc = client_obd_cleanup(obd, flags);
2932 struct obd_ops osc_obd_ops = {
2933 o_owner: THIS_MODULE,
2934 o_attach: osc_attach,
2935 o_detach: osc_detach,
2937 o_cleanup: osc_cleanup,
2938 o_connect: osc_connect,
2939 o_disconnect: osc_disconnect,
2940 o_statfs: osc_statfs,
2941 o_packmd: osc_packmd,
2942 o_unpackmd: osc_unpackmd,
2943 o_create: osc_create,
2944 o_destroy: osc_destroy,
2945 o_getattr: osc_getattr,
2946 o_getattr_async:osc_getattr_async,
2947 o_setattr: osc_setattr,
2949 o_brw_async: osc_brw_async,
2950 .o_prep_async_page = osc_prep_async_page,
2951 .o_queue_async_io = osc_queue_async_io,
2952 .o_set_async_flags = osc_set_async_flags,
2953 .o_queue_group_io = osc_queue_group_io,
2954 .o_trigger_group_io = osc_trigger_group_io,
2955 .o_teardown_async_page = osc_teardown_async_page,
2958 o_enqueue: osc_enqueue,
2960 o_change_cbdata:osc_change_cbdata,
2961 o_cancel: osc_cancel,
2962 o_cancel_unused:osc_cancel_unused,
2963 o_iocontrol: osc_iocontrol,
2964 o_get_info: osc_get_info,
2965 o_set_info: osc_set_info,
2966 o_import_event: osc_import_event,
2967 o_llog_init: osc_llog_init,
2968 o_llog_finish: osc_llog_finish,
2971 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2972 struct obd_ops sanosc_obd_ops = {
2973 o_owner: THIS_MODULE,
2974 o_attach: osc_attach,
2975 o_detach: osc_detach,
2976 o_cleanup: client_obd_cleanup,
2977 o_connect: osc_connect,
2978 o_disconnect: client_disconnect_export,
2979 o_statfs: osc_statfs,
2980 o_packmd: osc_packmd,
2981 o_unpackmd: osc_unpackmd,
2982 o_create: osc_real_create,
2983 o_destroy: osc_destroy,
2984 o_getattr: osc_getattr,
2985 o_getattr_async:osc_getattr_async,
2986 o_setattr: osc_setattr,
2987 o_setup: client_sanobd_setup,
2991 o_enqueue: osc_enqueue,
2993 o_change_cbdata:osc_change_cbdata,
2994 o_cancel: osc_cancel,
2995 o_cancel_unused:osc_cancel_unused,
2996 o_iocontrol: osc_iocontrol,
2997 o_import_event: osc_import_event,
2998 o_llog_init: osc_llog_init,
2999 o_llog_finish: osc_llog_finish,
3003 int __init osc_init(void)
3005 struct lprocfs_static_vars lvars, sanlvars;
3009 lprocfs_init_vars(osc, &lvars);
3010 lprocfs_init_vars(osc, &sanlvars);
3012 rc = class_register_type(&osc_obd_ops, lvars.module_vars,
3017 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3018 rc = class_register_type(&sanosc_obd_ops, sanlvars.module_vars,
3019 LUSTRE_SANOSC_NAME);
3021 class_unregister_type(LUSTRE_OSC_NAME);
3027 static void /*__exit*/ osc_exit(void)
3029 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3030 class_unregister_type(LUSTRE_SANOSC_NAME);
3032 class_unregister_type(LUSTRE_OSC_NAME);
3036 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3037 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3038 MODULE_LICENSE("GPL");
3040 module_init(osc_init);
3041 module_exit(osc_exit);