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_interpret(struct ptlrpc_request *req,
501 struct obd_import *imp = req->rq_import;
502 struct client_obd *cli = &imp->imp_obd->u.cli;
504 spin_lock(&cli->cl_loi_list_lock);
505 cli->cl_dstr_in_flight--;
506 LASSERT(cli->cl_dstr_in_flight >= 0);
507 LASSERT(cli->cl_dstr_in_flight <= cli->cl_dstr_in_flight);
508 spin_unlock(&cli->cl_loi_list_lock);
509 wake_up(&cli->cl_wait_for_destroy_slot);
513 static void osc_grab_destroy_slot(struct client_obd *cli)
515 spin_lock(&cli->cl_loi_list_lock);
517 struct l_wait_info lwi = { 0 };
519 if (cli->cl_dstr_in_flight < cli->cl_max_dstr_in_flight) {
520 cli->cl_dstr_in_flight++;
524 spin_unlock(&cli->cl_loi_list_lock);
525 l_wait_event(cli->cl_wait_for_destroy_slot,
526 cli->cl_dstr_in_flight < cli->cl_max_dstr_in_flight,
528 spin_lock(&cli->cl_loi_list_lock);
530 spin_unlock(&cli->cl_loi_list_lock);
533 static int osc_destroy(struct obd_export *exp, struct obdo *oa,
534 struct lov_stripe_md *ea, struct obd_trans_info *oti)
536 struct obd_import *imp = class_exp2cliimp(exp);
537 struct client_obd *cli = &imp->imp_obd->u.cli;
538 struct ptlrpc_request *request;
539 struct ost_body *body;
540 int rc, size = sizeof(*body);
548 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
549 OST_DESTROY, 1, &size, NULL);
552 request->rq_request_portal = OST_DESTROY_PORTAL;
554 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
556 if (oti != NULL && oa->o_valid & OBD_MD_FLCOOKIE) {
557 memcpy(obdo_logcookie(oa), oti->oti_logcookies,
558 sizeof(*oti->oti_logcookies));
559 oti->oti_logcookies++;
562 memcpy(&body->oa, oa, sizeof(*oa));
563 request->rq_replen = lustre_msg_size(1, &size);
565 if (oti != NULL && (oti->oti_flags & OBD_MODE_ASYNC)) {
566 osc_grab_destroy_slot(cli);
567 request->rq_interpret_reply = osc_destroy_interpret;
568 ptlrpcd_add_req(request);
571 rc = ptlrpc_queue_wait(request);
577 ptlrpc_req_finished(request);
581 body = lustre_swab_repbuf(request, 0, sizeof(*body),
582 lustre_swab_ost_body);
584 CERROR ("Can't unpack body\n");
585 ptlrpc_req_finished(request);
589 memcpy(oa, &body->oa, sizeof(*oa));
590 ptlrpc_req_finished(request);
595 static void osc_announce_cached(struct client_obd *cli, struct obdo *oa,
598 obd_valid bits = OBD_MD_FLBLOCKS|OBD_MD_FLGRANT;
600 LASSERT(!(oa->o_valid & bits));
603 spin_lock(&cli->cl_loi_list_lock);
604 oa->o_dirty = cli->cl_dirty;
605 oa->o_undirty = cli->cl_dirty_max - oa->o_dirty;
606 oa->o_grant = cli->cl_avail_grant;
607 oa->o_dropped = cli->cl_lost_grant;
608 cli->cl_lost_grant = 0;
609 spin_unlock(&cli->cl_loi_list_lock);
610 CDEBUG(D_CACHE,"dirty: "LPU64" undirty: %u dropped %u grant: "LPU64"\n",
611 oa->o_dirty, oa->o_undirty, oa->o_dropped, oa->o_grant);
614 /* caller must hold loi_list_lock */
615 static void osc_consume_write_grant(struct client_obd *cli,
616 struct osc_async_page *oap)
618 cli->cl_dirty += PAGE_SIZE;
619 cli->cl_avail_grant -= PAGE_SIZE;
620 oap->oap_brw_flags |= OBD_BRW_FROM_GRANT;
621 CDEBUG(D_CACHE, "using %lu grant credits for oap %p\n", PAGE_SIZE, oap);
622 LASSERT(cli->cl_avail_grant >= 0);
625 static unsigned long rpcs_in_flight(struct client_obd *cli)
627 return cli->cl_r_in_flight + cli->cl_w_in_flight;
630 /* caller must hold loi_list_lock */
631 void osc_wake_cache_waiters(struct client_obd *cli)
633 struct list_head *l, *tmp;
634 struct osc_cache_waiter *ocw;
636 list_for_each_safe(l, tmp, &cli->cl_cache_waiters) {
637 /* if we can't dirty more, we must wait until some is written */
638 if (cli->cl_dirty + PAGE_SIZE > cli->cl_dirty_max) {
639 CDEBUG(D_CACHE, "no dirty room: dirty: %ld max %ld\n",
640 cli->cl_dirty, cli->cl_dirty_max);
644 /* if still dirty cache but no grant wait for pending RPCs that
645 * may yet return us some grant before doing sync writes */
646 if (cli->cl_w_in_flight && cli->cl_avail_grant < PAGE_SIZE) {
647 CDEBUG(D_CACHE, "%u BRW writes in flight, no grant\n",
648 cli->cl_w_in_flight);
650 ocw = list_entry(l, struct osc_cache_waiter, ocw_entry);
651 list_del_init(&ocw->ocw_entry);
652 if (cli->cl_avail_grant < PAGE_SIZE) {
653 /* no more RPCs in flight to return grant, do sync IO */
654 ocw->ocw_rc = -EDQUOT;
655 CDEBUG(D_INODE, "wake oap %p for sync\n", ocw->ocw_oap);
657 osc_consume_write_grant(cli, ocw->ocw_oap);
660 wake_up(&ocw->ocw_waitq);
666 static void osc_update_grant(struct client_obd *cli, struct ost_body *body)
668 spin_lock(&cli->cl_loi_list_lock);
669 CDEBUG(D_CACHE, "got "LPU64" extra grant\n", body->oa.o_grant);
670 cli->cl_avail_grant += body->oa.o_grant;
671 /* waiters are woken in brw_interpret_oap */
672 spin_unlock(&cli->cl_loi_list_lock);
675 /* We assume that the reason this OSC got a short read is because it read
676 * beyond the end of a stripe file; i.e. lustre is reading a sparse file
677 * via the LOV, and it _knows_ it's reading inside the file, it's just that
678 * this stripe never got written at or beyond this stripe offset yet. */
679 static void handle_short_read(int nob_read, obd_count page_count,
680 struct brw_page *pga)
684 /* skip bytes read OK */
685 while (nob_read > 0) {
686 LASSERT (page_count > 0);
688 if (pga->count > nob_read) {
689 /* EOF inside this page */
690 ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
691 memset(ptr + nob_read, 0, pga->count - nob_read);
698 nob_read -= pga->count;
703 /* zero remaining pages */
704 while (page_count-- > 0) {
705 ptr = kmap(pga->pg) + (pga->page_offset & ~PAGE_MASK);
706 memset(ptr, 0, pga->count);
712 static int check_write_rcs(struct ptlrpc_request *request,
713 int requested_nob, int niocount,
714 obd_count page_count, struct brw_page *pga)
718 /* return error if any niobuf was in error */
719 remote_rcs = lustre_swab_repbuf(request, 1,
720 sizeof(*remote_rcs) * niocount, NULL);
721 if (remote_rcs == NULL) {
722 CERROR("Missing/short RC vector on BRW_WRITE reply\n");
725 if (lustre_msg_swabbed(request->rq_repmsg))
726 for (i = 0; i < niocount; i++)
727 __swab32s((__u32 *)&remote_rcs[i]);
729 for (i = 0; i < niocount; i++) {
730 if (remote_rcs[i] < 0)
731 return(remote_rcs[i]);
733 if (remote_rcs[i] != 0) {
734 CERROR("rc[%d] invalid (%d) req %p\n",
735 i, remote_rcs[i], request);
740 if (request->rq_bulk->bd_nob_transferred != requested_nob) {
741 CERROR("Unexpected # bytes transferred: %d (requested %d)\n",
742 requested_nob, request->rq_bulk->bd_nob_transferred);
749 static inline int can_merge_pages(struct brw_page *p1, struct brw_page *p2)
751 if (p1->flag != p2->flag) {
752 unsigned mask = ~OBD_BRW_FROM_GRANT;
754 /* warn if we try to combine flags that we don't know to be
756 if ((p1->flag & mask) != (p2->flag & mask))
757 CERROR("is it ok to have flags 0x%x and 0x%x in the "
758 "same brw?\n", p1->flag, p2->flag);
762 return (p1->disk_offset + p1->count == p2->disk_offset);
766 static obd_count cksum_pages(int nob, obd_count page_count,
767 struct brw_page *pga)
773 LASSERT (page_count > 0);
776 ost_checksum(&cksum, ptr + (pga->off & (PAGE_SIZE - 1)),
777 pga->count > nob ? nob : pga->count);
789 #define osc_encrypt_page(page, off, count) \
790 osc_crypt_page(page, off, count, ENCRYPT_DATA)
791 #define osc_decrypt_page(page, off, count) \
792 osc_crypt_page(page, off, count, DECRYPT_DATA)
793 /*Put a global call back var here is Ugly, but put it to client_obd
794 *also seems not a good idea, WangDi*/
795 crypt_cb_t osc_crypt_cb = NULL;
797 static int osc_crypt_page(struct page *page, obd_off page_off, obd_off count,
803 if (osc_crypt_cb != NULL)
804 rc = osc_crypt_cb(page, page_off, count, flags);
806 CERROR("crypt page error %d \n", rc);
810 static int osc_decrypt_pages(struct brw_page *pga, int page_count)
815 for (i = 0; i < page_count; i++) {
816 struct brw_page *pg = &pga[i];
817 osc_decrypt_page(pg->pg, pg->page_offset, pg->count);
822 static int osc_brw_prep_request(int cmd, struct obd_import *imp,struct obdo *oa,
823 struct lov_stripe_md *lsm, obd_count page_count,
824 struct brw_page *pga, int *requested_nobp,
825 int *niocountp, struct ptlrpc_request **reqp)
827 struct ptlrpc_request *req;
828 struct ptlrpc_bulk_desc *desc;
829 struct client_obd *cli = &imp->imp_obd->u.cli;
830 struct ost_body *body;
831 struct lustre_id *raw_id = obdo_id(oa);
832 struct obd_capa *ocapa = NULL;
833 struct lustre_capa *capa = NULL;
834 struct obd_ioobj *ioobj;
835 struct niobuf_remote *niobuf;
844 opc = ((cmd & OBD_BRW_WRITE) != 0) ? OST_WRITE : OST_READ;
846 for (niocount = i = 1; i < page_count; i++)
847 if (!can_merge_pages(&pga[i - 1], &pga[i]))
850 /* TODO: this could be optimized: thie capability can be
851 * found from ll_inode_info->lli_capas. */
852 /* partial write might cause read, both CAPA_READ and CAPA_WRITE
853 * capability could be used here */
854 capa_op = (opc == OST_WRITE) ? CAPA_WRITE : CAPA_READ | CAPA_WRITE;
855 ocapa = capa_get(oa->o_fsuid, capa_op, id_group(raw_id),
856 id_ino(raw_id), id_gen(raw_id), CLIENT_CAPA);
858 size[bufcnt++] = sizeof(*body);
859 size[bufcnt++] = sizeof(*ioobj);
861 size[bufcnt++] = sizeof(*capa);
862 size[bufcnt++] = niocount * sizeof(*niobuf);
864 req = ptlrpc_prep_req(imp, LUSTRE_OBD_VERSION, opc, bufcnt, size, NULL);
868 if (opc == OST_WRITE)
869 desc = ptlrpc_prep_bulk_imp (req, page_count,
870 BULK_GET_SOURCE, OST_BULK_PORTAL);
872 desc = ptlrpc_prep_bulk_imp (req, page_count,
873 BULK_PUT_SINK, OST_BULK_PORTAL);
875 GOTO(out, rc = -ENOMEM);
876 /* NB request now owns desc and will free it when it gets freed */
879 body = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*body));
880 memcpy(&body->oa, oa, sizeof(*oa));
881 ioobj = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*ioobj));
883 capa = lustre_msg_buf(req->rq_reqmsg, bufcnt++, sizeof(*capa));
884 capa_dup(capa, ocapa);
885 body->oa.o_valid |= OBD_MD_CAPA;
887 niobuf = lustre_msg_buf(req->rq_reqmsg, bufcnt++,
888 niocount * sizeof(*niobuf));
890 obdo_to_ioobj(oa, ioobj);
891 ioobj->ioo_bufcnt = niocount;
893 LASSERT (page_count > 0);
895 for (requested_nob = i = 0; i < page_count; i++, niobuf++) {
896 struct brw_page *pg = &pga[i];
897 struct brw_page *pg_prev = pg - 1;
899 LASSERT(pg->count > 0);
900 LASSERTF((pg->page_offset & ~PAGE_MASK)+ pg->count <= PAGE_SIZE,
901 "i: %d pg: %p pg_off: "LPU64", count: %u\n", i, pg,
902 pg->page_offset, pg->count);
903 LASSERTF(i == 0 || pg->disk_offset > pg_prev->disk_offset,
904 "i %d p_c %u pg %p [pri %lu ind %lu] off "LPU64
905 " prev_pg %p [pri %lu ind %lu] off "LPU64"\n",
907 pg->pg, pg->pg->private, pg->pg->index, pg->disk_offset,
908 pg_prev->pg, pg_prev->pg->private, pg_prev->pg->index,
909 pg_prev->disk_offset);
911 if (opc == OST_WRITE) {
912 rc = osc_encrypt_page(pg->pg, pg->page_offset, pg->count);
917 ptlrpc_prep_bulk_page(desc, pg->pg,
918 pg->page_offset & ~PAGE_MASK, pg->count);
919 requested_nob += pg->count;
921 if (i > 0 && can_merge_pages(pg_prev, pg)) {
923 niobuf->len += pg->count;
925 niobuf->offset = pg->disk_offset;
926 niobuf->len = pg->count;
927 niobuf->flags = pg->flag;
931 LASSERT((void *)(niobuf - niocount) ==
932 lustre_msg_buf(req->rq_reqmsg, bufcnt - 1,
933 niocount * sizeof(*niobuf)));
934 osc_announce_cached(cli, &body->oa, opc == OST_WRITE ? requested_nob:0);
936 /* size[0] still sizeof (*body) */
937 if (opc == OST_WRITE) {
939 body->oa.o_valid |= OBD_MD_FLCKSUM;
940 body->oa.o_cksum = cksum_pages(requested_nob, page_count, pga);
942 /* 1 RC per niobuf */
943 size[1] = sizeof(__u32) * niocount;
944 req->rq_replen = lustre_msg_size(2, size);
946 /* 1 RC for the whole I/O */
947 req->rq_replen = lustre_msg_size(1, size);
950 *niocountp = niocount;
951 *requested_nobp = requested_nob;
956 ptlrpc_req_finished (req);
960 static int osc_brw_fini_request(struct ptlrpc_request *req, struct obdo *oa,
961 int requested_nob, int niocount,
962 obd_count page_count, struct brw_page *pga,
965 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
966 struct ost_body *body;
972 body = lustre_swab_repbuf(req, 0, sizeof(*body), lustre_swab_ost_body);
974 CERROR ("Can't unpack body\n");
978 osc_update_grant(cli, body);
979 memcpy(oa, &body->oa, sizeof(*oa));
981 if (req->rq_reqmsg->opc == OST_WRITE) {
983 CERROR ("Unexpected +ve rc %d\n", rc);
986 LASSERT (req->rq_bulk->bd_nob == requested_nob);
987 osc_decrypt_pages(pga, page_count);
988 RETURN(check_write_rcs(req, requested_nob, niocount,
992 if (rc > requested_nob) {
993 CERROR("Unexpected rc %d (%d requested)\n", rc, requested_nob);
997 if (rc != req->rq_bulk->bd_nob_transferred) {
998 CERROR ("Unexpected rc %d (%d transferred)\n",
999 rc, req->rq_bulk->bd_nob_transferred);
1003 if (rc < requested_nob)
1004 handle_short_read(rc, page_count, pga);
1007 if (oa->o_valid & OBD_MD_FLCKSUM) {
1008 const struct ptlrpc_peer *peer =
1009 &req->rq_import->imp_connection->c_peer;
1010 static int cksum_counter;
1011 obd_count server_cksum = oa->o_cksum;
1012 obd_count cksum = cksum_pages(rc, page_count, pga);
1013 char str[PTL_NALFMT_SIZE];
1015 ptlrpc_peernid2str(peer, str);
1018 if (server_cksum != cksum) {
1019 CERROR("Bad checksum: server %x, client %x, server NID "
1020 LPX64" (%s)\n", server_cksum, cksum,
1021 peer->peer_id.nid, str);
1023 oa->o_cksum = cksum;
1024 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
1025 CWARN("Checksum %u from "LPX64" (%s) OK: %x\n",
1026 cksum_counter, peer->peer_id.nid, str, cksum);
1029 static int cksum_missed;
1032 if ((cksum_missed & (-cksum_missed)) == cksum_missed)
1033 CERROR("Request checksum %u from "LPX64", no reply\n",
1035 req->rq_import->imp_connection->c_peer.peer_id.nid);
1038 osc_decrypt_pages(pga, page_count);
1042 static int osc_brw_internal(int cmd, struct obd_export *exp,struct obdo *oa,
1043 struct lov_stripe_md *lsm,
1044 obd_count page_count, struct brw_page *pga)
1048 struct ptlrpc_request *request;
1053 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
1054 page_count, pga, &requested_nob, &niocount,
1059 rc = ptlrpc_queue_wait(request);
1061 if (rc == -ETIMEDOUT && request->rq_resend) {
1062 DEBUG_REQ(D_HA, request, "BULK TIMEOUT");
1063 ptlrpc_req_finished(request);
1067 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
1068 page_count, pga, rc);
1070 ptlrpc_req_finished(request);
1074 static int brw_interpret(struct ptlrpc_request *request,
1075 struct osc_brw_async_args *aa, int rc)
1077 struct obdo *oa = aa->aa_oa;
1078 int requested_nob = aa->aa_requested_nob;
1079 int niocount = aa->aa_nio_count;
1080 obd_count page_count = aa->aa_page_count;
1081 struct brw_page *pga = aa->aa_pga;
1084 rc = osc_brw_fini_request(request, oa, requested_nob, niocount,
1085 page_count, pga, rc);
1089 static int async_internal(int cmd, struct obd_export *exp, struct obdo *oa,
1090 struct lov_stripe_md *lsm, obd_count page_count,
1091 struct brw_page *pga, struct ptlrpc_request_set *set)
1093 struct ptlrpc_request *request;
1096 struct osc_brw_async_args *aa;
1100 rc = osc_brw_prep_request(cmd, class_exp2cliimp(exp), oa, lsm,
1101 page_count, pga, &requested_nob, &nio_count,
1104 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1105 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1107 aa->aa_requested_nob = requested_nob;
1108 aa->aa_nio_count = nio_count;
1109 aa->aa_page_count = page_count;
1112 request->rq_interpret_reply = brw_interpret;
1113 ptlrpc_set_add_req(set, request);
1119 #define min_t(type,x,y) \
1120 ({ type __x = (x); type __y = (y); __x < __y ? __x: __y; })
1124 * ugh, we want disk allocation on the target to happen in offset order. we'll
1125 * follow sedgewicks advice and stick to the dead simple shellsort -- it'll do
1126 * fine for our small page arrays and doesn't require allocation. its an
1127 * insertion sort that swaps elements that are strides apart, shrinking the
1128 * stride down until its '1' and the array is sorted.
1130 static void sort_brw_pages(struct brw_page *array, int num)
1133 struct brw_page tmp;
1137 for (stride = 1; stride < num ; stride = (stride * 3) + 1)
1142 for (i = stride ; i < num ; i++) {
1145 while (j >= stride && array[j - stride].disk_offset >
1147 array[j] = array[j - stride];
1152 } while (stride > 1);
1155 /* make sure we the regions we're passing to elan don't violate its '4
1156 * fragments' constraint. portal headers are a fragment, all full
1157 * PAGE_SIZE long pages count as 1 fragment, and each partial page
1158 * counts as a fragment. I think. see bug 934. */
1159 static obd_count check_elan_limit(struct brw_page *pg, obd_count pages)
1162 int saw_whole_frag = 0;
1165 for (i = 0 ; frags_left && i < pages ; pg++, i++) {
1166 if (pg->count == PAGE_SIZE) {
1167 if (!saw_whole_frag) {
1178 static int osc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
1179 struct lov_stripe_md *lsm, obd_count page_count,
1180 struct brw_page *pga, struct obd_trans_info *oti)
1184 if (cmd == OBD_BRW_CHECK) {
1185 /* The caller just wants to know if there's a chance that this
1186 * I/O can succeed */
1187 struct obd_import *imp = class_exp2cliimp(exp);
1189 if (imp == NULL || imp->imp_invalid)
1194 while (page_count) {
1195 obd_count pages_per_brw;
1198 if (page_count > PTLRPC_MAX_BRW_PAGES)
1199 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1201 pages_per_brw = page_count;
1203 sort_brw_pages(pga, pages_per_brw);
1204 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1206 rc = osc_brw_internal(cmd, exp, oa, lsm, pages_per_brw, pga);
1211 page_count -= pages_per_brw;
1212 pga += pages_per_brw;
1217 static int osc_brw_async(int cmd, struct obd_export *exp, struct obdo *oa,
1218 struct lov_stripe_md *lsm, obd_count page_count,
1219 struct brw_page *pga, struct ptlrpc_request_set *set,
1220 struct obd_trans_info *oti)
1224 if (cmd == OBD_BRW_CHECK) {
1225 /* The caller just wants to know if there's a chance that this
1226 * I/O can succeed */
1227 struct obd_import *imp = class_exp2cliimp(exp);
1229 if (imp == NULL || imp->imp_invalid)
1234 while (page_count) {
1235 obd_count pages_per_brw;
1238 if (page_count > PTLRPC_MAX_BRW_PAGES)
1239 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
1241 pages_per_brw = page_count;
1243 sort_brw_pages(pga, pages_per_brw);
1244 pages_per_brw = check_elan_limit(pga, pages_per_brw);
1246 rc = async_internal(cmd, exp, oa, lsm, pages_per_brw, pga, set);
1251 page_count -= pages_per_brw;
1252 pga += pages_per_brw;
1257 static void osc_check_rpcs(struct client_obd *cli);
1258 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1260 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi);
1261 static void lop_update_pending(struct client_obd *cli,
1262 struct loi_oap_pages *lop, int cmd, int delta);
1264 /* this is called when a sync waiter receives an interruption. Its job is to
1265 * get the caller woken as soon as possible. If its page hasn't been put in an
1266 * rpc yet it can dequeue immediately. Otherwise it has to mark the rpc as
1267 * desiring interruption which will forcefully complete the rpc once the rpc
1269 static void osc_occ_interrupted(struct oig_callback_context *occ)
1271 struct osc_async_page *oap;
1272 struct loi_oap_pages *lop;
1273 struct lov_oinfo *loi;
1276 /* XXX member_of() */
1277 oap = list_entry(occ, struct osc_async_page, oap_occ);
1279 spin_lock(&oap->oap_cli->cl_loi_list_lock);
1281 oap->oap_interrupted = 1;
1283 /* ok, it's been put in an rpc. */
1284 if (oap->oap_request != NULL) {
1285 ptlrpc_mark_interrupted(oap->oap_request);
1286 ptlrpcd_wake(oap->oap_request);
1290 /* we don't get interruption callbacks until osc_trigger_sync_io()
1291 * has been called and put the sync oaps in the pending/urgent lists.*/
1292 if (!list_empty(&oap->oap_pending_item)) {
1293 list_del_init(&oap->oap_pending_item);
1294 if (oap->oap_async_flags & ASYNC_URGENT)
1295 list_del_init(&oap->oap_urgent_item);
1298 lop = (oap->oap_cmd == OBD_BRW_WRITE) ?
1299 &loi->loi_write_lop : &loi->loi_read_lop;
1300 lop_update_pending(oap->oap_cli, lop, oap->oap_cmd, -1);
1301 loi_list_maint(oap->oap_cli, oap->oap_loi);
1303 oig_complete_one(oap->oap_oig, &oap->oap_occ, 0);
1304 oap->oap_oig = NULL;
1308 spin_unlock(&oap->oap_cli->cl_loi_list_lock);
1311 /* this must be called holding the loi list lock to give coverage to exit_cache,
1312 * async_flag maintenance, and oap_request */
1313 static void osc_ap_completion(struct client_obd *cli, struct obdo *oa,
1314 struct osc_async_page *oap, int sent, int rc)
1316 osc_exit_cache(cli, oap, sent);
1317 oap->oap_async_flags = 0;
1318 oap->oap_interrupted = 0;
1320 if (oap->oap_request != NULL) {
1321 ptlrpc_req_finished(oap->oap_request);
1322 oap->oap_request = NULL;
1325 if (rc == 0 && oa != NULL)
1326 oap->oap_loi->loi_blocks = oa->o_blocks;
1329 oig_complete_one(oap->oap_oig, &oap->oap_occ, rc);
1330 oap->oap_oig = NULL;
1335 oap->oap_caller_ops->ap_completion(oap->oap_caller_data, oap->oap_cmd,
1339 static int brw_interpret_oap(struct ptlrpc_request *request,
1340 struct osc_brw_async_args *aa, int rc)
1342 struct osc_async_page *oap;
1343 struct client_obd *cli;
1344 struct list_head *pos, *n;
1348 do_gettimeofday(&now);
1349 rc = osc_brw_fini_request(request, aa->aa_oa, aa->aa_requested_nob,
1350 aa->aa_nio_count, aa->aa_page_count,
1353 CDEBUG(D_INODE, "request %p aa %p rc %d\n", request, aa, rc);
1357 /* in failout recovery we ignore writeback failure and want
1358 * to just tell llite to unlock the page and continue */
1359 if (request->rq_reqmsg->opc == OST_WRITE &&
1360 (cli->cl_import == NULL || cli->cl_import->imp_invalid)) {
1361 CDEBUG(D_INODE, "flipping to rc 0 imp %p inv %d\n",
1363 cli->cl_import ? cli->cl_import->imp_invalid : -1);
1367 spin_lock(&cli->cl_loi_list_lock);
1369 if (request->rq_reqmsg->opc == OST_WRITE)
1370 lprocfs_stime_record(&cli->cl_write_stime, &now,
1371 &request->rq_rpcd_start);
1373 lprocfs_stime_record(&cli->cl_read_stime, &now,
1374 &request->rq_rpcd_start);
1376 /* We need to decrement before osc_ap_completion->osc_wake_cache_waiters
1377 * is called so we know whether to go to sync BRWs or wait for more
1378 * RPCs to complete */
1379 if (request->rq_reqmsg->opc == OST_WRITE)
1380 cli->cl_w_in_flight--;
1382 cli->cl_r_in_flight--;
1384 /* the caller may re-use the oap after the completion call so
1385 * we need to clean it up a little */
1386 list_for_each_safe(pos, n, &aa->aa_oaps) {
1387 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1389 //CDEBUG(D_INODE, "page %p index %lu oap %p\n",
1390 //oap->oap_page, oap->oap_page->index, oap);
1392 list_del_init(&oap->oap_rpc_item);
1393 osc_ap_completion(cli, aa->aa_oa, oap, 1, rc);
1396 /* no write RPCs in flight, reset the time */
1397 if (request->rq_reqmsg->opc == OST_WRITE && cli->cl_w_in_flight == 0)
1398 do_gettimeofday(&cli->cl_last_write_time);
1400 osc_wake_cache_waiters(cli);
1401 osc_check_rpcs(cli);
1402 spin_unlock(&cli->cl_loi_list_lock);
1404 obdo_free(aa->aa_oa);
1405 OBD_FREE(aa->aa_pga, aa->aa_page_count * sizeof(struct brw_page));
1410 static struct ptlrpc_request *osc_build_req(struct client_obd *cli,
1411 struct list_head *rpc_list,
1412 int page_count, int cmd)
1414 struct ptlrpc_request *req;
1415 struct brw_page *pga = NULL;
1416 int requested_nob, nio_count;
1417 struct osc_brw_async_args *aa;
1418 struct obdo *oa = NULL;
1419 struct obd_async_page_ops *ops = NULL;
1420 void *caller_data = NULL;
1421 struct list_head *pos;
1424 LASSERT(!list_empty(rpc_list));
1426 OBD_ALLOC(pga, sizeof(*pga) * page_count);
1428 RETURN(ERR_PTR(-ENOMEM));
1432 GOTO(out, req = ERR_PTR(-ENOMEM));
1435 list_for_each(pos, rpc_list) {
1436 struct osc_async_page *oap;
1438 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1440 ops = oap->oap_caller_ops;
1441 caller_data = oap->oap_caller_data;
1443 pga[i].disk_offset = oap->oap_obj_off + oap->oap_page_off;
1444 pga[i].page_offset = pga[i].disk_offset;
1445 pga[i].pg = oap->oap_page;
1446 pga[i].count = oap->oap_count;
1447 pga[i].flag = oap->oap_brw_flags;
1448 CDEBUG(0, "put page %p index %lu oap %p flg %x to pga\n",
1449 pga[i].pg, oap->oap_page->index, oap, pga[i].flag);
1453 /* always get the data for the obdo for the rpc */
1454 LASSERT(ops != NULL);
1455 ops->ap_fill_obdo(caller_data, cmd, oa);
1457 sort_brw_pages(pga, page_count);
1458 rc = osc_brw_prep_request(cmd, cli->cl_import, oa, NULL, page_count,
1459 pga, &requested_nob, &nio_count, &req);
1461 CERROR("prep_req failed: %d\n", rc);
1462 GOTO(out, req = ERR_PTR(rc));
1465 LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
1466 aa = (struct osc_brw_async_args *)&req->rq_async_args;
1468 aa->aa_requested_nob = requested_nob;
1469 aa->aa_nio_count = nio_count;
1470 aa->aa_page_count = page_count;
1479 OBD_FREE(pga, sizeof(*pga) * page_count);
1484 /* strange write gap too long (15s) */
1485 #define CLI_ODD_WRITE_GAP 15000000
1487 static void lop_update_pending(struct client_obd *cli,
1488 struct loi_oap_pages *lop, int cmd, int delta)
1490 lop->lop_num_pending += delta;
1491 if (cmd == OBD_BRW_WRITE)
1492 cli->cl_pending_w_pages += delta;
1494 cli->cl_pending_r_pages += delta;
1497 static long timeval_sub(struct timeval *large, struct timeval *small)
1499 return (large->tv_sec - small->tv_sec) * 1000000 +
1500 (large->tv_usec - small->tv_usec);
1503 /* the loi lock is held across this function but it's allowed to release
1504 * and reacquire it during its work */
1505 static int osc_send_oap_rpc(struct client_obd *cli, struct lov_oinfo *loi,
1506 int cmd, struct loi_oap_pages *lop)
1508 struct ptlrpc_request *request;
1509 obd_count page_count = 0;
1510 struct list_head *tmp, *pos;
1511 struct osc_async_page *oap = NULL;
1512 struct osc_brw_async_args *aa;
1513 struct obd_async_page_ops *ops;
1514 LIST_HEAD(rpc_list);
1517 LASSERT(lop != LP_POISON);
1518 LASSERT(lop->lop_pending.next != LP_POISON);
1520 /* first we find the pages we're allowed to work with */
1521 list_for_each_safe(pos, tmp, &lop->lop_pending) {
1522 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
1523 ops = oap->oap_caller_ops;
1525 LASSERT(oap->oap_magic == OAP_MAGIC);
1527 /* in llite being 'ready' equates to the page being locked
1528 * until completion unlocks it. commit_write submits a page
1529 * as not ready because its unlock will happen unconditionally
1530 * as the call returns. if we race with commit_write giving
1531 * us that page we dont' want to create a hole in the page
1532 * stream, so we stop and leave the rpc to be fired by
1533 * another dirtier or kupdated interval (the not ready page
1534 * will still be on the dirty list). we could call in
1535 * at the end of ll_file_write to process the queue again. */
1536 if (!(oap->oap_async_flags & ASYNC_READY)) {
1537 int rc = ops->ap_make_ready(oap->oap_caller_data, cmd);
1539 CDEBUG(D_INODE, "oap %p page %p returned %d "
1540 "instead of ready\n", oap,
1544 /* llite is telling us that the page is still
1545 * in commit_write and that we should try
1546 * and put it in an rpc again later. we
1547 * break out of the loop so we don't create
1548 * a hole in the sequence of pages in the rpc
1553 /* the io isn't needed.. tell the checks
1554 * below to complete the rpc with EINTR */
1555 oap->oap_async_flags |= ASYNC_COUNT_STABLE;
1556 oap->oap_count = -EINTR;
1559 oap->oap_async_flags |= ASYNC_READY;
1562 LASSERTF(0, "oap %p page %p returned %d "
1563 "from make_ready\n", oap,
1571 * Page submitted for IO has to be locked. Either by
1572 * ->ap_make_ready() or by higher layers.
1574 * XXX nikita: this assertion should be adjusted when lustre
1575 * starts using PG_writeback for pages being written out.
1577 LASSERT(PageLocked(oap->oap_page));
1579 /* take the page out of our book-keeping */
1580 list_del_init(&oap->oap_pending_item);
1581 lop_update_pending(cli, lop, cmd, -1);
1582 list_del_init(&oap->oap_urgent_item);
1584 /* ask the caller for the size of the io as the rpc leaves. */
1585 if (!(oap->oap_async_flags & ASYNC_COUNT_STABLE))
1587 ops->ap_refresh_count(oap->oap_caller_data,cmd);
1588 if (oap->oap_count <= 0) {
1589 CDEBUG(D_CACHE, "oap %p count %d, completing\n", oap,
1591 osc_ap_completion(cli, NULL, oap, 0, oap->oap_count);
1595 /* now put the page back in our accounting */
1596 list_add_tail(&oap->oap_rpc_item, &rpc_list);
1597 if (++page_count >= cli->cl_max_pages_per_rpc)
1601 osc_wake_cache_waiters(cli);
1603 if (page_count == 0)
1606 loi_list_maint(cli, loi);
1607 spin_unlock(&cli->cl_loi_list_lock);
1609 request = osc_build_req(cli, &rpc_list, page_count, cmd);
1610 if (IS_ERR(request)) {
1611 /* this should happen rarely and is pretty bad, it makes the
1612 * pending list not follow the dirty order */
1613 spin_lock(&cli->cl_loi_list_lock);
1614 list_for_each_safe(pos, tmp, &rpc_list) {
1615 oap = list_entry(pos, struct osc_async_page,
1617 list_del_init(&oap->oap_rpc_item);
1619 /* queued sync pages can be torn down while the pages
1620 * were between the pending list and the rpc */
1621 if (oap->oap_interrupted) {
1622 CDEBUG(D_INODE, "oap %p interrupted\n", oap);
1623 osc_ap_completion(cli, NULL, oap, 0,
1628 /* put the page back in the loi/lop lists */
1629 list_add_tail(&oap->oap_pending_item,
1631 lop_update_pending(cli, lop, cmd, 1);
1632 if (oap->oap_async_flags & ASYNC_URGENT)
1633 list_add(&oap->oap_urgent_item,
1636 loi_list_maint(cli, loi);
1637 RETURN(PTR_ERR(request));
1640 LASSERT(sizeof(*aa) <= sizeof(request->rq_async_args));
1641 aa = (struct osc_brw_async_args *)&request->rq_async_args;
1642 INIT_LIST_HEAD(&aa->aa_oaps);
1643 list_splice(&rpc_list, &aa->aa_oaps);
1644 INIT_LIST_HEAD(&rpc_list);
1647 if (cmd == OBD_BRW_READ) {
1648 lprocfs_oh_tally_log2(&cli->cl_read_page_hist, page_count);
1649 lprocfs_oh_tally(&cli->cl_read_rpc_hist, cli->cl_r_in_flight);
1651 lprocfs_oh_tally_log2(&cli->cl_write_page_hist, page_count);
1652 lprocfs_oh_tally(&cli->cl_write_rpc_hist,
1653 cli->cl_w_in_flight);
1657 spin_lock(&cli->cl_loi_list_lock);
1659 /* collect write gaps and sum of them */
1660 if (cmd == OBD_BRW_WRITE && cli->cl_w_in_flight == 0) {
1664 do_gettimeofday(&now);
1666 if (cli->cl_last_write_time.tv_sec) {
1667 diff = timeval_sub(&now, &cli->cl_last_write_time);
1668 if (diff < CLI_ODD_WRITE_GAP) {
1669 cli->cl_write_gap_sum += diff;
1670 cli->cl_write_gaps++;
1675 if (cmd == OBD_BRW_READ) {
1676 cli->cl_r_in_flight++;
1679 cli->cl_w_in_flight++;
1680 cli->cl_write_num++;
1683 /* queued sync pages can be torn down while the pages
1684 * were between the pending list and the rpc */
1685 list_for_each(pos, &aa->aa_oaps) {
1686 oap = list_entry(pos, struct osc_async_page, oap_rpc_item);
1687 if (oap->oap_interrupted) {
1688 CDEBUG(D_INODE, "oap %p in req %p interrupted\n",
1690 ptlrpc_mark_interrupted(request);
1695 CDEBUG(D_INODE, "req %p: %d pages, aa %p. now %dr/%dw in flight\n",
1696 request, page_count, aa, cli->cl_r_in_flight,
1697 cli->cl_w_in_flight);
1699 oap->oap_request = ptlrpc_request_addref(request);
1700 request->rq_interpret_reply = brw_interpret_oap;
1702 ptlrpcd_add_req(request);
1706 static int lop_makes_rpc(struct client_obd *cli, struct loi_oap_pages *lop,
1712 if (lop->lop_num_pending == 0)
1715 /* if we have an invalid import we want to drain the queued pages
1716 * by forcing them through rpcs that immediately fail and complete
1717 * the pages. recovery relies on this to empty the queued pages
1718 * before canceling the locks and evicting down the llite pages */
1719 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
1722 /* stream rpcs in queue order as long as as there is an urgent page
1723 * queued. this is our cheap solution for good batching in the case
1724 * where writepage marks some random page in the middle of the file as
1725 * urgent because of, say, memory pressure */
1726 if (!list_empty(&lop->lop_urgent))
1729 /* fire off rpcs when we have 'optimal' rpcs as tuned for the wire. */
1730 optimal = cli->cl_max_pages_per_rpc;
1731 if (cmd == OBD_BRW_WRITE) {
1732 /* trigger a write rpc stream as long as there are dirtiers
1733 * waiting for space. as they're waiting, they're not going to
1734 * create more pages to coallesce with what's waiting.. */
1735 if (!list_empty(&cli->cl_cache_waiters))
1738 /* *2 to avoid triggering rpcs that would want to include pages
1739 * that are being queued but which can't be made ready until
1740 * the queuer finishes with the page. this is a wart for
1741 * llite::commit_write() */
1744 if (lop->lop_num_pending >= optimal)
1750 static void on_list(struct list_head *item, struct list_head *list,
1753 if (list_empty(item) && should_be_on)
1754 list_add_tail(item, list);
1755 else if (!list_empty(item) && !should_be_on)
1756 list_del_init(item);
1759 /* maintain the loi's cli list membership invariants so that osc_send_oap_rpc
1760 * can find pages to build into rpcs quickly */
1761 static void loi_list_maint(struct client_obd *cli, struct lov_oinfo *loi)
1763 on_list(&loi->loi_cli_item, &cli->cl_loi_ready_list,
1764 lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE) ||
1765 lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ));
1767 on_list(&loi->loi_write_item, &cli->cl_loi_write_list,
1768 loi->loi_write_lop.lop_num_pending);
1770 on_list(&loi->loi_read_item, &cli->cl_loi_read_list,
1771 loi->loi_read_lop.lop_num_pending);
1774 #define LOI_DEBUG(LOI, STR, args...) \
1775 CDEBUG(D_INODE, "loi ready %d wr %d:%d rd %d:%d " STR, \
1776 !list_empty(&(LOI)->loi_cli_item), \
1777 (LOI)->loi_write_lop.lop_num_pending, \
1778 !list_empty(&(LOI)->loi_write_lop.lop_urgent), \
1779 (LOI)->loi_read_lop.lop_num_pending, \
1780 !list_empty(&(LOI)->loi_read_lop.lop_urgent), \
1783 struct lov_oinfo *osc_next_loi(struct client_obd *cli)
1786 /* first return all objects which we already know to have
1787 * pages ready to be stuffed into rpcs */
1788 if (!list_empty(&cli->cl_loi_ready_list))
1789 RETURN(list_entry(cli->cl_loi_ready_list.next,
1790 struct lov_oinfo, loi_cli_item));
1792 /* then if we have cache waiters, return all objects with queued
1793 * writes. This is especially important when many small files
1794 * have filled up the cache and not been fired into rpcs because
1795 * they don't pass the nr_pending/object threshhold */
1796 if (!list_empty(&cli->cl_cache_waiters) &&
1797 !list_empty(&cli->cl_loi_write_list))
1798 RETURN(list_entry(cli->cl_loi_write_list.next,
1799 struct lov_oinfo, loi_write_item));
1801 /* then return all queued objects when we have an invalid import
1802 * so that they get flushed */
1803 if (cli->cl_import == NULL || cli->cl_import->imp_invalid) {
1804 if (!list_empty(&cli->cl_loi_write_list))
1805 RETURN(list_entry(cli->cl_loi_write_list.next,
1806 struct lov_oinfo, loi_write_item));
1807 if (!list_empty(&cli->cl_loi_read_list))
1808 RETURN(list_entry(cli->cl_loi_read_list.next,
1809 struct lov_oinfo, loi_read_item));
1814 /* called with the loi list lock held */
1815 static void osc_check_rpcs(struct client_obd *cli)
1817 struct lov_oinfo *loi;
1818 int rc = 0, race_counter = 0;
1821 while ((loi = osc_next_loi(cli)) != NULL) {
1822 LOI_DEBUG(loi, "%lu in flight\n", rpcs_in_flight(cli));
1823 LASSERT(loi->loi_ost_idx != LL_POISON);
1825 if (rpcs_in_flight(cli) >= cli->cl_max_rpcs_in_flight)
1828 /* attempt some read/write balancing by alternating between
1829 * reads and writes in an object. The makes_rpc checks here
1830 * would be redundant if we were getting read/write work items
1831 * instead of objects. we don't want send_oap_rpc to drain a
1832 * partial read pending queue when we're given this object to
1833 * do io on writes while there are cache waiters */
1834 if (lop_makes_rpc(cli, &loi->loi_write_lop, OBD_BRW_WRITE)) {
1835 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_WRITE,
1836 &loi->loi_write_lop);
1844 if (lop_makes_rpc(cli, &loi->loi_read_lop, OBD_BRW_READ)) {
1845 rc = osc_send_oap_rpc(cli, loi, OBD_BRW_READ,
1846 &loi->loi_read_lop);
1855 /* attempt some inter-object balancing by issueing rpcs
1856 * for each object in turn */
1857 if (!list_empty(&loi->loi_cli_item))
1858 list_del_init(&loi->loi_cli_item);
1859 if (!list_empty(&loi->loi_write_item))
1860 list_del_init(&loi->loi_write_item);
1861 if (!list_empty(&loi->loi_read_item))
1862 list_del_init(&loi->loi_read_item);
1864 loi_list_maint(cli, loi);
1866 /* send_oap_rpc fails with 0 when make_ready tells it to
1867 * back off. llite's make_ready does this when it tries
1868 * to lock a page queued for write that is already locked.
1869 * we want to try sending rpcs from many objects, but we
1870 * don't want to spin failing with 0. */
1871 if (race_counter == 10)
1877 /* we're trying to queue a page in the osc so we're subject to the
1878 * 'cl_dirty_max' limit on the number of pages that can be queued in the osc.
1879 * If the osc's queued pages are already at that limit, then we want to sleep
1880 * until there is space in the osc's queue for us. We also may be waiting for
1881 * write credits from the OST if there are RPCs in flight that may return some
1882 * before we fall back to sync writes.
1884 * We need this know our allocation was granted in the presence of signals */
1885 static int ocw_granted(struct client_obd *cli, struct osc_cache_waiter *ocw)
1889 spin_lock(&cli->cl_loi_list_lock);
1890 rc = list_empty(&ocw->ocw_entry) || rpcs_in_flight(cli) == 0;
1891 spin_unlock(&cli->cl_loi_list_lock);
1895 /* Caller must hold loi_list_lock - we drop/regain it if we need to wait for
1896 * grant or cache space. */
1897 static int osc_enter_cache(struct client_obd *cli, struct lov_oinfo *loi,
1898 struct osc_async_page *oap)
1900 struct osc_cache_waiter ocw;
1901 struct l_wait_info lwi = { 0 };
1902 struct timeval start, stop;
1905 CDEBUG(D_CACHE, "dirty: %ld dirty_max: %ld dropped: %lu grant: %lu\n",
1906 cli->cl_dirty, cli->cl_dirty_max, cli->cl_lost_grant,
1907 cli->cl_avail_grant);
1909 if (cli->cl_dirty_max < PAGE_SIZE)
1912 if (~0ul - cli->cl_dirty_sum <= cli->cl_dirty) {
1913 cli->cl_dirty_av = (cli->cl_dirty_av +
1914 (cli->cl_dirty_sum / cli->cl_dirty_num)) / 2;
1915 cli->cl_dirty_num = 0;
1916 cli->cl_dirty_sum = 0;
1918 if (cli->cl_dirty_num)
1919 cli->cl_dirty_av = (cli->cl_dirty_sum / cli->cl_dirty_num);
1922 cli->cl_dirty_num++;
1923 cli->cl_dirty_sum += cli->cl_dirty;
1925 if (cli->cl_dirty > cli->cl_dirty_dmax)
1926 cli->cl_dirty_dmax = cli->cl_dirty;
1927 if (cli->cl_dirty < cli->cl_dirty_dmin || !cli->cl_dirty_dmin)
1928 cli->cl_dirty_dmin = cli->cl_dirty;
1930 /* Hopefully normal case - cache space and write credits available */
1931 if (cli->cl_dirty + PAGE_SIZE <= cli->cl_dirty_max &&
1932 cli->cl_avail_grant >= PAGE_SIZE) {
1933 /* account for ourselves */
1934 osc_consume_write_grant(cli, oap);
1938 /* Make sure that there are write rpcs in flight to wait for. This
1939 * is a little silly as this object may not have any pending but
1940 * other objects sure might. */
1941 if (cli->cl_w_in_flight) {
1942 list_add_tail(&ocw.ocw_entry, &cli->cl_cache_waiters);
1943 init_waitqueue_head(&ocw.ocw_waitq);
1947 loi_list_maint(cli, loi);
1948 osc_check_rpcs(cli);
1949 spin_unlock(&cli->cl_loi_list_lock);
1951 CDEBUG(0, "sleeping for cache space\n");
1952 do_gettimeofday(&start);
1953 l_wait_event(ocw.ocw_waitq, ocw_granted(cli, &ocw), &lwi);
1954 do_gettimeofday(&stop);
1956 cli->cl_cache_wait_num++;
1957 cli->cl_cache_wait_sum += timeval_sub(&stop, &start);
1959 spin_lock(&cli->cl_loi_list_lock);
1960 lprocfs_stime_record(&cli->cl_enter_stime, &stop, &start);
1961 if (!list_empty(&ocw.ocw_entry)) {
1962 list_del(&ocw.ocw_entry);
1970 cli->cl_sync_rpcs++;
1974 /* the companion to enter_cache, called when an oap is no longer part of the
1975 * dirty accounting.. so writeback completes or truncate happens before writing
1976 * starts. must be called with the loi lock held. */
1977 static void osc_exit_cache(struct client_obd *cli, struct osc_async_page *oap,
1982 if (!(oap->oap_brw_flags & OBD_BRW_FROM_GRANT)) {
1987 if (~0ul - cli->cl_dirty_sum <= cli->cl_dirty) {
1988 cli->cl_dirty_av = (cli->cl_dirty_av +
1989 (cli->cl_dirty_sum / cli->cl_dirty_num)) / 2;
1990 cli->cl_dirty_num = 0;
1991 cli->cl_dirty_sum = 0;
1993 if (cli->cl_dirty_num)
1994 cli->cl_dirty_av = (cli->cl_dirty_sum / cli->cl_dirty_num);
1997 cli->cl_dirty_num++;
1998 cli->cl_dirty_sum += cli->cl_dirty;
2000 if (cli->cl_dirty > cli->cl_dirty_dmax)
2001 cli->cl_dirty_dmax = cli->cl_dirty;
2002 if (cli->cl_dirty < cli->cl_dirty_dmin || !cli->cl_dirty_dmin)
2003 cli->cl_dirty_dmin = cli->cl_dirty;
2005 oap->oap_brw_flags &= ~OBD_BRW_FROM_GRANT;
2006 cli->cl_dirty -= PAGE_SIZE;
2008 cli->cl_lost_grant += PAGE_SIZE;
2009 CDEBUG(D_CACHE, "lost grant: %lu avail grant: %lu dirty: %lu\n",
2010 cli->cl_lost_grant, cli->cl_avail_grant, cli->cl_dirty);
2016 int osc_prep_async_page(struct obd_export *exp, struct lov_stripe_md *lsm,
2017 struct lov_oinfo *loi, struct page *page,
2018 obd_off offset, struct obd_async_page_ops *ops,
2019 void *data, void **res)
2021 struct osc_async_page *oap;
2024 OBD_ALLOC(oap, sizeof(*oap));
2028 oap->oap_magic = OAP_MAGIC;
2029 oap->oap_cli = &exp->exp_obd->u.cli;
2032 oap->oap_caller_ops = ops;
2033 oap->oap_caller_data = data;
2035 oap->oap_page = page;
2036 oap->oap_obj_off = offset;
2038 INIT_LIST_HEAD(&oap->oap_pending_item);
2039 INIT_LIST_HEAD(&oap->oap_urgent_item);
2040 INIT_LIST_HEAD(&oap->oap_rpc_item);
2042 oap->oap_occ.occ_interrupted = osc_occ_interrupted;
2044 CDEBUG(D_CACHE, "oap %p page %p obj off "LPU64"\n", oap, page, offset);
2049 static int osc_queue_async_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2050 struct lov_oinfo *loi, void *cookie,
2051 int cmd, obd_off off, int count,
2052 obd_flags brw_flags, enum async_flags async_flags)
2054 struct client_obd *cli = &exp->exp_obd->u.cli;
2055 struct osc_async_page *oap;
2056 struct loi_oap_pages *lop;
2060 oap = OAP_FROM_COOKIE(cookie);
2062 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2065 if (!list_empty(&oap->oap_pending_item) ||
2066 !list_empty(&oap->oap_urgent_item) ||
2067 !list_empty(&oap->oap_rpc_item))
2071 loi = &lsm->lsm_oinfo[0];
2073 spin_lock(&cli->cl_loi_list_lock);
2076 oap->oap_async_flags = async_flags;
2077 oap->oap_page_off = off;
2078 oap->oap_count = count;
2079 oap->oap_brw_flags = brw_flags;
2081 if (cmd == OBD_BRW_WRITE) {
2082 rc = osc_enter_cache(cli, loi, oap);
2084 spin_unlock(&cli->cl_loi_list_lock);
2087 lop = &loi->loi_write_lop;
2089 lop = &loi->loi_read_lop;
2092 if (oap->oap_async_flags & ASYNC_URGENT)
2093 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2094 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2095 lop_update_pending(cli, lop, cmd, 1);
2097 loi_list_maint(cli, loi);
2099 LOI_DEBUG(loi, "oap %p page %p added for cmd %d\n", oap, oap->oap_page,
2102 osc_check_rpcs(cli);
2103 spin_unlock(&cli->cl_loi_list_lock);
2108 /* aka (~was & now & flag), but this is more clear :) */
2109 #define SETTING(was, now, flag) (!(was & flag) && (now & flag))
2111 static int osc_set_async_flags(struct obd_export *exp,
2112 struct lov_stripe_md *lsm,
2113 struct lov_oinfo *loi, void *cookie,
2114 obd_flags async_flags)
2116 struct client_obd *cli = &exp->exp_obd->u.cli;
2117 struct loi_oap_pages *lop;
2118 struct osc_async_page *oap;
2122 oap = OAP_FROM_COOKIE(cookie);
2124 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2128 loi = &lsm->lsm_oinfo[0];
2130 if (oap->oap_cmd == OBD_BRW_WRITE) {
2131 lop = &loi->loi_write_lop;
2133 lop = &loi->loi_read_lop;
2136 spin_lock(&cli->cl_loi_list_lock);
2138 if (list_empty(&oap->oap_pending_item))
2139 GOTO(out, rc = -EINVAL);
2141 if ((oap->oap_async_flags & async_flags) == async_flags)
2144 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_READY))
2145 oap->oap_async_flags |= ASYNC_READY;
2147 if (SETTING(oap->oap_async_flags, async_flags, ASYNC_URGENT)) {
2148 if (list_empty(&oap->oap_rpc_item)) {
2149 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2150 loi_list_maint(cli, loi);
2154 LOI_DEBUG(loi, "oap %p page %p has flags %x\n", oap, oap->oap_page,
2155 oap->oap_async_flags);
2157 osc_check_rpcs(cli);
2158 spin_unlock(&cli->cl_loi_list_lock);
2162 static int osc_queue_group_io(struct obd_export *exp, struct lov_stripe_md *lsm,
2163 struct lov_oinfo *loi,
2164 struct obd_io_group *oig, void *cookie,
2165 int cmd, obd_off off, int count,
2166 obd_flags brw_flags,
2167 obd_flags async_flags)
2169 struct client_obd *cli = &exp->exp_obd->u.cli;
2170 struct osc_async_page *oap;
2171 struct loi_oap_pages *lop;
2174 oap = OAP_FROM_COOKIE(cookie);
2176 if (cli->cl_import == NULL || cli->cl_import->imp_invalid)
2179 if (!list_empty(&oap->oap_pending_item) ||
2180 !list_empty(&oap->oap_urgent_item) ||
2181 !list_empty(&oap->oap_rpc_item))
2185 loi = &lsm->lsm_oinfo[0];
2187 spin_lock(&cli->cl_loi_list_lock);
2190 oap->oap_page_off = off;
2191 oap->oap_count = count;
2192 oap->oap_brw_flags = brw_flags;
2193 oap->oap_async_flags = async_flags;
2195 if (cmd == OBD_BRW_WRITE)
2196 lop = &loi->loi_write_lop;
2198 lop = &loi->loi_read_lop;
2200 list_add_tail(&oap->oap_pending_item, &lop->lop_pending_group);
2201 if (oap->oap_async_flags & ASYNC_GROUP_SYNC) {
2203 oig_add_one(oig, &oap->oap_occ);
2206 LOI_DEBUG(loi, "oap %p page %p on group pending\n", oap, oap->oap_page);
2208 spin_unlock(&cli->cl_loi_list_lock);
2213 static void osc_group_to_pending(struct client_obd *cli, struct lov_oinfo *loi,
2214 struct loi_oap_pages *lop, int cmd)
2216 struct list_head *pos, *tmp;
2217 struct osc_async_page *oap;
2219 list_for_each_safe(pos, tmp, &lop->lop_pending_group) {
2220 oap = list_entry(pos, struct osc_async_page, oap_pending_item);
2221 list_del(&oap->oap_pending_item);
2222 list_add_tail(&oap->oap_pending_item, &lop->lop_pending);
2223 list_add(&oap->oap_urgent_item, &lop->lop_urgent);
2224 lop_update_pending(cli, lop, cmd, 1);
2226 loi_list_maint(cli, loi);
2229 static int osc_trigger_group_io(struct obd_export *exp,
2230 struct lov_stripe_md *lsm,
2231 struct lov_oinfo *loi,
2232 struct obd_io_group *oig)
2234 struct client_obd *cli = &exp->exp_obd->u.cli;
2238 loi = &lsm->lsm_oinfo[0];
2240 spin_lock(&cli->cl_loi_list_lock);
2242 osc_group_to_pending(cli, loi, &loi->loi_write_lop, OBD_BRW_WRITE);
2243 osc_group_to_pending(cli, loi, &loi->loi_read_lop, OBD_BRW_READ);
2245 osc_check_rpcs(cli);
2246 spin_unlock(&cli->cl_loi_list_lock);
2251 static int osc_teardown_async_page(struct obd_export *exp,
2252 struct lov_stripe_md *lsm,
2253 struct lov_oinfo *loi, void *cookie)
2255 struct client_obd *cli = &exp->exp_obd->u.cli;
2256 struct loi_oap_pages *lop;
2257 struct osc_async_page *oap;
2261 oap = OAP_FROM_COOKIE(cookie);
2264 loi = &lsm->lsm_oinfo[0];
2266 if (oap->oap_cmd == OBD_BRW_WRITE) {
2267 lop = &loi->loi_write_lop;
2269 lop = &loi->loi_read_lop;
2272 spin_lock(&cli->cl_loi_list_lock);
2274 if (!list_empty(&oap->oap_rpc_item))
2275 GOTO(out, rc = -EBUSY);
2277 osc_exit_cache(cli, oap, 0);
2278 osc_wake_cache_waiters(cli);
2280 if (!list_empty(&oap->oap_urgent_item)) {
2281 list_del_init(&oap->oap_urgent_item);
2282 oap->oap_async_flags &= ~ASYNC_URGENT;
2284 if (!list_empty(&oap->oap_pending_item)) {
2285 list_del_init(&oap->oap_pending_item);
2286 lop_update_pending(cli, lop, oap->oap_cmd, -1);
2288 loi_list_maint(cli, loi);
2290 LOI_DEBUG(loi, "oap %p page %p torn down\n", oap, oap->oap_page);
2292 spin_unlock(&cli->cl_loi_list_lock);
2294 OBD_FREE(oap, sizeof(*oap));
2299 /* Note: caller will lock/unlock, and set uptodate on the pages */
2300 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2301 static int sanosc_brw_read(struct obd_export *exp, struct obdo *oa,
2302 struct lov_stripe_md *lsm, obd_count page_count,
2303 struct brw_page *pga)
2305 struct ptlrpc_request *request = NULL;
2306 struct ost_body *body;
2307 struct niobuf_remote *nioptr;
2308 struct obd_ioobj *iooptr;
2309 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2313 /* XXX does not handle 'new' brw protocol */
2315 size[1] = sizeof(struct obd_ioobj);
2316 size[2] = page_count * sizeof(*nioptr);
2318 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2319 OST_SAN_READ, 3, size, NULL);
2323 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof(*body));
2324 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof(*iooptr));
2325 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2326 sizeof(*nioptr) * page_count);
2328 memcpy(&body->oa, oa, sizeof(body->oa));
2330 obdo_to_ioobj(oa, iooptr);
2331 iooptr->ioo_bufcnt = page_count;
2333 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2334 LASSERT(PageLocked(pga[mapped].pg));
2335 LASSERT(mapped == 0 ||
2336 pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2338 nioptr->offset = pga[mapped].disk_offset;
2339 nioptr->len = pga[mapped].count;
2340 nioptr->flags = pga[mapped].flag;
2343 size[1] = page_count * sizeof(*nioptr);
2344 request->rq_replen = lustre_msg_size(2, size);
2346 rc = ptlrpc_queue_wait(request);
2350 body = lustre_swab_repbuf(request, 0, sizeof(*body),
2351 lustre_swab_ost_body);
2353 CERROR("Can't unpack body\n");
2354 GOTO(out_req, rc = -EPROTO);
2357 memcpy(oa, &body->oa, sizeof(*oa));
2359 swab = lustre_msg_swabbed(request->rq_repmsg);
2360 LASSERT_REPSWAB(request, 1);
2361 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2363 /* nioptr missing or short */
2364 GOTO(out_req, rc = -EPROTO);
2368 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2369 struct page *page = pga[mapped].pg;
2370 struct buffer_head *bh;
2374 lustre_swab_niobuf_remote (nioptr);
2376 /* got san device associated */
2377 LASSERT(exp->exp_obd != NULL);
2378 dev = exp->exp_obd->u.cli.cl_sandev;
2381 if (!nioptr->offset) {
2382 CDEBUG(D_PAGE, "hole at ino %lu; index %ld\n",
2383 page->mapping->host->i_ino,
2385 memset(page_address(page), 0, PAGE_SIZE);
2389 if (!page->buffers) {
2390 create_empty_buffers(page, dev, PAGE_SIZE);
2393 clear_bit(BH_New, &bh->b_state);
2394 set_bit(BH_Mapped, &bh->b_state);
2395 bh->b_blocknr = (unsigned long)nioptr->offset;
2397 clear_bit(BH_Uptodate, &bh->b_state);
2399 ll_rw_block(READ, 1, &bh);
2403 /* if buffer already existed, it must be the
2404 * one we mapped before, check it */
2405 LASSERT(!test_bit(BH_New, &bh->b_state));
2406 LASSERT(test_bit(BH_Mapped, &bh->b_state));
2407 LASSERT(bh->b_blocknr == (unsigned long)nioptr->offset);
2409 /* wait it's io completion */
2410 if (test_bit(BH_Lock, &bh->b_state))
2413 if (!test_bit(BH_Uptodate, &bh->b_state))
2414 ll_rw_block(READ, 1, &bh);
2418 /* must do syncronous write here */
2420 if (!buffer_uptodate(bh)) {
2428 ptlrpc_req_finished(request);
2432 static int sanosc_brw_write(struct obd_export *exp, struct obdo *oa,
2433 struct lov_stripe_md *lsm, obd_count page_count,
2434 struct brw_page *pga)
2436 struct ptlrpc_request *request = NULL;
2437 struct ost_body *body;
2438 struct niobuf_remote *nioptr;
2439 struct obd_ioobj *iooptr;
2440 int rc, size[3] = {sizeof(*body)}, mapped = 0;
2444 size[1] = sizeof(struct obd_ioobj);
2445 size[2] = page_count * sizeof(*nioptr);
2447 request = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
2448 OST_SAN_WRITE, 3, size, NULL);
2452 body = lustre_msg_buf(request->rq_reqmsg, 0, sizeof (*body));
2453 iooptr = lustre_msg_buf(request->rq_reqmsg, 1, sizeof (*iooptr));
2454 nioptr = lustre_msg_buf(request->rq_reqmsg, 2,
2455 sizeof (*nioptr) * page_count);
2457 memcpy(&body->oa, oa, sizeof(body->oa));
2459 obdo_to_ioobj(oa, iooptr);
2460 iooptr->ioo_bufcnt = page_count;
2463 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2464 LASSERT(PageLocked(pga[mapped].pg));
2465 LASSERT(mapped == 0 ||
2466 pga[mapped].disk_offset > pga[mapped - 1].disk_offset);
2468 nioptr->offset = pga[mapped].disk_offset;
2469 nioptr->len = pga[mapped].count;
2470 nioptr->flags = pga[mapped].flag;
2473 size[1] = page_count * sizeof(*nioptr);
2474 request->rq_replen = lustre_msg_size(2, size);
2476 rc = ptlrpc_queue_wait(request);
2480 swab = lustre_msg_swabbed (request->rq_repmsg);
2481 LASSERT_REPSWAB (request, 1);
2482 nioptr = lustre_msg_buf(request->rq_repmsg, 1, size[1]);
2484 CERROR("absent/short niobuf array\n");
2485 GOTO(out_req, rc = -EPROTO);
2489 for (mapped = 0; mapped < page_count; mapped++, nioptr++) {
2490 struct page *page = pga[mapped].pg;
2491 struct buffer_head *bh;
2495 lustre_swab_niobuf_remote (nioptr);
2497 /* got san device associated */
2498 LASSERT(exp->exp_obd != NULL);
2499 dev = exp->exp_obd->u.cli.cl_sandev;
2501 if (!page->buffers) {
2502 create_empty_buffers(page, dev, PAGE_SIZE);
2505 LASSERT(!test_bit(BH_New, &page->buffers->b_state));
2506 LASSERT(test_bit(BH_Mapped, &page->buffers->b_state));
2507 LASSERT(page->buffers->b_blocknr ==
2508 (unsigned long)nioptr->offset);
2514 /* if buffer locked, wait it's io completion */
2515 if (test_bit(BH_Lock, &bh->b_state))
2518 clear_bit(BH_New, &bh->b_state);
2519 set_bit(BH_Mapped, &bh->b_state);
2521 /* override the block nr */
2522 bh->b_blocknr = (unsigned long)nioptr->offset;
2524 /* we are about to write it, so set it
2526 * page lock should garentee no race condition here */
2527 set_bit(BH_Uptodate, &bh->b_state);
2528 set_bit(BH_Dirty, &bh->b_state);
2530 ll_rw_block(WRITE, 1, &bh);
2532 /* must do syncronous write here */
2534 if (!buffer_uptodate(bh) || test_bit(BH_Dirty, &bh->b_state)) {
2542 ptlrpc_req_finished(request);
2546 static int sanosc_brw(int cmd, struct obd_export *exp, struct obdo *oa,
2547 struct lov_stripe_md *lsm, obd_count page_count,
2548 struct brw_page *pga, struct obd_trans_info *oti)
2552 while (page_count) {
2553 obd_count pages_per_brw;
2556 if (page_count > PTLRPC_MAX_BRW_PAGES)
2557 pages_per_brw = PTLRPC_MAX_BRW_PAGES;
2559 pages_per_brw = page_count;
2561 if (cmd & OBD_BRW_WRITE)
2562 rc = sanosc_brw_write(exp, oa, lsm, pages_per_brw,pga);
2564 rc = sanosc_brw_read(exp, oa, lsm, pages_per_brw, pga);
2569 page_count -= pages_per_brw;
2570 pga += pages_per_brw;
2577 static void osc_set_data_with_check(struct lustre_handle *lockh, void *data)
2579 struct ldlm_lock *lock = ldlm_handle2lock(lockh);
2582 CERROR("lockh %p, data %p - client evicted?\n", lockh, data);
2586 lock_res_and_lock(lock);
2588 if (lock->l_ast_data && lock->l_ast_data != data) {
2589 struct inode *new_inode = data;
2590 struct inode *old_inode = lock->l_ast_data;
2591 if (!(old_inode->i_state & I_FREEING))
2592 LDLM_ERROR(lock, "inconsistent l_ast_data found");
2593 LASSERTF(old_inode->i_state & I_FREEING,
2594 "Found existing inode %p/%lu/%u state %lu in lock: "
2595 "setting data to %p/%lu/%u\n", old_inode,
2596 old_inode->i_ino, old_inode->i_generation,
2598 new_inode, new_inode->i_ino, new_inode->i_generation);
2601 lock->l_ast_data = data;
2602 unlock_res_and_lock(lock);
2603 LDLM_LOCK_PUT(lock);
2606 static int osc_change_cbdata(struct obd_export *exp, struct lov_stripe_md *lsm,
2607 ldlm_iterator_t replace, void *data)
2609 struct ldlm_res_id res_id = { .name = {0} };
2610 struct obd_device *obd = class_exp2obd(exp);
2612 res_id.name[0] = lsm->lsm_object_id;
2613 res_id.name[2] = lsm->lsm_object_gr;
2614 ldlm_change_cbdata(obd->obd_namespace, &res_id, replace, data);
2618 static int osc_enqueue(struct obd_export *exp, struct lov_stripe_md *lsm,
2619 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2620 int *flags, void *bl_cb, void *cp_cb, void *gl_cb,
2621 void *data, __u32 lvb_len, void *lvb_swabber,
2622 struct lustre_handle *lockh)
2624 struct obd_device *obd = exp->exp_obd;
2625 struct ldlm_res_id res_id = { .name = {0} };
2627 struct ldlm_reply *rep;
2628 struct ptlrpc_request *req = NULL;
2632 res_id.name[0] = lsm->lsm_object_id;
2633 res_id.name[2] = lsm->lsm_object_gr;
2635 /* Filesystem lock extents are extended to page boundaries so that
2636 * dealing with the page cache is a little smoother. */
2637 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2638 policy->l_extent.end |= ~PAGE_MASK;
2640 if (lsm->lsm_oinfo->loi_kms_valid == 0)
2643 /* Next, search for already existing extent locks that will cover us */
2644 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type, policy, mode,
2647 if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2648 /* return immediately if no credential held */
2649 ldlm_lock_decref(lockh, mode);
2653 osc_set_data_with_check(lockh, data);
2654 if (*flags & LDLM_FL_HAS_INTENT) {
2655 /* I would like to be able to ASSERT here that rss <=
2656 * kms, but I can't, for reasons which are explained in
2659 /* We already have a lock, and it's referenced */
2663 /* If we're trying to read, we also search for an existing PW lock. The
2664 * VFS and page cache already protect us locally, so lots of readers/
2665 * writers can share a single PW lock.
2667 * There are problems with conversion deadlocks, so instead of
2668 * converting a read lock to a write lock, we'll just enqueue a new
2671 * At some point we should cancel the read lock instead of making them
2672 * send us a blocking callback, but there are problems with canceling
2673 * locks out from other users right now, too. */
2675 if (mode == LCK_PR) {
2676 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2677 policy, LCK_PW, lockh);
2679 if (ptlrpcs_check_cred(obd->u.cli.cl_import)) {
2680 /* return immediately if no credential held */
2681 ldlm_lock_decref(lockh, LCK_PW);
2685 /* FIXME: This is not incredibly elegant, but it might
2686 * be more elegant than adding another parameter to
2687 * lock_match. I want a second opinion. */
2688 ldlm_lock_addref(lockh, LCK_PR);
2689 ldlm_lock_decref(lockh, LCK_PW);
2690 osc_set_data_with_check(lockh, data);
2694 if (mode == LCK_PW) {
2695 rc = ldlm_lock_match(obd->obd_namespace, 0, &res_id, type,
2696 policy, LCK_PR, lockh);
2698 rc = ldlm_cli_convert(lockh, mode, flags);
2700 /* Update readers/writers accounting */
2701 ldlm_lock_addref(lockh, LCK_PW);
2702 ldlm_lock_decref(lockh, LCK_PR);
2703 osc_set_data_with_check(lockh, data);
2706 /* If the conversion failed, we need to drop refcount
2707 on matched lock before we get new one */
2708 /* XXX Won't it save us some efforts if we cancel PR
2709 lock here? We are going to take PW lock anyway and it
2710 will invalidate PR lock */
2711 ldlm_lock_decref(lockh, LCK_PR);
2712 if (rc != EDEADLOCK) {
2719 if (*flags & LDLM_FL_HAS_INTENT) {
2720 int size[2] = {0, sizeof(struct ldlm_request)};
2722 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_DLM_VERSION,
2723 LDLM_ENQUEUE, 2, size, NULL);
2727 size[0] = sizeof(*rep);
2728 size[1] = sizeof(lvb);
2729 req->rq_replen = lustre_msg_size(2, size);
2731 rc = ldlm_cli_enqueue(exp, req, obd->obd_namespace, res_id, type,
2732 policy, mode, flags, bl_cb, cp_cb, gl_cb, data,
2733 &lvb, sizeof(lvb), lustre_swab_ost_lvb, lockh);
2735 if (rc == ELDLM_LOCK_ABORTED) {
2736 /* swabbed by ldlm_cli_enqueue() */
2737 LASSERT_REPSWABBED(req, 0);
2738 rep = lustre_msg_buf(req->rq_repmsg, 0, sizeof(*rep));
2739 LASSERT(rep != NULL);
2740 if (rep->lock_policy_res1)
2741 rc = rep->lock_policy_res1;
2743 ptlrpc_req_finished(req);
2746 if ((*flags & LDLM_FL_HAS_INTENT && rc == ELDLM_LOCK_ABORTED) || !rc) {
2747 CDEBUG(D_INODE, "received kms == "LPU64", blocks == "LPU64"\n",
2748 lvb.lvb_size, lvb.lvb_blocks);
2749 lsm->lsm_oinfo->loi_rss = lvb.lvb_size;
2750 lsm->lsm_oinfo->loi_mtime = lvb.lvb_mtime;
2751 lsm->lsm_oinfo->loi_blocks = lvb.lvb_blocks;
2757 static int osc_match(struct obd_export *exp, struct lov_stripe_md *lsm,
2758 __u32 type, ldlm_policy_data_t *policy, __u32 mode,
2759 int *flags, void *data, struct lustre_handle *lockh)
2761 struct ldlm_res_id res_id = { .name = {0} };
2762 struct obd_device *obd = exp->exp_obd;
2766 res_id.name[0] = lsm->lsm_object_id;
2767 res_id.name[2] = lsm->lsm_object_gr;
2769 OBD_FAIL_RETURN(OBD_FAIL_OSC_MATCH, -EIO);
2771 /* Filesystem lock extents are extended to page boundaries so that
2772 * dealing with the page cache is a little smoother */
2773 policy->l_extent.start -= policy->l_extent.start & ~PAGE_MASK;
2774 policy->l_extent.end |= ~PAGE_MASK;
2776 /* Next, search for already existing extent locks that will cover us */
2777 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2778 policy, mode, lockh);
2780 // if (!(*flags & LDLM_FL_TEST_LOCK))
2781 osc_set_data_with_check(lockh, data);
2784 /* If we're trying to read, we also search for an existing PW lock. The
2785 * VFS and page cache already protect us locally, so lots of readers/
2786 * writers can share a single PW lock. */
2787 if (mode == LCK_PR) {
2788 rc = ldlm_lock_match(obd->obd_namespace, *flags, &res_id, type,
2789 policy, LCK_PW, lockh);
2790 if (rc == 1 && !(*flags & LDLM_FL_TEST_LOCK)) {
2791 /* FIXME: This is not incredibly elegant, but it might
2792 * be more elegant than adding another parameter to
2793 * lock_match. I want a second opinion. */
2794 osc_set_data_with_check(lockh, data);
2795 ldlm_lock_addref(lockh, LCK_PR);
2796 ldlm_lock_decref(lockh, LCK_PW);
2802 static int osc_cancel(struct obd_export *exp, struct lov_stripe_md *md,
2803 __u32 mode, struct lustre_handle *lockh)
2807 ldlm_lock_decref(lockh, mode);
2811 static int osc_cancel_unused(struct obd_export *exp,
2812 struct lov_stripe_md *lsm,
2813 int flags, void *opaque)
2815 struct obd_device *obd = class_exp2obd(exp);
2816 struct ldlm_res_id res_id = { .name = {0} }, *resp = NULL;
2819 res_id.name[0] = lsm->lsm_object_id;
2820 res_id.name[2] = lsm->lsm_object_gr;
2824 return ldlm_cli_cancel_unused(obd->obd_namespace, resp, flags, opaque);
2827 static int osc_statfs(struct obd_device *obd, struct obd_statfs *osfs,
2828 unsigned long max_age)
2830 struct obd_statfs *msfs;
2831 struct ptlrpc_request *request;
2832 int rc, size = sizeof(*osfs);
2835 /* We could possibly pass max_age in the request (as an absolute
2836 * timestamp or a "seconds.usec ago") so the target can avoid doing
2837 * extra calls into the filesystem if that isn't necessary (e.g.
2838 * during mount that would help a bit). Having relative timestamps
2839 * is not so great if request processing is slow, while absolute
2840 * timestamps are not ideal because they need time synchronization. */
2841 request = ptlrpc_prep_req(obd->u.cli.cl_import, LUSTRE_OBD_VERSION,
2842 OST_STATFS, 0, NULL, NULL);
2846 request->rq_replen = lustre_msg_size(1, &size);
2847 request->rq_request_portal = OST_CREATE_PORTAL; //XXX FIXME bug 249
2849 rc = ptlrpc_queue_wait(request);
2853 msfs = lustre_swab_repbuf(request, 0, sizeof(*msfs),
2854 lustre_swab_obd_statfs);
2856 CERROR("Can't unpack obd_statfs\n");
2857 GOTO(out, rc = -EPROTO);
2860 memcpy(osfs, msfs, sizeof(*osfs));
2864 ptlrpc_req_finished(request);
2868 /* Retrieve object striping information.
2870 * @lmmu is a pointer to an in-core struct with lmm_ost_count indicating
2871 * the maximum number of OST indices which will fit in the user buffer.
2872 * lmm_magic must be LOV_MAGIC (we only use 1 slot here).
2874 static int osc_getstripe(struct lov_stripe_md *lsm, struct lov_user_md *lump)
2876 struct lov_user_md lum, *lumk;
2883 rc = copy_from_user(&lum, lump, sizeof(lum));
2887 if (lum.lmm_magic != LOV_USER_MAGIC)
2890 if (lum.lmm_stripe_count > 0) {
2891 lum_size = sizeof(lum) + sizeof(lum.lmm_objects[0]);
2892 OBD_ALLOC(lumk, lum_size);
2896 lumk->lmm_objects[0].l_object_id = lsm->lsm_object_id;
2897 lumk->lmm_objects[0].l_object_gr = lsm->lsm_object_gr;
2899 lum_size = sizeof(lum);
2903 lumk->lmm_object_id = lsm->lsm_object_id;
2904 lumk->lmm_object_gr = lsm->lsm_object_gr;
2905 lumk->lmm_stripe_count = 1;
2907 if (copy_to_user(lump, lumk, lum_size))
2911 OBD_FREE(lumk, lum_size);
2916 static int osc_iocontrol(unsigned int cmd, struct obd_export *exp, int len,
2917 void *karg, void *uarg)
2919 struct obd_device *obd = exp->exp_obd;
2920 struct obd_ioctl_data *data = karg;
2924 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
2927 if (!try_module_get(THIS_MODULE)) {
2928 CERROR("Can't get module. Is it alive?");
2933 case OBD_IOC_LOV_GET_CONFIG: {
2935 struct lov_desc *desc;
2936 struct obd_uuid uuid;
2940 if (obd_ioctl_getdata(&buf, &len, (void *)uarg))
2941 GOTO(out, err = -EINVAL);
2943 data = (struct obd_ioctl_data *)buf;
2945 if (sizeof(*desc) > data->ioc_inllen1) {
2947 GOTO(out, err = -EINVAL);
2950 if (data->ioc_inllen2 < sizeof(uuid)) {
2952 GOTO(out, err = -EINVAL);
2955 if (data->ioc_inllen3 < sizeof(__u32)) {
2957 GOTO(out, err = -EINVAL);
2960 desc = (struct lov_desc *)data->ioc_inlbuf1;
2961 desc->ld_tgt_count = 1;
2962 desc->ld_active_tgt_count = 1;
2963 desc->ld_default_stripe_count = 1;
2964 desc->ld_default_stripe_size = 0;
2965 desc->ld_default_stripe_offset = 0;
2966 desc->ld_pattern = 0;
2967 memcpy(&desc->ld_uuid, &obd->obd_uuid, sizeof(uuid));
2968 memcpy(data->ioc_inlbuf2, &obd->obd_uuid, sizeof(uuid));
2969 *((__u32 *)data->ioc_inlbuf3) = 1;
2971 err = copy_to_user((void *)uarg, buf, len);
2974 obd_ioctl_freedata(buf, len);
2977 case LL_IOC_LOV_SETSTRIPE:
2978 err = obd_alloc_memmd(exp, karg);
2982 case LL_IOC_LOV_GETSTRIPE:
2983 err = osc_getstripe(karg, uarg);
2985 case OBD_IOC_CLIENT_RECOVER:
2986 err = ptlrpc_recover_import(obd->u.cli.cl_import,
2991 case IOC_OSC_SET_ACTIVE:
2992 err = ptlrpc_set_import_active(obd->u.cli.cl_import,
2995 case IOC_OSC_CTL_RECOVERY:
2996 err = ptlrpc_import_control_recovery(obd->u.cli.cl_import,
3000 CDEBUG(D_INODE, "unrecognised ioctl %#x by %s\n", cmd, current->comm);
3001 GOTO(out, err = -ENOTTY);
3004 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3007 module_put(THIS_MODULE);
3012 static int osc_get_info(struct obd_export *exp, __u32 keylen,
3013 void *key, __u32 *vallen, void *val)
3016 if (!vallen || !val)
3019 if (keylen > strlen("lock_to_stripe") &&
3020 strcmp(key, "lock_to_stripe") == 0) {
3021 __u32 *stripe = val;
3022 *vallen = sizeof(*stripe);
3025 } else if (keylen >= strlen("last_id") && strcmp(key, "last_id") == 0) {
3026 struct ptlrpc_request *req;
3028 char *bufs[1] = {key};
3030 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3031 OST_GET_INFO, 1, (int *)&keylen, bufs);
3035 req->rq_replen = lustre_msg_size(1, (int *)vallen);
3036 rc = ptlrpc_queue_wait(req);
3040 reply = lustre_swab_repbuf(req, 0, sizeof(*reply),
3041 lustre_swab_ost_last_id);
3042 if (reply == NULL) {
3043 CERROR("Can't unpack OST last ID\n");
3044 GOTO(out, rc = -EPROTO);
3046 *((obd_id *)val) = *reply;
3048 ptlrpc_req_finished(req);
3050 } else if (keylen == 10 && strcmp(key, "client_nid") == 0) {
3051 struct ptlrpc_connection * conn;
3052 ptl_nid_t * nid = val;
3053 ptl_process_id_t id;
3056 *vallen = sizeof(*nid);
3057 conn = class_exp2cliimp(exp)->imp_connection;
3059 if (!conn || !conn->c_peer.peer_ni)
3062 rc = PtlGetId(conn->c_peer.peer_ni->pni_ni_h, &id);
3071 static int osc_set_info(struct obd_export *exp, obd_count keylen,
3072 void *key, obd_count vallen, void *val)
3074 struct obd_device *obd = exp->exp_obd;
3075 struct obd_import *imp = class_exp2cliimp(exp);
3076 struct llog_ctxt *ctxt;
3080 if (keylen == strlen("unlinked") &&
3081 memcmp(key, "unlinked", keylen) == 0) {
3082 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3083 spin_lock(&oscc->oscc_lock);
3084 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3085 spin_unlock(&oscc->oscc_lock);
3089 if (keylen == strlen("unrecovery") &&
3090 memcmp(key, "unrecovery", keylen) == 0) {
3091 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3092 spin_lock(&oscc->oscc_lock);
3093 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3094 spin_unlock(&oscc->oscc_lock);
3098 if (keylen == strlen("initial_recov") &&
3099 memcmp(key, "initial_recov", strlen("initial_recov")) == 0) {
3100 struct obd_import *imp = class_exp2cliimp(exp);
3101 if (vallen != sizeof(int))
3103 imp->imp_initial_recov = *(int *)val;
3104 CDEBUG(D_HA, "%s: set imp_no_init_recov = %d\n",
3105 exp->exp_obd->obd_name,
3106 imp->imp_initial_recov);
3110 if (keylen == strlen("async") &&
3111 memcmp(key, "async", keylen) == 0) {
3112 struct client_obd *cl = &obd->u.cli;
3113 if (vallen != sizeof(int))
3115 cl->cl_async = *(int *)val;
3116 CDEBUG(D_HA, "%s: set async = %d\n",
3117 obd->obd_name, cl->cl_async);
3121 if (keylen == 5 && strcmp(key, "audit") == 0) {
3122 struct ptlrpc_request *req;
3123 char *bufs[2] = {key, val};
3124 int size[2] = {keylen, vallen};
3126 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3127 OST_SET_INFO, 2, size, bufs);
3131 req->rq_replen = lustre_msg_size(0, size);
3132 lustre_swab_reqbuf(req, 1, sizeof(struct audit_attr_msg),
3133 lustre_swab_audit_attr);
3134 rc = ptlrpc_queue_wait(req);
3136 ptlrpc_req_finished(req);
3140 if (keylen == 9 && strcmp(key, "audit_obj") == 0) {
3141 struct ptlrpc_request *req;
3142 char *bufs[2] = {key, val};
3143 int size[2] = {keylen, vallen};
3145 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3146 OST_SET_INFO, 2, size, bufs);
3150 req->rq_replen = lustre_msg_size(0, size);
3151 lustre_swab_reqbuf(req, 1, sizeof(struct obdo),
3153 rc = ptlrpc_queue_wait(req);
3155 ptlrpc_req_finished(req);
3159 if (keylen == 8 && memcmp(key, "auditlog", 8) == 0) {
3160 struct ptlrpc_request *req;
3161 char *bufs[2] = {key, val};
3162 int size[2] = {keylen, vallen};
3164 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3165 OST_SET_INFO, 2, size, bufs);
3169 req->rq_replen = lustre_msg_size(0, size);
3170 lustre_swab_reqbuf(req, 1, sizeof(struct audit_msg),
3171 lustre_swab_audit_msg);
3172 rc = ptlrpc_queue_wait(req);
3174 ptlrpc_req_finished(req);
3178 if (keylen == strlen("sec") && memcmp(key, "sec", keylen) == 0) {
3179 struct client_obd *cli = &exp->exp_obd->u.cli;
3181 cli->cl_sec_flavor = ptlrpcs_name2flavor(val);
3182 if (cli->cl_sec_flavor == PTLRPCS_FLVR_INVALID) {
3183 CERROR("unrecognized security flavor %s\n", (char*) val);
3190 if (keylen == strlen("sec_flags") &&
3191 memcmp(key, "sec_flags", keylen) == 0) {
3192 struct client_obd *cli = &exp->exp_obd->u.cli;
3194 cli->cl_sec_flags = *((unsigned long *) val);
3198 if (keylen == strlen("flush_cred") &&
3199 memcmp(key, "flush_cred", keylen) == 0) {
3200 struct client_obd *cli = &exp->exp_obd->u.cli;
3203 ptlrpcs_import_flush_current_creds(cli->cl_import);
3206 if (keylen == strlen("crypto_cb") &&
3207 memcmp(key, "crypto_cb", keylen) == 0) {
3208 LASSERT(vallen == sizeof(crypt_cb_t));
3209 osc_crypt_cb = (crypt_cb_t)val;
3213 if (keylen == 8 && memcmp(key, "capa_key", 8) == 0) {
3214 struct ptlrpc_request *req;
3215 char *bufs[2] = {key, val};
3216 unsigned long irqflags;
3217 int rc, size[2] = {keylen, vallen};
3219 LASSERT(vallen == sizeof(struct lustre_capa_key));
3221 req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OBD_VERSION,
3222 OST_SET_INFO, 2, size, bufs);
3226 spin_lock_irqsave (&req->rq_lock, irqflags);
3228 spin_unlock_irqrestore (&req->rq_lock, irqflags);
3230 req->rq_replen = lustre_msg_size(0, NULL);
3231 rc = ptlrpc_queue_wait(req);
3232 ptlrpc_req_finished(req);
3236 if (keylen == strlen("setext") &&
3237 memcmp(key, "setext", keylen) == 0) {
3238 struct client_obd *cli = &exp->exp_obd->u.cli;
3239 struct osc_creator *oscc = &cli->cl_oscc;
3240 struct fid_extent *ext = val;
3242 oscc->oscc_next_id = (obd_id)ext->fe_start;
3246 if (keylen < strlen("mds_conn") ||
3247 memcmp(key, "mds_conn", keylen) != 0)
3250 ctxt = llog_get_context(&exp->exp_obd->obd_llogs,
3251 LLOG_UNLINK_ORIG_CTXT);
3254 rc = llog_initiator_connect(ctxt);
3256 CERROR("cannot establish the connect for "
3257 "ctxt %p: %d\n", ctxt, rc);
3260 imp->imp_server_timeout = 1;
3261 CDEBUG(D_HA, "pinging OST %s\n", imp->imp_target_uuid.uuid);
3262 imp->imp_pingable = 1;
3268 static struct llog_operations osc_size_repl_logops = {
3269 lop_cancel: llog_obd_repl_cancel
3272 static struct llog_operations osc_unlink_orig_logops;
3274 static int osc_llog_init(struct obd_device *obd, struct obd_llogs *llogs,
3275 struct obd_device *tgt, int count,
3276 struct llog_catid *catid)
3281 osc_unlink_orig_logops = llog_lvfs_ops;
3282 osc_unlink_orig_logops.lop_setup = llog_obd_origin_setup;
3283 osc_unlink_orig_logops.lop_cleanup = llog_catalog_cleanup;
3284 osc_unlink_orig_logops.lop_add = llog_catalog_add;
3285 osc_unlink_orig_logops.lop_connect = llog_origin_connect;
3287 rc = obd_llog_setup(obd, llogs, LLOG_UNLINK_ORIG_CTXT, tgt, count,
3288 &catid->lci_logid, &osc_unlink_orig_logops);
3292 rc = obd_llog_setup(obd, llogs, LLOG_SIZE_REPL_CTXT, tgt, count, NULL,
3293 &osc_size_repl_logops);
3297 static int osc_llog_finish(struct obd_device *obd,
3298 struct obd_llogs *llogs, int count)
3303 rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_UNLINK_ORIG_CTXT));
3307 rc = obd_llog_cleanup(llog_get_context(llogs, LLOG_SIZE_REPL_CTXT));
3311 static int osc_connect(struct lustre_handle *exph,
3312 struct obd_device *obd, struct obd_uuid *cluuid,
3313 struct obd_connect_data *data,
3314 unsigned long connect_flags)
3318 rc = client_connect_import(exph, obd, cluuid, data, connect_flags);
3322 static int osc_disconnect(struct obd_export *exp, unsigned long flags)
3324 struct obd_device *obd = class_exp2obd(exp);
3325 struct llog_ctxt *ctxt;
3329 ctxt = llog_get_context(&obd->obd_llogs, LLOG_SIZE_REPL_CTXT);
3330 if (obd->u.cli.cl_conn_count == 1)
3331 /* flush any remaining cancel messages out to the target */
3332 llog_sync(ctxt, exp);
3334 rc = client_disconnect_export(exp, flags);
3338 static int osc_import_event(struct obd_device *obd,
3339 struct obd_import *imp,
3340 enum obd_import_event event)
3342 struct client_obd *cli;
3346 LASSERT(imp->imp_obd == obd);
3349 case IMP_EVENT_DISCON: {
3350 /* Only do this on the MDS OSC's */
3351 if (imp->imp_server_timeout) {
3352 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3354 spin_lock(&oscc->oscc_lock);
3355 oscc->oscc_flags |= OSCC_FLAG_RECOVERING;
3356 spin_unlock(&oscc->oscc_lock);
3360 case IMP_EVENT_INACTIVE: {
3361 if (obd->obd_observer)
3362 rc = obd_notify(obd->obd_observer, obd, 0, 0);
3365 case IMP_EVENT_INVALIDATE: {
3366 struct ldlm_namespace *ns = obd->obd_namespace;
3370 spin_lock(&cli->cl_loi_list_lock);
3371 cli->cl_avail_grant = 0;
3372 cli->cl_lost_grant = 0;
3373 /* all pages go to failing rpcs due to the invalid import */
3374 osc_check_rpcs(cli);
3375 spin_unlock(&cli->cl_loi_list_lock);
3377 ldlm_namespace_cleanup(ns, LDLM_FL_LOCAL_ONLY);
3381 case IMP_EVENT_ACTIVE: {
3382 /* Only do this on the MDS OSC's */
3383 if (imp->imp_server_timeout) {
3384 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3386 spin_lock(&oscc->oscc_lock);
3387 oscc->oscc_flags &= ~OSCC_FLAG_NOSPC;
3388 spin_unlock(&oscc->oscc_lock);
3391 if (obd->obd_observer)
3392 rc = obd_notify(obd->obd_observer, obd, 1, 0);
3396 CERROR("Unknown import event %d\n", event);
3402 static int osc_attach(struct obd_device *dev, obd_count len, void *data)
3404 struct lprocfs_static_vars lvars;
3408 lprocfs_init_vars(osc,&lvars);
3409 rc = lprocfs_obd_attach(dev, lvars.obd_vars);
3413 rc = lproc_osc_attach_seqstat(dev);
3415 lprocfs_obd_detach(dev);
3419 ptlrpc_lprocfs_register_obd(dev);
3423 static int osc_detach(struct obd_device *dev)
3425 ptlrpc_lprocfs_unregister_obd(dev);
3426 return lprocfs_obd_detach(dev);
3429 static int osc_setup(struct obd_device *obd, obd_count len, void *buf)
3433 rc = ptlrpcd_addref();
3437 rc = client_obd_setup(obd, len, buf);
3446 static int osc_cleanup(struct obd_device *obd, int flags)
3448 struct osc_creator *oscc = &obd->u.cli.cl_oscc;
3451 rc = ldlm_cli_cancel_unused(obd->obd_namespace, NULL,
3452 LDLM_FL_CONFIG_CHANGE, NULL);
3456 spin_lock(&oscc->oscc_lock);
3457 oscc->oscc_flags &= ~OSCC_FLAG_RECOVERING;
3458 oscc->oscc_flags |= OSCC_FLAG_EXITING;
3459 spin_unlock(&oscc->oscc_lock);
3461 rc = client_obd_cleanup(obd, flags);
3467 struct obd_ops osc_obd_ops = {
3468 .o_owner = THIS_MODULE,
3469 .o_attach = osc_attach,
3470 .o_detach = osc_detach,
3471 .o_setup = osc_setup,
3472 .o_cleanup = osc_cleanup,
3473 .o_add_conn = client_import_add_conn,
3474 .o_del_conn = client_import_del_conn,
3475 .o_connect = osc_connect,
3476 .o_disconnect = osc_disconnect,
3477 .o_statfs = osc_statfs,
3478 .o_packmd = osc_packmd,
3479 .o_unpackmd = osc_unpackmd,
3480 .o_create = osc_create,
3481 .o_destroy = osc_destroy,
3482 .o_getattr = osc_getattr,
3483 .o_getattr_async = osc_getattr_async,
3484 .o_setattr = osc_setattr,
3486 .o_brw_async = osc_brw_async,
3487 .o_prep_async_page = osc_prep_async_page,
3488 .o_queue_async_io = osc_queue_async_io,
3489 .o_set_async_flags = osc_set_async_flags,
3490 .o_queue_group_io = osc_queue_group_io,
3491 .o_trigger_group_io = osc_trigger_group_io,
3492 .o_teardown_async_page = osc_teardown_async_page,
3493 .o_punch = osc_punch,
3495 .o_enqueue = osc_enqueue,
3496 .o_match = osc_match,
3497 .o_change_cbdata = osc_change_cbdata,
3498 .o_cancel = osc_cancel,
3499 .o_cancel_unused = osc_cancel_unused,
3500 .o_iocontrol = osc_iocontrol,
3501 .o_get_info = osc_get_info,
3502 .o_set_info = osc_set_info,
3503 .o_import_event = osc_import_event,
3504 .o_llog_init = osc_llog_init,
3505 .o_llog_finish = osc_llog_finish,
3508 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3509 struct obd_ops sanosc_obd_ops = {
3510 .o_owner = THIS_MODULE,
3511 .o_attach = osc_attach,
3512 .o_detach = osc_detach,
3513 .o_cleanup = client_obd_cleanup,
3514 .o_add_conn = client_import_add_conn,
3515 .o_del_conn = client_import_del_conn,
3516 .o_connect = osc_connect,
3517 .o_disconnect = client_disconnect_export,
3518 .o_statfs = osc_statfs,
3519 .o_packmd = osc_packmd,
3520 .o_unpackmd = osc_unpackmd,
3521 .o_create = osc_real_create,
3522 .o_destroy = osc_destroy,
3523 .o_getattr = osc_getattr,
3524 .o_getattr_async = osc_getattr_async,
3525 .o_setattr = osc_setattr,
3526 .o_setup = client_sanobd_setup,
3527 .o_brw = sanosc_brw,
3528 .o_punch = osc_punch,
3530 .o_enqueue = osc_enqueue,
3531 .o_match = osc_match,
3532 .o_change_cbdata = osc_change_cbdata,
3533 .o_cancel = osc_cancel,
3534 .o_cancel_unused = osc_cancel_unused,
3535 .o_iocontrol = osc_iocontrol,
3536 .o_import_event = osc_import_event,
3537 .o_llog_init = osc_llog_init,
3538 .o_llog_finish = osc_llog_finish,
3542 int __init osc_init(void)
3544 struct lprocfs_static_vars lvars;
3545 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3546 struct lprocfs_static_vars sanlvars;
3551 lprocfs_init_vars(osc, &lvars);
3552 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3553 lprocfs_init_vars(osc, &sanlvars);
3556 rc = class_register_type(&osc_obd_ops, NULL, lvars.module_vars,
3557 OBD_OSC_DEVICENAME);
3561 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3562 rc = class_register_type(&sanosc_obd_ops, NULL, sanlvars.module_vars,
3563 OBD_SANOSC_DEVICENAME);
3565 class_unregister_type(OBD_OSC_DEVICENAME);
3572 static void /*__exit*/ osc_exit(void)
3574 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
3575 class_unregister_type(OBD_SANOSC_DEVICENAME);
3577 class_unregister_type(OBD_OSC_DEVICENAME);
3580 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
3581 MODULE_DESCRIPTION("Lustre Object Storage Client (OSC)");
3582 MODULE_LICENSE("GPL");
3584 module_init(osc_init);
3585 module_exit(osc_exit);