1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * Implementation of cl_io for OSC layer.
41 * Author: Nikita Danilov <nikita.danilov@sun.com>
44 #define DEBUG_SUBSYSTEM S_OSC
46 #include "osc_cl_internal.h"
52 /*****************************************************************************
58 static struct osc_req *cl2osc_req(const struct cl_req_slice *slice)
60 LINVRNT(slice->crs_dev->cd_lu_dev.ld_type == &osc_device_type);
61 return container_of0(slice, struct osc_req, or_cl);
64 static struct osc_io *cl2osc_io(const struct lu_env *env,
65 const struct cl_io_slice *slice)
67 struct osc_io *oio = container_of0(slice, struct osc_io, oi_cl);
68 LINVRNT(oio == osc_env_io(env));
72 static struct osc_page *osc_cl_page_osc(struct cl_page *page)
74 const struct cl_page_slice *slice;
76 slice = cl_page_at(page, &osc_device_type);
77 LASSERT(slice != NULL);
79 return cl2osc_page(slice);
83 /*****************************************************************************
89 static void osc_io_fini(const struct lu_env *env, const struct cl_io_slice *io)
93 struct cl_page *osc_oap2cl_page(struct osc_async_page *oap)
95 return container_of(oap, struct osc_page, ops_oap)->ops_cl.cpl_page;
98 static void osc_io_unplug(const struct lu_env *env, struct osc_object *osc,
99 struct client_obd *cli)
101 loi_list_maint(cli, osc->oo_oinfo);
102 osc_check_rpcs(env, cli);
106 * An implementation of cl_io_operations::cio_io_submit() method for osc
107 * layer. Iterates over pages in the in-queue, prepares each for io by calling
108 * cl_page_prep() and then either submits them through osc_io_submit_page()
109 * or, if page is already submitted, changes osc flags through
110 * osc_set_async_flags_base().
112 static int osc_io_submit(const struct lu_env *env,
113 const struct cl_io_slice *ios,
114 enum cl_req_type crt, struct cl_2queue *queue,
115 enum cl_req_priority priority)
117 struct cl_page *page;
119 struct osc_object *osc0 = NULL;
120 struct client_obd *cli = NULL;
121 struct osc_object *osc = NULL; /* to keep gcc happy */
122 struct osc_page *opg;
125 struct cl_page_list *qin = &queue->c2_qin;
126 struct cl_page_list *qout = &queue->c2_qout;
130 LASSERT(qin->pl_nr > 0);
132 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, crt);
134 * NOTE: here @page is a top-level page. This is done to avoid
135 * creation of sub-page-list.
137 cl_page_list_for_each_safe(page, tmp, qin) {
138 struct osc_async_page *oap;
139 struct obd_export *exp;
145 opg = osc_cl_page_osc(page);
147 osc = cl2osc(opg->ops_cl.cpl_obj);
148 exp = osc_export(osc);
150 if (priority > CRP_NORMAL) {
151 cfs_spin_lock(&oap->oap_lock);
152 oap->oap_async_flags |= ASYNC_HP;
153 cfs_spin_unlock(&oap->oap_lock);
156 if (osc0 == NULL) { /* first iteration */
157 cli = &exp->exp_obd->u.cli;
159 client_obd_list_lock(&cli->cl_loi_list_lock);
160 } else /* check that all pages are against the same object
162 LASSERT(osc == osc0);
164 if (!cfs_list_empty(&oap->oap_urgent_item) ||
165 !cfs_list_empty(&oap->oap_rpc_item)) {
170 result = cl_page_prep(env, io, page, crt);
173 cl_page_list_move(qout, qin, page);
174 if (cfs_list_empty(&oap->oap_pending_item)) {
175 osc_io_submit_page(env, cl2osc_io(env, ios),
178 result = osc_set_async_flags_base(cli,
183 * bug 18881: we can't just break out here when
184 * error occurs after cl_page_prep has been
185 * called against the page. The correct
186 * way is to call page's completion routine,
187 * as in osc_oap_interrupted. For simplicity,
188 * we just force osc_set_async_flags_base() to
191 LASSERT(result == 0);
193 opg->ops_submit_time = cfs_time_current();
196 if (result != -EALREADY)
199 * Handle -EALREADY error: for read case, the page is
200 * already in UPTODATE state; for write, the page
207 * We might hold client_obd_list_lock() for too long and cause
208 * soft-lockups (see bug 16651). But on the other hand, pages
209 * are queued here with ASYNC_URGENT flag, thus will be sent
210 * out immediately once osc_io_unplug() be called, possibly
211 * resulting sub-optimal RPCs.
213 * We think creating optimal-sized RPCs is more important than
214 * avoiding the transient soft-lockups, plus I believe the
215 * soft-locks only happen in full debug testing.
219 LASSERT(ergo(result == 0, cli != NULL));
220 LASSERT(ergo(result == 0, osc == osc0));
223 osc_io_unplug(env, osc, cli);
225 client_obd_list_unlock(&cli->cl_loi_list_lock);
226 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
227 return qout->pl_nr > 0 ? 0 : result;
230 static void osc_page_touch_at(const struct lu_env *env,
231 struct cl_object *obj, pgoff_t idx, unsigned to)
233 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
234 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
238 /* offset within stripe */
239 kms = cl_offset(obj, idx) + to;
241 cl_object_attr_lock(obj);
245 * ll_inode_size_lock(inode, 0); lov_stripe_lock(lsm);
249 CDEBUG(D_INODE, "stripe KMS %sincreasing "LPU64"->"LPU64" "LPU64"\n",
250 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
251 loi->loi_lvb.lvb_size);
254 if (kms > loi->loi_kms) {
258 if (kms > loi->loi_lvb.lvb_size) {
259 attr->cat_size = kms;
262 cl_object_attr_set(env, obj, attr, valid);
263 cl_object_attr_unlock(obj);
267 * This is called when a page is accessed within file in a way that creates
268 * new page, if one were missing (i.e., if there were a hole at that place in
269 * the file, or accessed page is beyond the current file size). Examples:
270 * ->commit_write() and ->nopage() methods.
272 * Expand stripe KMS if necessary.
274 static void osc_page_touch(const struct lu_env *env,
275 struct osc_page *opage, unsigned to)
277 struct cl_page *page = opage->ops_cl.cpl_page;
278 struct cl_object *obj = opage->ops_cl.cpl_obj;
280 osc_page_touch_at(env, obj, page->cp_index, to);
284 * Implements cl_io_operations::cio_prepare_write() method for osc layer.
286 * \retval -EIO transfer initiated against this osc will most likely fail
287 * \retval 0 transfer initiated against this osc will most likely succeed.
289 * The reason for this check is to immediately return an error to the caller
290 * in the case of a deactivated import. Note, that import can be deactivated
291 * later, while pages, dirtied by this IO, are still in the cache, but this is
292 * irrelevant, because that would still return an error to the application (if
293 * it does fsync), but many applications don't do fsync because of performance
294 * issues, and we wanted to return an -EIO at write time to notify the
297 static int osc_io_prepare_write(const struct lu_env *env,
298 const struct cl_io_slice *ios,
299 const struct cl_page_slice *slice,
300 unsigned from, unsigned to)
302 struct osc_device *dev = lu2osc_dev(slice->cpl_obj->co_lu.lo_dev);
303 struct obd_import *imp = class_exp2cliimp(dev->od_exp);
304 struct osc_io *oio = cl2osc_io(env, ios);
309 * This implements OBD_BRW_CHECK logic from old client.
312 if (imp == NULL || imp->imp_invalid)
314 if (result == 0 && oio->oi_lockless)
315 /* this page contains `invalid' data, but who cares?
316 * nobody can access the invalid data.
317 * in osc_io_commit_write(), we're going to write exact
318 * [from, to) bytes of this page to OST. -jay */
319 cl_page_export(env, slice->cpl_page, 1);
324 static int osc_io_commit_write(const struct lu_env *env,
325 const struct cl_io_slice *ios,
326 const struct cl_page_slice *slice,
327 unsigned from, unsigned to)
329 struct osc_io *oio = cl2osc_io(env, ios);
330 struct osc_page *opg = cl2osc_page(slice);
331 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
332 struct osc_async_page *oap = &opg->ops_oap;
337 * XXX instead of calling osc_page_touch() here and in
338 * osc_io_fault_start() it might be more logical to introduce
339 * cl_page_touch() method, that generic cl_io_commit_write() and page
342 osc_page_touch(env, cl2osc_page(slice), to);
343 if (!client_is_remote(osc_export(obj)) &&
344 cfs_capable(CFS_CAP_SYS_RESOURCE))
345 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
347 if (oio->oi_lockless)
348 /* see osc_io_prepare_write() for lockless io handling. */
349 cl_page_clip(env, slice->cpl_page, from, to);
354 static int osc_io_fault_start(const struct lu_env *env,
355 const struct cl_io_slice *ios)
358 struct cl_fault_io *fio;
363 fio = &io->u.ci_fault;
364 CDEBUG(D_INFO, "%lu %d %d\n",
365 fio->ft_index, fio->ft_writable, fio->ft_nob);
367 * If mapping is writeable, adjust kms to cover this page,
368 * but do not extend kms beyond actual file size.
371 if (fio->ft_writable)
372 osc_page_touch_at(env, ios->cis_obj,
373 fio->ft_index, fio->ft_nob);
377 static int osc_setattr_upcall(void *a, int rc)
379 struct osc_setattr_cbargs *args = a;
382 cfs_complete(&args->opc_sync);
386 /* Disable osc_trunc_check() because it is naturally race between read and
387 * truncate. See bug 20645 for details.
389 #if 0 && defined(__KERNEL__)
391 * Checks that there are no pages being written in the extent being truncated.
393 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
394 struct osc_io *oio, size_t size)
397 struct osc_object *obj;
398 struct cl_object *clob;
399 struct cl_page *page;
400 struct cl_page_list *list;
404 clob = oio->oi_cl.cis_obj;
406 start = cl_index(clob, size);
407 partial = cl_offset(clob, start) < size;
408 list = &osc_env_info(env)->oti_plist;
411 * Complain if there are pages in the truncated region.
413 * XXX this is quite expensive check.
415 cl_page_list_init(list);
416 cl_page_gang_lookup(env, clob, io, start + partial, CL_PAGE_EOF, list);
418 cl_page_list_for_each(page, list)
419 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %lu\n", start);
421 cl_page_list_disown(env, io, list);
422 cl_page_list_fini(env, list);
424 cfs_spin_lock(&obj->oo_seatbelt);
425 cfs_list_for_each_entry(cp, &obj->oo_inflight[CRT_WRITE],
427 page = cp->ops_cl.cpl_page;
428 if (page->cp_index >= start + partial) {
429 cfs_task_t *submitter;
431 submitter = cp->ops_submitter;
433 * XXX Linux specific debugging stuff.
435 CL_PAGE_DEBUG(D_ERROR, env, page, "%s/%d %lu\n",
436 submitter->comm, submitter->pid, start);
437 libcfs_debug_dumpstack(submitter);
440 cfs_spin_unlock(&obj->oo_seatbelt);
442 #else /* __KERNEL__ */
443 # define osc_trunc_check(env, io, oio, size) do {;} while (0)
446 static int osc_io_setattr_start(const struct lu_env *env,
447 const struct cl_io_slice *slice)
449 struct cl_io *io = slice->cis_io;
450 struct osc_io *oio = cl2osc_io(env, slice);
451 struct cl_object *obj = slice->cis_obj;
452 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
453 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
454 struct obdo *oa = &oio->oi_oa;
455 struct osc_setattr_cbargs *cbargs = &oio->oi_setattr_cbarg;
456 loff_t size = io->u.ci_setattr.sa_attr.lvb_size;
457 unsigned int ia_valid = io->u.ci_setattr.sa_valid;
459 struct obd_info oinfo = { { { 0 } } };
461 if (ia_valid & ATTR_SIZE)
462 osc_trunc_check(env, io, oio, size);
464 if (oio->oi_lockless == 0) {
465 cl_object_attr_lock(obj);
466 result = cl_object_attr_get(env, obj, attr);
468 unsigned int cl_valid = 0;
470 if (ia_valid & ATTR_SIZE) {
471 attr->cat_size = attr->cat_kms = size;
472 cl_valid = (CAT_SIZE | CAT_KMS);
474 if (ia_valid & ATTR_MTIME_SET) {
475 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
476 cl_valid |= CAT_MTIME;
478 if (ia_valid & ATTR_ATIME_SET) {
479 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
480 cl_valid |= CAT_ATIME;
482 if (ia_valid & ATTR_CTIME_SET) {
483 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
484 cl_valid |= CAT_CTIME;
486 result = cl_object_attr_set(env, obj, attr, cl_valid);
488 cl_object_attr_unlock(obj);
490 memset(oa, 0, sizeof(*oa));
492 oa->o_id = loi->loi_id;
493 oa->o_seq = loi->loi_seq;
494 oa->o_mtime = attr->cat_mtime;
495 oa->o_atime = attr->cat_atime;
496 oa->o_ctime = attr->cat_ctime;
497 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLATIME |
498 OBD_MD_FLCTIME | OBD_MD_FLMTIME;
499 if (ia_valid & ATTR_SIZE) {
501 oa->o_blocks = OBD_OBJECT_EOF;
502 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
504 if (oio->oi_lockless) {
505 oa->o_flags = OBD_FL_SRVLOCK;
506 oa->o_valid |= OBD_MD_FLFLAGS;
509 LASSERT(oio->oi_lockless == 0);
513 oinfo.oi_capa = io->u.ci_setattr.sa_capa;
514 cfs_init_completion(&cbargs->opc_sync);
516 if (ia_valid & ATTR_SIZE)
517 result = osc_punch_base(osc_export(cl2osc(obj)),
518 &oinfo, osc_setattr_upcall,
519 cbargs, PTLRPCD_SET);
521 result = osc_setattr_async_base(osc_export(cl2osc(obj)),
524 cbargs, PTLRPCD_SET);
529 static void osc_io_setattr_end(const struct lu_env *env,
530 const struct cl_io_slice *slice)
532 struct cl_io *io = slice->cis_io;
533 struct osc_io *oio = cl2osc_io(env, slice);
534 struct osc_setattr_cbargs *cbargs = &oio->oi_setattr_cbarg;
537 cfs_wait_for_completion(&cbargs->opc_sync);
539 result = io->ci_result = cbargs->opc_rc;
541 struct cl_object *obj = slice->cis_obj;
542 if (oio->oi_lockless) {
543 /* lockless truncate */
544 struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
546 LASSERT(cl_io_is_trunc(io));
547 /* XXX: Need a lock. */
548 osd->od_stats.os_lockless_truncates++;
553 static int osc_io_read_start(const struct lu_env *env,
554 const struct cl_io_slice *slice)
556 struct osc_io *oio = cl2osc_io(env, slice);
557 struct cl_object *obj = slice->cis_obj;
558 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
562 if (oio->oi_lockless == 0) {
563 cl_object_attr_lock(obj);
564 result = cl_object_attr_get(env, obj, attr);
566 attr->cat_atime = LTIME_S(CFS_CURRENT_TIME);
567 result = cl_object_attr_set(env, obj, attr,
570 cl_object_attr_unlock(obj);
575 static int osc_io_write_start(const struct lu_env *env,
576 const struct cl_io_slice *slice)
578 struct osc_io *oio = cl2osc_io(env, slice);
579 struct cl_object *obj = slice->cis_obj;
580 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
584 if (oio->oi_lockless == 0) {
585 cl_object_attr_lock(obj);
586 result = cl_object_attr_get(env, obj, attr);
588 attr->cat_mtime = attr->cat_ctime =
589 LTIME_S(CFS_CURRENT_TIME);
590 result = cl_object_attr_set(env, obj, attr,
591 CAT_MTIME | CAT_CTIME);
593 cl_object_attr_unlock(obj);
598 static const struct cl_io_operations osc_io_ops = {
601 .cio_start = osc_io_read_start,
602 .cio_fini = osc_io_fini
605 .cio_start = osc_io_write_start,
606 .cio_fini = osc_io_fini
609 .cio_start = osc_io_setattr_start,
610 .cio_end = osc_io_setattr_end
613 .cio_fini = osc_io_fini,
614 .cio_start = osc_io_fault_start
617 .cio_fini = osc_io_fini
622 .cio_submit = osc_io_submit
625 .cio_submit = osc_io_submit
628 .cio_prepare_write = osc_io_prepare_write,
629 .cio_commit_write = osc_io_commit_write
632 /*****************************************************************************
634 * Transfer operations.
638 static int osc_req_prep(const struct lu_env *env,
639 const struct cl_req_slice *slice)
644 static void osc_req_completion(const struct lu_env *env,
645 const struct cl_req_slice *slice, int ioret)
649 or = cl2osc_req(slice);
650 OBD_SLAB_FREE_PTR(or, osc_req_kmem);
654 * Implementation of struct cl_req_operations::cro_attr_set() for osc
655 * layer. osc is responsible for struct obdo::o_id and struct obdo::o_seq
658 static void osc_req_attr_set(const struct lu_env *env,
659 const struct cl_req_slice *slice,
660 const struct cl_object *obj,
661 struct cl_req_attr *attr, obd_valid flags)
663 struct lov_oinfo *oinfo;
664 struct cl_req *clerq;
665 struct cl_page *apage; /* _some_ page in @clerq */
666 struct cl_lock *lock; /* _some_ lock protecting @apage */
667 struct osc_lock *olck;
668 struct osc_page *opg;
672 oinfo = cl2osc(obj)->oo_oinfo;
673 if (flags & OBD_MD_FLID) {
674 oa->o_id = oinfo->loi_id;
675 oa->o_valid |= OBD_MD_FLID;
677 if (flags & OBD_MD_FLGROUP) {
678 oa->o_seq = oinfo->loi_seq;
679 oa->o_valid |= OBD_MD_FLGROUP;
681 if (flags & OBD_MD_FLHANDLE) {
682 clerq = slice->crs_req;
683 LASSERT(!cfs_list_empty(&clerq->crq_pages));
684 apage = container_of(clerq->crq_pages.next,
685 struct cl_page, cp_flight);
686 opg = osc_cl_page_osc(apage);
687 apage = opg->ops_cl.cpl_page; /* now apage is a sub-page */
688 lock = cl_lock_at_page(env, apage->cp_obj, apage, NULL, 1, 1);
690 struct cl_object_header *head;
691 struct cl_lock *scan;
693 head = cl_object_header(apage->cp_obj);
694 cfs_list_for_each_entry(scan, &head->coh_locks,
696 CL_LOCK_DEBUG(D_ERROR, env, scan,
698 CL_PAGE_DEBUG(D_ERROR, env, apage,
699 "dump uncover page!\n");
700 libcfs_debug_dumpstack(NULL);
704 olck = osc_lock_at(lock);
705 LASSERT(olck != NULL);
706 LASSERT(ergo(opg->ops_srvlock, olck->ols_lock == NULL));
707 /* check for lockless io. */
708 if (olck->ols_lock != NULL) {
709 oa->o_handle = olck->ols_lock->l_remote_handle;
710 oa->o_valid |= OBD_MD_FLHANDLE;
712 cl_lock_put(env, lock);
716 static const struct cl_req_operations osc_req_ops = {
717 .cro_prep = osc_req_prep,
718 .cro_attr_set = osc_req_attr_set,
719 .cro_completion = osc_req_completion
723 int osc_io_init(const struct lu_env *env,
724 struct cl_object *obj, struct cl_io *io)
726 struct osc_io *oio = osc_env_io(env);
728 CL_IO_SLICE_CLEAN(oio, oi_cl);
729 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);
733 int osc_req_init(const struct lu_env *env, struct cl_device *dev,
739 OBD_SLAB_ALLOC_PTR_GFP(or, osc_req_kmem, CFS_ALLOC_IO);
741 cl_req_slice_add(req, &or->or_cl, dev, &osc_req_ops);