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, Intel Corporation.
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);
103 client_obd_list_unlock(&cli->cl_loi_list_lock);
107 * An implementation of cl_io_operations::cio_io_submit() method for osc
108 * layer. Iterates over pages in the in-queue, prepares each for io by calling
109 * cl_page_prep() and then either submits them through osc_io_submit_page()
110 * or, if page is already submitted, changes osc flags through
111 * osc_set_async_flags_base().
113 static int osc_io_submit(const struct lu_env *env,
114 const struct cl_io_slice *ios,
115 enum cl_req_type crt, struct cl_2queue *queue,
116 enum cl_req_priority priority)
118 struct cl_page *page;
120 struct osc_object *osc0 = NULL;
121 struct client_obd *cli = NULL;
122 struct osc_object *osc = NULL; /* to keep gcc happy */
123 struct osc_page *opg;
126 struct cl_page_list *qin = &queue->c2_qin;
127 struct cl_page_list *qout = &queue->c2_qout;
131 LASSERT(qin->pl_nr > 0);
133 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, crt);
135 * NOTE: here @page is a top-level page. This is done to avoid
136 * creation of sub-page-list.
138 cl_page_list_for_each_safe(page, tmp, qin) {
139 struct osc_async_page *oap;
140 struct obd_export *exp;
146 opg = osc_cl_page_osc(page);
148 osc = cl2osc(opg->ops_cl.cpl_obj);
149 exp = osc_export(osc);
151 if (priority > CRP_NORMAL) {
152 cfs_spin_lock(&oap->oap_lock);
153 oap->oap_async_flags |= ASYNC_HP;
154 cfs_spin_unlock(&oap->oap_lock);
157 * This can be checked without cli->cl_loi_list_lock, because
158 * ->oap_*_item are always manipulated when the page is owned.
160 if (!cfs_list_empty(&oap->oap_urgent_item) ||
161 !cfs_list_empty(&oap->oap_rpc_item)) {
166 if (osc0 == NULL) { /* first iteration */
167 cli = &exp->exp_obd->u.cli;
169 } else /* check that all pages are against the same object
171 LASSERT(osc == osc0);
173 client_obd_list_lock(&cli->cl_loi_list_lock);
174 result = cl_page_prep(env, io, page, crt);
176 cl_page_list_move(qout, qin, page);
177 if (cfs_list_empty(&oap->oap_pending_item)) {
178 osc_io_submit_page(env, cl2osc_io(env, ios),
181 result = osc_set_async_flags_base(cli,
186 * bug 18881: we can't just break out here when
187 * error occurs after cl_page_prep has been
188 * called against the page. The correct
189 * way is to call page's completion routine,
190 * as in osc_oap_interrupted. For simplicity,
191 * we just force osc_set_async_flags_base() to
194 LASSERT(result == 0);
196 opg->ops_submit_time = cfs_time_current();
199 if (result != -EALREADY)
202 * Handle -EALREADY error: for read case, the page is
203 * already in UPTODATE state; for write, the page
210 * We might hold client_obd_list_lock() for too long and cause
211 * soft-lockups (see bug 16651). But on the other hand, pages
212 * are queued here with ASYNC_URGENT flag, thus will be sent
213 * out immediately once osc_io_unplug() be called, possibly
214 * resulting sub-optimal RPCs.
216 * We think creating optimal-sized RPCs is more important than
217 * avoiding the transient soft-lockups, plus I believe the
218 * soft-locks only happen in full debug testing.
222 LASSERT(ergo(result == 0, cli != NULL));
223 LASSERT(ergo(result == 0, osc == osc0));
226 osc_io_unplug(env, osc, cli);
227 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
228 return qout->pl_nr > 0 ? 0 : result;
231 static void osc_page_touch_at(const struct lu_env *env,
232 struct cl_object *obj, pgoff_t idx, unsigned to)
234 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
235 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
239 /* offset within stripe */
240 kms = cl_offset(obj, idx) + to;
242 cl_object_attr_lock(obj);
246 * ll_inode_size_lock(inode, 0); lov_stripe_lock(lsm);
250 CDEBUG(D_INODE, "stripe KMS %sincreasing "LPU64"->"LPU64" "LPU64"\n",
251 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
252 loi->loi_lvb.lvb_size);
255 if (kms > loi->loi_kms) {
259 if (kms > loi->loi_lvb.lvb_size) {
260 attr->cat_size = kms;
263 cl_object_attr_set(env, obj, attr, valid);
264 cl_object_attr_unlock(obj);
268 * This is called when a page is accessed within file in a way that creates
269 * new page, if one were missing (i.e., if there were a hole at that place in
270 * the file, or accessed page is beyond the current file size). Examples:
271 * ->commit_write() and ->nopage() methods.
273 * Expand stripe KMS if necessary.
275 static void osc_page_touch(const struct lu_env *env,
276 struct osc_page *opage, unsigned to)
278 struct cl_page *page = opage->ops_cl.cpl_page;
279 struct cl_object *obj = opage->ops_cl.cpl_obj;
281 osc_page_touch_at(env, obj, page->cp_index, to);
285 * Implements cl_io_operations::cio_prepare_write() method for osc layer.
287 * \retval -EIO transfer initiated against this osc will most likely fail
288 * \retval 0 transfer initiated against this osc will most likely succeed.
290 * The reason for this check is to immediately return an error to the caller
291 * in the case of a deactivated import. Note, that import can be deactivated
292 * later, while pages, dirtied by this IO, are still in the cache, but this is
293 * irrelevant, because that would still return an error to the application (if
294 * it does fsync), but many applications don't do fsync because of performance
295 * issues, and we wanted to return an -EIO at write time to notify the
298 static int osc_io_prepare_write(const struct lu_env *env,
299 const struct cl_io_slice *ios,
300 const struct cl_page_slice *slice,
301 unsigned from, unsigned to)
303 struct osc_device *dev = lu2osc_dev(slice->cpl_obj->co_lu.lo_dev);
304 struct obd_import *imp = class_exp2cliimp(dev->od_exp);
305 struct osc_io *oio = cl2osc_io(env, ios);
310 * This implements OBD_BRW_CHECK logic from old client.
313 if (imp == NULL || imp->imp_invalid)
315 if (result == 0 && oio->oi_lockless)
316 /* this page contains `invalid' data, but who cares?
317 * nobody can access the invalid data.
318 * in osc_io_commit_write(), we're going to write exact
319 * [from, to) bytes of this page to OST. -jay */
320 cl_page_export(env, slice->cpl_page, 1);
325 static int osc_io_commit_write(const struct lu_env *env,
326 const struct cl_io_slice *ios,
327 const struct cl_page_slice *slice,
328 unsigned from, unsigned to)
330 struct osc_io *oio = cl2osc_io(env, ios);
331 struct osc_page *opg = cl2osc_page(slice);
332 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
333 struct osc_async_page *oap = &opg->ops_oap;
338 * XXX instead of calling osc_page_touch() here and in
339 * osc_io_fault_start() it might be more logical to introduce
340 * cl_page_touch() method, that generic cl_io_commit_write() and page
343 osc_page_touch(env, cl2osc_page(slice), to);
344 if (!client_is_remote(osc_export(obj)) &&
345 cfs_capable(CFS_CAP_SYS_RESOURCE))
346 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
348 if (oio->oi_lockless)
349 /* see osc_io_prepare_write() for lockless io handling. */
350 cl_page_clip(env, slice->cpl_page, from, to);
355 static int osc_io_fault_start(const struct lu_env *env,
356 const struct cl_io_slice *ios)
359 struct cl_fault_io *fio;
364 fio = &io->u.ci_fault;
365 CDEBUG(D_INFO, "%lu %d %d\n",
366 fio->ft_index, fio->ft_writable, fio->ft_nob);
368 * If mapping is writeable, adjust kms to cover this page,
369 * but do not extend kms beyond actual file size.
372 if (fio->ft_writable)
373 osc_page_touch_at(env, ios->cis_obj,
374 fio->ft_index, fio->ft_nob);
378 static int osc_setattr_upcall(void *a, int rc)
380 struct osc_setattr_cbargs *args = a;
383 cfs_complete(&args->opc_sync);
387 /* Disable osc_trunc_check() because it is naturally race between read and
388 * truncate. See bug 20645 for details.
390 #if 0 && defined(__KERNEL__)
392 * Checks that there are no pages being written in the extent being truncated.
394 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
395 struct osc_io *oio, size_t size)
398 struct osc_object *obj;
399 struct cl_object *clob;
400 struct cl_page *page;
401 struct cl_page_list *list;
405 clob = oio->oi_cl.cis_obj;
407 start = cl_index(clob, size);
408 partial = cl_offset(clob, start) < size;
409 list = &osc_env_info(env)->oti_plist;
412 * Complain if there are pages in the truncated region.
414 * XXX this is quite expensive check.
416 cl_page_list_init(list);
417 cl_page_gang_lookup(env, clob, io, start + partial, CL_PAGE_EOF, list);
419 cl_page_list_for_each(page, list)
420 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %lu\n", start);
422 cl_page_list_disown(env, io, list);
423 cl_page_list_fini(env, list);
425 cfs_spin_lock(&obj->oo_seatbelt);
426 cfs_list_for_each_entry(cp, &obj->oo_inflight[CRT_WRITE],
428 page = cp->ops_cl.cpl_page;
429 if (page->cp_index >= start + partial) {
430 cfs_task_t *submitter;
432 submitter = cp->ops_submitter;
434 * XXX Linux specific debugging stuff.
436 CL_PAGE_DEBUG(D_ERROR, env, page, "%s/%d %lu\n",
437 submitter->comm, submitter->pid, start);
438 libcfs_debug_dumpstack(submitter);
441 cfs_spin_unlock(&obj->oo_seatbelt);
443 #else /* __KERNEL__ */
444 # define osc_trunc_check(env, io, oio, size) do {;} while (0)
447 static int osc_io_setattr_start(const struct lu_env *env,
448 const struct cl_io_slice *slice)
450 struct cl_io *io = slice->cis_io;
451 struct osc_io *oio = cl2osc_io(env, slice);
452 struct cl_object *obj = slice->cis_obj;
453 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
454 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
455 struct obdo *oa = &oio->oi_oa;
456 struct osc_setattr_cbargs *cbargs = &oio->oi_setattr_cbarg;
457 loff_t size = io->u.ci_setattr.sa_attr.lvb_size;
458 unsigned int ia_valid = io->u.ci_setattr.sa_valid;
460 struct obd_info oinfo = { { { 0 } } };
462 if (ia_valid & ATTR_SIZE)
463 osc_trunc_check(env, io, oio, size);
465 if (oio->oi_lockless == 0) {
466 cl_object_attr_lock(obj);
467 result = cl_object_attr_get(env, obj, attr);
469 unsigned int cl_valid = 0;
471 if (ia_valid & ATTR_SIZE) {
472 attr->cat_size = attr->cat_kms = size;
473 cl_valid = (CAT_SIZE | CAT_KMS);
475 if (ia_valid & ATTR_MTIME_SET) {
476 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
477 cl_valid |= CAT_MTIME;
479 if (ia_valid & ATTR_ATIME_SET) {
480 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
481 cl_valid |= CAT_ATIME;
483 if (ia_valid & ATTR_CTIME_SET) {
484 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
485 cl_valid |= CAT_CTIME;
487 result = cl_object_attr_set(env, obj, attr, cl_valid);
489 cl_object_attr_unlock(obj);
491 memset(oa, 0, sizeof(*oa));
493 oa->o_id = loi->loi_id;
494 oa->o_seq = loi->loi_seq;
495 oa->o_mtime = attr->cat_mtime;
496 oa->o_atime = attr->cat_atime;
497 oa->o_ctime = attr->cat_ctime;
498 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLATIME |
499 OBD_MD_FLCTIME | OBD_MD_FLMTIME;
500 if (ia_valid & ATTR_SIZE) {
502 oa->o_blocks = OBD_OBJECT_EOF;
503 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
505 if (oio->oi_lockless) {
506 oa->o_flags = OBD_FL_SRVLOCK;
507 oa->o_valid |= OBD_MD_FLFLAGS;
510 LASSERT(oio->oi_lockless == 0);
514 oinfo.oi_capa = io->u.ci_setattr.sa_capa;
515 cfs_init_completion(&cbargs->opc_sync);
517 if (ia_valid & ATTR_SIZE)
518 result = osc_punch_base(osc_export(cl2osc(obj)),
519 &oinfo, osc_setattr_upcall,
520 cbargs, PTLRPCD_SET);
522 result = osc_setattr_async_base(osc_export(cl2osc(obj)),
525 cbargs, PTLRPCD_SET);
530 static void osc_io_setattr_end(const struct lu_env *env,
531 const struct cl_io_slice *slice)
533 struct cl_io *io = slice->cis_io;
534 struct osc_io *oio = cl2osc_io(env, slice);
535 struct osc_setattr_cbargs *cbargs = &oio->oi_setattr_cbarg;
538 cfs_wait_for_completion(&cbargs->opc_sync);
540 result = io->ci_result = cbargs->opc_rc;
542 struct cl_object *obj = slice->cis_obj;
543 if (oio->oi_lockless) {
544 /* lockless truncate */
545 struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
547 LASSERT(cl_io_is_trunc(io));
548 /* XXX: Need a lock. */
549 osd->od_stats.os_lockless_truncates++;
554 static int osc_io_read_start(const struct lu_env *env,
555 const struct cl_io_slice *slice)
557 struct osc_io *oio = cl2osc_io(env, slice);
558 struct cl_object *obj = slice->cis_obj;
559 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
563 if (oio->oi_lockless == 0) {
564 cl_object_attr_lock(obj);
565 result = cl_object_attr_get(env, obj, attr);
567 attr->cat_atime = LTIME_S(CFS_CURRENT_TIME);
568 result = cl_object_attr_set(env, obj, attr,
571 cl_object_attr_unlock(obj);
576 static int osc_io_write_start(const struct lu_env *env,
577 const struct cl_io_slice *slice)
579 struct osc_io *oio = cl2osc_io(env, slice);
580 struct cl_object *obj = slice->cis_obj;
581 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
585 if (oio->oi_lockless == 0) {
586 cl_object_attr_lock(obj);
587 result = cl_object_attr_get(env, obj, attr);
589 attr->cat_mtime = attr->cat_ctime =
590 LTIME_S(CFS_CURRENT_TIME);
591 result = cl_object_attr_set(env, obj, attr,
592 CAT_MTIME | CAT_CTIME);
594 cl_object_attr_unlock(obj);
599 static const struct cl_io_operations osc_io_ops = {
602 .cio_start = osc_io_read_start,
603 .cio_fini = osc_io_fini
606 .cio_start = osc_io_write_start,
607 .cio_fini = osc_io_fini
610 .cio_start = osc_io_setattr_start,
611 .cio_end = osc_io_setattr_end
614 .cio_fini = osc_io_fini,
615 .cio_start = osc_io_fault_start
618 .cio_fini = osc_io_fini
623 .cio_submit = osc_io_submit
626 .cio_submit = osc_io_submit
629 .cio_prepare_write = osc_io_prepare_write,
630 .cio_commit_write = osc_io_commit_write
633 /*****************************************************************************
635 * Transfer operations.
639 static int osc_req_prep(const struct lu_env *env,
640 const struct cl_req_slice *slice)
645 static void osc_req_completion(const struct lu_env *env,
646 const struct cl_req_slice *slice, int ioret)
650 or = cl2osc_req(slice);
651 OBD_SLAB_FREE_PTR(or, osc_req_kmem);
655 * Implementation of struct cl_req_operations::cro_attr_set() for osc
656 * layer. osc is responsible for struct obdo::o_id and struct obdo::o_seq
659 static void osc_req_attr_set(const struct lu_env *env,
660 const struct cl_req_slice *slice,
661 const struct cl_object *obj,
662 struct cl_req_attr *attr, obd_valid flags)
664 struct lov_oinfo *oinfo;
665 struct cl_req *clerq;
666 struct cl_page *apage; /* _some_ page in @clerq */
667 struct cl_lock *lock; /* _some_ lock protecting @apage */
668 struct osc_lock *olck;
669 struct osc_page *opg;
673 oinfo = cl2osc(obj)->oo_oinfo;
674 if (flags & OBD_MD_FLID) {
675 oa->o_id = oinfo->loi_id;
676 oa->o_valid |= OBD_MD_FLID;
678 if (flags & OBD_MD_FLGROUP) {
679 oa->o_seq = oinfo->loi_seq;
680 oa->o_valid |= OBD_MD_FLGROUP;
682 if (flags & OBD_MD_FLHANDLE) {
683 clerq = slice->crs_req;
684 LASSERT(!cfs_list_empty(&clerq->crq_pages));
685 apage = container_of(clerq->crq_pages.next,
686 struct cl_page, cp_flight);
687 opg = osc_cl_page_osc(apage);
688 apage = opg->ops_cl.cpl_page; /* now apage is a sub-page */
689 lock = cl_lock_at_page(env, apage->cp_obj, apage, NULL, 1, 1);
691 struct cl_object_header *head;
692 struct cl_lock *scan;
694 head = cl_object_header(apage->cp_obj);
695 cfs_list_for_each_entry(scan, &head->coh_locks,
697 CL_LOCK_DEBUG(D_ERROR, env, scan,
699 CL_PAGE_DEBUG(D_ERROR, env, apage,
700 "dump uncover page!\n");
701 libcfs_debug_dumpstack(NULL);
705 olck = osc_lock_at(lock);
706 LASSERT(olck != NULL);
707 LASSERT(ergo(opg->ops_srvlock, olck->ols_lock == NULL));
708 /* check for lockless io. */
709 if (olck->ols_lock != NULL) {
710 oa->o_handle = olck->ols_lock->l_remote_handle;
711 oa->o_valid |= OBD_MD_FLHANDLE;
713 cl_lock_put(env, lock);
717 static const struct cl_req_operations osc_req_ops = {
718 .cro_prep = osc_req_prep,
719 .cro_attr_set = osc_req_attr_set,
720 .cro_completion = osc_req_completion
724 int osc_io_init(const struct lu_env *env,
725 struct cl_object *obj, struct cl_io *io)
727 struct osc_io *oio = osc_env_io(env);
729 CL_IO_SLICE_CLEAN(oio, oi_cl);
730 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);
734 int osc_req_init(const struct lu_env *env, struct cl_device *dev,
740 OBD_SLAB_ALLOC_PTR_GFP(or, osc_req_kmem, CFS_ALLOC_IO);
742 cl_req_slice_add(req, &or->or_cl, dev, &osc_req_ops);