4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_io for OSC layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
41 #define DEBUG_SUBSYSTEM S_OSC
43 #include "osc_cl_internal.h"
49 /*****************************************************************************
55 static struct osc_req *cl2osc_req(const struct cl_req_slice *slice)
57 LINVRNT(slice->crs_dev->cd_lu_dev.ld_type == &osc_device_type);
58 return container_of0(slice, struct osc_req, or_cl);
61 static struct osc_io *cl2osc_io(const struct lu_env *env,
62 const struct cl_io_slice *slice)
64 struct osc_io *oio = container_of0(slice, struct osc_io, oi_cl);
65 LINVRNT(oio == osc_env_io(env));
69 static struct osc_page *osc_cl_page_osc(struct cl_page *page)
71 const struct cl_page_slice *slice;
73 slice = cl_page_at(page, &osc_device_type);
74 LASSERT(slice != NULL);
76 return cl2osc_page(slice);
80 /*****************************************************************************
86 static void osc_io_fini(const struct lu_env *env, const struct cl_io_slice *io)
90 struct cl_page *osc_oap2cl_page(struct osc_async_page *oap)
92 return container_of(oap, struct osc_page, ops_oap)->ops_cl.cpl_page;
95 static void osc_io_unplug(const struct lu_env *env, struct osc_object *osc,
96 struct client_obd *cli)
98 loi_list_maint(cli, osc->oo_oinfo);
99 osc_check_rpcs(env, cli);
103 * An implementation of cl_io_operations::cio_io_submit() method for osc
104 * layer. Iterates over pages in the in-queue, prepares each for io by calling
105 * cl_page_prep() and then either submits them through osc_io_submit_page()
106 * or, if page is already submitted, changes osc flags through
107 * osc_set_async_flags_base().
109 static int osc_io_submit(const struct lu_env *env,
110 const struct cl_io_slice *ios,
111 enum cl_req_type crt, struct cl_2queue *queue,
112 enum cl_req_priority priority)
114 struct cl_page *page;
116 struct osc_object *osc0 = NULL;
117 struct client_obd *cli = NULL;
118 struct osc_object *osc = NULL; /* to keep gcc happy */
119 struct osc_page *opg;
122 struct cl_page_list *qin = &queue->c2_qin;
123 struct cl_page_list *qout = &queue->c2_qout;
127 LASSERT(qin->pl_nr > 0);
129 CDEBUG(D_INFO, "%d %d\n", qin->pl_nr, crt);
131 * NOTE: here @page is a top-level page. This is done to avoid
132 * creation of sub-page-list.
134 cl_page_list_for_each_safe(page, tmp, qin) {
135 struct osc_async_page *oap;
136 struct obd_export *exp;
142 opg = osc_cl_page_osc(page);
144 osc = cl2osc(opg->ops_cl.cpl_obj);
145 exp = osc_export(osc);
147 if (priority > CRP_NORMAL) {
148 cfs_spin_lock(&oap->oap_lock);
149 oap->oap_async_flags |= ASYNC_HP;
150 cfs_spin_unlock(&oap->oap_lock);
153 if (osc0 == NULL) { /* first iteration */
154 cli = &exp->exp_obd->u.cli;
156 client_obd_list_lock(&cli->cl_loi_list_lock);
157 } else /* check that all pages are against the same object
159 LASSERT(osc == osc0);
161 if (!cfs_list_empty(&oap->oap_urgent_item) ||
162 !cfs_list_empty(&oap->oap_rpc_item)) {
167 result = cl_page_prep(env, io, page, crt);
170 cl_page_list_move(qout, qin, page);
171 if (cfs_list_empty(&oap->oap_pending_item)) {
172 osc_io_submit_page(env, cl2osc_io(env, ios),
175 result = osc_set_async_flags_base(cli,
180 * bug 18881: we can't just break out here when
181 * error occurs after cl_page_prep has been
182 * called against the page. The correct
183 * way is to call page's completion routine,
184 * as in osc_oap_interrupted. For simplicity,
185 * we just force osc_set_async_flags_base() to
188 LASSERT(result == 0);
190 opg->ops_submit_time = cfs_time_current();
193 if (result != -EALREADY)
196 * Handle -EALREADY error: for read case, the page is
197 * already in UPTODATE state; for write, the page
204 * We might hold client_obd_list_lock() for too long and cause
205 * soft-lockups (see bug 16651). But on the other hand, pages
206 * are queued here with ASYNC_URGENT flag, thus will be sent
207 * out immediately once osc_io_unplug() be called, possibly
208 * resulting sub-optimal RPCs.
210 * We think creating optimal-sized RPCs is more important than
211 * avoiding the transient soft-lockups, plus I believe the
212 * soft-locks only happen in full debug testing.
216 LASSERT(ergo(result == 0, cli != NULL));
217 LASSERT(ergo(result == 0, osc == osc0));
220 osc_io_unplug(env, osc, cli);
222 client_obd_list_unlock(&cli->cl_loi_list_lock);
223 CDEBUG(D_INFO, "%d/%d %d\n", qin->pl_nr, qout->pl_nr, result);
224 return qout->pl_nr > 0 ? 0 : result;
227 static void osc_page_touch_at(const struct lu_env *env,
228 struct cl_object *obj, pgoff_t idx, unsigned to)
230 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
231 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
235 /* offset within stripe */
236 kms = cl_offset(obj, idx) + to;
238 cl_object_attr_lock(obj);
242 * ll_inode_size_lock(inode, 0); lov_stripe_lock(lsm);
246 CDEBUG(D_INODE, "stripe KMS %sincreasing "LPU64"->"LPU64" "LPU64"\n",
247 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
248 loi->loi_lvb.lvb_size);
251 if (kms > loi->loi_kms) {
255 if (kms > loi->loi_lvb.lvb_size) {
256 attr->cat_size = kms;
259 cl_object_attr_set(env, obj, attr, valid);
260 cl_object_attr_unlock(obj);
264 * This is called when a page is accessed within file in a way that creates
265 * new page, if one were missing (i.e., if there were a hole at that place in
266 * the file, or accessed page is beyond the current file size). Examples:
267 * ->commit_write() and ->nopage() methods.
269 * Expand stripe KMS if necessary.
271 static void osc_page_touch(const struct lu_env *env,
272 struct osc_page *opage, unsigned to)
274 struct cl_page *page = opage->ops_cl.cpl_page;
275 struct cl_object *obj = opage->ops_cl.cpl_obj;
277 osc_page_touch_at(env, obj, page->cp_index, to);
281 * Implements cl_io_operations::cio_prepare_write() method for osc layer.
283 * \retval -EIO transfer initiated against this osc will most likely fail
284 * \retval 0 transfer initiated against this osc will most likely succeed.
286 * The reason for this check is to immediately return an error to the caller
287 * in the case of a deactivated import. Note, that import can be deactivated
288 * later, while pages, dirtied by this IO, are still in the cache, but this is
289 * irrelevant, because that would still return an error to the application (if
290 * it does fsync), but many applications don't do fsync because of performance
291 * issues, and we wanted to return an -EIO at write time to notify the
294 static int osc_io_prepare_write(const struct lu_env *env,
295 const struct cl_io_slice *ios,
296 const struct cl_page_slice *slice,
297 unsigned from, unsigned to)
299 struct osc_device *dev = lu2osc_dev(slice->cpl_obj->co_lu.lo_dev);
300 struct obd_import *imp = class_exp2cliimp(dev->od_exp);
301 struct osc_io *oio = cl2osc_io(env, ios);
306 * This implements OBD_BRW_CHECK logic from old client.
309 if (imp == NULL || imp->imp_invalid)
311 if (result == 0 && oio->oi_lockless)
312 /* this page contains `invalid' data, but who cares?
313 * nobody can access the invalid data.
314 * in osc_io_commit_write(), we're going to write exact
315 * [from, to) bytes of this page to OST. -jay */
316 cl_page_export(env, slice->cpl_page, 1);
321 static int osc_io_commit_write(const struct lu_env *env,
322 const struct cl_io_slice *ios,
323 const struct cl_page_slice *slice,
324 unsigned from, unsigned to)
326 struct osc_io *oio = cl2osc_io(env, ios);
327 struct osc_page *opg = cl2osc_page(slice);
328 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
329 struct osc_async_page *oap = &opg->ops_oap;
334 * XXX instead of calling osc_page_touch() here and in
335 * osc_io_fault_start() it might be more logical to introduce
336 * cl_page_touch() method, that generic cl_io_commit_write() and page
339 osc_page_touch(env, cl2osc_page(slice), to);
340 if (!client_is_remote(osc_export(obj)) &&
341 cfs_capable(CFS_CAP_SYS_RESOURCE))
342 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
344 if (oio->oi_lockless)
345 /* see osc_io_prepare_write() for lockless io handling. */
346 cl_page_clip(env, slice->cpl_page, from, to);
351 static int osc_io_fault_start(const struct lu_env *env,
352 const struct cl_io_slice *ios)
355 struct cl_fault_io *fio;
360 fio = &io->u.ci_fault;
361 CDEBUG(D_INFO, "%lu %d %d\n",
362 fio->ft_index, fio->ft_writable, fio->ft_nob);
364 * If mapping is writeable, adjust kms to cover this page,
365 * but do not extend kms beyond actual file size.
368 if (fio->ft_writable)
369 osc_page_touch_at(env, ios->cis_obj,
370 fio->ft_index, fio->ft_nob);
374 static int osc_setattr_upcall(void *a, int rc)
376 struct osc_setattr_cbargs *args = a;
379 cfs_complete(&args->opc_sync);
383 /* Disable osc_trunc_check() because it is naturally race between read and
384 * truncate. See bug 20645 for details.
386 #if 0 && defined(__KERNEL__)
388 * Checks that there are no pages being written in the extent being truncated.
390 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
391 struct osc_io *oio, size_t size)
394 struct osc_object *obj;
395 struct cl_object *clob;
396 struct cl_page *page;
397 struct cl_page_list *list;
401 clob = oio->oi_cl.cis_obj;
403 start = cl_index(clob, size);
404 partial = cl_offset(clob, start) < size;
405 list = &osc_env_info(env)->oti_plist;
408 * Complain if there are pages in the truncated region.
410 * XXX this is quite expensive check.
412 cl_page_list_init(list);
413 cl_page_gang_lookup(env, clob, io, start + partial, CL_PAGE_EOF, list);
415 cl_page_list_for_each(page, list)
416 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %lu\n", start);
418 cl_page_list_disown(env, io, list);
419 cl_page_list_fini(env, list);
421 cfs_spin_lock(&obj->oo_seatbelt);
422 cfs_list_for_each_entry(cp, &obj->oo_inflight[CRT_WRITE],
424 page = cp->ops_cl.cpl_page;
425 if (page->cp_index >= start + partial) {
426 cfs_task_t *submitter;
428 submitter = cp->ops_submitter;
430 * XXX Linux specific debugging stuff.
432 CL_PAGE_DEBUG(D_ERROR, env, page, "%s/%d %lu\n",
433 submitter->comm, submitter->pid, start);
434 libcfs_debug_dumpstack(submitter);
437 cfs_spin_unlock(&obj->oo_seatbelt);
439 #else /* __KERNEL__ */
440 # define osc_trunc_check(env, io, oio, size) do {;} while (0)
443 static int osc_io_setattr_start(const struct lu_env *env,
444 const struct cl_io_slice *slice)
446 struct cl_io *io = slice->cis_io;
447 struct osc_io *oio = cl2osc_io(env, slice);
448 struct cl_object *obj = slice->cis_obj;
449 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
450 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
451 struct obdo *oa = &oio->oi_oa;
452 struct osc_setattr_cbargs *cbargs = &oio->oi_setattr_cbarg;
453 loff_t size = io->u.ci_setattr.sa_attr.lvb_size;
454 unsigned int ia_valid = io->u.ci_setattr.sa_valid;
456 struct obd_info oinfo = { { { 0 } } };
458 if (ia_valid & ATTR_SIZE)
459 osc_trunc_check(env, io, oio, size);
461 if (oio->oi_lockless == 0) {
462 cl_object_attr_lock(obj);
463 result = cl_object_attr_get(env, obj, attr);
465 unsigned int cl_valid = 0;
467 if (ia_valid & ATTR_SIZE) {
468 attr->cat_size = attr->cat_kms = size;
469 cl_valid = (CAT_SIZE | CAT_KMS);
471 if (ia_valid & ATTR_MTIME_SET) {
472 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
473 cl_valid |= CAT_MTIME;
475 if (ia_valid & ATTR_ATIME_SET) {
476 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
477 cl_valid |= CAT_ATIME;
479 if (ia_valid & ATTR_CTIME_SET) {
480 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
481 cl_valid |= CAT_CTIME;
483 result = cl_object_attr_set(env, obj, attr, cl_valid);
485 cl_object_attr_unlock(obj);
487 memset(oa, 0, sizeof(*oa));
489 oa->o_id = loi->loi_id;
490 oa->o_seq = loi->loi_seq;
491 oa->o_mtime = attr->cat_mtime;
492 oa->o_atime = attr->cat_atime;
493 oa->o_ctime = attr->cat_ctime;
494 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLATIME |
495 OBD_MD_FLCTIME | OBD_MD_FLMTIME;
496 if (ia_valid & ATTR_SIZE) {
498 oa->o_blocks = OBD_OBJECT_EOF;
499 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
501 if (oio->oi_lockless) {
502 oa->o_flags = OBD_FL_SRVLOCK;
503 oa->o_valid |= OBD_MD_FLFLAGS;
506 LASSERT(oio->oi_lockless == 0);
510 oinfo.oi_capa = io->u.ci_setattr.sa_capa;
511 cfs_init_completion(&cbargs->opc_sync);
513 if (ia_valid & ATTR_SIZE)
514 result = osc_punch_base(osc_export(cl2osc(obj)),
515 &oinfo, osc_setattr_upcall,
516 cbargs, PTLRPCD_SET);
518 result = osc_setattr_async_base(osc_export(cl2osc(obj)),
521 cbargs, PTLRPCD_SET);
526 static void osc_io_setattr_end(const struct lu_env *env,
527 const struct cl_io_slice *slice)
529 struct cl_io *io = slice->cis_io;
530 struct osc_io *oio = cl2osc_io(env, slice);
531 struct osc_setattr_cbargs *cbargs = &oio->oi_setattr_cbarg;
534 cfs_wait_for_completion(&cbargs->opc_sync);
536 result = io->ci_result = cbargs->opc_rc;
538 struct cl_object *obj = slice->cis_obj;
539 if (oio->oi_lockless) {
540 /* lockless truncate */
541 struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
543 LASSERT(cl_io_is_trunc(io));
544 /* XXX: Need a lock. */
545 osd->od_stats.os_lockless_truncates++;
550 static int osc_io_read_start(const struct lu_env *env,
551 const struct cl_io_slice *slice)
553 struct osc_io *oio = cl2osc_io(env, slice);
554 struct cl_object *obj = slice->cis_obj;
555 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
559 if (oio->oi_lockless == 0) {
560 cl_object_attr_lock(obj);
561 result = cl_object_attr_get(env, obj, attr);
563 attr->cat_atime = LTIME_S(CFS_CURRENT_TIME);
564 result = cl_object_attr_set(env, obj, attr,
567 cl_object_attr_unlock(obj);
572 static int osc_io_write_start(const struct lu_env *env,
573 const struct cl_io_slice *slice)
575 struct osc_io *oio = cl2osc_io(env, slice);
576 struct cl_object *obj = slice->cis_obj;
577 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
581 if (oio->oi_lockless == 0) {
582 cl_object_attr_lock(obj);
583 result = cl_object_attr_get(env, obj, attr);
585 attr->cat_mtime = attr->cat_ctime =
586 LTIME_S(CFS_CURRENT_TIME);
587 result = cl_object_attr_set(env, obj, attr,
588 CAT_MTIME | CAT_CTIME);
590 cl_object_attr_unlock(obj);
595 static const struct cl_io_operations osc_io_ops = {
598 .cio_start = osc_io_read_start,
599 .cio_fini = osc_io_fini
602 .cio_start = osc_io_write_start,
603 .cio_fini = osc_io_fini
606 .cio_start = osc_io_setattr_start,
607 .cio_end = osc_io_setattr_end
610 .cio_fini = osc_io_fini,
611 .cio_start = osc_io_fault_start
614 .cio_fini = osc_io_fini
619 .cio_submit = osc_io_submit
622 .cio_submit = osc_io_submit
625 .cio_prepare_write = osc_io_prepare_write,
626 .cio_commit_write = osc_io_commit_write
629 /*****************************************************************************
631 * Transfer operations.
635 static int osc_req_prep(const struct lu_env *env,
636 const struct cl_req_slice *slice)
641 static void osc_req_completion(const struct lu_env *env,
642 const struct cl_req_slice *slice, int ioret)
646 or = cl2osc_req(slice);
647 OBD_SLAB_FREE_PTR(or, osc_req_kmem);
651 * Implementation of struct cl_req_operations::cro_attr_set() for osc
652 * layer. osc is responsible for struct obdo::o_id and struct obdo::o_seq
655 static void osc_req_attr_set(const struct lu_env *env,
656 const struct cl_req_slice *slice,
657 const struct cl_object *obj,
658 struct cl_req_attr *attr, obd_valid flags)
660 struct lov_oinfo *oinfo;
661 struct cl_req *clerq;
662 struct cl_page *apage; /* _some_ page in @clerq */
663 struct cl_lock *lock; /* _some_ lock protecting @apage */
664 struct osc_lock *olck;
665 struct osc_page *opg;
669 oinfo = cl2osc(obj)->oo_oinfo;
670 if (flags & OBD_MD_FLID) {
671 oa->o_id = oinfo->loi_id;
672 oa->o_valid |= OBD_MD_FLID;
674 if (flags & OBD_MD_FLGROUP) {
675 oa->o_seq = oinfo->loi_seq;
676 oa->o_valid |= OBD_MD_FLGROUP;
678 if (flags & OBD_MD_FLHANDLE) {
679 clerq = slice->crs_req;
680 LASSERT(!cfs_list_empty(&clerq->crq_pages));
681 apage = container_of(clerq->crq_pages.next,
682 struct cl_page, cp_flight);
683 opg = osc_cl_page_osc(apage);
684 apage = opg->ops_cl.cpl_page; /* now apage is a sub-page */
685 lock = cl_lock_at_page(env, apage->cp_obj, apage, NULL, 1, 1);
687 struct cl_object_header *head;
688 struct cl_lock *scan;
690 head = cl_object_header(apage->cp_obj);
691 cfs_list_for_each_entry(scan, &head->coh_locks,
693 CL_LOCK_DEBUG(D_ERROR, env, scan,
695 CL_PAGE_DEBUG(D_ERROR, env, apage,
696 "dump uncover page!\n");
697 libcfs_debug_dumpstack(NULL);
701 olck = osc_lock_at(lock);
702 LASSERT(olck != NULL);
703 LASSERT(ergo(opg->ops_srvlock, olck->ols_lock == NULL));
704 /* check for lockless io. */
705 if (olck->ols_lock != NULL) {
706 oa->o_handle = olck->ols_lock->l_remote_handle;
707 oa->o_valid |= OBD_MD_FLHANDLE;
709 cl_lock_put(env, lock);
713 static const struct cl_req_operations osc_req_ops = {
714 .cro_prep = osc_req_prep,
715 .cro_attr_set = osc_req_attr_set,
716 .cro_completion = osc_req_completion
720 int osc_io_init(const struct lu_env *env,
721 struct cl_object *obj, struct cl_io *io)
723 struct osc_io *oio = osc_env_io(env);
725 CL_IO_SLICE_CLEAN(oio, oi_cl);
726 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);
730 int osc_req_init(const struct lu_env *env, struct cl_device *dev,
736 OBD_SLAB_ALLOC_PTR_GFP(or, osc_req_kmem, CFS_ALLOC_IO);
738 cl_req_slice_add(req, &or->or_cl, dev, &osc_req_ops);