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 2008 Sun Microsystems, Inc. All rights reserved.
30 * Use is subject to license terms.
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 /** \addtogroup osc osc @{ */
43 #define DEBUG_SUBSYSTEM S_OSC
45 #include "osc_cl_internal.h"
47 /*****************************************************************************
53 static struct osc_req *cl2osc_req(const struct cl_req_slice *slice)
55 LINVRNT(slice->crs_dev->cd_lu_dev.ld_type == &osc_device_type);
56 return container_of0(slice, struct osc_req, or_cl);
59 static struct osc_io *cl2osc_io(const struct lu_env *env,
60 const struct cl_io_slice *slice)
62 struct osc_io *oio = container_of0(slice, struct osc_io, oi_cl);
63 LINVRNT(oio == osc_env_io(env));
67 static struct osc_page *osc_cl_page_osc(struct cl_page *page)
69 const struct cl_page_slice *slice;
71 slice = cl_page_at(page, &osc_device_type);
72 LASSERT(slice != NULL);
74 return cl2osc_page(slice);
78 /*****************************************************************************
84 static void osc_io_fini(const struct lu_env *env, const struct cl_io_slice *io)
88 struct cl_page *osc_oap2cl_page(struct osc_async_page *oap)
90 return container_of(oap, struct osc_page, ops_oap)->ops_cl.cpl_page;
93 static void osc_io_unplug(const struct lu_env *env, struct osc_object *osc,
94 struct client_obd *cli)
96 loi_list_maint(cli, osc->oo_oinfo);
97 osc_check_rpcs(env, cli);
98 client_obd_list_unlock(&cli->cl_loi_list_lock);
102 * How many pages osc_io_submit() queues before checking whether an RPC is
105 #define OSC_QUEUE_GRAIN (32)
108 * An implementation of cl_io_operations::cio_io_submit() method for osc
109 * layer. Iterates over pages in the in-queue, prepares each for io by calling
110 * cl_page_prep() and then either submits them through osc_io_submit_page()
111 * or, if page is already submitted, changes osc flags through
112 * osc_set_async_flags_base().
114 static int osc_io_submit(const struct lu_env *env,
115 const struct cl_io_slice *ios,
116 enum cl_req_type crt, struct cl_2queue *queue)
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, "%i %i\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);
152 * This can be checked without cli->cl_loi_list_lock, because
153 * ->oap_*_item are always manipulated when the page is owned.
155 if (!list_empty(&oap->oap_urgent_item) ||
156 !list_empty(&oap->oap_rpc_item)) {
161 if (osc0 == NULL) { /* first iteration */
162 cli = &exp->exp_obd->u.cli;
164 } else /* check that all pages are against the same object
166 LASSERT(osc == osc0);
168 client_obd_list_lock(&cli->cl_loi_list_lock);
169 result = cl_page_prep(env, io, page, crt);
171 cl_page_list_move(qout, qin, page);
172 if (list_empty(&oap->oap_pending_item)) {
173 osc_io_submit_page(env, cl2osc_io(env, ios),
176 result = osc_set_async_flags_base(cli,
185 if (result != -EALREADY)
188 * Handle -EALREADY error: for read case, the page is
189 * already in UPTODATE state; for write, the page
195 * Don't keep client_obd_list_lock() for too long.
197 * XXX lock_need_resched() should be used here, but it is not
198 * available in the older of supported kernels.
200 if (queued > OSC_QUEUE_GRAIN || cfs_need_resched()) {
202 osc_io_unplug(env, osc, cli);
207 LASSERT(ergo(result == 0, cli != NULL));
208 LASSERT(ergo(result == 0, osc == osc0));
211 osc_io_unplug(env, osc, cli);
212 CDEBUG(D_INFO, "%i/%i %i\n", qin->pl_nr, qout->pl_nr, result);
213 return qout->pl_nr > 0 ? 0 : result;
216 static void osc_page_touch_at(const struct lu_env *env,
217 struct cl_object *obj, pgoff_t idx, unsigned to)
219 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
220 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
224 /* offset within stripe */
225 kms = cl_offset(obj, idx) + to;
227 cl_object_attr_lock(obj);
231 * ll_inode_size_lock(inode, 0); lov_stripe_lock(lsm);
235 CDEBUG(D_INODE, "stripe KMS %sincreasing "LPU64"->"LPU64" "LPU64"\n",
236 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
237 loi->loi_lvb.lvb_size);
240 if (kms > loi->loi_kms) {
244 if (kms > loi->loi_lvb.lvb_size) {
245 attr->cat_size = kms;
248 cl_object_attr_set(env, obj, attr, valid);
249 cl_object_attr_unlock(obj);
253 * This is called when a page is accessed within file in a way that creates
254 * new page, if one were missing (i.e., if there were a hole at that place in
255 * the file, or accessed page is beyond the current file size). Examples:
256 * ->commit_write() and ->nopage() methods.
258 * Expand stripe KMS if necessary.
260 static void osc_page_touch(const struct lu_env *env,
261 struct osc_page *opage, unsigned to)
263 struct cl_page *page = opage->ops_cl.cpl_page;
264 struct cl_object *obj = opage->ops_cl.cpl_obj;
266 osc_page_touch_at(env, obj, page->cp_index, to);
270 * Implements cl_io_operations::cio_prepare_write() method for osc layer.
272 * \retval -EIO transfer initiated against this osc will most likely fail
273 * \retval 0 transfer initiated against this osc will most likely succeed.
275 * The reason for this check is to immediately return an error to the caller
276 * in the case of a deactivated import. Note, that import can be deactivated
277 * later, while pages, dirtied by this IO, are still in the cache, but this is
278 * irrelevant, because that would still return an error to the application (if
279 * it does fsync), but many applications don't do fsync because of performance
280 * issues, and we wanted to return an -EIO at write time to notify the
283 static int osc_io_prepare_write(const struct lu_env *env,
284 const struct cl_io_slice *ios,
285 const struct cl_page_slice *slice,
286 unsigned from, unsigned to)
288 struct osc_device *dev = lu2osc_dev(slice->cpl_obj->co_lu.lo_dev);
289 struct obd_import *imp = class_exp2cliimp(dev->od_exp);
294 * This implements OBD_BRW_CHECK logic from old client.
297 RETURN(imp == NULL || imp->imp_invalid ? -EIO : 0);
300 static int osc_io_commit_write(const struct lu_env *env,
301 const struct cl_io_slice *ios,
302 const struct cl_page_slice *slice,
303 unsigned from, unsigned to)
309 * XXX instead of calling osc_page_touch() here and in
310 * osc_io_fault_start() it might be more logical to introduce
311 * cl_page_touch() method, that generic cl_io_commit_write() and page
314 osc_page_touch(env, cl2osc_page(slice), to);
318 static int osc_io_fault_start(const struct lu_env *env,
319 const struct cl_io_slice *ios)
322 struct cl_fault_io *fio;
327 fio = &io->u.ci_fault;
328 CDEBUG(D_INFO, "%lu %i %i\n",
329 fio->ft_index, fio->ft_writable, fio->ft_nob);
331 * If mapping is writeable, adjust kms to cover this page,
332 * but do not extend kms beyond actual file size.
335 if (fio->ft_writable)
336 osc_page_touch_at(env, ios->cis_obj,
337 fio->ft_index, fio->ft_nob);
341 static int osc_punch_upcall(void *a, int rc)
343 struct osc_punch_cbargs *args = a;
346 complete(&args->opc_sync);
352 * Checks that there are no pages being written in the extent being truncated.
354 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
355 struct osc_io *oio, size_t size)
358 struct osc_object *obj;
359 struct cl_object *clob;
360 struct cl_page *page;
361 struct cl_page_list *list;
365 clob = oio->oi_cl.cis_obj;
367 start = cl_index(clob, size);
368 partial = cl_offset(clob, start) < size;
369 list = &osc_env_info(env)->oti_plist;
372 * Complain if there are pages in the truncated region.
374 * XXX this is quite expensive check.
376 cl_page_list_init(list);
377 cl_page_gang_lookup(env, clob, io, start + partial, CL_PAGE_EOF, list);
379 cl_page_list_for_each(page, list)
380 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %lu\n", start);
382 cl_page_list_disown(env, io, list);
383 cl_page_list_fini(env, list);
385 spin_lock(&obj->oo_seatbelt);
386 list_for_each_entry(cp, &obj->oo_inflight[CRT_WRITE], ops_inflight) {
387 page = cp->ops_cl.cpl_page;
388 if (page->cp_index >= start + partial) {
389 cfs_task_t *submitter;
391 submitter = cp->ops_submitter;
393 * XXX Linux specific debugging stuff.
395 CL_PAGE_DEBUG(D_ERROR, env, page, "%s/%i %lu\n",
396 submitter->comm, submitter->pid, start);
397 libcfs_debug_dumpstack(submitter);
400 spin_unlock(&obj->oo_seatbelt);
402 #else /* __KERNEL__ */
403 # define osc_trunc_check(env, io, oio, size) do {;} while (0)
406 static int osc_io_trunc_start(const struct lu_env *env,
407 const struct cl_io_slice *slice)
409 struct cl_io *io = slice->cis_io;
410 struct osc_io *oio = cl2osc_io(env, slice);
411 struct cl_object *obj = slice->cis_obj;
412 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
413 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
414 struct obdo *oa = &oio->oi_oa;
415 struct osc_punch_cbargs *cbargs = &oio->oi_punch_cbarg;
416 struct obd_capa *capa;
417 loff_t size = io->u.ci_truncate.tr_size;
420 memset(oa, 0, sizeof(*oa));
422 osc_trunc_check(env, io, oio, size);
424 cl_object_attr_lock(obj);
425 result = cl_object_attr_get(env, obj, attr);
427 attr->cat_size = attr->cat_kms = size;
428 result = cl_object_attr_set(env, obj, attr, CAT_SIZE|CAT_KMS);
430 cl_object_attr_unlock(obj);
433 oa->o_id = loi->loi_id;
434 oa->o_gr = loi->loi_gr;
435 oa->o_mtime = attr->cat_mtime;
436 oa->o_atime = attr->cat_atime;
437 oa->o_ctime = attr->cat_ctime;
438 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLATIME |
439 OBD_MD_FLCTIME | OBD_MD_FLMTIME;
440 if (oio->oi_lockless) {
441 oa->o_flags = OBD_FL_TRUNCLOCK;
442 oa->o_valid |= OBD_MD_FLFLAGS;
445 oa->o_blocks = OBD_OBJECT_EOF;
446 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
448 capa = io->u.ci_truncate.tr_capa;
449 init_completion(&cbargs->opc_sync);
450 result = osc_punch_base(osc_export(cl2osc(obj)), oa, capa,
451 osc_punch_upcall, cbargs, PTLRPCD_SET);
456 static void osc_io_trunc_end(const struct lu_env *env,
457 const struct cl_io_slice *slice)
459 struct cl_io *io = slice->cis_io;
460 struct osc_io *oio = cl2osc_io(env, slice);
461 struct osc_punch_cbargs *cbargs = &oio->oi_punch_cbarg;
462 struct obdo *oa = &oio->oi_oa;
465 wait_for_completion(&cbargs->opc_sync);
467 result = io->ci_result = cbargs->opc_rc;
469 struct cl_object *obj = slice->cis_obj;
470 if (oio->oi_lockless == 0) {
471 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
474 /* Update kms & size */
475 if (oa->o_valid & OBD_MD_FLSIZE) {
476 attr->cat_size = oa->o_size;
477 attr->cat_kms = oa->o_size;
478 valid |= CAT_KMS|CAT_SIZE;
480 if (oa->o_valid & OBD_MD_FLBLOCKS) {
481 attr->cat_blocks = oa->o_blocks;
484 if (oa->o_valid & OBD_MD_FLMTIME) {
485 attr->cat_mtime = oa->o_mtime;
488 if (oa->o_valid & OBD_MD_FLCTIME) {
489 attr->cat_ctime = oa->o_ctime;
492 if (oa->o_valid & OBD_MD_FLATIME) {
493 attr->cat_atime = oa->o_atime;
496 cl_object_attr_lock(obj);
497 result = cl_object_attr_set(env, obj, attr, valid);
498 cl_object_attr_unlock(obj);
499 } else { /* lockless truncate */
500 struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
501 /* XXX: Need a lock. */
502 osd->od_stats.os_lockless_truncates++;
509 static const struct cl_io_operations osc_io_ops = {
512 .cio_fini = osc_io_fini
515 .cio_fini = osc_io_fini
518 .cio_start = osc_io_trunc_start,
519 .cio_end = osc_io_trunc_end
522 .cio_fini = osc_io_fini,
523 .cio_start = osc_io_fault_start
526 .cio_fini = osc_io_fini
531 .cio_submit = osc_io_submit
534 .cio_submit = osc_io_submit
537 .cio_prepare_write = osc_io_prepare_write,
538 .cio_commit_write = osc_io_commit_write
541 /*****************************************************************************
543 * Transfer operations.
547 static int osc_req_prep(const struct lu_env *env,
548 const struct cl_req_slice *slice)
553 static void osc_req_completion(const struct lu_env *env,
554 const struct cl_req_slice *slice, int ioret)
558 or = cl2osc_req(slice);
559 OBD_SLAB_FREE_PTR(or, osc_req_kmem);
563 * Implementation of struct cl_req_operations::cro_attr_set() for osc
564 * layer. osc is responsible for struct obdo::o_id and struct obdo::o_gr
567 static void osc_req_attr_set(const struct lu_env *env,
568 const struct cl_req_slice *slice,
569 const struct cl_object *obj,
570 struct cl_req_attr *attr, obd_valid flags)
572 struct lov_oinfo *oinfo;
573 struct cl_req *clerq;
574 struct cl_page *apage; /* _some_ page in @clerq */
575 struct cl_lock *lock; /* _some_ lock protecting @apage */
576 struct osc_lock *olck;
577 struct osc_page *opg;
581 oinfo = cl2osc(obj)->oo_oinfo;
582 if (flags & OBD_MD_FLID) {
583 oa->o_id = oinfo->loi_id;
584 oa->o_valid |= OBD_MD_FLID;
586 if (flags & OBD_MD_FLGROUP) {
587 oa->o_gr = oinfo->loi_gr;
588 oa->o_valid |= OBD_MD_FLGROUP;
590 if (flags & OBD_MD_FLHANDLE) {
591 clerq = slice->crs_req;
592 LASSERT(!list_empty(&clerq->crq_pages));
593 apage = container_of(clerq->crq_pages.next,
594 struct cl_page, cp_flight);
595 opg = osc_cl_page_osc(apage);
596 apage = opg->ops_cl.cpl_page; /* now apage is a sub-page */
597 lock = cl_lock_at_page(env, apage->cp_obj, apage, NULL, 1, 1);
599 olck = osc_lock_at(lock);
600 LASSERT(olck != NULL);
601 /* check for lockless io. */
602 if (olck->ols_lock != NULL) {
603 oa->o_handle = olck->ols_lock->l_remote_handle;
604 oa->o_valid |= OBD_MD_FLHANDLE;
606 cl_lock_put(env, lock);
608 /* Should only be possible with liblustre */
609 LASSERT(LIBLUSTRE_CLIENT);
614 static const struct cl_req_operations osc_req_ops = {
615 .cro_prep = osc_req_prep,
616 .cro_attr_set = osc_req_attr_set,
617 .cro_completion = osc_req_completion
621 int osc_io_init(const struct lu_env *env,
622 struct cl_object *obj, struct cl_io *io)
624 struct osc_io *oio = osc_env_io(env);
626 CL_IO_SLICE_CLEAN(oio, oi_cl);
627 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);
631 int osc_req_init(const struct lu_env *env, struct cl_device *dev,
637 OBD_SLAB_ALLOC_PTR(or, osc_req_kmem);
639 cl_req_slice_add(req, &or->or_cl, dev, &osc_req_ops);