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,
117 enum cl_req_priority priority)
119 struct cl_page *page;
121 struct osc_object *osc0 = NULL;
122 struct client_obd *cli = NULL;
123 struct osc_object *osc = NULL; /* to keep gcc happy */
124 struct osc_page *opg;
127 struct cl_page_list *qin = &queue->c2_qin;
128 struct cl_page_list *qout = &queue->c2_qout;
132 LASSERT(qin->pl_nr > 0);
134 CDEBUG(D_INFO, "%i %i\n", qin->pl_nr, crt);
136 * NOTE: here @page is a top-level page. This is done to avoid
137 * creation of sub-page-list.
139 cl_page_list_for_each_safe(page, tmp, qin) {
140 struct osc_async_page *oap;
141 struct obd_export *exp;
147 opg = osc_cl_page_osc(page);
149 osc = cl2osc(opg->ops_cl.cpl_obj);
150 exp = osc_export(osc);
152 if (priority > CRP_NORMAL) {
153 spin_lock(&oap->oap_lock);
154 oap->oap_async_flags |= ASYNC_HP;
155 spin_unlock(&oap->oap_lock);
158 * This can be checked without cli->cl_loi_list_lock, because
159 * ->oap_*_item are always manipulated when the page is owned.
161 if (!list_empty(&oap->oap_urgent_item) ||
162 !list_empty(&oap->oap_rpc_item)) {
167 if (osc0 == NULL) { /* first iteration */
168 cli = &exp->exp_obd->u.cli;
170 } else /* check that all pages are against the same object
172 LASSERT(osc == osc0);
174 client_obd_list_lock(&cli->cl_loi_list_lock);
175 result = cl_page_prep(env, io, page, crt);
177 cl_page_list_move(qout, qin, page);
178 if (list_empty(&oap->oap_pending_item)) {
179 osc_io_submit_page(env, cl2osc_io(env, ios),
182 result = osc_set_async_flags_base(cli,
187 * bug 18881: we can't just break out here when
188 * error occurrs after cl_page_prep has been
189 * called against the page. The correct
190 * way is to call page's completion routine,
191 * as in osc_oap_interrupted. For simplicity,
192 * we just force osc_set_async_flags_base() to
195 LASSERT(result == 0);
197 opg->ops_submit_time = cfs_time_current();
200 if (result != -EALREADY)
203 * Handle -EALREADY error: for read case, the page is
204 * already in UPTODATE state; for write, the page
210 * Don't keep client_obd_list_lock() for too long.
212 * XXX client_obd_list lock has to be unlocked periodically to
213 * avoid soft-lockups that tend to happen otherwise (see bug
214 * 16651). On the other hand, osc_io_submit_page() queues a
215 * page with ASYNC_URGENT flag and so all pages queued up
216 * until this point are sent out immediately by
217 * osc_io_unplug() resulting in sub-optimal RPCs (sub-optimal
218 * RPCs only happen during `warm up' phase when less than
219 * cl_max_rpcs_in_flight RPCs are in flight). To balance these
220 * conflicting requirements, one might unplug once enough
221 * pages to form a large RPC were queued (i.e., use
222 * cli->cl_max_pages_per_rpc as OSC_QUEUE_GRAIN, see
223 * lop_makes_rpc()), or ignore soft-lockup issue altogether.
225 * XXX lock_need_resched() should be used here, but it is not
226 * available in the older of supported kernels.
228 if (queued > OSC_QUEUE_GRAIN || cfs_need_resched()) {
230 osc_io_unplug(env, osc, cli);
235 LASSERT(ergo(result == 0, cli != NULL));
236 LASSERT(ergo(result == 0, osc == osc0));
239 osc_io_unplug(env, osc, cli);
240 CDEBUG(D_INFO, "%i/%i %i\n", qin->pl_nr, qout->pl_nr, result);
241 return qout->pl_nr > 0 ? 0 : result;
244 static void osc_page_touch_at(const struct lu_env *env,
245 struct cl_object *obj, pgoff_t idx, unsigned to)
247 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
248 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
252 /* offset within stripe */
253 kms = cl_offset(obj, idx) + to;
255 cl_object_attr_lock(obj);
259 * ll_inode_size_lock(inode, 0); lov_stripe_lock(lsm);
263 CDEBUG(D_INODE, "stripe KMS %sincreasing "LPU64"->"LPU64" "LPU64"\n",
264 kms > loi->loi_kms ? "" : "not ", loi->loi_kms, kms,
265 loi->loi_lvb.lvb_size);
268 if (kms > loi->loi_kms) {
272 if (kms > loi->loi_lvb.lvb_size) {
273 attr->cat_size = kms;
276 cl_object_attr_set(env, obj, attr, valid);
277 cl_object_attr_unlock(obj);
281 * This is called when a page is accessed within file in a way that creates
282 * new page, if one were missing (i.e., if there were a hole at that place in
283 * the file, or accessed page is beyond the current file size). Examples:
284 * ->commit_write() and ->nopage() methods.
286 * Expand stripe KMS if necessary.
288 static void osc_page_touch(const struct lu_env *env,
289 struct osc_page *opage, unsigned to)
291 struct cl_page *page = opage->ops_cl.cpl_page;
292 struct cl_object *obj = opage->ops_cl.cpl_obj;
294 osc_page_touch_at(env, obj, page->cp_index, to);
298 * Implements cl_io_operations::cio_prepare_write() method for osc layer.
300 * \retval -EIO transfer initiated against this osc will most likely fail
301 * \retval 0 transfer initiated against this osc will most likely succeed.
303 * The reason for this check is to immediately return an error to the caller
304 * in the case of a deactivated import. Note, that import can be deactivated
305 * later, while pages, dirtied by this IO, are still in the cache, but this is
306 * irrelevant, because that would still return an error to the application (if
307 * it does fsync), but many applications don't do fsync because of performance
308 * issues, and we wanted to return an -EIO at write time to notify the
311 static int osc_io_prepare_write(const struct lu_env *env,
312 const struct cl_io_slice *ios,
313 const struct cl_page_slice *slice,
314 unsigned from, unsigned to)
316 struct osc_device *dev = lu2osc_dev(slice->cpl_obj->co_lu.lo_dev);
317 struct obd_import *imp = class_exp2cliimp(dev->od_exp);
322 * This implements OBD_BRW_CHECK logic from old client.
325 RETURN(imp == NULL || imp->imp_invalid ? -EIO : 0);
328 static int osc_io_commit_write(const struct lu_env *env,
329 const struct cl_io_slice *ios,
330 const struct cl_page_slice *slice,
331 unsigned from, unsigned to)
333 struct osc_page *opg = cl2osc_page(slice);
334 struct osc_object *obj = cl2osc(opg->ops_cl.cpl_obj);
335 struct osc_async_page *oap = &opg->ops_oap;
340 * XXX instead of calling osc_page_touch() here and in
341 * osc_io_fault_start() it might be more logical to introduce
342 * cl_page_touch() method, that generic cl_io_commit_write() and page
345 osc_page_touch(env, cl2osc_page(slice), to);
346 if (!client_is_remote(osc_export(obj)) &&
347 cfs_capable(CFS_CAP_SYS_RESOURCE))
348 oap->oap_brw_flags |= OBD_BRW_NOQUOTA;
353 static int osc_io_fault_start(const struct lu_env *env,
354 const struct cl_io_slice *ios)
357 struct cl_fault_io *fio;
362 fio = &io->u.ci_fault;
363 CDEBUG(D_INFO, "%lu %i %i\n",
364 fio->ft_index, fio->ft_writable, fio->ft_nob);
366 * If mapping is writeable, adjust kms to cover this page,
367 * but do not extend kms beyond actual file size.
370 if (fio->ft_writable)
371 osc_page_touch_at(env, ios->cis_obj,
372 fio->ft_index, fio->ft_nob);
376 static int osc_punch_upcall(void *a, int rc)
378 struct osc_punch_cbargs *args = a;
381 complete(&args->opc_sync);
387 * Checks that there are no pages being written in the extent being truncated.
389 static void osc_trunc_check(const struct lu_env *env, struct cl_io *io,
390 struct osc_io *oio, size_t size)
393 struct osc_object *obj;
394 struct cl_object *clob;
395 struct cl_page *page;
396 struct cl_page_list *list;
400 clob = oio->oi_cl.cis_obj;
402 start = cl_index(clob, size);
403 partial = cl_offset(clob, start) < size;
404 list = &osc_env_info(env)->oti_plist;
407 * Complain if there are pages in the truncated region.
409 * XXX this is quite expensive check.
411 cl_page_list_init(list);
412 cl_page_gang_lookup(env, clob, io, start + partial, CL_PAGE_EOF, list);
414 cl_page_list_for_each(page, list)
415 CL_PAGE_DEBUG(D_ERROR, env, page, "exists %lu\n", start);
417 cl_page_list_disown(env, io, list);
418 cl_page_list_fini(env, list);
420 spin_lock(&obj->oo_seatbelt);
421 list_for_each_entry(cp, &obj->oo_inflight[CRT_WRITE], ops_inflight) {
422 page = cp->ops_cl.cpl_page;
423 if (page->cp_index >= start + partial) {
424 cfs_task_t *submitter;
426 submitter = cp->ops_submitter;
428 * XXX Linux specific debugging stuff.
430 CL_PAGE_DEBUG(D_ERROR, env, page, "%s/%i %lu\n",
431 submitter->comm, submitter->pid, start);
432 libcfs_debug_dumpstack(submitter);
435 spin_unlock(&obj->oo_seatbelt);
437 #else /* __KERNEL__ */
438 # define osc_trunc_check(env, io, oio, size) do {;} while (0)
441 static int osc_io_trunc_start(const struct lu_env *env,
442 const struct cl_io_slice *slice)
444 struct cl_io *io = slice->cis_io;
445 struct osc_io *oio = cl2osc_io(env, slice);
446 struct cl_object *obj = slice->cis_obj;
447 struct lov_oinfo *loi = cl2osc(obj)->oo_oinfo;
448 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
449 struct obdo *oa = &oio->oi_oa;
450 struct osc_punch_cbargs *cbargs = &oio->oi_punch_cbarg;
451 struct obd_capa *capa;
452 loff_t size = io->u.ci_truncate.tr_size;
456 memset(oa, 0, sizeof(*oa));
458 osc_trunc_check(env, io, oio, size);
460 if (oio->oi_lockless == 0) {
461 cl_object_attr_lock(obj);
462 result = cl_object_attr_get(env, obj, attr);
464 attr->cat_size = attr->cat_kms = size;
465 result = cl_object_attr_set(env, obj, attr,
468 cl_object_attr_unlock(obj);
472 oa->o_id = loi->loi_id;
473 oa->o_gr = loi->loi_gr;
474 oa->o_mtime = attr->cat_mtime;
475 oa->o_atime = attr->cat_atime;
476 oa->o_ctime = attr->cat_ctime;
477 oa->o_valid = OBD_MD_FLID | OBD_MD_FLGROUP | OBD_MD_FLATIME |
478 OBD_MD_FLCTIME | OBD_MD_FLMTIME;
479 if (oio->oi_lockless) {
480 oa->o_flags = OBD_FL_TRUNCLOCK;
481 oa->o_valid |= OBD_MD_FLFLAGS;
484 oa->o_blocks = OBD_OBJECT_EOF;
485 oa->o_valid |= OBD_MD_FLSIZE | OBD_MD_FLBLOCKS;
487 capa = io->u.ci_truncate.tr_capa;
488 init_completion(&cbargs->opc_sync);
489 result = osc_punch_base(osc_export(cl2osc(obj)), oa, capa,
490 osc_punch_upcall, cbargs, PTLRPCD_SET);
495 static void osc_io_trunc_end(const struct lu_env *env,
496 const struct cl_io_slice *slice)
498 struct cl_io *io = slice->cis_io;
499 struct osc_io *oio = cl2osc_io(env, slice);
500 struct osc_punch_cbargs *cbargs = &oio->oi_punch_cbarg;
501 struct obdo *oa = &oio->oi_oa;
504 wait_for_completion(&cbargs->opc_sync);
506 result = io->ci_result = cbargs->opc_rc;
508 struct cl_object *obj = slice->cis_obj;
509 if (oio->oi_lockless == 0) {
510 struct cl_attr *attr = &osc_env_info(env)->oti_attr;
513 /* Update kms & size */
514 if (oa->o_valid & OBD_MD_FLSIZE) {
515 attr->cat_size = oa->o_size;
516 attr->cat_kms = oa->o_size;
517 valid |= CAT_KMS|CAT_SIZE;
519 if (oa->o_valid & OBD_MD_FLBLOCKS) {
520 attr->cat_blocks = oa->o_blocks;
523 if (oa->o_valid & OBD_MD_FLMTIME) {
524 attr->cat_mtime = oa->o_mtime;
527 if (oa->o_valid & OBD_MD_FLCTIME) {
528 attr->cat_ctime = oa->o_ctime;
531 if (oa->o_valid & OBD_MD_FLATIME) {
532 attr->cat_atime = oa->o_atime;
535 cl_object_attr_lock(obj);
536 result = cl_object_attr_set(env, obj, attr, valid);
537 cl_object_attr_unlock(obj);
538 } else { /* lockless truncate */
539 struct osc_device *osd = lu2osc_dev(obj->co_lu.lo_dev);
540 /* XXX: Need a lock. */
541 osd->od_stats.os_lockless_truncates++;
548 static const struct cl_io_operations osc_io_ops = {
551 .cio_fini = osc_io_fini
554 .cio_fini = osc_io_fini
557 .cio_start = osc_io_trunc_start,
558 .cio_end = osc_io_trunc_end
561 .cio_fini = osc_io_fini,
562 .cio_start = osc_io_fault_start
565 .cio_fini = osc_io_fini
570 .cio_submit = osc_io_submit
573 .cio_submit = osc_io_submit
576 .cio_prepare_write = osc_io_prepare_write,
577 .cio_commit_write = osc_io_commit_write
580 /*****************************************************************************
582 * Transfer operations.
586 static int osc_req_prep(const struct lu_env *env,
587 const struct cl_req_slice *slice)
592 static void osc_req_completion(const struct lu_env *env,
593 const struct cl_req_slice *slice, int ioret)
597 or = cl2osc_req(slice);
598 OBD_SLAB_FREE_PTR(or, osc_req_kmem);
602 * Implementation of struct cl_req_operations::cro_attr_set() for osc
603 * layer. osc is responsible for struct obdo::o_id and struct obdo::o_gr
606 static void osc_req_attr_set(const struct lu_env *env,
607 const struct cl_req_slice *slice,
608 const struct cl_object *obj,
609 struct cl_req_attr *attr, obd_valid flags)
611 struct lov_oinfo *oinfo;
612 struct cl_req *clerq;
613 struct cl_page *apage; /* _some_ page in @clerq */
614 struct cl_lock *lock; /* _some_ lock protecting @apage */
615 struct osc_lock *olck;
616 struct osc_page *opg;
620 oinfo = cl2osc(obj)->oo_oinfo;
621 if (flags & OBD_MD_FLID) {
622 oa->o_id = oinfo->loi_id;
623 oa->o_valid |= OBD_MD_FLID;
625 if (flags & OBD_MD_FLGROUP) {
626 oa->o_gr = oinfo->loi_gr;
627 oa->o_valid |= OBD_MD_FLGROUP;
629 if (flags & OBD_MD_FLHANDLE) {
630 clerq = slice->crs_req;
631 LASSERT(!list_empty(&clerq->crq_pages));
632 apage = container_of(clerq->crq_pages.next,
633 struct cl_page, cp_flight);
634 opg = osc_cl_page_osc(apage);
635 apage = opg->ops_cl.cpl_page; /* now apage is a sub-page */
636 lock = cl_lock_at_page(env, apage->cp_obj, apage, NULL, 1, 1);
637 LASSERT(lock != NULL);
638 olck = osc_lock_at(lock);
639 LASSERT(olck != NULL);
640 /* check for lockless io. */
641 if (olck->ols_lock != NULL) {
642 oa->o_handle = olck->ols_lock->l_remote_handle;
643 oa->o_valid |= OBD_MD_FLHANDLE;
645 cl_lock_put(env, lock);
649 static const struct cl_req_operations osc_req_ops = {
650 .cro_prep = osc_req_prep,
651 .cro_attr_set = osc_req_attr_set,
652 .cro_completion = osc_req_completion
656 int osc_io_init(const struct lu_env *env,
657 struct cl_object *obj, struct cl_io *io)
659 struct osc_io *oio = osc_env_io(env);
661 CL_IO_SLICE_CLEAN(oio, oi_cl);
662 cl_io_slice_add(io, &oio->oi_cl, obj, &osc_io_ops);
666 int osc_req_init(const struct lu_env *env, struct cl_device *dev,
672 OBD_SLAB_ALLOC_PTR_GFP(or, osc_req_kmem, CFS_ALLOC_IO);
674 cl_req_slice_add(req, &or->or_cl, dev, &osc_req_ops);