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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2012, 2017 Intel Corporation.
28 * Author: Mikhail Pershin <mike.pershin@intel.com>
31 #define DEBUG_SUBSYSTEM S_FILTER
33 #include <dt_object.h>
34 #include "mdt_internal.h"
36 /* --------------- MDT grant code ---------------- */
38 long mdt_grant_connect(const struct lu_env *env,
39 struct obd_export *exp,
40 u64 want, bool conservative)
42 struct mdt_device *mdt = mdt_exp2dev(exp);
48 dt_statfs(env, mdt->mdt_bottom, &mdt->mdt_osfs);
50 left = (mdt->mdt_osfs.os_bavail * mdt->mdt_osfs.os_bsize) / 2;
54 CDEBUG(D_CACHE, "%s: cli %s/%p ocd_grant: %ld want: %llu left: %llu\n",
55 exp->exp_obd->obd_name, exp->exp_client_uuid.uuid,
56 exp, grant, want, left);
61 void mdt_grant_prepare_write(const struct lu_env *env,
62 struct obd_export *exp, struct obdo *oa,
63 struct niobuf_remote *rnb, int niocount)
65 struct mdt_device *mdt = mdt_exp2dev(exp);
70 left = (mdt->mdt_osfs.os_bavail * mdt->mdt_osfs.os_bsize) / 2;
72 /* grant more space back to the client if possible */
75 /* ---------------- end of MDT grant code ---------------- */
77 /* functions below are stubs for now, they will be implemented with
78 * grant support on MDT */
79 static inline void mdt_io_counter_incr(struct obd_export *exp, int opcode,
80 char *jobid, long amount)
85 void mdt_grant_prepare_read(const struct lu_env *env,
86 struct obd_export *exp, struct obdo *oa)
91 void mdt_grant_commit(struct obd_export *exp, unsigned long pending,
98 static inline void mdt_dom_read_lock(struct mdt_object *mo)
100 down_read(&mo->mot_dom_sem);
103 static inline void mdt_dom_read_unlock(struct mdt_object *mo)
105 up_read(&mo->mot_dom_sem);
108 static inline void mdt_dom_write_lock(struct mdt_object *mo)
110 down_write(&mo->mot_dom_sem);
113 static inline void mdt_dom_write_unlock(struct mdt_object *mo)
115 up_write(&mo->mot_dom_sem);
118 static int mdt_preprw_read(const struct lu_env *env, struct obd_export *exp,
119 struct mdt_device *mdt, struct mdt_object *mo,
120 struct lu_attr *la, int niocount,
121 struct niobuf_remote *rnb, int *nr_local,
122 struct niobuf_local *lnb, char *jobid)
124 struct dt_object *dob;
125 int i, j, rc, tot_bytes = 0;
129 mdt_dom_read_lock(mo);
130 if (!mdt_object_exists(mo))
131 GOTO(unlock, rc = -ENOENT);
133 dob = mdt_obj2dt(mo);
134 /* parse remote buffers to local buffers and prepare the latter */
136 for (i = 0, j = 0; i < niocount; i++) {
137 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 0);
138 if (unlikely(rc < 0))
140 /* correct index for local buffers to continue with */
143 tot_bytes += rnb[i].rnb_len;
146 rc = dt_attr_get(env, dob, la);
150 rc = dt_read_prep(env, dob, lnb, *nr_local);
154 mdt_io_counter_incr(exp, LPROC_MDT_IO_READ, jobid, tot_bytes);
157 dt_bufs_put(env, dob, lnb, *nr_local);
159 mdt_dom_read_unlock(mo);
163 static int mdt_preprw_write(const struct lu_env *env, struct obd_export *exp,
164 struct mdt_device *mdt, struct mdt_object *mo,
165 struct lu_attr *la, struct obdo *oa,
166 int objcount, struct obd_ioobj *obj,
167 struct niobuf_remote *rnb, int *nr_local,
168 struct niobuf_local *lnb, char *jobid)
170 struct dt_object *dob;
171 int i, j, k, rc = 0, tot_bytes = 0;
175 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
176 * space back if possible */
177 mdt_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
179 mdt_dom_read_lock(mo);
180 if (!mdt_object_exists(mo)) {
181 CDEBUG(D_ERROR, "%s: BRW to missing obj "DFID"\n",
182 exp->exp_obd->obd_name, PFID(mdt_object_fid(mo)));
183 GOTO(unlock, rc = -ENOENT);
186 dob = mdt_obj2dt(mo);
187 /* parse remote buffers to local buffers and prepare the latter */
189 for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
190 rc = dt_bufs_get(env, dob, rnb + i, lnb + j, 1);
191 if (unlikely(rc < 0))
193 /* correct index for local buffers to continue with */
194 for (k = 0; k < rc; k++)
195 lnb[j+k].lnb_flags = rnb[i].rnb_flags;
198 tot_bytes += rnb[i].rnb_len;
201 rc = dt_write_prep(env, dob, lnb, *nr_local);
205 mdt_io_counter_incr(exp, LPROC_MDT_IO_WRITE, jobid, tot_bytes);
208 dt_bufs_put(env, dob, lnb, *nr_local);
210 mdt_dom_read_unlock(mo);
211 /* tgt_grant_prepare_write() was called, so we must commit */
212 mdt_grant_commit(exp, oa->o_grant_used, rc);
213 /* let's still process incoming grant information packed in the oa,
214 * but without enforcing grant since we won't proceed with the write.
215 * Just like a read request actually. */
216 mdt_grant_prepare_read(env, exp, oa);
220 int mdt_obd_preprw(const struct lu_env *env, int cmd, struct obd_export *exp,
221 struct obdo *oa, int objcount, struct obd_ioobj *obj,
222 struct niobuf_remote *rnb, int *nr_local,
223 struct niobuf_local *lnb)
225 struct tgt_session_info *tsi = tgt_ses_info(env);
226 struct mdt_thread_info *info = tsi2mdt_info(tsi);
227 struct lu_attr *la = &info->mti_attr.ma_attr;
228 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
229 struct mdt_object *mo;
233 /* The default value PTLRPC_MAX_BRW_PAGES is set in tgt_brw_write()
234 * but for MDT it is different, correct it here. */
235 if (*nr_local > MD_MAX_BRW_PAGES)
236 *nr_local = MD_MAX_BRW_PAGES;
238 jobid = tsi->tsi_jobid;
240 if (!oa || objcount != 1 || obj->ioo_bufcnt == 0) {
241 CERROR("%s: bad parameters %p/%i/%i\n",
242 exp->exp_obd->obd_name, oa, objcount, obj->ioo_bufcnt);
246 mo = mdt_object_find(env, mdt, &tsi->tsi_fid);
248 GOTO(out, rc = PTR_ERR(mo));
250 LASSERT(info->mti_object == NULL);
251 info->mti_object = mo;
253 if (cmd == OBD_BRW_WRITE) {
254 la_from_obdo(la, oa, OBD_MD_FLGETATTR);
255 rc = mdt_preprw_write(env, exp, mdt, mo, la, oa,
256 objcount, obj, rnb, nr_local, lnb,
258 } else if (cmd == OBD_BRW_READ) {
259 mdt_grant_prepare_read(env, exp, oa);
260 rc = mdt_preprw_read(env, exp, mdt, mo, la,
261 obj->ioo_bufcnt, rnb, nr_local, lnb,
263 obdo_from_la(oa, la, LA_ATIME);
265 CERROR("%s: wrong cmd %d received!\n",
266 exp->exp_obd->obd_name, cmd);
270 lu_object_put(env, &mo->mot_obj);
271 info->mti_object = NULL;
277 static int mdt_commitrw_read(const struct lu_env *env, struct mdt_device *mdt,
278 struct mdt_object *mo, int objcount, int niocount,
279 struct niobuf_local *lnb)
281 struct dt_object *dob;
286 LASSERT(niocount > 0);
288 dob = mdt_obj2dt(mo);
290 dt_bufs_put(env, dob, lnb, niocount);
292 mdt_dom_read_unlock(mo);
296 static int mdt_commitrw_write(const struct lu_env *env, struct obd_export *exp,
297 struct mdt_device *mdt, struct mdt_object *mo,
298 struct lu_attr *la, int objcount, int niocount,
299 struct niobuf_local *lnb, unsigned long granted,
302 struct dt_device *dt = mdt->mdt_bottom;
303 struct dt_object *dob;
311 dob = mdt_obj2dt(mo);
314 GOTO(out, rc = old_rc);
316 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
318 if (!dt_object_exists(dob))
319 GOTO(out, rc = -ENOENT);
321 th = dt_trans_create(env, dt);
323 GOTO(out, rc = PTR_ERR(th));
325 for (i = 0; i < niocount; i++) {
326 if (!(lnb[i].lnb_flags & OBD_BRW_ASYNC)) {
332 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
333 GOTO(out_stop, rc = -EINPROGRESS);
335 rc = dt_declare_write_commit(env, dob, lnb, niocount, th);
340 /* update [mac]time if needed */
341 rc = dt_declare_attr_set(env, dob, la, th);
346 rc = dt_trans_start(env, dt, th);
350 dt_write_lock(env, dob, 0);
351 rc = dt_write_commit(env, dob, lnb, niocount, th);
356 rc = dt_attr_set(env, dob, la, th);
360 /* get attr to return */
361 rc = dt_attr_get(env, dob, la);
363 dt_write_unlock(env, dob);
366 /* Force commit to make the just-deleted blocks
367 * reusable. LU-456 */
372 dt_trans_stop(env, dt, th);
373 if (rc == -ENOSPC && retries++ < 3) {
374 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
380 dt_bufs_put(env, dob, lnb, niocount);
381 mdt_dom_read_unlock(mo);
382 mdt_grant_commit(exp, granted, old_rc);
386 int mdt_obd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
387 struct obdo *oa, int objcount, struct obd_ioobj *obj,
388 struct niobuf_remote *rnb, int npages,
389 struct niobuf_local *lnb, int old_rc)
391 struct mdt_thread_info *info = mdt_th_info(env);
392 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
393 struct mdt_object *mo = info->mti_object;
394 struct lu_attr *la = &info->mti_attr.ma_attr;
399 CERROR("%s: no pages to commit\n",
400 exp->exp_obd->obd_name);
406 if (cmd == OBD_BRW_WRITE) {
407 /* Don't update timestamps if this write is older than a
408 * setattr which modifies the timestamps. b=10150 */
410 /* XXX when we start having persistent reservations this needs
411 * to be changed to ofd_fmd_get() to create the fmd if it
412 * doesn't already exist so we can store the reservation handle
414 valid = OBD_MD_FLUID | OBD_MD_FLGID;
415 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME | OBD_MD_FLCTIME;
417 la_from_obdo(la, oa, valid);
419 rc = mdt_commitrw_write(env, exp, mdt, mo, la, objcount,
420 npages, lnb, oa->o_grant_used, old_rc);
422 obdo_from_la(oa, la, VALID_FLAGS | LA_GID | LA_UID);
424 obdo_from_la(oa, la, LA_GID | LA_UID);
426 /* don't report overquota flag if we failed before reaching
428 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
429 /* return the overquota flags to client */
430 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
431 if (oa->o_valid & OBD_MD_FLFLAGS)
432 oa->o_flags |= OBD_FL_NO_USRQUOTA;
434 oa->o_flags = OBD_FL_NO_USRQUOTA;
437 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
438 if (oa->o_valid & OBD_MD_FLFLAGS)
439 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
441 oa->o_flags = OBD_FL_NO_GRPQUOTA;
444 oa->o_valid |= OBD_MD_FLFLAGS | OBD_MD_FLUSRQUOTA |
447 } else if (cmd == OBD_BRW_READ) {
448 rc = mdt_commitrw_read(env, mdt, mo, objcount, npages, lnb);
454 /* this put is pair to object_get in ofd_preprw_write */
455 mdt_thread_info_fini(info);
459 int mdt_object_punch(const struct lu_env *env, struct dt_device *dt,
460 struct dt_object *dob, __u64 start, __u64 end,
468 /* we support truncate, not punch yet */
469 LASSERT(end == OBD_OBJECT_EOF);
471 if (!dt_object_exists(dob))
474 th = dt_trans_create(env, dt);
478 rc = dt_declare_attr_set(env, dob, la, th);
482 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
486 tgt_vbr_obj_set(env, dob);
487 rc = dt_trans_start(env, dt, th);
491 dt_write_lock(env, dob, 0);
492 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
495 rc = dt_attr_set(env, dob, la, th);
499 dt_write_unlock(env, dob);
502 dt_trans_stop(env, dt, th);
506 int mdt_punch_hdl(struct tgt_session_info *tsi)
508 const struct obdo *oa = &tsi->tsi_ost_body->oa;
509 struct ost_body *repbody;
510 struct mdt_thread_info *info;
512 struct ldlm_namespace *ns = tsi->tsi_tgt->lut_obd->obd_namespace;
513 struct obd_export *exp = tsi->tsi_exp;
514 struct mdt_device *mdt = mdt_dev(exp->exp_obd->obd_lu_dev);
515 struct mdt_object *mo;
516 struct dt_object *dob;
518 struct lustre_handle lh = { 0, };
525 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
526 CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
528 if ((oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
529 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
530 RETURN(err_serious(-EPROTO));
532 repbody = req_capsule_server_get(tsi->tsi_pill, &RMF_OST_BODY);
534 RETURN(err_serious(-ENOMEM));
536 /* punch start,end are passed in o_size,o_blocks throught wire */
540 if (end != OBD_OBJECT_EOF) /* Only truncate is supported */
543 info = tsi2mdt_info(tsi);
544 la = &info->mti_attr.ma_attr;
545 /* standard truncate optimization: if file body is completely
546 * destroyed, don't send data back to the server. */
548 flags |= LDLM_FL_AST_DISCARD_DATA;
550 repbody->oa.o_oi = oa->o_oi;
551 repbody->oa.o_valid = OBD_MD_FLID;
553 srvlock = (exp_connect_flags(exp) & OBD_CONNECT_SRVLOCK) &&
554 oa->o_valid & OBD_MD_FLFLAGS &&
555 oa->o_flags & OBD_FL_SRVLOCK;
558 rc = tgt_mdt_data_lock(ns, &tsi->tsi_resid, &lh, LCK_PW,
564 CDEBUG(D_INODE, "calling punch for object "DFID", valid = %#llx"
565 ", start = %lld, end = %lld\n", PFID(&tsi->tsi_fid),
566 oa->o_valid, start, end);
568 mo = mdt_object_find(tsi->tsi_env, mdt, &tsi->tsi_fid);
570 GOTO(out_unlock, rc = PTR_ERR(mo));
572 mdt_dom_write_lock(mo);
573 if (!mdt_object_exists(mo))
574 GOTO(out_put, rc = -ENOENT);
575 dob = mdt_obj2dt(mo);
577 la_from_obdo(la, oa, OBD_MD_FLMTIME | OBD_MD_FLATIME | OBD_MD_FLCTIME);
579 la->la_valid |= LA_SIZE;
581 rc = mdt_object_punch(tsi->tsi_env, mdt->mdt_bottom, dob,
583 mdt_dom_write_unlock(mo);
587 mdt_io_counter_incr(tsi->tsi_exp, LPROC_MDT_IO_PUNCH,
591 lu_object_put(tsi->tsi_env, &mo->mot_obj);
594 mdt_save_lock(info, &lh, LCK_PW, rc);
596 mdt_thread_info_fini(info);
598 struct ldlm_resource *res;
600 /* we do not call this before to avoid lu_object_find() in
601 * ->lvbo_update() holding another reference on the object.
602 * otherwise concurrent destroy can make the object unavailable
603 * for 2nd lu_object_find() waiting for the first reference
604 * to go... deadlock! */
605 res = ldlm_resource_get(ns, NULL, &tsi->tsi_resid,
608 ldlm_res_lvbo_update(res, NULL, 0);
609 ldlm_resource_putref(res);