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) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2010, 2011, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/obdecho/echo.c
38 * Author: Peter Braam <braam@clusterfs.com>
39 * Author: Andreas Dilger <adilger@clusterfs.com>
42 #define DEBUG_SUBSYSTEM S_ECHO
44 #include <obd_support.h>
45 #include <obd_class.h>
46 #include <lustre_debug.h>
47 #include <lustre_dlm.h>
48 #include <lprocfs_status.h>
50 #include "echo_internal.h"
52 /* The echo objid needs to be below 2^32, because regular FID numbers are
53 * limited to 2^32 objects in f_oid for the FID_SEQ_ECHO range. b=23335 */
54 #define ECHO_INIT_OID 0x10000000ULL
55 #define ECHO_HANDLE_MAGIC 0xabcd0123fedc9876ULL
57 #define ECHO_PERSISTENT_PAGES (ECHO_PERSISTENT_SIZE >> CFS_PAGE_SHIFT)
58 static cfs_page_t *echo_persistent_pages[ECHO_PERSISTENT_PAGES];
61 LPROC_ECHO_READ_BYTES = 1,
62 LPROC_ECHO_WRITE_BYTES = 2,
63 LPROC_ECHO_LAST = LPROC_ECHO_WRITE_BYTES +1
66 static int echo_connect(const struct lu_env *env,
67 struct obd_export **exp, struct obd_device *obd,
68 struct obd_uuid *cluuid, struct obd_connect_data *data,
71 struct lustre_handle conn = { 0 };
74 data->ocd_connect_flags &= ECHO_CONNECT_SUPPORTED;
75 rc = class_connect(&conn, obd, cluuid);
77 CERROR("can't connect %d\n", rc);
80 *exp = class_conn2export(&conn);
85 static int echo_disconnect(struct obd_export *exp)
87 LASSERT (exp != NULL);
89 return server_disconnect_export(exp);
92 static int echo_init_export(struct obd_export *exp)
94 return ldlm_init_export(exp);
97 static int echo_destroy_export(struct obd_export *exp)
101 target_destroy_export(exp);
102 ldlm_destroy_export(exp);
107 static __u64 echo_next_id(struct obd_device *obddev)
111 spin_lock(&obddev->u.echo.eo_lock);
112 id = ++obddev->u.echo.eo_lastino;
113 spin_unlock(&obddev->u.echo.eo_lock);
118 static int echo_create(const struct lu_env *env, struct obd_export *exp,
119 struct obdo *oa, struct lov_stripe_md **ea,
120 struct obd_trans_info *oti)
122 struct obd_device *obd = class_exp2obd(exp);
125 CERROR("invalid client cookie "LPX64"\n",
126 exp->exp_handle.h_cookie);
130 if (!(oa->o_mode && S_IFMT)) {
131 CERROR("echo obd: no type!\n");
135 if (!(oa->o_valid & OBD_MD_FLTYPE)) {
136 CERROR("invalid o_valid "LPX64"\n", oa->o_valid);
140 oa->o_id = echo_next_id(obd);
141 oa->o_valid = OBD_MD_FLID;
146 static int echo_destroy(const struct lu_env *env, struct obd_export *exp,
147 struct obdo *oa, struct lov_stripe_md *ea,
148 struct obd_trans_info *oti, struct obd_export *md_exp,
151 struct obd_device *obd = class_exp2obd(exp);
155 CERROR("invalid client cookie "LPX64"\n",
156 exp->exp_handle.h_cookie);
160 if (!(oa->o_valid & OBD_MD_FLID)) {
161 CERROR("obdo missing FLID valid flag: "LPX64"\n", oa->o_valid);
165 if (oa->o_id > obd->u.echo.eo_lastino || oa->o_id < ECHO_INIT_OID) {
166 CERROR("bad destroy objid: "LPX64"\n", oa->o_id);
173 static int echo_getattr(const struct lu_env *env, struct obd_export *exp,
174 struct obd_info *oinfo)
176 struct obd_device *obd = class_exp2obd(exp);
177 obd_id id = oinfo->oi_oa->o_id;
181 CERROR("invalid client cookie "LPX64"\n",
182 exp->exp_handle.h_cookie);
186 if (!(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
187 CERROR("obdo missing FLID valid flag: "LPX64"\n",
188 oinfo->oi_oa->o_valid);
192 obdo_cpy_md(oinfo->oi_oa, &obd->u.echo.eo_oa, oinfo->oi_oa->o_valid);
193 oinfo->oi_oa->o_id = id;
198 static int echo_setattr(const struct lu_env *env, struct obd_export *exp,
199 struct obd_info *oinfo, struct obd_trans_info *oti)
201 struct obd_device *obd = class_exp2obd(exp);
205 CERROR("invalid client cookie "LPX64"\n",
206 exp->exp_handle.h_cookie);
210 if (!(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
211 CERROR("obdo missing FLID valid flag: "LPX64"\n",
212 oinfo->oi_oa->o_valid);
216 memcpy(&obd->u.echo.eo_oa, oinfo->oi_oa, sizeof(*oinfo->oi_oa));
218 if (oinfo->oi_oa->o_id & 4) {
219 /* Save lock to force ACKed reply */
220 ldlm_lock_addref (&obd->u.echo.eo_nl_lock, LCK_NL);
221 oti->oti_ack_locks[0].mode = LCK_NL;
222 oti->oti_ack_locks[0].lock = obd->u.echo.eo_nl_lock;
229 echo_page_debug_setup(cfs_page_t *page, int rw, obd_id id,
230 __u64 offset, int len)
232 int page_offset = offset & ~CFS_PAGE_MASK;
233 char *addr = ((char *)cfs_kmap(page)) + page_offset;
235 if (len % OBD_ECHO_BLOCK_SIZE != 0)
236 CERROR("Unexpected block size %d\n", len);
239 if (rw & OBD_BRW_READ)
240 block_debug_setup(addr, OBD_ECHO_BLOCK_SIZE,
243 block_debug_setup(addr, OBD_ECHO_BLOCK_SIZE,
244 0xecc0ecc0ecc0ecc0ULL,
245 0xecc0ecc0ecc0ecc0ULL);
247 addr += OBD_ECHO_BLOCK_SIZE;
248 offset += OBD_ECHO_BLOCK_SIZE;
249 len -= OBD_ECHO_BLOCK_SIZE;
256 echo_page_debug_check(cfs_page_t *page, obd_id id,
257 __u64 offset, int len)
259 int page_offset = offset & ~CFS_PAGE_MASK;
260 char *addr = ((char *)cfs_kmap(page)) + page_offset;
264 if (len % OBD_ECHO_BLOCK_SIZE != 0)
265 CERROR("Unexpected block size %d\n", len);
268 rc2 = block_debug_check("echo", addr, OBD_ECHO_BLOCK_SIZE,
271 if (rc2 != 0 && rc == 0)
274 addr += OBD_ECHO_BLOCK_SIZE;
275 offset += OBD_ECHO_BLOCK_SIZE;
276 len -= OBD_ECHO_BLOCK_SIZE;
284 /* This allows us to verify that desc_private is passed unmolested */
285 #define DESC_PRIV 0x10293847
287 static int echo_map_nb_to_lb(struct obdo *oa, struct obd_ioobj *obj,
288 struct niobuf_remote *nb, int *pages,
289 struct niobuf_local *lb, int cmd, int *left)
291 int gfp_mask = (obj->ioo_id & 1) ? CFS_ALLOC_HIGHUSER : CFS_ALLOC_STD;
292 int ispersistent = obj->ioo_id == ECHO_PERSISTENT_OBJID;
293 int debug_setup = (!ispersistent &&
294 (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
295 (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0);
296 struct niobuf_local *res = lb;
297 obd_off offset = nb->offset;
301 int plen = CFS_PAGE_SIZE - (offset & (CFS_PAGE_SIZE-1));
305 /* check for local buf overflow */
309 res->lnb_file_offset = offset;
311 LASSERT((res->lnb_file_offset & ~CFS_PAGE_MASK) + res->len <=
315 ((res->lnb_file_offset >> CFS_PAGE_SHIFT) <
316 ECHO_PERSISTENT_PAGES)) {
318 echo_persistent_pages[res->lnb_file_offset >>
320 /* Take extra ref so __free_pages() can be called OK */
321 cfs_get_page (res->page);
323 OBD_PAGE_ALLOC(res->page, gfp_mask);
324 if (res->page == NULL) {
325 CERROR("can't get page for id " LPU64"\n",
331 CDEBUG(D_PAGE, "$$$$ get page %p @ "LPU64" for %d\n",
332 res->page, res->lnb_file_offset, res->len);
334 if (cmd & OBD_BRW_READ)
338 echo_page_debug_setup(res->page, cmd, obj->ioo_id,
339 res->lnb_file_offset, res->len);
352 static int echo_finalize_lb(struct obdo *oa, struct obd_ioobj *obj,
353 struct niobuf_remote *rb, int *pgs,
354 struct niobuf_local *lb, int verify)
356 struct niobuf_local *res = lb;
357 obd_off start = rb->offset >> CFS_PAGE_SHIFT;
358 obd_off end = (rb->offset + rb->len + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
359 int count = (int)(end - start);
363 for (i = 0; i < count; i++, (*pgs) ++, res++) {
364 cfs_page_t *page = res->page;
368 CERROR("null page objid "LPU64":%p, buf %d/%d\n",
369 obj->ioo_id, page, i, obj->ioo_bufcnt);
373 addr = cfs_kmap(page);
375 CDEBUG(D_PAGE, "$$$$ use page %p, addr %p@"LPU64"\n",
376 res->page, addr, res->lnb_file_offset);
379 int vrc = echo_page_debug_check(page, obj->ioo_id,
380 res->lnb_file_offset,
382 /* check all the pages always */
383 if (vrc != 0 && rc == 0)
388 /* NB see comment above regarding persistent pages */
395 static int echo_preprw(const struct lu_env *env, int cmd,
396 struct obd_export *export, struct obdo *oa,
397 int objcount, struct obd_ioobj *obj,
398 struct niobuf_remote *nb, int *pages,
399 struct niobuf_local *res, struct obd_trans_info *oti,
400 struct lustre_capa *unused)
402 struct obd_device *obd;
408 obd = export->exp_obd;
412 /* Temp fix to stop falling foul of osc_announce_cached() */
413 oa->o_valid &= ~(OBD_MD_FLBLOCKS | OBD_MD_FLGRANT);
415 memset(res, 0, sizeof(*res) * *pages);
417 CDEBUG(D_PAGE, "%s %d obdos with %d IOs\n",
418 cmd == OBD_BRW_READ ? "reading" : "writing", objcount, *pages);
421 oti->oti_handle = (void *)DESC_PRIV;
426 for (i = 0; i < objcount; i++, obj++) {
429 for (j = 0 ; j < obj->ioo_bufcnt ; j++, nb++) {
431 rc = echo_map_nb_to_lb(oa, obj, nb, pages,
432 res + *pages, cmd, &left);
434 GOTO(preprw_cleanup, rc);
436 tot_bytes += nb->len;
440 cfs_atomic_add(*pages, &obd->u.echo.eo_prep);
442 if (cmd & OBD_BRW_READ)
443 lprocfs_counter_add(obd->obd_stats, LPROC_ECHO_READ_BYTES,
446 lprocfs_counter_add(obd->obd_stats, LPROC_ECHO_WRITE_BYTES,
449 CDEBUG(D_PAGE, "%d pages allocated after prep\n",
450 cfs_atomic_read(&obd->u.echo.eo_prep));
455 /* It is possible that we would rather handle errors by allow
456 * any already-set-up pages to complete, rather than tearing them
457 * all down again. I believe that this is what the in-kernel
458 * prep/commit operations do.
460 CERROR("cleaning up %u pages (%d obdos)\n", *pages, objcount);
461 for (i = 0; i < *pages; i++) {
462 cfs_kunmap(res[i].page);
463 /* NB if this is a persistent page, __free_pages will just
464 * lose the extra ref gained above */
465 OBD_PAGE_FREE(res[i].page);
467 cfs_atomic_dec(&obd->u.echo.eo_prep);
473 static int echo_commitrw(const struct lu_env *env, int cmd,
474 struct obd_export *export, struct obdo *oa,
475 int objcount, struct obd_ioobj *obj,
476 struct niobuf_remote *rb, int niocount,
477 struct niobuf_local *res, struct obd_trans_info *oti,
480 struct obd_device *obd;
485 obd = export->exp_obd;
490 GOTO(commitrw_cleanup, rc);
492 if ((cmd & OBD_BRW_RWMASK) == OBD_BRW_READ) {
493 CDEBUG(D_PAGE, "reading %d obdos with %d IOs\n",
496 CDEBUG(D_PAGE, "writing %d obdos with %d IOs\n",
500 if (niocount && res == NULL) {
501 CERROR("NULL res niobuf with niocount %d\n", niocount);
505 LASSERT(oti == NULL || oti->oti_handle == (void *)DESC_PRIV);
507 for (i = 0; i < objcount; i++, obj++) {
508 int verify = (rc == 0 &&
509 obj->ioo_id != ECHO_PERSISTENT_OBJID &&
510 (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
511 (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0);
514 for (j = 0 ; j < obj->ioo_bufcnt ; j++, rb++) {
515 int vrc = echo_finalize_lb(oa, obj, rb, &pgs, &res[pgs], verify);
521 GOTO(commitrw_cleanup, rc = vrc);
529 cfs_atomic_sub(pgs, &obd->u.echo.eo_prep);
531 CDEBUG(D_PAGE, "%d pages remain after commit\n",
532 cfs_atomic_read(&obd->u.echo.eo_prep));
536 cfs_atomic_sub(pgs, &obd->u.echo.eo_prep);
538 CERROR("cleaning up %d pages (%d obdos)\n",
539 niocount - pgs - 1, objcount);
541 while (pgs < niocount) {
542 cfs_page_t *page = res[pgs++].page;
547 /* NB see comment above regarding persistent pages */
549 cfs_atomic_dec(&obd->u.echo.eo_prep);
554 static int echo_setup(struct obd_device *obd, struct lustre_cfg *lcfg)
556 struct lprocfs_static_vars lvars;
558 __u64 lock_flags = 0;
559 struct ldlm_res_id res_id = {.name = {1}};
563 obd->u.echo.eo_obt.obt_magic = OBT_MAGIC;
564 spin_lock_init(&obd->u.echo.eo_lock);
565 obd->u.echo.eo_lastino = ECHO_INIT_OID;
567 sprintf(ns_name, "echotgt-%s", obd->obd_uuid.uuid);
568 obd->obd_namespace = ldlm_namespace_new(obd, ns_name,
569 LDLM_NAMESPACE_SERVER,
570 LDLM_NAMESPACE_MODEST,
572 if (obd->obd_namespace == NULL) {
577 rc = ldlm_cli_enqueue_local(obd->obd_namespace, &res_id, LDLM_PLAIN,
578 NULL, LCK_NL, &lock_flags, NULL,
579 ldlm_completion_ast, NULL, NULL, 0,
580 LVB_T_NONE, NULL, &obd->u.echo.eo_nl_lock);
581 LASSERT (rc == ELDLM_OK);
583 lprocfs_echo_init_vars(&lvars);
584 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
585 lprocfs_alloc_obd_stats(obd, LPROC_ECHO_LAST) == 0) {
586 lprocfs_counter_init(obd->obd_stats, LPROC_ECHO_READ_BYTES,
587 LPROCFS_CNTR_AVGMINMAX,
588 "read_bytes", "bytes");
589 lprocfs_counter_init(obd->obd_stats, LPROC_ECHO_WRITE_BYTES,
590 LPROCFS_CNTR_AVGMINMAX,
591 "write_bytes", "bytes");
594 ptlrpc_init_client (LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
595 "echo_ldlm_cb_client", &obd->obd_ldlm_client);
599 static int echo_cleanup(struct obd_device *obd)
604 lprocfs_obd_cleanup(obd);
605 lprocfs_free_obd_stats(obd);
607 ldlm_lock_decref(&obd->u.echo.eo_nl_lock, LCK_NL);
609 /* XXX Bug 3413; wait for a bit to ensure the BL callback has
610 * happened before calling ldlm_namespace_free() */
611 cfs_schedule_timeout_and_set_state(CFS_TASK_UNINT, cfs_time_seconds(1));
613 ldlm_namespace_free(obd->obd_namespace, NULL, obd->obd_force);
614 obd->obd_namespace = NULL;
616 leaked = cfs_atomic_read(&obd->u.echo.eo_prep);
618 CERROR("%d prep/commitrw pages leaked\n", leaked);
623 struct obd_ops echo_obd_ops = {
624 .o_owner = THIS_MODULE,
625 .o_connect = echo_connect,
626 .o_disconnect = echo_disconnect,
627 .o_init_export = echo_init_export,
628 .o_destroy_export = echo_destroy_export,
629 .o_create = echo_create,
630 .o_destroy = echo_destroy,
631 .o_getattr = echo_getattr,
632 .o_setattr = echo_setattr,
633 .o_preprw = echo_preprw,
634 .o_commitrw = echo_commitrw,
635 .o_setup = echo_setup,
636 .o_cleanup = echo_cleanup
639 void echo_persistent_pages_fini(void)
643 for (i = 0; i < ECHO_PERSISTENT_PAGES; i++)
644 if (echo_persistent_pages[i] != NULL) {
645 OBD_PAGE_FREE(echo_persistent_pages[i]);
646 echo_persistent_pages[i] = NULL;
650 int echo_persistent_pages_init(void)
655 for (i = 0; i < ECHO_PERSISTENT_PAGES; i++) {
656 int gfp_mask = (i < ECHO_PERSISTENT_PAGES/2) ?
657 CFS_ALLOC_STD : CFS_ALLOC_HIGHUSER;
659 OBD_PAGE_ALLOC(pg, gfp_mask);
661 echo_persistent_pages_fini ();
665 memset (cfs_kmap (pg), 0, CFS_PAGE_SIZE);
668 echo_persistent_pages[i] = pg;