1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
5 * Author: Peter Braam <braam@clusterfs.com>
6 * Author: Andreas Dilger <adilger@clusterfs.com>
8 * This file is part of the Lustre file system, http://www.lustre.org
9 * Lustre is a trademark of Cluster File Systems, Inc.
11 * You may have signed or agreed to another license before downloading
12 * this software. If so, you are bound by the terms and conditions
13 * of that agreement, and the following does not apply to you. See the
14 * LICENSE file included with this distribution for more information.
16 * If you did not agree to a different license, then this copy of Lustre
17 * is open source software; you can redistribute it and/or modify it
18 * under the terms of version 2 of the GNU General Public License as
19 * published by the Free Software Foundation.
21 * In either case, Lustre is distributed in the hope that it will be
22 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
23 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
24 * license text for more details.
28 # define EXPORT_SYMTAB
31 #define DEBUG_SUBSYSTEM S_ECHO
33 #include <obd_support.h>
34 #include <obd_class.h>
36 #include <lustre_debug.h>
37 #include <lustre_dlm.h>
38 #include <lprocfs_status.h>
40 #define ECHO_INIT_OBJID 0x1000000000000000ULL
41 #define ECHO_HANDLE_MAGIC 0xabcd0123fedc9876ULL
43 #define ECHO_PERSISTENT_PAGES (ECHO_PERSISTENT_SIZE >> CFS_PAGE_SHIFT)
44 static cfs_page_t *echo_persistent_pages[ECHO_PERSISTENT_PAGES];
47 LPROC_ECHO_READ_BYTES = 1,
48 LPROC_ECHO_WRITE_BYTES = 2,
49 LPROC_ECHO_LAST = LPROC_ECHO_WRITE_BYTES +1
52 static int echo_connect(struct lustre_handle *conn, struct obd_device *obd,
53 struct obd_uuid *cluuid, struct obd_connect_data *data)
55 data->ocd_connect_flags &= ECHO_CONNECT_SUPPORTED;
56 return class_connect(conn, obd, cluuid);
59 static int echo_disconnect(struct obd_export *exp)
61 LASSERT (exp != NULL);
63 ldlm_cancel_locks_for_export(exp);
65 /* complete all outstanding replies */
66 spin_lock(&exp->exp_lock);
67 while (!list_empty(&exp->exp_outstanding_replies)) {
68 struct ptlrpc_reply_state *rs =
69 list_entry(exp->exp_outstanding_replies.next,
70 struct ptlrpc_reply_state, rs_exp_list);
71 struct ptlrpc_service *svc = rs->rs_service;
73 spin_lock(&svc->srv_lock);
74 list_del_init(&rs->rs_exp_list);
75 ptlrpc_schedule_difficult_reply(rs);
76 spin_unlock(&svc->srv_lock);
78 spin_unlock(&exp->exp_lock);
80 return class_disconnect(exp);
83 static int echo_destroy_export(struct obd_export *exp)
87 target_destroy_export(exp);
92 static __u64 echo_next_id(struct obd_device *obddev)
96 spin_lock(&obddev->u.echo.eo_lock);
97 id = ++obddev->u.echo.eo_lastino;
98 spin_unlock(&obddev->u.echo.eo_lock);
103 int echo_create(struct obd_export *exp, struct obdo *oa,
104 struct lov_stripe_md **ea, struct obd_trans_info *oti)
106 struct obd_device *obd = class_exp2obd(exp);
109 CERROR("invalid client cookie "LPX64"\n",
110 exp->exp_handle.h_cookie);
114 if (!(oa->o_mode && S_IFMT)) {
115 CERROR("echo obd: no type!\n");
119 if (!(oa->o_valid & OBD_MD_FLTYPE)) {
120 CERROR("invalid o_valid "LPX64"\n", oa->o_valid);
124 oa->o_id = echo_next_id(obd);
125 oa->o_valid = OBD_MD_FLID;
130 int echo_destroy(struct obd_export *exp, struct obdo *oa,
131 struct lov_stripe_md *ea, struct obd_trans_info *oti,
132 struct obd_export *md_exp)
134 struct obd_device *obd = class_exp2obd(exp);
138 CERROR("invalid client cookie "LPX64"\n",
139 exp->exp_handle.h_cookie);
143 if (!(oa->o_valid & OBD_MD_FLID)) {
144 CERROR("obdo missing FLID valid flag: "LPX64"\n", oa->o_valid);
148 if (oa->o_id > obd->u.echo.eo_lastino || oa->o_id < ECHO_INIT_OBJID) {
149 CERROR("bad destroy objid: "LPX64"\n", oa->o_id);
156 static int echo_getattr(struct obd_export *exp, struct obd_info *oinfo)
158 struct obd_device *obd = class_exp2obd(exp);
159 obd_id id = oinfo->oi_oa->o_id;
163 CERROR("invalid client cookie "LPX64"\n",
164 exp->exp_handle.h_cookie);
168 if (!(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
169 CERROR("obdo missing FLID valid flag: "LPX64"\n",
170 oinfo->oi_oa->o_valid);
174 obdo_cpy_md(oinfo->oi_oa, &obd->u.echo.eo_oa, oinfo->oi_oa->o_valid);
175 oinfo->oi_oa->o_id = id;
180 static int echo_setattr(struct obd_export *exp, struct obd_info *oinfo,
181 struct obd_trans_info *oti)
183 struct obd_device *obd = class_exp2obd(exp);
187 CERROR("invalid client cookie "LPX64"\n",
188 exp->exp_handle.h_cookie);
192 if (!(oinfo->oi_oa->o_valid & OBD_MD_FLID)) {
193 CERROR("obdo missing FLID valid flag: "LPX64"\n",
194 oinfo->oi_oa->o_valid);
198 memcpy(&obd->u.echo.eo_oa, oinfo->oi_oa, sizeof(*oinfo->oi_oa));
200 if (oinfo->oi_oa->o_id & 4) {
201 /* Save lock to force ACKed reply */
202 ldlm_lock_addref (&obd->u.echo.eo_nl_lock, LCK_NL);
203 oti->oti_ack_locks[0].mode = LCK_NL;
204 oti->oti_ack_locks[0].lock = obd->u.echo.eo_nl_lock;
211 echo_page_debug_setup(cfs_page_t *page, int rw, obd_id id,
212 __u64 offset, int len)
214 int page_offset = offset & ~CFS_PAGE_MASK;
215 char *addr = ((char *)cfs_kmap(page)) + page_offset;
217 if (len % OBD_ECHO_BLOCK_SIZE != 0)
218 CERROR("Unexpected block size %d\n", len);
221 if (rw & OBD_BRW_READ)
222 block_debug_setup(addr, OBD_ECHO_BLOCK_SIZE,
225 block_debug_setup(addr, OBD_ECHO_BLOCK_SIZE,
226 0xecc0ecc0ecc0ecc0ULL,
227 0xecc0ecc0ecc0ecc0ULL);
229 addr += OBD_ECHO_BLOCK_SIZE;
230 offset += OBD_ECHO_BLOCK_SIZE;
231 len -= OBD_ECHO_BLOCK_SIZE;
238 echo_page_debug_check(cfs_page_t *page, obd_id id,
239 __u64 offset, int len)
241 int page_offset = offset & ~CFS_PAGE_MASK;
242 char *addr = ((char *)cfs_kmap(page)) + page_offset;
246 if (len % OBD_ECHO_BLOCK_SIZE != 0)
247 CERROR("Unexpected block size %d\n", len);
250 rc2 = block_debug_check("echo", addr, OBD_ECHO_BLOCK_SIZE,
253 if (rc2 != 0 && rc == 0)
256 addr += OBD_ECHO_BLOCK_SIZE;
257 offset += OBD_ECHO_BLOCK_SIZE;
258 len -= OBD_ECHO_BLOCK_SIZE;
266 /* This allows us to verify that desc_private is passed unmolested */
267 #define DESC_PRIV 0x10293847
269 int echo_preprw(int cmd, struct obd_export *export, struct obdo *oa,
270 int objcount, struct obd_ioobj *obj, int niocount,
271 struct niobuf_remote *nb, struct niobuf_local *res,
272 struct obd_trans_info *oti)
274 struct obd_device *obd;
275 struct niobuf_local *r = res;
281 obd = export->exp_obd;
285 /* Temp fix to stop falling foul of osc_announce_cached() */
286 oa->o_valid &= ~(OBD_MD_FLBLOCKS | OBD_MD_FLGRANT);
288 memset(res, 0, sizeof(*res) * niocount);
290 CDEBUG(D_PAGE, "%s %d obdos with %d IOs\n",
291 cmd == OBD_BRW_READ ? "reading" : "writing", objcount, niocount);
294 oti->oti_handle = (void *)DESC_PRIV;
296 for (i = 0; i < objcount; i++, obj++) {
297 int gfp_mask = (obj->ioo_id & 1) ? CFS_ALLOC_HIGHUSER : CFS_ALLOC_STD;
298 int ispersistent = obj->ioo_id == ECHO_PERSISTENT_OBJID;
299 int debug_setup = (!ispersistent &&
300 (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
301 (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0);
304 for (j = 0 ; j < obj->ioo_bufcnt ; j++, nb++, r++) {
307 (nb->offset >> CFS_PAGE_SHIFT) < ECHO_PERSISTENT_PAGES) {
308 r->page = echo_persistent_pages[nb->offset >>
310 /* Take extra ref so __free_pages() can be called OK */
311 cfs_get_page (r->page);
313 r->page = cfs_alloc_page(gfp_mask);
314 if (r->page == NULL) {
315 CERROR("can't get page %u/%u for id "
317 j, obj->ioo_bufcnt, obj->ioo_id);
318 GOTO(preprw_cleanup, rc = -ENOMEM);
322 tot_bytes += nb->len;
324 atomic_inc(&obd->u.echo.eo_prep);
326 r->offset = nb->offset;
328 LASSERT((r->offset & ~CFS_PAGE_MASK) + r->len <= CFS_PAGE_SIZE);
330 CDEBUG(D_PAGE, "$$$$ get page %p @ "LPU64" for %d\n",
331 r->page, r->offset, r->len);
333 if (cmd & OBD_BRW_READ)
337 echo_page_debug_setup(r->page, cmd, obj->ioo_id,
341 if (cmd & OBD_BRW_READ)
342 lprocfs_counter_add(obd->obd_stats, LPROC_ECHO_READ_BYTES,
345 lprocfs_counter_add(obd->obd_stats, LPROC_ECHO_WRITE_BYTES,
348 CDEBUG(D_PAGE, "%d pages allocated after prep\n",
349 atomic_read(&obd->u.echo.eo_prep));
354 /* It is possible that we would rather handle errors by allow
355 * any already-set-up pages to complete, rather than tearing them
356 * all down again. I believe that this is what the in-kernel
357 * prep/commit operations do.
359 CERROR("cleaning up %ld pages (%d obdos)\n", (long)(r - res), objcount);
362 /* NB if this is a persistent page, __free_pages will just
363 * lose the extra ref gained above */
364 cfs_free_page(r->page);
365 atomic_dec(&obd->u.echo.eo_prep);
367 memset(res, 0, sizeof(*res) * niocount);
372 int echo_commitrw(int cmd, struct obd_export *export, struct obdo *oa,
373 int objcount, struct obd_ioobj *obj, int niocount,
374 struct niobuf_local *res, struct obd_trans_info *oti, int rc)
376 struct obd_device *obd;
377 struct niobuf_local *r = res;
381 obd = export->exp_obd;
386 GOTO(commitrw_cleanup, rc);
388 if ((cmd & OBD_BRW_RWMASK) == OBD_BRW_READ) {
389 CDEBUG(D_PAGE, "reading %d obdos with %d IOs\n",
392 CDEBUG(D_PAGE, "writing %d obdos with %d IOs\n",
396 if (niocount && !r) {
397 CERROR("NULL res niobuf with niocount %d\n", niocount);
401 LASSERT(oti == NULL || oti->oti_handle == (void *)DESC_PRIV);
403 for (i = 0; i < objcount; i++, obj++) {
404 int verify = (rc == 0 &&
405 obj->ioo_id != ECHO_PERSISTENT_OBJID &&
406 (oa->o_valid & OBD_MD_FLFLAGS) != 0 &&
407 (oa->o_flags & OBD_FL_DEBUG_CHECK) != 0);
410 for (j = 0 ; j < obj->ioo_bufcnt ; j++, r++) {
411 cfs_page_t *page = r->page;
415 CERROR("null page objid "LPU64":%p, buf %d/%d\n",
416 obj->ioo_id, page, j, obj->ioo_bufcnt);
417 GOTO(commitrw_cleanup, rc = -EFAULT);
420 addr = cfs_kmap(page);
422 CDEBUG(D_PAGE, "$$$$ use page %p, addr %p@"LPU64"\n",
423 r->page, addr, r->offset);
426 vrc = echo_page_debug_check(page, obj->ioo_id,
428 /* check all the pages always */
429 if (vrc != 0 && rc == 0)
434 /* NB see comment above regarding persistent pages */
436 atomic_dec(&obd->u.echo.eo_prep);
439 CDEBUG(D_PAGE, "%d pages remain after commit\n",
440 atomic_read(&obd->u.echo.eo_prep));
444 CERROR("cleaning up %ld pages (%d obdos)\n",
445 niocount - (long)(r - res) - 1, objcount);
446 while (++r < res + niocount) {
447 cfs_page_t *page = r->page;
449 /* NB see comment above regarding persistent pages */
451 atomic_dec(&obd->u.echo.eo_prep);
456 static int echo_setup(struct obd_device *obd, obd_count len, void *buf)
458 struct lprocfs_static_vars lvars;
461 struct ldlm_res_id res_id = {.name = {1}};
464 spin_lock_init(&obd->u.echo.eo_lock);
465 obd->u.echo.eo_lastino = ECHO_INIT_OBJID;
467 obd->obd_namespace = ldlm_namespace_new("echo-tgt",
468 LDLM_NAMESPACE_SERVER);
469 if (obd->obd_namespace == NULL) {
474 rc = ldlm_cli_enqueue_local(obd->obd_namespace, res_id, LDLM_PLAIN,
475 NULL, LCK_NL, &lock_flags, NULL,
476 ldlm_completion_ast, NULL, NULL,
477 0, NULL, &obd->u.echo.eo_nl_lock);
478 LASSERT (rc == ELDLM_OK);
480 lprocfs_init_vars(echo, &lvars);
481 if (lprocfs_obd_setup(obd, lvars.obd_vars) == 0 &&
482 lprocfs_alloc_obd_stats(obd, LPROC_ECHO_LAST) == 0) {
483 lprocfs_counter_init(obd->obd_stats, LPROC_ECHO_READ_BYTES,
484 LPROCFS_CNTR_AVGMINMAX,
485 "read_bytes", "bytes");
486 lprocfs_counter_init(obd->obd_stats, LPROC_ECHO_WRITE_BYTES,
487 LPROCFS_CNTR_AVGMINMAX,
488 "write_bytes", "bytes");
491 ptlrpc_init_client (LDLM_CB_REQUEST_PORTAL, LDLM_CB_REPLY_PORTAL,
492 "echo_ldlm_cb_client", &obd->obd_ldlm_client);
496 static int echo_cleanup(struct obd_device *obd)
501 lprocfs_obd_cleanup(obd);
502 lprocfs_free_obd_stats(obd);
504 ldlm_lock_decref (&obd->u.echo.eo_nl_lock, LCK_NL);
506 /* XXX Bug 3413; wait for a bit to ensure the BL callback has
507 * happened before calling ldlm_namespace_free() */
508 set_current_state (TASK_UNINTERRUPTIBLE);
509 cfs_schedule_timeout (CFS_TASK_UNINT, cfs_time_seconds(1));
511 ldlm_namespace_free(obd->obd_namespace, obd->obd_force);
513 leaked = atomic_read(&obd->u.echo.eo_prep);
515 CERROR("%d prep/commitrw pages leaked\n", leaked);
520 static struct obd_ops echo_obd_ops = {
521 .o_owner = THIS_MODULE,
522 .o_connect = echo_connect,
523 .o_disconnect = echo_disconnect,
524 .o_destroy_export = echo_destroy_export,
525 .o_create = echo_create,
526 .o_destroy = echo_destroy,
527 .o_getattr = echo_getattr,
528 .o_setattr = echo_setattr,
529 .o_preprw = echo_preprw,
530 .o_commitrw = echo_commitrw,
531 .o_setup = echo_setup,
532 .o_cleanup = echo_cleanup
535 extern int echo_client_init(void);
536 extern void echo_client_exit(void);
539 echo_persistent_pages_fini (void)
543 for (i = 0; i < ECHO_PERSISTENT_PAGES; i++)
544 if (echo_persistent_pages[i] != NULL) {
545 cfs_free_page (echo_persistent_pages[i]);
546 echo_persistent_pages[i] = NULL;
551 echo_persistent_pages_init (void)
556 for (i = 0; i < ECHO_PERSISTENT_PAGES; i++) {
557 int gfp_mask = (i < ECHO_PERSISTENT_PAGES/2) ?
558 CFS_ALLOC_STD : CFS_ALLOC_HIGHUSER;
560 pg = cfs_alloc_page (gfp_mask);
562 echo_persistent_pages_fini ();
566 memset (cfs_kmap (pg), 0, CFS_PAGE_SIZE);
569 echo_persistent_pages[i] = pg;
575 static int __init obdecho_init(void)
577 struct lprocfs_static_vars lvars;
581 printk(KERN_INFO "Lustre: Echo OBD driver; info@clusterfs.com\n");
583 LASSERT(CFS_PAGE_SIZE % OBD_ECHO_BLOCK_SIZE == 0);
585 lprocfs_init_vars(echo, &lvars);
587 rc = echo_persistent_pages_init ();
591 rc = class_register_type(&echo_obd_ops, lvars.module_vars,
596 rc = echo_client_init();
600 class_unregister_type(LUSTRE_ECHO_NAME);
602 echo_persistent_pages_fini ();
607 static void /*__exit*/ obdecho_exit(void)
610 class_unregister_type(LUSTRE_ECHO_NAME);
611 echo_persistent_pages_fini ();
614 MODULE_AUTHOR("Cluster File Systems, Inc. <info@clusterfs.com>");
615 MODULE_DESCRIPTION("Lustre Testing Echo OBD driver");
616 MODULE_LICENSE("GPL");
618 cfs_module(obdecho, "1.0.0", obdecho_init, obdecho_exit);