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 (c) 2002, 2010, Oracle and/or its affiliates. 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 * lustre/lov/lov_pack.c
38 * (Un)packing of OST/MDS requests
40 * Author: Andreas Dilger <adilger@clusterfs.com>
43 #define DEBUG_SUBSYSTEM S_LOV
45 #include <liblustre.h>
48 #include <lustre_net.h>
51 #include <obd_class.h>
52 #include <obd_support.h>
53 #include <lustre/lustre_user.h>
55 #include "lov_internal.h"
57 static void lov_dump_lmm_common(int level, void *lmmp)
59 struct lov_mds_md *lmm = lmmp;
61 CDEBUG(level, "objid "LPX64", magic 0x%08x, pattern %#x\n",
62 (__u64)le64_to_cpu(lmm->lmm_object_id),
63 le32_to_cpu(lmm->lmm_magic),
64 le32_to_cpu(lmm->lmm_pattern));
65 CDEBUG(level,"stripe_size %u, stripe_count %u\n",
66 le32_to_cpu(lmm->lmm_stripe_size),
67 le32_to_cpu(lmm->lmm_stripe_count));
70 static void lov_dump_lmm_objects(int level, struct lov_ost_data *lod,
75 if (stripe_count > LOV_V1_INSANE_STRIPE_COUNT) {
76 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
77 stripe_count, LOV_V1_INSANE_STRIPE_COUNT);
80 for (i = 0; i < stripe_count; ++i, ++lod) {
81 CDEBUG(level, "stripe %u idx %u subobj "LPX64"/"LPX64"\n", i,
82 le32_to_cpu(lod->l_ost_idx),
83 (__u64)le64_to_cpu(lod->l_object_seq),
84 (__u64)le64_to_cpu(lod->l_object_id));
88 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
90 lov_dump_lmm_common(level, lmm);
91 lov_dump_lmm_objects(level, lmm->lmm_objects,
92 le32_to_cpu(lmm->lmm_stripe_count));
95 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
97 lov_dump_lmm_common(level, lmm);
98 CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
99 lov_dump_lmm_objects(level, lmm->lmm_objects,
100 le32_to_cpu(lmm->lmm_stripe_count));
103 void lov_dump_lmm(int level, void *lmm)
107 magic = ((struct lov_mds_md_v1 *)(lmm))->lmm_magic;
110 return lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)(lmm));
112 return lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)(lmm));
114 CERROR("Cannot recognize lmm_magic %x", magic);
119 #define LMM_ASSERT(test) \
121 if (!(test)) lov_dump_lmm(D_ERROR, lmm); \
122 LASSERT(test); /* so we know what assertion failed */ \
125 /* Pack LOV object metadata for disk storage. It is packed in LE byte
126 * order and is opaque to the networking layer.
128 * XXX In the future, this will be enhanced to get the EA size from the
129 * underlying OSC device(s) to get their EA sizes so we can stack
130 * LOVs properly. For now lov_mds_md_size() just assumes one obd_id
133 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
134 struct lov_stripe_md *lsm)
136 struct obd_device *obd = class_exp2obd(exp);
137 struct lov_obd *lov = &obd->u.lov;
138 struct lov_mds_md_v1 *lmmv1;
139 struct lov_mds_md_v3 *lmmv3;
141 struct lov_ost_data_v1 *lmm_objects;
142 int lmm_size, lmm_magic;
147 lmm_magic = lsm->lsm_magic;
149 /* If we are just sizing the EA, limit the stripe count
150 * to the actual number of OSTs in this filesystem. */
152 stripe_count = lov_get_stripecnt(lov,
153 lsm->lsm_stripe_count);
154 lsm->lsm_stripe_count = stripe_count;
156 stripe_count = lsm->lsm_stripe_count;
159 /* No needs to allocated more than LOV_MAX_STRIPE_COUNT.
160 * Anyway, this is pretty inaccurate since ld_tgt_count now
161 * represents max index and we should rely on the actual number
163 stripe_count = min((__u32)LOV_MAX_STRIPE_COUNT,
164 lov->desc.ld_tgt_count);
167 lmm_magic = le32_to_cpu((*lmmp)->lmm_magic);
169 /* lsm == NULL and lmmp == NULL */
170 lmm_magic = LOV_MAGIC;
173 if ((lmm_magic != LOV_MAGIC_V1) &&
174 (lmm_magic != LOV_MAGIC_V3)) {
175 CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
176 lmm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
181 /* XXX LOV STACKING call into osc for sizes */
182 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
188 stripe_count = le32_to_cpu((*lmmp)->lmm_stripe_count);
189 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
190 OBD_FREE(*lmmp, lmm_size);
196 OBD_ALLOC(*lmmp, lmm_size);
201 CDEBUG(D_INFO, "lov_packmd: LOV_MAGIC 0x%08X, lmm_size = %d \n",
202 lmm_magic, lmm_size);
205 lmmv3 = (struct lov_mds_md_v3 *)*lmmp;
206 if (lmm_magic == LOV_MAGIC_V3)
207 lmmv3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3);
209 lmmv1->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
214 /* lmmv1 and lmmv3 point to the same struct and have the
217 lmmv1->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
218 lmmv1->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
219 lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
220 lmmv1->lmm_stripe_count = cpu_to_le32(stripe_count);
221 lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
222 if (lsm->lsm_magic == LOV_MAGIC_V3) {
223 strncpy(lmmv3->lmm_pool_name, lsm->lsm_pool_name,
225 lmm_objects = lmmv3->lmm_objects;
227 lmm_objects = lmmv1->lmm_objects;
230 for (i = 0; i < stripe_count; i++) {
231 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
232 /* XXX LOV STACKING call down to osc_packmd() to do packing */
233 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
234 lmmv1->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
235 lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
236 lmm_objects[i].l_object_seq = cpu_to_le64(loi->loi_seq);
237 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
238 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
244 /* Find the max stripecount we should use */
245 int lov_get_stripecnt(struct lov_obd *lov, __u32 stripe_count)
248 stripe_count = lov->desc.ld_default_stripe_count;
249 if (stripe_count > lov->desc.ld_active_tgt_count)
250 stripe_count = lov->desc.ld_active_tgt_count;
253 /* for now, we limit the stripe count directly, when bug 4424 is
254 * fixed this needs to be somewhat dynamic based on whether ext3
255 * can handle larger EA sizes. */
256 if (stripe_count > LOV_MAX_STRIPE_COUNT)
257 stripe_count = LOV_MAX_STRIPE_COUNT;
263 static int lov_verify_lmm(void *lmm, int lmm_bytes, int *stripe_count)
267 if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
271 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
272 le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
273 sz = lmm_bytes * 2 + 1;
274 OBD_ALLOC(buffer, sz);
275 if (buffer != NULL) {
278 for (i = 0; i < lmm_bytes; i++)
279 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
281 CERROR("%s\n", buffer);
282 OBD_FREE(buffer, sz);
286 rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
287 lmm_bytes, stripe_count);
291 int lov_alloc_memmd(struct lov_stripe_md **lsmp, int stripe_count,
292 int pattern, int magic)
297 CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
299 *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
301 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
305 cfs_spin_lock_init(&(*lsmp)->lsm_lock);
306 (*lsmp)->lsm_magic = magic;
307 (*lsmp)->lsm_stripe_count = stripe_count;
308 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
309 (*lsmp)->lsm_pattern = pattern;
310 (*lsmp)->lsm_pool_name[0] = '\0';
311 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
313 for (i = 0; i < stripe_count; i++)
314 loi_init((*lsmp)->lsm_oinfo[i]);
319 void lov_free_memmd(struct lov_stripe_md **lsmp)
321 struct lov_stripe_md *lsm = *lsmp;
323 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
324 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
330 /* Unpack LOV object metadata from disk storage. It is packed in LE byte
331 * order and is opaque to the networking layer.
333 int lov_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
334 struct lov_mds_md *lmm, int lmm_bytes)
336 struct obd_device *obd = class_exp2obd(exp);
337 struct lov_obd *lov = &obd->u.lov;
338 int rc = 0, stripe_count, lsm_size;
342 /* If passed an MDS struct use values from there, otherwise defaults */
344 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
347 magic = le32_to_cpu(lmm->lmm_magic);
349 stripe_count = lov_get_stripecnt(lov, 0);
353 /* If we aren't passed an lsmp struct, we just want the size */
355 /* XXX LOV STACKING call into osc for sizes */
357 RETURN(lov_stripe_md_size(stripe_count));
359 /* If we are passed an allocated struct but nothing to unpack, free */
361 lov_free_memmd(lsmp);
365 lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
370 /* If we are passed a pointer but nothing to unpack, we only alloc */
374 LASSERT(lsm_op_find(magic) != NULL);
375 rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
377 lov_free_memmd(lsmp);
384 static int __lov_setstripe(struct obd_export *exp, int max_lmm_size,
385 struct lov_stripe_md **lsmp,
386 struct lov_user_md *lump)
388 struct obd_device *obd = class_exp2obd(exp);
389 struct lov_obd *lov = &obd->u.lov;
390 char buffer[sizeof(struct lov_user_md_v3)];
391 struct lov_user_md_v3 *lumv3 = (struct lov_user_md_v3 *)&buffer[0];
392 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&buffer[0];
398 if (cfs_copy_from_user(lumv3, lump, sizeof(struct lov_user_md_v1)))
401 lmm_magic = lumv1->lmm_magic;
403 if (lmm_magic == __swab32(LOV_USER_MAGIC_V1)) {
404 lustre_swab_lov_user_md_v1(lumv1);
405 lmm_magic = LOV_USER_MAGIC_V1;
406 } else if (lmm_magic == LOV_USER_MAGIC_V3) {
407 if (cfs_copy_from_user(lumv3, lump, sizeof(*lumv3)))
409 } else if (lmm_magic == __swab32(LOV_USER_MAGIC_V3)) {
410 if (cfs_copy_from_user(lumv3, lump, sizeof(*lumv3)))
412 lustre_swab_lov_user_md_v3(lumv3);
413 lmm_magic = LOV_USER_MAGIC_V3;
414 } else if (lmm_magic != LOV_USER_MAGIC_V1) {
416 "bad userland LOV MAGIC: %#08x != %#08x nor %#08x\n",
417 lmm_magic, LOV_USER_MAGIC_V1, LOV_USER_MAGIC_V3);
421 /* in the rest of the tests, as *lumv1 and lumv3 have the same
422 * fields, we use lumv1 to avoid code duplication */
424 if (lumv1->lmm_pattern == 0) {
425 lumv1->lmm_pattern = lov->desc.ld_pattern ?
426 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
429 if (lumv1->lmm_pattern != LOV_PATTERN_RAID0) {
430 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
435 /* 64kB is the largest common page size we see (ia64), and matches the
437 if (lumv1->lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
438 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
439 lumv1->lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
440 lumv1->lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
443 if ((lumv1->lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
444 (lumv1->lmm_stripe_offset !=
445 (typeof(lumv1->lmm_stripe_offset))(-1))) {
446 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
447 lumv1->lmm_stripe_offset, lov->desc.ld_tgt_count);
450 stripe_count = lov_get_stripecnt(lov, lumv1->lmm_stripe_count);
453 int max_stripes = (max_lmm_size -
454 lov_mds_md_size(0, lmm_magic)) /
455 sizeof(struct lov_ost_data_v1);
456 if (unlikely(max_stripes < stripe_count)) {
457 CDEBUG(D_IOCTL, "stripe count reset from %d to %d\n",
458 stripe_count, max_stripes);
459 stripe_count = max_stripes;
463 if (lmm_magic == LOV_USER_MAGIC_V3) {
464 struct pool_desc *pool;
466 pool = lov_find_pool(lov, lumv3->lmm_pool_name);
468 if (lumv3->lmm_stripe_offset !=
469 (typeof(lumv3->lmm_stripe_offset))(-1)) {
470 rc = lov_check_index_in_pool(
471 lumv3->lmm_stripe_offset, pool);
473 lov_pool_putref(pool);
478 if (stripe_count > pool_tgt_count(pool))
479 stripe_count = pool_tgt_count(pool);
481 lov_pool_putref(pool);
485 rc = lov_alloc_memmd(lsmp, stripe_count, lumv1->lmm_pattern, lmm_magic);
488 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lumv1->lmm_stripe_offset;
489 (*lsmp)->lsm_stripe_size = lumv1->lmm_stripe_size;
490 if (lmm_magic == LOV_USER_MAGIC_V3)
491 strncpy((*lsmp)->lsm_pool_name, lumv3->lmm_pool_name,
499 /* Configure object striping information on a new file.
501 * @lmmu is a pointer to a user struct with one or more of the fields set to
502 * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
503 * lmm_stripe_offset, and lmm_stripe_pattern. lmm_magic must be LOV_MAGIC.
504 * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
506 int lov_setstripe(struct obd_export *exp, int max_lmm_size,
507 struct lov_stripe_md **lsmp, struct lov_user_md *lump)
515 rc = __lov_setstripe(exp, max_lmm_size, lsmp, lump);
520 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
521 struct lov_user_md *lump)
525 struct obd_export *oexp;
526 struct lov_obd *lov = &exp->exp_obd->u.lov;
528 struct lov_user_ost_data_v1 *lmm_objects;
532 if (lump->lmm_magic == LOV_USER_MAGIC_V3)
533 lmm_objects = ((struct lov_user_md_v3 *)lump)->lmm_objects;
535 lmm_objects = lump->lmm_objects;
537 for (i = 0; i < lump->lmm_stripe_count; i++) {
538 __u32 len = sizeof(last_id);
539 oexp = lov->lov_tgts[lmm_objects[i].l_ost_idx]->ltd_exp;
540 rc = obd_get_info(oexp, sizeof(KEY_LAST_ID), KEY_LAST_ID,
541 &len, &last_id, NULL);
544 if (lmm_objects[i].l_object_id > last_id) {
545 CERROR("Setting EA for object > than last id on "
546 "ost idx %d "LPD64" > "LPD64" \n",
547 lmm_objects[i].l_ost_idx,
548 lmm_objects[i].l_object_id, last_id);
553 rc = lov_setstripe(exp, 0, lsmp, lump);
557 for (i = 0; i < lump->lmm_stripe_count; i++) {
558 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
559 lmm_objects[i].l_ost_idx;
560 (*lsmp)->lsm_oinfo[i]->loi_id = lmm_objects[i].l_object_id;
561 (*lsmp)->lsm_oinfo[i]->loi_seq = lmm_objects[i].l_object_seq;
567 /* Retrieve object striping information.
569 * @lump is a pointer to an in-core struct with lmm_ost_count indicating
570 * the maximum number of OST indices which will fit in the user buffer.
571 * lmm_magic must be LOV_USER_MAGIC.
573 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
574 struct lov_user_md *lump)
577 * XXX huge struct allocated on stack.
579 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
580 struct lov_user_md_v3 lum;
581 struct lov_mds_md *lmmk = NULL;
591 * "Switch to kernel segment" to allow copying from kernel space by
592 * copy_{to,from}_user().
597 /* we only need the header part from user space to get lmm_magic and
598 * lmm_stripe_count, (the header part is common to v1 and v3) */
599 lum_size = sizeof(struct lov_user_md_v1);
600 if (cfs_copy_from_user(&lum, lump, lum_size))
601 GOTO(out_set, rc = -EFAULT);
602 else if ((lum.lmm_magic != LOV_USER_MAGIC) &&
603 (lum.lmm_magic != LOV_USER_MAGIC_V3))
604 GOTO(out_set, rc = -EINVAL);
606 if (lum.lmm_stripe_count &&
607 (lum.lmm_stripe_count < lsm->lsm_stripe_count)) {
608 /* Return right size of stripe to user */
609 lum.lmm_stripe_count = lsm->lsm_stripe_count;
610 rc = cfs_copy_to_user(lump, &lum, lum_size);
611 GOTO(out_set, rc = -EOVERFLOW);
613 rc = lov_packmd(exp, &lmmk, lsm);
619 /* FIXME: Bug 1185 - copy fields properly when structs change */
620 /* struct lov_user_md_v3 and struct lov_mds_md_v3 must be the same */
621 CLASSERT(sizeof(lum) == sizeof(struct lov_mds_md_v3));
622 CLASSERT(sizeof lum.lmm_objects[0] == sizeof lmmk->lmm_objects[0]);
624 if ((cpu_to_le32(LOV_MAGIC) != LOV_MAGIC) &&
625 ((lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V1)) ||
626 (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)))) {
627 lustre_swab_lov_mds_md(lmmk);
628 lustre_swab_lov_user_md_objects(
629 (struct lov_user_ost_data*)lmmk->lmm_objects,
630 lmmk->lmm_stripe_count);
632 if (lum.lmm_magic == LOV_USER_MAGIC) {
633 /* User request for v1, we need skip lmm_pool_name */
634 if (lmmk->lmm_magic == LOV_MAGIC_V3) {
635 memmove((char*)(&lmmk->lmm_stripe_count) +
636 sizeof(lmmk->lmm_stripe_count),
637 ((struct lov_mds_md_v3*)lmmk)->lmm_objects,
638 lmmk->lmm_stripe_count *
639 sizeof(struct lov_ost_data_v1));
640 lmm_size -= LOV_MAXPOOLNAME;
643 /* if v3 we just have to update the lum_size */
644 lum_size = sizeof(struct lov_user_md_v3);
647 /* User wasn't expecting this many OST entries */
648 if (lum.lmm_stripe_count == 0)
650 else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count)
651 GOTO(out_set, rc = -EOVERFLOW);
653 * Have a difference between lov_mds_md & lov_user_md.
654 * So we have to re-order the data before copy to user.
656 lum.lmm_stripe_count = lmmk->lmm_stripe_count;
657 ((struct lov_user_md*)lmmk)->lmm_stripe_offset = 0;
658 ((struct lov_user_md*)lmmk)->lmm_stripe_count = lum.lmm_stripe_count;
659 if (cfs_copy_to_user(lump, lmmk, lmm_size))
662 obd_free_diskmd(exp, &lmmk);