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) 2011, 2013, Intel Corporation.
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 void lov_dump_lmm_common(int level, void *lmmp)
59 struct lov_mds_md *lmm = lmmp;
62 lmm_oi_le_to_cpu(&oi, &lmm->lmm_oi);
63 CDEBUG(level, "objid "DOSTID", magic 0x%08x, pattern %#x\n",
64 POSTID(&oi), le32_to_cpu(lmm->lmm_magic),
65 le32_to_cpu(lmm->lmm_pattern));
66 CDEBUG(level, "stripe_size %u, stripe_count %u, layout_gen %u\n",
67 le32_to_cpu(lmm->lmm_stripe_size),
68 le16_to_cpu(lmm->lmm_stripe_count),
69 le16_to_cpu(lmm->lmm_layout_gen));
72 static void lov_dump_lmm_objects(int level, struct lov_ost_data *lod,
77 if (stripe_count > LOV_V1_INSANE_STRIPE_COUNT) {
78 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
79 stripe_count, LOV_V1_INSANE_STRIPE_COUNT);
83 for (i = 0; i < stripe_count; ++i, ++lod) {
86 ostid_le_to_cpu(&lod->l_ost_oi, &oi);
87 CDEBUG(level, "stripe %u idx %u subobj "DOSTID"\n", i,
88 le32_to_cpu(lod->l_ost_idx), POSTID(&oi));
92 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
94 lov_dump_lmm_common(level, lmm);
95 lov_dump_lmm_objects(level, lmm->lmm_objects,
96 le16_to_cpu(lmm->lmm_stripe_count));
99 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
101 lov_dump_lmm_common(level, lmm);
102 CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
103 lov_dump_lmm_objects(level, lmm->lmm_objects,
104 le16_to_cpu(lmm->lmm_stripe_count));
107 void lov_dump_lmm(int level, void *lmm)
111 magic = ((struct lov_mds_md_v1 *)(lmm))->lmm_magic;
114 return lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)(lmm));
116 return lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)(lmm));
118 CERROR("Cannot recognize lmm_magic %x", magic);
123 #define LMM_ASSERT(test) \
125 if (!(test)) lov_dump_lmm(D_ERROR, lmm); \
126 LASSERT(test); /* so we know what assertion failed */ \
129 /* Pack LOV object metadata for disk storage. It is packed in LE byte
130 * order and is opaque to the networking layer.
132 * XXX In the future, this will be enhanced to get the EA size from the
133 * underlying OSC device(s) to get their EA sizes so we can stack
134 * LOVs properly. For now lov_mds_md_size() just assumes one obd_id
137 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
138 struct lov_stripe_md *lsm)
140 struct obd_device *obd = class_exp2obd(exp);
141 struct lov_obd *lov = &obd->u.lov;
142 struct lov_mds_md_v1 *lmmv1;
143 struct lov_mds_md_v3 *lmmv3;
145 struct lov_ost_data_v1 *lmm_objects;
146 int lmm_size, lmm_magic;
152 lmm_magic = lsm->lsm_magic;
155 lmm_magic = le32_to_cpu((*lmmp)->lmm_magic);
157 /* lsm == NULL and lmmp == NULL */
158 lmm_magic = LOV_MAGIC;
161 if ((lmm_magic != LOV_MAGIC_V1) &&
162 (lmm_magic != LOV_MAGIC_V3)) {
163 CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
164 lmm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
170 /* If we are just sizing the EA, limit the stripe count
171 * to the actual number of OSTs in this filesystem. */
173 stripe_count = lov_get_stripecnt(lov, lmm_magic,
174 lsm->lsm_stripe_count);
175 lsm->lsm_stripe_count = stripe_count;
176 } else if (!lsm_is_released(lsm)) {
177 stripe_count = lsm->lsm_stripe_count;
182 /* No need to allocate more than maximum supported stripes.
183 * Anyway, this is pretty inaccurate since ld_tgt_count now
184 * represents max index and we should rely on the actual number
186 stripe_count = lov_mds_md_stripecnt(lov->lov_ocd.ocd_max_easize,
188 if (stripe_count > lov->desc.ld_tgt_count)
189 stripe_count = lov->desc.ld_tgt_count;
192 /* XXX LOV STACKING call into osc for sizes */
193 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
199 stripe_count = le16_to_cpu((*lmmp)->lmm_stripe_count);
200 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
201 OBD_FREE_LARGE(*lmmp, lmm_size);
207 OBD_ALLOC_LARGE(*lmmp, lmm_size);
212 CDEBUG(D_INFO, "lov_packmd: LOV_MAGIC 0x%08X, lmm_size = %d \n",
213 lmm_magic, lmm_size);
216 lmmv3 = (struct lov_mds_md_v3 *)*lmmp;
217 if (lmm_magic == LOV_MAGIC_V3)
218 lmmv3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3);
220 lmmv1->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
225 /* lmmv1 and lmmv3 point to the same struct and have the
228 lmm_oi_cpu_to_le(&lmmv1->lmm_oi, &lsm->lsm_oi);
229 lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
230 lmmv1->lmm_stripe_count = cpu_to_le16(stripe_count);
231 lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
232 lmmv1->lmm_layout_gen = cpu_to_le16(lsm->lsm_layout_gen);
233 if (lsm->lsm_magic == LOV_MAGIC_V3) {
234 cplen = strlcpy(lmmv3->lmm_pool_name, lsm->lsm_pool_name,
235 sizeof(lmmv3->lmm_pool_name));
236 if (cplen >= sizeof(lmmv3->lmm_pool_name))
238 lmm_objects = lmmv3->lmm_objects;
240 lmm_objects = lmmv1->lmm_objects;
243 for (i = 0; i < stripe_count; i++) {
244 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
245 /* XXX LOV STACKING call down to osc_packmd() to do packing */
246 LASSERTF(ostid_id(&loi->loi_oi) != 0, "lmm_oi "DOSTID
247 " stripe %u/%u idx %u\n", POSTID(&lmmv1->lmm_oi),
248 i, stripe_count, loi->loi_ost_idx);
249 ostid_cpu_to_le(&loi->loi_oi, &lmm_objects[i].l_ost_oi);
250 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
251 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
257 /* Find the max stripecount we should use */
258 __u16 lov_get_stripecnt(struct lov_obd *lov, __u32 magic, __u16 stripe_count)
260 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
263 stripe_count = lov->desc.ld_default_stripe_count;
264 if (stripe_count > lov->desc.ld_active_tgt_count)
265 stripe_count = lov->desc.ld_active_tgt_count;
269 /* stripe count is based on whether ldiskfs can handle
271 if (lov->lov_ocd.ocd_connect_flags & OBD_CONNECT_MAX_EASIZE &&
272 lov->lov_ocd.ocd_max_easize)
273 max_stripes = lov_mds_md_stripecnt(lov->lov_ocd.ocd_max_easize,
276 if (stripe_count > max_stripes)
277 stripe_count = max_stripes;
283 static int lov_verify_lmm(void *lmm, int lmm_bytes, __u16 *stripe_count)
287 if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
291 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
292 le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
293 sz = lmm_bytes * 2 + 1;
294 OBD_ALLOC_LARGE(buffer, sz);
295 if (buffer != NULL) {
298 for (i = 0; i < lmm_bytes; i++)
299 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
300 buffer[sz - 1] = '\0';
301 CERROR("%s\n", buffer);
302 OBD_FREE_LARGE(buffer, sz);
306 rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
307 lmm_bytes, stripe_count);
311 int lov_alloc_memmd(struct lov_stripe_md **lsmp, __u16 stripe_count,
312 int pattern, int magic)
317 CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
319 *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
321 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
325 cfs_atomic_set(&(*lsmp)->lsm_refc, 1);
326 spin_lock_init(&(*lsmp)->lsm_lock);
327 (*lsmp)->lsm_magic = magic;
328 (*lsmp)->lsm_stripe_count = stripe_count;
329 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
330 (*lsmp)->lsm_pattern = pattern;
331 (*lsmp)->lsm_pool_name[0] = '\0';
332 (*lsmp)->lsm_layout_gen = 0;
333 if (stripe_count > 0)
334 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
336 for (i = 0; i < stripe_count; i++)
337 loi_init((*lsmp)->lsm_oinfo[i]);
342 int lov_free_memmd(struct lov_stripe_md **lsmp)
344 struct lov_stripe_md *lsm = *lsmp;
348 LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 0);
349 if ((refc = cfs_atomic_dec_return(&lsm->lsm_refc)) == 0) {
350 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
351 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
357 /* Unpack LOV object metadata from disk storage. It is packed in LE byte
358 * order and is opaque to the networking layer.
360 int lov_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
361 struct lov_mds_md *lmm, int lmm_bytes)
363 struct obd_device *obd = class_exp2obd(exp);
364 struct lov_obd *lov = &obd->u.lov;
365 int rc = 0, lsm_size;
371 /* If passed an MDS struct use values from there, otherwise defaults */
373 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
376 magic = le32_to_cpu(lmm->lmm_magic);
379 stripe_count = lov_get_stripecnt(lov, magic, 0);
382 /* If we aren't passed an lsmp struct, we just want the size */
384 /* XXX LOV STACKING call into osc for sizes */
386 RETURN(lov_stripe_md_size(stripe_count));
388 /* If we are passed an allocated struct but nothing to unpack, free */
390 lov_free_memmd(lsmp);
394 pattern = le32_to_cpu(lmm->lmm_pattern);
395 lsm_size = lov_alloc_memmd(lsmp, stripe_count, pattern, magic);
399 /* If we are passed a pointer but nothing to unpack, we only alloc */
403 LASSERT(lsm_op_find(magic) != NULL);
404 rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
406 lov_free_memmd(lsmp);
413 static int __lov_setstripe(struct obd_export *exp, int max_lmm_size,
414 struct lov_stripe_md **lsmp,
415 struct lov_user_md *lump)
417 struct obd_device *obd = class_exp2obd(exp);
418 struct lov_obd *lov = &obd->u.lov;
419 char buffer[sizeof(struct lov_user_md_v3)];
420 struct lov_user_md_v3 *lumv3 = (struct lov_user_md_v3 *)&buffer[0];
421 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&buffer[0];
428 rc = lov_lum_swab_if_needed(lumv3, &lmm_magic, lump);
432 /* in the rest of the tests, as *lumv1 and lumv3 have the same
433 * fields, we use lumv1 to avoid code duplication */
435 if (lumv1->lmm_pattern == 0) {
436 lumv1->lmm_pattern = lov->desc.ld_pattern ?
437 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
440 if (lov_pattern(lumv1->lmm_pattern) != LOV_PATTERN_RAID0) {
441 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
446 /* 64kB is the largest common page size we see (ia64), and matches the
448 if (lumv1->lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
449 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
450 lumv1->lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
451 lumv1->lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
454 if ((lumv1->lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
455 (lumv1->lmm_stripe_offset !=
456 (typeof(lumv1->lmm_stripe_offset))(-1))) {
457 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
458 lumv1->lmm_stripe_offset, lov->desc.ld_tgt_count);
461 stripe_count = lov_get_stripecnt(lov, lmm_magic,
462 lumv1->lmm_stripe_count);
465 int max_stripes = (max_lmm_size -
466 lov_mds_md_size(0, lmm_magic)) /
467 sizeof(struct lov_ost_data_v1);
468 if (unlikely(max_stripes < stripe_count)) {
469 CDEBUG(D_IOCTL, "stripe count reset from %d to %d\n",
470 stripe_count, max_stripes);
471 stripe_count = max_stripes;
475 if (lmm_magic == LOV_USER_MAGIC_V3) {
476 struct pool_desc *pool;
478 /* In the function below, .hs_keycmp resolves to
479 * pool_hashkey_keycmp() */
480 /* coverity[overrun-buffer-val] */
481 pool = lov_find_pool(lov, lumv3->lmm_pool_name);
483 if (lumv3->lmm_stripe_offset !=
484 (typeof(lumv3->lmm_stripe_offset))(-1)) {
485 rc = lov_check_index_in_pool(
486 lumv3->lmm_stripe_offset, pool);
488 lov_pool_putref(pool);
493 if (stripe_count > pool_tgt_count(pool))
494 stripe_count = pool_tgt_count(pool);
496 lov_pool_putref(pool);
500 if (lumv1->lmm_pattern & LOV_PATTERN_F_RELEASED)
503 rc = lov_alloc_memmd(lsmp, stripe_count, lumv1->lmm_pattern, lmm_magic);
506 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lumv1->lmm_stripe_offset;
507 (*lsmp)->lsm_stripe_size = lumv1->lmm_stripe_size;
508 if (lmm_magic == LOV_USER_MAGIC_V3) {
509 cplen = strlcpy((*lsmp)->lsm_pool_name,
510 lumv3->lmm_pool_name,
511 sizeof((*lsmp)->lsm_pool_name));
512 if (cplen >= sizeof((*lsmp)->lsm_pool_name))
521 /* Configure object striping information on a new file.
523 * @lmmu is a pointer to a user struct with one or more of the fields set to
524 * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
525 * lmm_stripe_offset, and lmm_stripe_pattern. lmm_magic must be LOV_MAGIC.
526 * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
528 int lov_setstripe(struct obd_export *exp, int max_lmm_size,
529 struct lov_stripe_md **lsmp, struct lov_user_md *lump)
537 rc = __lov_setstripe(exp, max_lmm_size, lsmp, lump);
542 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
543 struct lov_user_md *lump)
547 struct obd_export *oexp;
548 struct lov_obd *lov = &exp->exp_obd->u.lov;
550 struct lov_user_ost_data_v1 *lmm_objects;
554 if (lump->lmm_magic == LOV_USER_MAGIC_V3)
555 lmm_objects = ((struct lov_user_md_v3 *)lump)->lmm_objects;
557 lmm_objects = lump->lmm_objects;
559 for (i = 0; i < lump->lmm_stripe_count; i++) {
560 __u32 len = sizeof(last_id);
561 oexp = lov->lov_tgts[lmm_objects[i].l_ost_idx]->ltd_exp;
562 rc = obd_get_info(NULL, oexp, sizeof(KEY_LAST_ID), KEY_LAST_ID,
563 &len, &last_id, NULL);
566 if (ostid_id(&lmm_objects[i].l_ost_oi) > last_id) {
567 CERROR("Setting EA for object > than last id on"
568 " ost idx %d "DOSTID" > "LPD64" \n",
569 lmm_objects[i].l_ost_idx,
570 POSTID(&lmm_objects[i].l_ost_oi), last_id);
575 rc = lov_setstripe(exp, 0, lsmp, lump);
579 for (i = 0; i < lump->lmm_stripe_count; i++) {
580 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
581 lmm_objects[i].l_ost_idx;
582 (*lsmp)->lsm_oinfo[i]->loi_oi = lmm_objects[i].l_ost_oi;
588 /* Retrieve object striping information.
590 * @lump is a pointer to an in-core struct with lmm_ost_count indicating
591 * the maximum number of OST indices which will fit in the user buffer.
592 * lmm_magic must be LOV_USER_MAGIC.
594 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
595 struct lov_user_md *lump)
598 * XXX huge struct allocated on stack.
600 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
601 struct lov_user_md_v3 lum;
602 struct lov_mds_md *lmmk = NULL;
610 /* we only need the header part from user space to get lmm_magic and
611 * lmm_stripe_count, (the header part is common to v1 and v3) */
612 lum_size = sizeof(struct lov_user_md_v1);
613 if (copy_from_user(&lum, lump, lum_size))
614 GOTO(out_set, rc = -EFAULT);
615 else if ((lum.lmm_magic != LOV_USER_MAGIC) &&
616 (lum.lmm_magic != LOV_USER_MAGIC_V3))
617 GOTO(out_set, rc = -EINVAL);
619 if (lum.lmm_stripe_count &&
620 (lum.lmm_stripe_count < lsm->lsm_stripe_count)) {
621 /* Return right size of stripe to user */
622 lum.lmm_stripe_count = lsm->lsm_stripe_count;
623 rc = copy_to_user(lump, &lum, lum_size);
624 GOTO(out_set, rc = -EOVERFLOW);
626 rc = lov_packmd(exp, &lmmk, lsm);
632 /* FIXME: Bug 1185 - copy fields properly when structs change */
633 /* struct lov_user_md_v3 and struct lov_mds_md_v3 must be the same */
634 CLASSERT(sizeof(lum) == sizeof(struct lov_mds_md_v3));
635 CLASSERT(sizeof lum.lmm_objects[0] == sizeof lmmk->lmm_objects[0]);
637 if ((cpu_to_le32(LOV_MAGIC) != LOV_MAGIC) &&
638 ((lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V1)) ||
639 (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)))) {
640 lustre_swab_lov_mds_md(lmmk);
641 lustre_swab_lov_user_md_objects(
642 (struct lov_user_ost_data*)lmmk->lmm_objects,
643 lmmk->lmm_stripe_count);
645 if (lum.lmm_magic == LOV_USER_MAGIC) {
646 /* User request for v1, we need skip lmm_pool_name */
647 if (lmmk->lmm_magic == LOV_MAGIC_V3) {
648 memmove((char*)(&lmmk->lmm_stripe_count) +
649 sizeof(lmmk->lmm_stripe_count),
650 ((struct lov_mds_md_v3*)lmmk)->lmm_objects,
651 lmmk->lmm_stripe_count *
652 sizeof(struct lov_ost_data_v1));
653 lmm_size -= LOV_MAXPOOLNAME;
656 /* if v3 we just have to update the lum_size */
657 lum_size = sizeof(struct lov_user_md_v3);
660 /* User wasn't expecting this many OST entries */
661 if (lum.lmm_stripe_count == 0)
663 else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count)
664 GOTO(out_set, rc = -EOVERFLOW);
666 * Have a difference between lov_mds_md & lov_user_md.
667 * So we have to re-order the data before copy to user.
669 lum.lmm_stripe_count = lmmk->lmm_stripe_count;
670 lum.lmm_layout_gen = lmmk->lmm_layout_gen;
671 ((struct lov_user_md *)lmmk)->lmm_layout_gen = lum.lmm_layout_gen;
672 ((struct lov_user_md *)lmmk)->lmm_stripe_count = lum.lmm_stripe_count;
673 if (copy_to_user(lump, lmmk, lmm_size))
676 obd_free_diskmd(exp, &lmmk);