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, 2012, 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 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, layout_gen %u\n",
66 le32_to_cpu(lmm->lmm_stripe_size),
67 le16_to_cpu(lmm->lmm_stripe_count),
68 le16_to_cpu(lmm->lmm_layout_gen));
71 static void lov_dump_lmm_objects(int level, struct lov_ost_data *lod,
76 if (stripe_count > LOV_V1_INSANE_STRIPE_COUNT) {
77 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
78 stripe_count, LOV_V1_INSANE_STRIPE_COUNT);
81 for (i = 0; i < stripe_count; ++i, ++lod) {
82 CDEBUG(level, "stripe %u idx %u subobj "LPX64"/"LPX64"\n", i,
83 le32_to_cpu(lod->l_ost_idx),
84 (__u64)le64_to_cpu(lod->l_object_seq),
85 (__u64)le64_to_cpu(lod->l_object_id));
89 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
91 lov_dump_lmm_common(level, lmm);
92 lov_dump_lmm_objects(level, lmm->lmm_objects,
93 le16_to_cpu(lmm->lmm_stripe_count));
96 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
98 lov_dump_lmm_common(level, lmm);
99 CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
100 lov_dump_lmm_objects(level, lmm->lmm_objects,
101 le16_to_cpu(lmm->lmm_stripe_count));
104 void lov_dump_lmm(int level, void *lmm)
108 magic = ((struct lov_mds_md_v1 *)(lmm))->lmm_magic;
111 return lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)(lmm));
113 return lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)(lmm));
115 CERROR("Cannot recognize lmm_magic %x", magic);
120 #define LMM_ASSERT(test) \
122 if (!(test)) lov_dump_lmm(D_ERROR, lmm); \
123 LASSERT(test); /* so we know what assertion failed */ \
126 /* Pack LOV object metadata for disk storage. It is packed in LE byte
127 * order and is opaque to the networking layer.
129 * XXX In the future, this will be enhanced to get the EA size from the
130 * underlying OSC device(s) to get their EA sizes so we can stack
131 * LOVs properly. For now lov_mds_md_size() just assumes one obd_id
134 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
135 struct lov_stripe_md *lsm)
137 struct obd_device *obd = class_exp2obd(exp);
138 struct lov_obd *lov = &obd->u.lov;
139 struct lov_mds_md_v1 *lmmv1;
140 struct lov_mds_md_v3 *lmmv3;
142 struct lov_ost_data_v1 *lmm_objects;
143 int lmm_size, lmm_magic;
149 lmm_magic = lsm->lsm_magic;
152 lmm_magic = le32_to_cpu((*lmmp)->lmm_magic);
154 /* lsm == NULL and lmmp == NULL */
155 lmm_magic = LOV_MAGIC;
158 if ((lmm_magic != LOV_MAGIC_V1) &&
159 (lmm_magic != LOV_MAGIC_V3)) {
160 CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
161 lmm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
167 /* If we are just sizing the EA, limit the stripe count
168 * to the actual number of OSTs in this filesystem. */
170 stripe_count = lov_get_stripecnt(lov, lmm_magic,
171 lsm->lsm_stripe_count);
172 lsm->lsm_stripe_count = stripe_count;
174 stripe_count = lsm->lsm_stripe_count;
177 /* No need to allocate more than maximum supported stripes.
178 * Anyway, this is pretty inaccurate since ld_tgt_count now
179 * represents max index and we should rely on the actual number
181 stripe_count = lov_mds_md_stripecnt(lov->lov_ocd.ocd_max_easize,
183 if (stripe_count > lov->desc.ld_tgt_count)
184 stripe_count = lov->desc.ld_tgt_count;
187 /* XXX LOV STACKING call into osc for sizes */
188 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
194 stripe_count = le16_to_cpu((*lmmp)->lmm_stripe_count);
195 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
196 OBD_FREE_LARGE(*lmmp, lmm_size);
202 OBD_ALLOC_LARGE(*lmmp, lmm_size);
207 CDEBUG(D_INFO, "lov_packmd: LOV_MAGIC 0x%08X, lmm_size = %d \n",
208 lmm_magic, lmm_size);
211 lmmv3 = (struct lov_mds_md_v3 *)*lmmp;
212 if (lmm_magic == LOV_MAGIC_V3)
213 lmmv3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3);
215 lmmv1->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
220 /* lmmv1 and lmmv3 point to the same struct and have the
223 lmmv1->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
224 lmmv1->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
225 lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
226 lmmv1->lmm_stripe_count = cpu_to_le16(stripe_count);
227 lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
228 lmmv1->lmm_layout_gen = cpu_to_le16(lsm->lsm_layout_gen);
229 if (lsm->lsm_magic == LOV_MAGIC_V3) {
230 cplen = strlcpy(lmmv3->lmm_pool_name, lsm->lsm_pool_name,
231 sizeof(lmmv3->lmm_pool_name));
232 if (cplen >= sizeof(lmmv3->lmm_pool_name))
234 lmm_objects = lmmv3->lmm_objects;
236 lmm_objects = lmmv1->lmm_objects;
239 for (i = 0; i < stripe_count; i++) {
240 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
241 /* XXX LOV STACKING call down to osc_packmd() to do packing */
242 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
243 lmmv1->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
244 lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
245 lmm_objects[i].l_object_seq = cpu_to_le64(loi->loi_seq);
246 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
247 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
253 /* Find the max stripecount we should use */
254 __u16 lov_get_stripecnt(struct lov_obd *lov, __u32 magic, __u16 stripe_count)
256 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
259 stripe_count = lov->desc.ld_default_stripe_count;
260 if (stripe_count > lov->desc.ld_active_tgt_count)
261 stripe_count = lov->desc.ld_active_tgt_count;
265 /* stripe count is based on whether ldiskfs can handle
267 if (lov->lov_ocd.ocd_connect_flags & OBD_CONNECT_MAX_EASIZE &&
268 lov->lov_ocd.ocd_max_easize)
269 max_stripes = lov_mds_md_stripecnt(lov->lov_ocd.ocd_max_easize,
272 if (stripe_count > max_stripes)
273 stripe_count = max_stripes;
279 static int lov_verify_lmm(void *lmm, int lmm_bytes, __u16 *stripe_count)
283 if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
287 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
288 le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
289 sz = lmm_bytes * 2 + 1;
290 OBD_ALLOC_LARGE(buffer, sz);
291 if (buffer != NULL) {
294 for (i = 0; i < lmm_bytes; i++)
295 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
296 buffer[sz - 1] = '\0';
297 CERROR("%s\n", buffer);
298 OBD_FREE_LARGE(buffer, sz);
302 rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
303 lmm_bytes, stripe_count);
307 int lov_alloc_memmd(struct lov_stripe_md **lsmp, __u16 stripe_count,
308 int pattern, int magic)
313 CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
315 *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
317 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
321 cfs_atomic_set(&(*lsmp)->lsm_refc, 1);
322 spin_lock_init(&(*lsmp)->lsm_lock);
323 (*lsmp)->lsm_magic = magic;
324 (*lsmp)->lsm_stripe_count = stripe_count;
325 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
326 (*lsmp)->lsm_pattern = pattern;
327 (*lsmp)->lsm_pool_name[0] = '\0';
328 (*lsmp)->lsm_layout_gen = 0;
329 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
331 for (i = 0; i < stripe_count; i++)
332 loi_init((*lsmp)->lsm_oinfo[i]);
337 int lov_free_memmd(struct lov_stripe_md **lsmp)
339 struct lov_stripe_md *lsm = *lsmp;
343 LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 0);
344 if ((refc = cfs_atomic_dec_return(&lsm->lsm_refc)) == 0) {
345 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
346 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
352 /* Unpack LOV object metadata from disk storage. It is packed in LE byte
353 * order and is opaque to the networking layer.
355 int lov_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
356 struct lov_mds_md *lmm, int lmm_bytes)
358 struct obd_device *obd = class_exp2obd(exp);
359 struct lov_obd *lov = &obd->u.lov;
360 int rc = 0, lsm_size;
365 /* If passed an MDS struct use values from there, otherwise defaults */
367 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
370 magic = le32_to_cpu(lmm->lmm_magic);
373 stripe_count = lov_get_stripecnt(lov, magic, 0);
376 /* If we aren't passed an lsmp struct, we just want the size */
378 /* XXX LOV STACKING call into osc for sizes */
380 RETURN(lov_stripe_md_size(stripe_count));
382 /* If we are passed an allocated struct but nothing to unpack, free */
384 lov_free_memmd(lsmp);
388 lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
393 /* If we are passed a pointer but nothing to unpack, we only alloc */
397 LASSERT(lsm_op_find(magic) != NULL);
398 rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
400 lov_free_memmd(lsmp);
407 static int __lov_setstripe(struct obd_export *exp, int max_lmm_size,
408 struct lov_stripe_md **lsmp,
409 struct lov_user_md *lump)
411 struct obd_device *obd = class_exp2obd(exp);
412 struct lov_obd *lov = &obd->u.lov;
413 char buffer[sizeof(struct lov_user_md_v3)];
414 struct lov_user_md_v3 *lumv3 = (struct lov_user_md_v3 *)&buffer[0];
415 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&buffer[0];
422 rc = lov_lum_swab_if_needed(lumv3, &lmm_magic, lump);
426 /* in the rest of the tests, as *lumv1 and lumv3 have the same
427 * fields, we use lumv1 to avoid code duplication */
429 if (lumv1->lmm_pattern == 0) {
430 lumv1->lmm_pattern = lov->desc.ld_pattern ?
431 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
434 if (lumv1->lmm_pattern != LOV_PATTERN_RAID0) {
435 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
440 /* 64kB is the largest common page size we see (ia64), and matches the
442 if (lumv1->lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
443 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
444 lumv1->lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
445 lumv1->lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
448 if ((lumv1->lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
449 (lumv1->lmm_stripe_offset !=
450 (typeof(lumv1->lmm_stripe_offset))(-1))) {
451 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
452 lumv1->lmm_stripe_offset, lov->desc.ld_tgt_count);
455 stripe_count = lov_get_stripecnt(lov, lmm_magic,
456 lumv1->lmm_stripe_count);
459 int max_stripes = (max_lmm_size -
460 lov_mds_md_size(0, lmm_magic)) /
461 sizeof(struct lov_ost_data_v1);
462 if (unlikely(max_stripes < stripe_count)) {
463 CDEBUG(D_IOCTL, "stripe count reset from %d to %d\n",
464 stripe_count, max_stripes);
465 stripe_count = max_stripes;
469 if (lmm_magic == LOV_USER_MAGIC_V3) {
470 struct pool_desc *pool;
472 /* In the function below, .hs_keycmp resolves to
473 * pool_hashkey_keycmp() */
474 /* coverity[overrun-buffer-val] */
475 pool = lov_find_pool(lov, lumv3->lmm_pool_name);
477 if (lumv3->lmm_stripe_offset !=
478 (typeof(lumv3->lmm_stripe_offset))(-1)) {
479 rc = lov_check_index_in_pool(
480 lumv3->lmm_stripe_offset, pool);
482 lov_pool_putref(pool);
487 if (stripe_count > pool_tgt_count(pool))
488 stripe_count = pool_tgt_count(pool);
490 lov_pool_putref(pool);
494 rc = lov_alloc_memmd(lsmp, stripe_count, lumv1->lmm_pattern, lmm_magic);
497 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lumv1->lmm_stripe_offset;
498 (*lsmp)->lsm_stripe_size = lumv1->lmm_stripe_size;
499 if (lmm_magic == LOV_USER_MAGIC_V3) {
500 cplen = strlcpy((*lsmp)->lsm_pool_name,
501 lumv3->lmm_pool_name,
502 sizeof((*lsmp)->lsm_pool_name));
503 if (cplen >= sizeof((*lsmp)->lsm_pool_name))
512 /* Configure object striping information on a new file.
514 * @lmmu is a pointer to a user struct with one or more of the fields set to
515 * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
516 * lmm_stripe_offset, and lmm_stripe_pattern. lmm_magic must be LOV_MAGIC.
517 * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
519 int lov_setstripe(struct obd_export *exp, int max_lmm_size,
520 struct lov_stripe_md **lsmp, struct lov_user_md *lump)
528 rc = __lov_setstripe(exp, max_lmm_size, lsmp, lump);
533 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
534 struct lov_user_md *lump)
538 struct obd_export *oexp;
539 struct lov_obd *lov = &exp->exp_obd->u.lov;
541 struct lov_user_ost_data_v1 *lmm_objects;
545 if (lump->lmm_magic == LOV_USER_MAGIC_V3)
546 lmm_objects = ((struct lov_user_md_v3 *)lump)->lmm_objects;
548 lmm_objects = lump->lmm_objects;
550 for (i = 0; i < lump->lmm_stripe_count; i++) {
551 __u32 len = sizeof(last_id);
552 oexp = lov->lov_tgts[lmm_objects[i].l_ost_idx]->ltd_exp;
553 rc = obd_get_info(NULL, oexp, sizeof(KEY_LAST_ID), KEY_LAST_ID,
554 &len, &last_id, NULL);
557 if (lmm_objects[i].l_object_id > last_id) {
558 CERROR("Setting EA for object > than last id on "
559 "ost idx %d "LPD64" > "LPD64" \n",
560 lmm_objects[i].l_ost_idx,
561 lmm_objects[i].l_object_id, last_id);
566 rc = lov_setstripe(exp, 0, lsmp, lump);
570 for (i = 0; i < lump->lmm_stripe_count; i++) {
571 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
572 lmm_objects[i].l_ost_idx;
573 (*lsmp)->lsm_oinfo[i]->loi_id = lmm_objects[i].l_object_id;
574 (*lsmp)->lsm_oinfo[i]->loi_seq = lmm_objects[i].l_object_seq;
580 /* Retrieve object striping information.
582 * @lump is a pointer to an in-core struct with lmm_ost_count indicating
583 * the maximum number of OST indices which will fit in the user buffer.
584 * lmm_magic must be LOV_USER_MAGIC.
586 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
587 struct lov_user_md *lump)
590 * XXX huge struct allocated on stack.
592 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
593 struct lov_user_md_v3 lum;
594 struct lov_mds_md *lmmk = NULL;
604 * "Switch to kernel segment" to allow copying from kernel space by
605 * copy_{to,from}_user().
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 (cfs_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 = cfs_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.u.lum_layout_gen = lmmk->lmm_layout_gen;
671 ((struct lov_user_md*)lmmk)->u.lum_layout_gen = lum.u.lum_layout_gen;
672 ((struct lov_user_md*)lmmk)->lmm_stripe_count = lum.lmm_stripe_count;
673 if (cfs_copy_to_user(lump, lmmk, lmm_size))
676 obd_free_diskmd(exp, &lmmk);