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;
62 ostid_le_to_cpu(&lmm->lmm_oi, &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);
82 for (i = 0; i < stripe_count; ++i, ++lod) {
85 ostid_le_to_cpu(&lod->l_ost_oi, &oi);
86 CDEBUG(level, "stripe %u idx %u subobj "DOSTID"\n", i,
87 le32_to_cpu(lod->l_ost_idx), POSTID(&oi));
91 void lov_dump_lmm_v1(int level, struct lov_mds_md_v1 *lmm)
93 lov_dump_lmm_common(level, lmm);
94 lov_dump_lmm_objects(level, lmm->lmm_objects,
95 le16_to_cpu(lmm->lmm_stripe_count));
98 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
100 lov_dump_lmm_common(level, lmm);
101 CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
102 lov_dump_lmm_objects(level, lmm->lmm_objects,
103 le16_to_cpu(lmm->lmm_stripe_count));
106 void lov_dump_lmm(int level, void *lmm)
110 magic = ((struct lov_mds_md_v1 *)(lmm))->lmm_magic;
113 return lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)(lmm));
115 return lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)(lmm));
117 CERROR("Cannot recognize lmm_magic %x", magic);
122 #define LMM_ASSERT(test) \
124 if (!(test)) lov_dump_lmm(D_ERROR, lmm); \
125 LASSERT(test); /* so we know what assertion failed */ \
128 /* Pack LOV object metadata for disk storage. It is packed in LE byte
129 * order and is opaque to the networking layer.
131 * XXX In the future, this will be enhanced to get the EA size from the
132 * underlying OSC device(s) to get their EA sizes so we can stack
133 * LOVs properly. For now lov_mds_md_size() just assumes one obd_id
136 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
137 struct lov_stripe_md *lsm)
139 struct obd_device *obd = class_exp2obd(exp);
140 struct lov_obd *lov = &obd->u.lov;
141 struct lov_mds_md_v1 *lmmv1;
142 struct lov_mds_md_v3 *lmmv3;
144 struct lov_ost_data_v1 *lmm_objects;
145 int lmm_size, lmm_magic;
151 lmm_magic = lsm->lsm_magic;
154 lmm_magic = le32_to_cpu((*lmmp)->lmm_magic);
156 /* lsm == NULL and lmmp == NULL */
157 lmm_magic = LOV_MAGIC;
160 if ((lmm_magic != LOV_MAGIC_V1) &&
161 (lmm_magic != LOV_MAGIC_V3)) {
162 CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
163 lmm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
169 /* If we are just sizing the EA, limit the stripe count
170 * to the actual number of OSTs in this filesystem. */
172 stripe_count = lov_get_stripecnt(lov, lmm_magic,
173 lsm->lsm_stripe_count);
174 lsm->lsm_stripe_count = stripe_count;
176 stripe_count = lsm->lsm_stripe_count;
179 /* No need to allocate more than maximum supported stripes.
180 * Anyway, this is pretty inaccurate since ld_tgt_count now
181 * represents max index and we should rely on the actual number
183 stripe_count = lov_mds_md_stripecnt(lov->lov_ocd.ocd_max_easize,
185 if (stripe_count > lov->desc.ld_tgt_count)
186 stripe_count = lov->desc.ld_tgt_count;
189 /* XXX LOV STACKING call into osc for sizes */
190 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
196 stripe_count = le16_to_cpu((*lmmp)->lmm_stripe_count);
197 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
198 OBD_FREE_LARGE(*lmmp, lmm_size);
204 OBD_ALLOC_LARGE(*lmmp, lmm_size);
209 CDEBUG(D_INFO, "lov_packmd: LOV_MAGIC 0x%08X, lmm_size = %d \n",
210 lmm_magic, lmm_size);
213 lmmv3 = (struct lov_mds_md_v3 *)*lmmp;
214 if (lmm_magic == LOV_MAGIC_V3)
215 lmmv3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3);
217 lmmv1->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
222 /* lmmv1 and lmmv3 point to the same struct and have the
225 ostid_cpu_to_le(&lsm->lsm_oi, &lmmv1->lmm_oi);
226 lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
227 lmmv1->lmm_stripe_count = cpu_to_le16(stripe_count);
228 lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
229 lmmv1->lmm_layout_gen = cpu_to_le16(lsm->lsm_layout_gen);
230 if (lsm->lsm_magic == LOV_MAGIC_V3) {
231 cplen = strlcpy(lmmv3->lmm_pool_name, lsm->lsm_pool_name,
232 sizeof(lmmv3->lmm_pool_name));
233 if (cplen >= sizeof(lmmv3->lmm_pool_name))
235 lmm_objects = lmmv3->lmm_objects;
237 lmm_objects = lmmv1->lmm_objects;
240 for (i = 0; i < stripe_count; i++) {
241 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
242 /* XXX LOV STACKING call down to osc_packmd() to do packing */
243 LASSERTF(ostid_id(&loi->loi_oi) != 0, "lmm_oi "DOSTID
244 " stripe %u/%u idx %u\n", POSTID(&lmmv1->lmm_oi),
245 i, stripe_count, loi->loi_ost_idx);
246 ostid_cpu_to_le(&loi->loi_oi, &lmm_objects[i].l_ost_oi);
247 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
248 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
254 /* Find the max stripecount we should use */
255 __u16 lov_get_stripecnt(struct lov_obd *lov, __u32 magic, __u16 stripe_count)
257 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
260 stripe_count = lov->desc.ld_default_stripe_count;
261 if (stripe_count > lov->desc.ld_active_tgt_count)
262 stripe_count = lov->desc.ld_active_tgt_count;
266 /* stripe count is based on whether ldiskfs can handle
268 if (lov->lov_ocd.ocd_connect_flags & OBD_CONNECT_MAX_EASIZE &&
269 lov->lov_ocd.ocd_max_easize)
270 max_stripes = lov_mds_md_stripecnt(lov->lov_ocd.ocd_max_easize,
273 if (stripe_count > max_stripes)
274 stripe_count = max_stripes;
280 static int lov_verify_lmm(void *lmm, int lmm_bytes, __u16 *stripe_count)
284 if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
288 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
289 le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
290 sz = lmm_bytes * 2 + 1;
291 OBD_ALLOC_LARGE(buffer, sz);
292 if (buffer != NULL) {
295 for (i = 0; i < lmm_bytes; i++)
296 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
297 buffer[sz - 1] = '\0';
298 CERROR("%s\n", buffer);
299 OBD_FREE_LARGE(buffer, sz);
303 rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
304 lmm_bytes, stripe_count);
308 int lov_alloc_memmd(struct lov_stripe_md **lsmp, __u16 stripe_count,
309 int pattern, int magic)
314 CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
316 *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
318 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
322 cfs_atomic_set(&(*lsmp)->lsm_refc, 1);
323 spin_lock_init(&(*lsmp)->lsm_lock);
324 (*lsmp)->lsm_magic = magic;
325 (*lsmp)->lsm_stripe_count = stripe_count;
326 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
327 (*lsmp)->lsm_pattern = pattern;
328 (*lsmp)->lsm_pool_name[0] = '\0';
329 (*lsmp)->lsm_layout_gen = 0;
330 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
332 for (i = 0; i < stripe_count; i++)
333 loi_init((*lsmp)->lsm_oinfo[i]);
338 int lov_free_memmd(struct lov_stripe_md **lsmp)
340 struct lov_stripe_md *lsm = *lsmp;
344 LASSERT(cfs_atomic_read(&lsm->lsm_refc) > 0);
345 if ((refc = cfs_atomic_dec_return(&lsm->lsm_refc)) == 0) {
346 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
347 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
353 /* Unpack LOV object metadata from disk storage. It is packed in LE byte
354 * order and is opaque to the networking layer.
356 int lov_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
357 struct lov_mds_md *lmm, int lmm_bytes)
359 struct obd_device *obd = class_exp2obd(exp);
360 struct lov_obd *lov = &obd->u.lov;
361 int rc = 0, lsm_size;
366 /* If passed an MDS struct use values from there, otherwise defaults */
368 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
371 magic = le32_to_cpu(lmm->lmm_magic);
374 stripe_count = lov_get_stripecnt(lov, magic, 0);
377 /* If we aren't passed an lsmp struct, we just want the size */
379 /* XXX LOV STACKING call into osc for sizes */
381 RETURN(lov_stripe_md_size(stripe_count));
383 /* If we are passed an allocated struct but nothing to unpack, free */
385 lov_free_memmd(lsmp);
389 lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
394 /* If we are passed a pointer but nothing to unpack, we only alloc */
398 LASSERT(lsm_op_find(magic) != NULL);
399 rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
401 lov_free_memmd(lsmp);
408 static int __lov_setstripe(struct obd_export *exp, int max_lmm_size,
409 struct lov_stripe_md **lsmp,
410 struct lov_user_md *lump)
412 struct obd_device *obd = class_exp2obd(exp);
413 struct lov_obd *lov = &obd->u.lov;
414 char buffer[sizeof(struct lov_user_md_v3)];
415 struct lov_user_md_v3 *lumv3 = (struct lov_user_md_v3 *)&buffer[0];
416 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&buffer[0];
423 rc = lov_lum_swab_if_needed(lumv3, &lmm_magic, lump);
427 /* in the rest of the tests, as *lumv1 and lumv3 have the same
428 * fields, we use lumv1 to avoid code duplication */
430 if (lumv1->lmm_pattern == 0) {
431 lumv1->lmm_pattern = lov->desc.ld_pattern ?
432 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
435 if (lumv1->lmm_pattern != LOV_PATTERN_RAID0) {
436 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
441 /* 64kB is the largest common page size we see (ia64), and matches the
443 if (lumv1->lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
444 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
445 lumv1->lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
446 lumv1->lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
449 if ((lumv1->lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
450 (lumv1->lmm_stripe_offset !=
451 (typeof(lumv1->lmm_stripe_offset))(-1))) {
452 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
453 lumv1->lmm_stripe_offset, lov->desc.ld_tgt_count);
456 stripe_count = lov_get_stripecnt(lov, lmm_magic,
457 lumv1->lmm_stripe_count);
460 int max_stripes = (max_lmm_size -
461 lov_mds_md_size(0, lmm_magic)) /
462 sizeof(struct lov_ost_data_v1);
463 if (unlikely(max_stripes < stripe_count)) {
464 CDEBUG(D_IOCTL, "stripe count reset from %d to %d\n",
465 stripe_count, max_stripes);
466 stripe_count = max_stripes;
470 if (lmm_magic == LOV_USER_MAGIC_V3) {
471 struct pool_desc *pool;
473 /* In the function below, .hs_keycmp resolves to
474 * pool_hashkey_keycmp() */
475 /* coverity[overrun-buffer-val] */
476 pool = lov_find_pool(lov, lumv3->lmm_pool_name);
478 if (lumv3->lmm_stripe_offset !=
479 (typeof(lumv3->lmm_stripe_offset))(-1)) {
480 rc = lov_check_index_in_pool(
481 lumv3->lmm_stripe_offset, pool);
483 lov_pool_putref(pool);
488 if (stripe_count > pool_tgt_count(pool))
489 stripe_count = pool_tgt_count(pool);
491 lov_pool_putref(pool);
495 rc = lov_alloc_memmd(lsmp, stripe_count, lumv1->lmm_pattern, lmm_magic);
498 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lumv1->lmm_stripe_offset;
499 (*lsmp)->lsm_stripe_size = lumv1->lmm_stripe_size;
500 if (lmm_magic == LOV_USER_MAGIC_V3) {
501 cplen = strlcpy((*lsmp)->lsm_pool_name,
502 lumv3->lmm_pool_name,
503 sizeof((*lsmp)->lsm_pool_name));
504 if (cplen >= sizeof((*lsmp)->lsm_pool_name))
513 /* Configure object striping information on a new file.
515 * @lmmu is a pointer to a user struct with one or more of the fields set to
516 * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
517 * lmm_stripe_offset, and lmm_stripe_pattern. lmm_magic must be LOV_MAGIC.
518 * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
520 int lov_setstripe(struct obd_export *exp, int max_lmm_size,
521 struct lov_stripe_md **lsmp, struct lov_user_md *lump)
529 rc = __lov_setstripe(exp, max_lmm_size, lsmp, lump);
534 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
535 struct lov_user_md *lump)
539 struct obd_export *oexp;
540 struct lov_obd *lov = &exp->exp_obd->u.lov;
542 struct lov_user_ost_data_v1 *lmm_objects;
546 if (lump->lmm_magic == LOV_USER_MAGIC_V3)
547 lmm_objects = ((struct lov_user_md_v3 *)lump)->lmm_objects;
549 lmm_objects = lump->lmm_objects;
551 for (i = 0; i < lump->lmm_stripe_count; i++) {
552 __u32 len = sizeof(last_id);
553 oexp = lov->lov_tgts[lmm_objects[i].l_ost_idx]->ltd_exp;
554 rc = obd_get_info(NULL, oexp, sizeof(KEY_LAST_ID), KEY_LAST_ID,
555 &len, &last_id, NULL);
558 if (ostid_id(&lmm_objects[i].l_ost_oi) > last_id) {
559 CERROR("Setting EA for object > than last id on"
560 " ost idx %d "DOSTID" > "LPD64" \n",
561 lmm_objects[i].l_ost_idx,
562 POSTID(&lmm_objects[i].l_ost_oi), last_id);
567 rc = lov_setstripe(exp, 0, lsmp, lump);
571 for (i = 0; i < lump->lmm_stripe_count; i++) {
572 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
573 lmm_objects[i].l_ost_idx;
574 (*lsmp)->lsm_oinfo[i]->loi_oi = lmm_objects[i].l_ost_oi;
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.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 (cfs_copy_to_user(lump, lmmk, lmm_size))
676 obd_free_diskmd(exp, &lmmk);