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 2008 Sun Microsystems, Inc. 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 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 le64_to_cpu(lod->l_object_gr),
84 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_join(int level, struct lov_mds_md_join *lmmj)
97 lov_dump_lmm_common(level, &lmmj->lmmj_md);
98 CDEBUG(level, "extent_count %u\n",
99 le32_to_cpu(lmmj->lmmj_extent_count));
102 void lov_dump_lmm_v3(int level, struct lov_mds_md_v3 *lmm)
104 lov_dump_lmm_common(level, lmm);
105 CDEBUG(level,"pool_name "LOV_POOLNAMEF"\n", lmm->lmm_pool_name);
106 lov_dump_lmm_objects(level, lmm->lmm_objects,
107 le32_to_cpu(lmm->lmm_stripe_count));
110 void lov_dump_lmm(int level, void *lmm)
114 magic = ((struct lov_mds_md_v1 *)(lmm))->lmm_magic;
117 return lov_dump_lmm_v1(level, (struct lov_mds_md_v1 *)(lmm));
119 return lov_dump_lmm_join(level, (struct lov_mds_md_join *)(lmm));
121 return lov_dump_lmm_v3(level, (struct lov_mds_md_v3 *)(lmm));
123 CERROR("Cannot recognize lmm_magic %x", magic);
128 #define LMM_ASSERT(test) \
130 if (!(test)) lov_dump_lmm(D_ERROR, lmm); \
131 LASSERT(test); /* so we know what assertion failed */ \
134 /* Pack LOV object metadata for disk storage. It is packed in LE byte
135 * order and is opaque to the networking layer.
137 * XXX In the future, this will be enhanced to get the EA size from the
138 * underlying OSC device(s) to get their EA sizes so we can stack
139 * LOVs properly. For now lov_mds_md_size() just assumes one obd_id
142 int lov_packmd(struct obd_export *exp, struct lov_mds_md **lmmp,
143 struct lov_stripe_md *lsm)
145 struct obd_device *obd = class_exp2obd(exp);
146 struct lov_obd *lov = &obd->u.lov;
147 struct lov_mds_md_v1 *lmmv1;
148 struct lov_mds_md_v3 *lmmv3;
150 struct lov_ost_data_v1 *lmm_objects;
151 int lmm_size, lmm_magic;
156 lmm_magic = lsm->lsm_magic;
158 /* If we are just sizing the EA, limit the stripe count
159 * to the actual number of OSTs in this filesystem. */
161 stripe_count = lov_get_stripecnt(lov,
162 lsm->lsm_stripe_count);
163 lsm->lsm_stripe_count = stripe_count;
165 stripe_count = lsm->lsm_stripe_count;
168 /* No needs to allocated more than LOV_MAX_STRIPE_COUNT.
169 * Anyway, this is pretty inaccurate since ld_tgt_count now
170 * represents max index and we should rely on the actual number
172 stripe_count = min((__u32)LOV_MAX_STRIPE_COUNT,
173 lov->desc.ld_tgt_count);
176 lmm_magic = le32_to_cpu((*lmmp)->lmm_magic);
178 /* lsm == NULL and lmmp == NULL */
179 lmm_magic = LOV_MAGIC;
182 if ((lmm_magic != LOV_MAGIC_V1) &&
183 (lmm_magic != LOV_MAGIC_V3)) {
184 CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
185 lmm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
190 /* XXX LOV STACKING call into osc for sizes */
191 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
197 stripe_count = le32_to_cpu((*lmmp)->lmm_stripe_count);
198 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
199 OBD_FREE(*lmmp, lmm_size);
205 OBD_ALLOC(*lmmp, lmm_size);
210 CDEBUG(D_INFO, "lov_packmd: LOV_MAGIC 0x%08X, lmm_size = %d \n",
211 lmm_magic, lmm_size);
214 lmmv3 = (struct lov_mds_md_v3 *)*lmmp;
215 if (lmm_magic == LOV_MAGIC_V3)
216 lmmv3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3);
218 lmmv1->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
223 /* lmmv1 and lmmv3 point to the same struct and have the
226 lmmv1->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
227 lmmv1->lmm_object_gr = cpu_to_le64(lsm->lsm_object_gr);
228 lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
229 lmmv1->lmm_stripe_count = cpu_to_le32(stripe_count);
230 lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
231 if (lsm->lsm_magic == LOV_MAGIC_V3) {
232 strncpy(lmmv3->lmm_pool_name, lsm->lsm_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];
242 /* XXX LOV STACKING call down to osc_packmd() to do packing */
243 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
244 lmmv1->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
245 lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
246 lmm_objects[i].l_object_gr = cpu_to_le64(loi->loi_gr);
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 int lov_get_stripecnt(struct lov_obd *lov, __u32 stripe_count)
258 stripe_count = lov->desc.ld_default_stripe_count;
259 if (stripe_count > lov->desc.ld_active_tgt_count)
260 stripe_count = lov->desc.ld_active_tgt_count;
263 /* for now, we limit the stripe count directly, when bug 4424 is
264 * fixed this needs to be somewhat dynamic based on whether ext3
265 * can handle larger EA sizes. */
266 if (stripe_count > LOV_MAX_STRIPE_COUNT)
267 stripe_count = LOV_MAX_STRIPE_COUNT;
273 static int lov_verify_lmm(void *lmm, int lmm_bytes, int *stripe_count)
277 if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
281 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
282 le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
283 sz = lmm_bytes * 2 + 1;
284 OBD_ALLOC(buffer, sz);
285 if (buffer != NULL) {
288 for (i = 0; i < lmm_bytes; i++)
289 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
291 CERROR("%s\n", buffer);
292 OBD_FREE(buffer, sz);
296 rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
297 lmm_bytes, stripe_count);
301 int lov_alloc_memmd(struct lov_stripe_md **lsmp, int stripe_count,
302 int pattern, int magic)
307 CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
309 *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
311 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
315 spin_lock_init(&(*lsmp)->lsm_lock);
316 (*lsmp)->lsm_magic = magic;
317 (*lsmp)->lsm_stripe_count = stripe_count;
318 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
319 (*lsmp)->lsm_pattern = pattern;
320 (*lsmp)->lsm_pool_name[0] = '\0';
321 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
323 for (i = 0; i < stripe_count; i++)
324 loi_init((*lsmp)->lsm_oinfo[i]);
329 void lov_free_memmd(struct lov_stripe_md **lsmp)
331 struct lov_stripe_md *lsm = *lsmp;
333 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
334 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
340 /* Unpack LOV object metadata from disk storage. It is packed in LE byte
341 * order and is opaque to the networking layer.
343 int lov_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
344 struct lov_mds_md *lmm, int lmm_bytes)
346 struct obd_device *obd = class_exp2obd(exp);
347 struct lov_obd *lov = &obd->u.lov;
348 int rc = 0, stripe_count, lsm_size;
352 /* If passed an MDS struct use values from there, otherwise defaults */
354 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
357 magic = le32_to_cpu(lmm->lmm_magic);
359 stripe_count = lov_get_stripecnt(lov, 0);
363 /* If we aren't passed an lsmp struct, we just want the size */
365 /* XXX LOV STACKING call into osc for sizes */
367 RETURN(lov_stripe_md_size(stripe_count));
369 /* If we are passed an allocated struct but nothing to unpack, free */
371 lov_free_memmd(lsmp);
375 lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
380 /* If we are passed a pointer but nothing to unpack, we only alloc */
384 LASSERT(lsm_op_find(magic) != NULL);
385 rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
387 lov_free_memmd(lsmp);
394 static int __lov_setstripe(struct obd_export *exp, int max_lmm_size,
395 struct lov_stripe_md **lsmp,
396 struct lov_user_md *lump)
398 struct obd_device *obd = class_exp2obd(exp);
399 struct lov_obd *lov = &obd->u.lov;
400 struct lov_user_md_v3 lumv3;
401 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&lumv3;
407 if (copy_from_user(&lumv3, lump, sizeof(struct lov_user_md_v1)))
410 lmm_magic = lumv1->lmm_magic;
412 if (lmm_magic == __swab32(LOV_USER_MAGIC_V1)) {
413 lustre_swab_lov_user_md_v1(lumv1);
414 lmm_magic = LOV_USER_MAGIC_V1;
415 } else if (lmm_magic == LOV_USER_MAGIC_V3) {
416 if (copy_from_user(&lumv3, lump, sizeof(lumv3)))
418 } else if (lmm_magic == __swab32(LOV_USER_MAGIC_V3)) {
419 if (copy_from_user(&lumv3, lump, sizeof(lumv3)))
421 lustre_swab_lov_user_md_v3(&lumv3);
422 lmm_magic = LOV_USER_MAGIC_V3;
423 } else if (lmm_magic != LOV_USER_MAGIC_V1) {
425 "bad userland LOV MAGIC: %#08x != %#08x nor %#08x\n",
426 lmm_magic, LOV_USER_MAGIC_V1, LOV_USER_MAGIC_V3);
430 /* in the rest of the tests, as *lumv1 and lumv3 have the same
431 * fields, we use lumv1 to avoid code duplication */
433 if (lumv1->lmm_pattern == 0) {
434 lumv1->lmm_pattern = lov->desc.ld_pattern ?
435 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
438 if (lumv1->lmm_pattern != LOV_PATTERN_RAID0) {
439 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
444 /* 64kB is the largest common page size we see (ia64), and matches the
446 if (lumv1->lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
447 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
448 lumv1->lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
449 lumv1->lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
452 if ((lumv1->lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
453 (lumv1->lmm_stripe_offset !=
454 (typeof(lumv1->lmm_stripe_offset))(-1))) {
455 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
456 lumv1->lmm_stripe_offset, lov->desc.ld_tgt_count);
459 stripe_count = lov_get_stripecnt(lov, lumv1->lmm_stripe_count);
462 int max_stripes = (max_lmm_size -
463 lov_mds_md_size(0, lmm_magic)) /
464 sizeof(struct lov_ost_data_v1);
465 if (unlikely(max_stripes < stripe_count)) {
466 CDEBUG(D_IOCTL, "stripe count reset from %d to %d\n",
467 stripe_count, max_stripes);
468 stripe_count = max_stripes;
472 if (lmm_magic == LOV_USER_MAGIC_V3) {
473 struct pool_desc *pool;
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 if ((__u64)lumv1->lmm_stripe_size * stripe_count > ~0UL) {
495 CDEBUG(D_IOCTL, "stripe width %ux%i exeeds %lu bytes\n",
496 lumv1->lmm_stripe_size, (int)lumv1->lmm_stripe_count,
501 rc = lov_alloc_memmd(lsmp, stripe_count, lumv1->lmm_pattern, lmm_magic);
504 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lumv1->lmm_stripe_offset;
505 (*lsmp)->lsm_stripe_size = lumv1->lmm_stripe_size;
506 if (lmm_magic == LOV_USER_MAGIC_V3)
507 strncpy((*lsmp)->lsm_pool_name, lumv3.lmm_pool_name,
515 /* Configure object striping information on a new file.
517 * @lmmu is a pointer to a user struct with one or more of the fields set to
518 * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
519 * lmm_stripe_offset, and lmm_stripe_pattern. lmm_magic must be LOV_MAGIC.
520 * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
522 int lov_setstripe(struct obd_export *exp, int max_lmm_size,
523 struct lov_stripe_md **lsmp, struct lov_user_md *lump)
531 rc = __lov_setstripe(exp, max_lmm_size, lsmp, lump);
536 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
537 struct lov_user_md *lump)
541 struct obd_export *oexp;
542 struct lov_obd *lov = &exp->exp_obd->u.lov;
544 struct lov_user_ost_data_v1 *lmm_objects;
548 if (lump->lmm_magic == LOV_USER_MAGIC_V3)
549 lmm_objects = ((struct lov_user_md_v3 *)lump)->lmm_objects;
551 lmm_objects = lump->lmm_objects;
553 for (i = 0; i < lump->lmm_stripe_count; i++) {
554 __u32 len = sizeof(last_id);
555 oexp = lov->lov_tgts[lmm_objects[i].l_ost_idx]->ltd_exp;
556 rc = obd_get_info(oexp, sizeof(KEY_LAST_ID), KEY_LAST_ID,
557 &len, &last_id, NULL);
560 if (lmm_objects[i].l_object_id > last_id) {
561 CERROR("Setting EA for object > than last id on "
562 "ost idx %d "LPD64" > "LPD64" \n",
563 lmm_objects[i].l_ost_idx,
564 lmm_objects[i].l_object_id, last_id);
569 rc = lov_setstripe(exp, 0, lsmp, lump);
573 for (i = 0; i < lump->lmm_stripe_count; i++) {
574 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
575 lmm_objects[i].l_ost_idx;
576 (*lsmp)->lsm_oinfo[i]->loi_id = lmm_objects[i].l_object_id;
577 (*lsmp)->lsm_oinfo[i]->loi_gr = lmm_objects[i].l_object_gr;
583 /* Retrieve object striping information.
585 * @lump is a pointer to an in-core struct with lmm_ost_count indicating
586 * the maximum number of OST indices which will fit in the user buffer.
587 * lmm_magic must be LOV_USER_MAGIC.
589 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
590 struct lov_user_md *lump)
593 * XXX huge struct allocated on stack.
595 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
596 struct lov_user_md_v3 lum;
597 struct lov_mds_md *lmmk = NULL;
607 * "Switch to kernel segment" to allow copying from kernel space by
608 * copy_{to,from}_user().
613 /* we only need the header part from user space to get lmm_magic and
614 * lmm_stripe_count, (the header part is common to v1 and v3) */
615 lum_size = sizeof(struct lov_user_md_v1);
616 if (copy_from_user(&lum, lump, lum_size))
617 GOTO(out_set, rc = -EFAULT);
618 else if ((lum.lmm_magic != LOV_USER_MAGIC) &&
619 (lum.lmm_magic != LOV_USER_MAGIC_V3))
620 GOTO(out_set, rc = -EINVAL);
622 if (lum.lmm_stripe_count &&
623 (lum.lmm_stripe_count < lsm->lsm_stripe_count)) {
624 /* Return right size of stripe to user */
625 lum.lmm_stripe_count = lsm->lsm_stripe_count;
626 rc = copy_to_user(lump, &lum, lum_size);
627 GOTO(out_set, rc = -EOVERFLOW);
629 rc = lov_packmd(exp, &lmmk, lsm);
635 /* FIXME: Bug 1185 - copy fields properly when structs change */
636 /* struct lov_user_md_v3 and struct lov_mds_md_v3 must be the same */
637 CLASSERT(sizeof(lum) == sizeof(struct lov_mds_md_v3));
638 CLASSERT(sizeof lum.lmm_objects[0] == sizeof lmmk->lmm_objects[0]);
640 if ((cpu_to_le32(LOV_MAGIC) != LOV_MAGIC) &&
641 ((lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V1)) ||
642 (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)))) {
643 lustre_swab_lov_mds_md(lmmk);
644 lustre_swab_lov_user_md_objects(
645 (struct lov_user_ost_data*)lmmk->lmm_objects,
646 lmmk->lmm_stripe_count);
648 if (lum.lmm_magic == LOV_USER_MAGIC) {
649 /* User request for v1, we need skip lmm_pool_name */
650 if (lmmk->lmm_magic == LOV_MAGIC_V3) {
651 memmove((char*)(&lmmk->lmm_stripe_count) +
652 sizeof(lmmk->lmm_stripe_count),
653 ((struct lov_mds_md_v3*)lmmk)->lmm_objects,
654 lmmk->lmm_stripe_count *
655 sizeof(struct lov_ost_data_v1));
656 lmm_size -= LOV_MAXPOOLNAME;
659 /* if v3 we just have to update the lum_size */
660 lum_size = sizeof(struct lov_user_md_v3);
663 /* User wasn't expecting this many OST entries */
664 if (lum.lmm_stripe_count == 0)
666 else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count)
667 GOTO(out_set, rc = -EOVERFLOW);
669 * Have a difference between lov_mds_md & lov_user_md.
670 * So we have to re-order the data before copy to user.
672 lum.lmm_stripe_count = lmmk->lmm_stripe_count;
673 ((struct lov_user_md*)lmmk)->lmm_stripe_offset = 0;
674 ((struct lov_user_md*)lmmk)->lmm_stripe_count = lum.lmm_stripe_count;
675 if (copy_to_user(lump, lmmk, lmm_size))
678 obd_free_diskmd(exp, &lmmk);