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 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * lustre/lov/lov_pack.c
41 * (Un)packing of OST/MDS requests
43 * Author: Andreas Dilger <adilger@clusterfs.com>
46 #define DEBUG_SUBSYSTEM S_LOV
48 #include <liblustre.h>
51 #include <lustre_net.h>
54 #include <obd_class.h>
55 #include <obd_support.h>
56 #include <lustre/lustre_user.h>
58 #include "lov_internal.h"
60 static void lov_dump_lmm_common(int level, void *lmmp)
62 struct lov_mds_md *lmm = lmmp;
64 CDEBUG(level, "objid "LPX64", magic 0x%08x, pattern %#x\n",
65 (__u64)le64_to_cpu(lmm->lmm_object_id),
66 le32_to_cpu(lmm->lmm_magic),
67 le32_to_cpu(lmm->lmm_pattern));
68 CDEBUG(level,"stripe_size %u, stripe_count %u\n",
69 le32_to_cpu(lmm->lmm_stripe_size),
70 le32_to_cpu(lmm->lmm_stripe_count));
73 static void lov_dump_lmm_objects(int level, struct lov_ost_data *lod,
78 if (stripe_count > LOV_V1_INSANE_STRIPE_COUNT) {
79 CDEBUG(level, "bad stripe_count %u > max_stripe_count %u\n",
80 stripe_count, LOV_V1_INSANE_STRIPE_COUNT);
83 for (i = 0; i < stripe_count; ++i, ++lod) {
84 CDEBUG(level, "stripe %u idx %u subobj "LPX64"/"LPX64"\n", i,
85 le32_to_cpu(lod->l_ost_idx),
86 (__u64)le64_to_cpu(lod->l_object_seq),
87 (__u64)le64_to_cpu(lod->l_object_id));
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 le32_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 le32_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;
150 lmm_magic = lsm->lsm_magic;
153 lmm_magic = le32_to_cpu((*lmmp)->lmm_magic);
155 /* lsm == NULL and lmmp == NULL */
156 lmm_magic = LOV_MAGIC;
159 if ((lmm_magic != LOV_MAGIC_V1) &&
160 (lmm_magic != LOV_MAGIC_V3)) {
161 CERROR("bad mem LOV MAGIC: 0x%08X != 0x%08X nor 0x%08X\n",
162 lmm_magic, LOV_MAGIC_V1, LOV_MAGIC_V3);
168 /* If we are just sizing the EA, limit the stripe count
169 * to the actual number of OSTs in this filesystem. */
171 stripe_count = lov_get_stripecnt(lov, lmm_magic,
172 lsm->lsm_stripe_count);
173 lsm->lsm_stripe_count = stripe_count;
175 stripe_count = lsm->lsm_stripe_count;
178 /* No need to allocate more than maximum supported stripes.
179 * Anyway, this is pretty inaccurate since ld_tgt_count now
180 * represents max index and we should rely on the actual number
182 stripe_count = lov_mds_md_stripecnt(lov->lov_ocd.ocd_max_easize,
184 if (stripe_count > lov->desc.ld_tgt_count)
185 stripe_count = lov->desc.ld_tgt_count;
188 /* XXX LOV STACKING call into osc for sizes */
189 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
195 stripe_count = le32_to_cpu((*lmmp)->lmm_stripe_count);
196 lmm_size = lov_mds_md_size(stripe_count, lmm_magic);
197 OBD_FREE_LARGE(*lmmp, lmm_size);
203 OBD_ALLOC_LARGE(*lmmp, lmm_size);
208 CDEBUG(D_INFO, "lov_packmd: LOV_MAGIC 0x%08X, lmm_size = %d \n",
209 lmm_magic, lmm_size);
212 lmmv3 = (struct lov_mds_md_v3 *)*lmmp;
213 if (lmm_magic == LOV_MAGIC_V3)
214 lmmv3->lmm_magic = cpu_to_le32(LOV_MAGIC_V3);
216 lmmv1->lmm_magic = cpu_to_le32(LOV_MAGIC_V1);
221 /* lmmv1 and lmmv3 point to the same struct and have the
224 lmmv1->lmm_object_id = cpu_to_le64(lsm->lsm_object_id);
225 lmmv1->lmm_object_seq = cpu_to_le64(lsm->lsm_object_seq);
226 lmmv1->lmm_stripe_size = cpu_to_le32(lsm->lsm_stripe_size);
227 lmmv1->lmm_stripe_count = cpu_to_le32(stripe_count);
228 lmmv1->lmm_pattern = cpu_to_le32(lsm->lsm_pattern);
229 if (lsm->lsm_magic == LOV_MAGIC_V3) {
230 strncpy(lmmv3->lmm_pool_name, lsm->lsm_pool_name,
232 lmm_objects = lmmv3->lmm_objects;
234 lmm_objects = lmmv1->lmm_objects;
237 for (i = 0; i < stripe_count; i++) {
238 struct lov_oinfo *loi = lsm->lsm_oinfo[i];
239 /* XXX LOV STACKING call down to osc_packmd() to do packing */
240 LASSERTF(loi->loi_id, "lmm_oid "LPU64" stripe %u/%u idx %u\n",
241 lmmv1->lmm_object_id, i, stripe_count, loi->loi_ost_idx);
242 lmm_objects[i].l_object_id = cpu_to_le64(loi->loi_id);
243 lmm_objects[i].l_object_seq = cpu_to_le64(loi->loi_seq);
244 lmm_objects[i].l_ost_gen = cpu_to_le32(loi->loi_ost_gen);
245 lmm_objects[i].l_ost_idx = cpu_to_le32(loi->loi_ost_idx);
251 /* Find the max stripecount we should use */
252 __u32 lov_get_stripecnt(struct lov_obd *lov, __u32 magic, __u32 stripe_count)
254 __u32 max_stripes = LOV_MAX_STRIPE_COUNT_OLD;
257 stripe_count = lov->desc.ld_default_stripe_count;
258 if (stripe_count > lov->desc.ld_active_tgt_count)
259 stripe_count = lov->desc.ld_active_tgt_count;
263 /* stripe count is based on whether ldiskfs can handle
265 if (lov->lov_ocd.ocd_connect_flags & OBD_CONNECT_MAX_EASIZE &&
266 lov->lov_ocd.ocd_max_easize)
267 max_stripes = lov_mds_md_stripecnt(lov->lov_ocd.ocd_max_easize,
270 if (stripe_count > max_stripes)
271 stripe_count = max_stripes;
277 static int lov_verify_lmm(void *lmm, int lmm_bytes, int *stripe_count)
281 if (lsm_op_find(le32_to_cpu(*(__u32 *)lmm)) == NULL) {
285 CERROR("bad disk LOV MAGIC: 0x%08X; dumping LMM (size=%d):\n",
286 le32_to_cpu(*(__u32 *)lmm), lmm_bytes);
287 sz = lmm_bytes * 2 + 1;
288 OBD_ALLOC_LARGE(buffer, sz);
289 if (buffer != NULL) {
292 for (i = 0; i < lmm_bytes; i++)
293 sprintf(buffer+2*i, "%.2X", ((char *)lmm)[i]);
295 CERROR("%s\n", buffer);
296 OBD_FREE_LARGE(buffer, sz);
300 rc = lsm_op_find(le32_to_cpu(*(__u32 *)lmm))->lsm_lmm_verify(lmm,
301 lmm_bytes, stripe_count);
305 int lov_alloc_memmd(struct lov_stripe_md **lsmp, int stripe_count,
306 int pattern, int magic)
311 CDEBUG(D_INFO, "alloc lsm, stripe_count %d\n", stripe_count);
313 *lsmp = lsm_alloc_plain(stripe_count, &lsm_size);
315 CERROR("can't allocate lsmp stripe_count %d\n", stripe_count);
319 cfs_spin_lock_init(&(*lsmp)->lsm_lock);
320 (*lsmp)->lsm_magic = magic;
321 (*lsmp)->lsm_stripe_count = stripe_count;
322 (*lsmp)->lsm_maxbytes = LUSTRE_STRIPE_MAXBYTES * stripe_count;
323 (*lsmp)->lsm_pattern = pattern;
324 (*lsmp)->lsm_pool_name[0] = '\0';
325 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = ~0;
327 for (i = 0; i < stripe_count; i++)
328 loi_init((*lsmp)->lsm_oinfo[i]);
333 void lov_free_memmd(struct lov_stripe_md **lsmp)
335 struct lov_stripe_md *lsm = *lsmp;
337 LASSERT(lsm_op_find(lsm->lsm_magic) != NULL);
338 lsm_op_find(lsm->lsm_magic)->lsm_free(lsm);
344 /* Unpack LOV object metadata from disk storage. It is packed in LE byte
345 * order and is opaque to the networking layer.
347 int lov_unpackmd(struct obd_export *exp, struct lov_stripe_md **lsmp,
348 struct lov_mds_md *lmm, int lmm_bytes)
350 struct obd_device *obd = class_exp2obd(exp);
351 struct lov_obd *lov = &obd->u.lov;
352 int rc = 0, stripe_count, lsm_size;
356 /* If passed an MDS struct use values from there, otherwise defaults */
358 rc = lov_verify_lmm(lmm, lmm_bytes, &stripe_count);
361 magic = le32_to_cpu(lmm->lmm_magic);
364 stripe_count = lov_get_stripecnt(lov, magic, 0);
367 /* If we aren't passed an lsmp struct, we just want the size */
369 /* XXX LOV STACKING call into osc for sizes */
371 RETURN(lov_stripe_md_size(stripe_count));
373 /* If we are passed an allocated struct but nothing to unpack, free */
375 lov_free_memmd(lsmp);
379 lsm_size = lov_alloc_memmd(lsmp, stripe_count, LOV_PATTERN_RAID0,
384 /* If we are passed a pointer but nothing to unpack, we only alloc */
388 LASSERT(lsm_op_find(magic) != NULL);
389 rc = lsm_op_find(magic)->lsm_unpackmd(lov, *lsmp, lmm);
391 lov_free_memmd(lsmp);
398 static int __lov_setstripe(struct obd_export *exp, int max_lmm_size,
399 struct lov_stripe_md **lsmp,
400 struct lov_user_md *lump)
402 struct obd_device *obd = class_exp2obd(exp);
403 struct lov_obd *lov = &obd->u.lov;
404 char buffer[sizeof(struct lov_user_md_v3)];
405 struct lov_user_md_v3 *lumv3 = (struct lov_user_md_v3 *)&buffer[0];
406 struct lov_user_md_v1 *lumv1 = (struct lov_user_md_v1 *)&buffer[0];
412 if (cfs_copy_from_user(lumv3, lump, sizeof(struct lov_user_md_v1)))
415 lmm_magic = lumv1->lmm_magic;
417 if (lmm_magic == __swab32(LOV_USER_MAGIC_V1)) {
418 lustre_swab_lov_user_md_v1(lumv1);
419 lmm_magic = LOV_USER_MAGIC_V1;
420 } else if (lmm_magic == LOV_USER_MAGIC_V3) {
421 if (cfs_copy_from_user(lumv3, lump, sizeof(*lumv3)))
423 } else if (lmm_magic == __swab32(LOV_USER_MAGIC_V3)) {
424 if (cfs_copy_from_user(lumv3, lump, sizeof(*lumv3)))
426 lustre_swab_lov_user_md_v3(lumv3);
427 lmm_magic = LOV_USER_MAGIC_V3;
428 } else if (lmm_magic != LOV_USER_MAGIC_V1) {
430 "bad userland LOV MAGIC: %#08x != %#08x nor %#08x\n",
431 lmm_magic, LOV_USER_MAGIC_V1, LOV_USER_MAGIC_V3);
435 /* in the rest of the tests, as *lumv1 and lumv3 have the same
436 * fields, we use lumv1 to avoid code duplication */
438 if (lumv1->lmm_pattern == 0) {
439 lumv1->lmm_pattern = lov->desc.ld_pattern ?
440 lov->desc.ld_pattern : LOV_PATTERN_RAID0;
443 if (lumv1->lmm_pattern != LOV_PATTERN_RAID0) {
444 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
449 /* 64kB is the largest common page size we see (ia64), and matches the
451 if (lumv1->lmm_stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
452 CDEBUG(D_IOCTL, "stripe size %u not multiple of %u, fixing\n",
453 lumv1->lmm_stripe_size, LOV_MIN_STRIPE_SIZE);
454 lumv1->lmm_stripe_size = LOV_MIN_STRIPE_SIZE;
457 if ((lumv1->lmm_stripe_offset >= lov->desc.ld_tgt_count) &&
458 (lumv1->lmm_stripe_offset !=
459 (typeof(lumv1->lmm_stripe_offset))(-1))) {
460 CDEBUG(D_IOCTL, "stripe offset %u > number of OSTs %u\n",
461 lumv1->lmm_stripe_offset, lov->desc.ld_tgt_count);
464 stripe_count = lov_get_stripecnt(lov, lmm_magic,
465 lumv1->lmm_stripe_count);
468 int max_stripes = (max_lmm_size -
469 lov_mds_md_size(0, lmm_magic)) /
470 sizeof(struct lov_ost_data_v1);
471 if (unlikely(max_stripes < stripe_count)) {
472 CDEBUG(D_IOCTL, "stripe count reset from %d to %d\n",
473 stripe_count, max_stripes);
474 stripe_count = max_stripes;
478 if (lmm_magic == LOV_USER_MAGIC_V3) {
479 struct pool_desc *pool;
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 rc = lov_alloc_memmd(lsmp, stripe_count, lumv1->lmm_pattern, lmm_magic);
503 (*lsmp)->lsm_oinfo[0]->loi_ost_idx = lumv1->lmm_stripe_offset;
504 (*lsmp)->lsm_stripe_size = lumv1->lmm_stripe_size;
505 if (lmm_magic == LOV_USER_MAGIC_V3)
506 strncpy((*lsmp)->lsm_pool_name, lumv3->lmm_pool_name,
514 /* Configure object striping information on a new file.
516 * @lmmu is a pointer to a user struct with one or more of the fields set to
517 * indicate the application preference: lmm_stripe_count, lmm_stripe_size,
518 * lmm_stripe_offset, and lmm_stripe_pattern. lmm_magic must be LOV_MAGIC.
519 * @lsmp is a pointer to an in-core stripe MD that needs to be filled in.
521 int lov_setstripe(struct obd_export *exp, int max_lmm_size,
522 struct lov_stripe_md **lsmp, struct lov_user_md *lump)
530 rc = __lov_setstripe(exp, max_lmm_size, lsmp, lump);
535 int lov_setea(struct obd_export *exp, struct lov_stripe_md **lsmp,
536 struct lov_user_md *lump)
540 struct obd_export *oexp;
541 struct lov_obd *lov = &exp->exp_obd->u.lov;
543 struct lov_user_ost_data_v1 *lmm_objects;
547 if (lump->lmm_magic == LOV_USER_MAGIC_V3)
548 lmm_objects = ((struct lov_user_md_v3 *)lump)->lmm_objects;
550 lmm_objects = lump->lmm_objects;
552 for (i = 0; i < lump->lmm_stripe_count; i++) {
553 __u32 len = sizeof(last_id);
554 oexp = lov->lov_tgts[lmm_objects[i].l_ost_idx]->ltd_exp;
555 rc = obd_get_info(oexp, sizeof(KEY_LAST_ID), KEY_LAST_ID,
556 &len, &last_id, NULL);
559 if (lmm_objects[i].l_object_id > last_id) {
560 CERROR("Setting EA for object > than last id on "
561 "ost idx %d "LPD64" > "LPD64" \n",
562 lmm_objects[i].l_ost_idx,
563 lmm_objects[i].l_object_id, last_id);
568 rc = lov_setstripe(exp, 0, lsmp, lump);
572 for (i = 0; i < lump->lmm_stripe_count; i++) {
573 (*lsmp)->lsm_oinfo[i]->loi_ost_idx =
574 lmm_objects[i].l_ost_idx;
575 (*lsmp)->lsm_oinfo[i]->loi_id = lmm_objects[i].l_object_id;
576 (*lsmp)->lsm_oinfo[i]->loi_seq = lmm_objects[i].l_object_seq;
582 /* Retrieve object striping information.
584 * @lump is a pointer to an in-core struct with lmm_ost_count indicating
585 * the maximum number of OST indices which will fit in the user buffer.
586 * lmm_magic must be LOV_USER_MAGIC.
588 int lov_getstripe(struct obd_export *exp, struct lov_stripe_md *lsm,
589 struct lov_user_md *lump)
592 * XXX huge struct allocated on stack.
594 /* we use lov_user_md_v3 because it is larger than lov_user_md_v1 */
595 struct lov_user_md_v3 lum;
596 struct lov_mds_md *lmmk = NULL;
606 * "Switch to kernel segment" to allow copying from kernel space by
607 * copy_{to,from}_user().
612 /* we only need the header part from user space to get lmm_magic and
613 * lmm_stripe_count, (the header part is common to v1 and v3) */
614 lum_size = sizeof(struct lov_user_md_v1);
615 if (cfs_copy_from_user(&lum, lump, lum_size))
616 GOTO(out_set, rc = -EFAULT);
617 else if ((lum.lmm_magic != LOV_USER_MAGIC) &&
618 (lum.lmm_magic != LOV_USER_MAGIC_V3))
619 GOTO(out_set, rc = -EINVAL);
621 if (lum.lmm_stripe_count &&
622 (lum.lmm_stripe_count < lsm->lsm_stripe_count)) {
623 /* Return right size of stripe to user */
624 lum.lmm_stripe_count = lsm->lsm_stripe_count;
625 rc = cfs_copy_to_user(lump, &lum, lum_size);
626 GOTO(out_set, rc = -EOVERFLOW);
628 rc = lov_packmd(exp, &lmmk, lsm);
634 /* FIXME: Bug 1185 - copy fields properly when structs change */
635 /* struct lov_user_md_v3 and struct lov_mds_md_v3 must be the same */
636 CLASSERT(sizeof(lum) == sizeof(struct lov_mds_md_v3));
637 CLASSERT(sizeof lum.lmm_objects[0] == sizeof lmmk->lmm_objects[0]);
639 if ((cpu_to_le32(LOV_MAGIC) != LOV_MAGIC) &&
640 ((lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V1)) ||
641 (lmmk->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)))) {
642 lustre_swab_lov_mds_md(lmmk);
643 lustre_swab_lov_user_md_objects(
644 (struct lov_user_ost_data*)lmmk->lmm_objects,
645 lmmk->lmm_stripe_count);
647 if (lum.lmm_magic == LOV_USER_MAGIC) {
648 /* User request for v1, we need skip lmm_pool_name */
649 if (lmmk->lmm_magic == LOV_MAGIC_V3) {
650 memmove((char*)(&lmmk->lmm_stripe_count) +
651 sizeof(lmmk->lmm_stripe_count),
652 ((struct lov_mds_md_v3*)lmmk)->lmm_objects,
653 lmmk->lmm_stripe_count *
654 sizeof(struct lov_ost_data_v1));
655 lmm_size -= LOV_MAXPOOLNAME;
658 /* if v3 we just have to update the lum_size */
659 lum_size = sizeof(struct lov_user_md_v3);
662 /* User wasn't expecting this many OST entries */
663 if (lum.lmm_stripe_count == 0)
665 else if (lum.lmm_stripe_count < lmmk->lmm_stripe_count)
666 GOTO(out_set, rc = -EOVERFLOW);
668 * Have a difference between lov_mds_md & lov_user_md.
669 * So we have to re-order the data before copy to user.
671 lum.lmm_stripe_count = lmmk->lmm_stripe_count;
672 ((struct lov_user_md*)lmmk)->lmm_stripe_offset = 0;
673 ((struct lov_user_md*)lmmk)->lmm_stripe_count = lum.lmm_stripe_count;
674 if (cfs_copy_to_user(lump, lmmk, lmm_size))
677 obd_free_diskmd(exp, &lmmk);