]> git.ipfire.org Git - thirdparty/mdadm.git/blobdiff - super0.c
Create: add support for RAID0 layouts.
[thirdparty/mdadm.git] / super0.c
index 3218377f74f4d72e12b4a6f4559d8041264ff0c6..6af140bbf72197223e54fd37b9d1fd1ac28167ad 100644 (file)
--- a/super0.c
+++ b/super0.c
@@ -76,8 +76,6 @@ static void super0_swap_endian(struct mdp_superblock_s *sb)
 
 }
 
-#ifndef MDASSEMBLE
-
 static void examine_super0(struct supertype *st, char *homehost)
 {
        mdp_super_t *sb = st->sb;
@@ -87,17 +85,17 @@ static void examine_super0(struct supertype *st, char *homehost)
        char *c;
 
        printf("          Magic : %08x\n", sb->md_magic);
-       printf("        Version : %d.%02d.%02d\n", sb->major_version, sb->minor_version,
-              sb->patch_version);
+       printf("        Version : %d.%02d.%02d\n",
+              sb->major_version, sb->minor_version, sb->patch_version);
        if (sb->minor_version >= 90) {
-               printf("           UUID : %08x:%08x:%08x:%08x", sb->set_uuid0, sb->set_uuid1,
-                      sb->set_uuid2, sb->set_uuid3);
+               printf("           UUID : %08x:%08x:%08x:%08x", sb->set_uuid0,
+                      sb->set_uuid1, sb->set_uuid2, sb->set_uuid3);
                if (homehost) {
                        char buf[20];
-                       void *hash = sha1_buffer(homehost,
-                                                strlen(homehost),
-                                                buf);
-                       if (memcmp(&sb->set_uuid2, hash, 8)==0)
+                       void *hash;
+
+                       hash = sha1_buffer(homehost, strlen(homehost), buf);
+                       if (memcmp(&sb->set_uuid2, hash, 8) == 0)
                                printf(" (local to host %s)", homehost);
                }
                printf("\n");
@@ -109,19 +107,27 @@ static void examine_super0(struct supertype *st, char *homehost)
 
        atime = sb->ctime;
        printf("  Creation Time : %.24s\n", ctime(&atime));
-       c=map_num(pers, sb->level);
+       c = map_num(pers, sb->level);
        printf("     Raid Level : %s\n", c?c:"-unknown-");
        if ((int)sb->level > 0) {
                int ddsks = 0, ddsks_denom = 1;
                printf("  Used Dev Size : %d%s\n", sb->size,
                       human_size((long long)sb->size<<10));
                switch(sb->level) {
-               case 1: ddsks=1;break;
+               case 1:
+                       ddsks=1;
+                       break;
                case 4:
-               case 5: ddsks = sb->raid_disks-1; break;
-               case 6: ddsks = sb->raid_disks-2; break;
-               case 10: ddsks = sb->raid_disks;
-                       ddsks_denom =  (sb->layout&255) * ((sb->layout>>8)&255);
+               case 5:
+                       ddsks = sb->raid_disks - 1;
+                       break;
+               case 6:
+                       ddsks = sb->raid_disks - 2;
+                       break;
+               case 10:
+                       ddsks = sb->raid_disks;
+                       ddsks_denom =
+                               (sb->layout & 255) * ((sb->layout >> 8) & 255);
                }
                if (ddsks) {
                        long long asize = sb->size;
@@ -134,11 +140,14 @@ static void examine_super0(struct supertype *st, char *homehost)
        printf("  Total Devices : %d\n", sb->nr_disks);
        printf("Preferred Minor : %d\n", sb->md_minor);
        printf("\n");
-       if (sb->minor_version > 90 && (sb->reshape_position+1) != 0) {
-               printf("  Reshape pos'n : %llu%s\n", (unsigned long long)sb->reshape_position/2, human_size((long long)sb->reshape_position<<9));
+       if (sb->minor_version > 90 && (sb->reshape_position + 1) != 0) {
+               printf("  Reshape pos'n : %llu%s\n",
+                      (unsigned long long)sb->reshape_position / 2,
+                      human_size((long long)sb->reshape_position << 9));
                if (sb->delta_disks) {
                        printf("  Delta Devices : %d", sb->delta_disks);
-                       printf(" (%d->%d)\n", sb->raid_disks-sb->delta_disks, sb->raid_disks);
+                       printf(" (%d->%d)\n", sb->raid_disks-sb->delta_disks,
+                              sb->raid_disks);
                        if (((int)sb->delta_disks) < 0)
                                delta_extra = - sb->delta_disks;
                }
@@ -149,11 +158,13 @@ static void examine_super0(struct supertype *st, char *homehost)
                if (sb->new_layout != sb->layout) {
                        if (sb->level == 5) {
                                c = map_num(r5layout, sb->new_layout);
-                               printf("     New Layout : %s\n", c?c:"-unknown-");
+                               printf("     New Layout : %s\n",
+                                      c?c:"-unknown-");
                        }
                        if (sb->level == 6) {
                                c = map_num(r6layout, sb->new_layout);
-                               printf("     New Layout : %s\n", c?c:"-unknown-");
+                               printf("     New Layout : %s\n",
+                                      c?c:"-unknown-");
                        }
                        if (sb->level == 10) {
                                printf("     New Layout : near=%d, %s=%d\n",
@@ -169,8 +180,8 @@ static void examine_super0(struct supertype *st, char *homehost)
        atime = sb->utime;
        printf("    Update Time : %.24s\n", ctime(&atime));
        printf("          State : %s\n",
-              (sb->state&(1<<MD_SB_CLEAN))?"clean":"active");
-       if (sb->state & (1<<MD_SB_BITMAP_PRESENT))
+              (sb->state&(1 << MD_SB_CLEAN)) ? "clean":"active");
+       if (sb->state & (1 << MD_SB_BITMAP_PRESENT))
                printf("Internal Bitmap : present\n");
        printf(" Active Devices : %d\n", sb->active_disks);
        printf("Working Devices : %d\n", sb->working_disks);
@@ -179,10 +190,10 @@ static void examine_super0(struct supertype *st, char *homehost)
        if (calc_sb0_csum(sb) == sb->sb_csum)
                printf("       Checksum : %x - correct\n", sb->sb_csum);
        else
-               printf("       Checksum : %x - expected %lx\n", sb->sb_csum, calc_sb0_csum(sb));
+               printf("       Checksum : %x - expected %lx\n",
+                      sb->sb_csum, calc_sb0_csum(sb));
        printf("         Events : %llu\n",
-              ((unsigned long long)sb->events_hi << 32)
-              + sb->events_lo);
+              ((unsigned long long)sb->events_hi << 32) + sb->events_lo);
        printf("\n");
        if (sb->level == 5) {
                c = map_num(r5layout, sb->layout);
@@ -203,37 +214,50 @@ static void examine_super0(struct supertype *st, char *homehost)
        case 5:
        case 6:
        case 10:
-               printf("     Chunk Size : %dK\n", sb->chunk_size/1024);
+               printf("     Chunk Size : %dK\n", sb->chunk_size / 1024);
                break;
        case -1:
-               printf("       Rounding : %dK\n", sb->chunk_size/1024);
+               printf("       Rounding : %dK\n", sb->chunk_size / 1024);
+               break;
+       default:
                break;
-       default: break;
        }
        printf("\n");
        printf("      Number   Major   Minor   RaidDevice State\n");
-       for (d= -1; d<(signed int)(sb->raid_disks+delta_extra + sb->spare_disks); d++) {
+       for (d = -1;
+            d < (signed int)(sb->raid_disks + delta_extra + sb->spare_disks);
+            d++) {
                mdp_disk_t *dp;
                char *dv;
-               char nb[5];
-               int wonly;
+               char nb[11];
+               int wonly, failfast;
                if (d>=0) dp = &sb->disks[d];
                else dp = &sb->this_disk;
                snprintf(nb, sizeof(nb), "%4d", d);
-               printf("%4s %5d   %5d    %5d    %5d     ", d < 0 ? "this" :  nb,
+               printf("%4s %5d   %5d    %5d    %5d     ", d < 0 ? "this" : nb,
                       dp->number, dp->major, dp->minor, dp->raid_disk);
-               wonly = dp->state & (1<<MD_DISK_WRITEMOSTLY);
-               dp->state &= ~(1<<MD_DISK_WRITEMOSTLY);
-               if (dp->state & (1<<MD_DISK_FAULTY)) printf(" faulty");
-               if (dp->state & (1<<MD_DISK_ACTIVE)) printf(" active");
-               if (dp->state & (1<<MD_DISK_SYNC)) printf(" sync");
-               if (dp->state & (1<<MD_DISK_REMOVED)) printf(" removed");
-               if (wonly) printf(" write-mostly");
-               if (dp->state == 0) printf(" spare");
-               if ((dv=map_dev(dp->major, dp->minor, 0)))
+               wonly = dp->state & (1 << MD_DISK_WRITEMOSTLY);
+               failfast = dp->state & (1<<MD_DISK_FAILFAST);
+               dp->state &= ~(wonly | failfast);
+               if (dp->state & (1 << MD_DISK_FAULTY))
+                       printf(" faulty");
+               if (dp->state & (1 << MD_DISK_ACTIVE))
+                       printf(" active");
+               if (dp->state & (1 << MD_DISK_SYNC))
+                       printf(" sync");
+               if (dp->state & (1 << MD_DISK_REMOVED))
+                       printf(" removed");
+               if (wonly)
+                       printf(" write-mostly");
+               if (failfast)
+                       printf(" failfast");
+               if (dp->state == 0)
+                       printf(" spare");
+               if ((dv = map_dev(dp->major, dp->minor, 0)))
                        printf("   %s", dv);
                printf("\n");
-               if (d == -1) printf("\n");
+               if (d == -1)
+                       printf("\n");
        }
 }
 
@@ -324,10 +348,10 @@ err:
        return 1;
 }
 
-static void detail_super0(struct supertype *st, char *homehost)
+static void detail_super0(struct supertype *st, char *homehost, char *subarray)
 {
        mdp_super_t *sb = st->sb;
-       printf("           UUID : ");
+       printf("              UUID : ");
        if (sb->minor_version >= 90)
                printf("%08x:%08x:%08x:%08x", sb->set_uuid0, sb->set_uuid1,
                       sb->set_uuid2, sb->set_uuid3);
@@ -341,10 +365,10 @@ static void detail_super0(struct supertype *st, char *homehost)
                if (memcmp(&sb->set_uuid2, hash, 8)==0)
                        printf(" (local to host %s)", homehost);
        }
-       printf("\n         Events : %d.%d\n\n", sb->events_hi, sb->events_lo);
+       printf("\n            Events : %d.%d\n\n", sb->events_hi, sb->events_lo);
 }
 
-static void brief_detail_super0(struct supertype *st)
+static void brief_detail_super0(struct supertype *st, char *subarray)
 {
        mdp_super_t *sb = st->sb;
        printf(" UUID=");
@@ -354,7 +378,6 @@ static void brief_detail_super0(struct supertype *st)
        else
                printf("%08x", sb->set_uuid0);
 }
-#endif
 
 static int match_home0(struct supertype *st, char *homehost)
 {
@@ -405,7 +428,8 @@ static void getinfo_super0(struct supertype *st, struct mdinfo *info, char *map)
        info->array.utime = sb->utime;
        info->array.chunk_size = sb->chunk_size;
        info->array.state = sb->state;
-       info->component_size = sb->size*2;
+       info->component_size = sb->size;
+       info->component_size *= 2;
 
        if (sb->state & (1<<MD_SB_BITMAP_PRESENT))
                info->bitmap_offset = 8;
@@ -496,9 +520,10 @@ static int update_super0(struct supertype *st, struct mdinfo *info,
                 * up 4 bytes before continuing
                 */
                __u32 *sb32 = (__u32*)sb;
-               memcpy(sb32+MD_SB_GENERIC_CONSTANT_WORDS+7,
-                      sb32+MD_SB_GENERIC_CONSTANT_WORDS+7+1,
-                      (MD_SB_WORDS - (MD_SB_GENERIC_CONSTANT_WORDS+7+1))*4);
+
+               memmove(sb32+MD_SB_GENERIC_CONSTANT_WORDS+7,
+                       sb32+MD_SB_GENERIC_CONSTANT_WORDS+7+1,
+                       (MD_SB_WORDS - (MD_SB_GENERIC_CONSTANT_WORDS+7+1))*4);
                if (verbose >= 0)
                        pr_err("adjusting superblock of %s for 2.2/sparc compatibility.\n",
                               devname);
@@ -557,16 +582,17 @@ static int update_super0(struct supertype *st, struct mdinfo *info,
        } else if (strcmp(update, "assemble")==0) {
                int d = info->disk.number;
                int wonly = sb->disks[d].state & (1<<MD_DISK_WRITEMOSTLY);
-               int mask = (1<<MD_DISK_WRITEMOSTLY);
+               int failfast = sb->disks[d].state & (1<<MD_DISK_FAILFAST);
+               int mask = (1<<MD_DISK_WRITEMOSTLY)|(1<<MD_DISK_FAILFAST);
                int add = 0;
                if (sb->minor_version >= 91)
                        /* During reshape we don't insist on everything
                         * being marked 'sync'
                         */
                        add = (1<<MD_DISK_SYNC);
-               if (((sb->disks[d].state & ~mask) | add)
-                   != (unsigned)info->disk.state) {
-                       sb->disks[d].state = info->disk.state | wonly;
+               if (((sb->disks[d].state & ~mask) | add) !=
+                   (unsigned)info->disk.state) {
+                       sb->disks[d].state = info->disk.state | wonly |failfast;
                        rv = 1;
                }
                if (info->reshape_active &&
@@ -692,16 +718,14 @@ static int update_super0(struct supertype *st, struct mdinfo *info,
 }
 
 /*
- * For verion-0 superblock, the homehost is 'stored' in the
- * uuid.  8 bytes for a hash of the host leaving 8 bytes
- * of random material.
- * We use the first 8 bytes (64bits) of the sha1 of the
- * host name
+ * For version-0 superblock, the homehost is 'stored' in the uuid.
+ * 8 bytes for a hash of the host leaving 8 bytes of random material.
+ * We use the first 8 bytes (64bits) of the sha1 of the host name
  */
-
 static int init_super0(struct supertype *st, mdu_array_info_t *info,
-                      unsigned long long size, char *ignored_name, char *homehost,
-                      int *uuid, unsigned long long data_offset)
+                      struct shape *s, char *ignored_name,
+                      char *homehost, int *uuid,
+                      unsigned long long data_offset)
 {
        mdp_super_t *sb;
        int spares;
@@ -713,7 +737,7 @@ static int init_super0(struct supertype *st, mdu_array_info_t *info,
 
        if (posix_memalign((void**)&sb, 4096,
                           MD_SB_BYTES + ROUND_UP(sizeof(bitmap_super_t), 4096)) != 0) {
-               pr_err("%s could not allocate superblock\n", __func__);
+               pr_err("could not allocate superblock\n");
                return 0;
        }
        memset(sb, 0, MD_SB_BYTES + sizeof(bitmap_super_t));
@@ -738,8 +762,8 @@ static int init_super0(struct supertype *st, mdu_array_info_t *info,
        sb->gvalid_words = 0; /* ignored */
        sb->ctime = time(0);
        sb->level = info->level;
-       sb->size = size;
-       if (size != (unsigned long long)sb->size)
+       sb->size = s->size;
+       if (s->size != (unsigned long long)sb->size)
                return 0;
        sb->nr_disks = info->nr_disks;
        sb->raid_disks = info->raid_disks;
@@ -751,16 +775,12 @@ static int init_super0(struct supertype *st, mdu_array_info_t *info,
                sb->set_uuid2 = uuid[2];
                sb->set_uuid3 = uuid[3];
        } else {
-               int rfd = open("/dev/urandom", O_RDONLY);
-               if (rfd < 0 || read(rfd, &sb->set_uuid0, 4) != 4)
-                       sb->set_uuid0 = random();
-               if (rfd < 0 || read(rfd, &sb->set_uuid1, 12) != 12) {
-                       sb->set_uuid1 = random();
-                       sb->set_uuid2 = random();
-                       sb->set_uuid3 = random();
-               }
-               if (rfd >= 0)
-                       close(rfd);
+               __u32 r[4];
+               random_uuid((__u8 *)r);
+               sb->set_uuid0 = r[0];
+               sb->set_uuid1 = r[1];
+               sb->set_uuid2 = r[2];
+               sb->set_uuid3 = r[3];
        }
        if (homehost && !uuid) {
                char buf[20];
@@ -792,7 +812,6 @@ struct devinfo {
        struct devinfo *next;
 };
 
-#ifndef MDASSEMBLE
 /* Add a device to the superblock being created */
 static int add_to_super0(struct supertype *st, mdu_disk_info_t *dinfo,
                         int fd, char *devname, unsigned long long data_offset)
@@ -805,9 +824,8 @@ static int add_to_super0(struct supertype *st, mdu_disk_info_t *dinfo,
        dk->major = dinfo->major;
        dk->minor = dinfo->minor;
        dk->raid_disk = dinfo->raid_disk;
-       dk->state = dinfo->state;
-       /* In case our source disk was writemostly, don't copy that bit */
-       dk->state &= ~(1<<MD_DISK_WRITEMOSTLY);
+       dk->state = dinfo->state & ((1<<MD_DISK_ACTIVE) |
+                                   (1<<MD_DISK_SYNC));
 
        sb->this_disk = sb->disks[dinfo->number];
        sb->sb_csum = calc_sb0_csum(sb);
@@ -824,7 +842,6 @@ static int add_to_super0(struct supertype *st, mdu_disk_info_t *dinfo,
 
        return 0;
 }
-#endif
 
 static int store_super0(struct supertype *st, int fd)
 {
@@ -878,7 +895,6 @@ static int store_super0(struct supertype *st, int fd)
        return 0;
 }
 
-#ifndef MDASSEMBLE
 static int write_init_super0(struct supertype *st)
 {
        mdp_super_t *sb = st->sb;
@@ -887,7 +903,7 @@ static int write_init_super0(struct supertype *st)
 
        for (di = st->info ; di && ! rv ; di = di->next) {
 
-               if (di->disk.state == 1)
+               if (di->disk.state & (1 << MD_DISK_FAULTY))
                        continue;
                if (di->fd == -1)
                        continue;
@@ -901,7 +917,7 @@ static int write_init_super0(struct supertype *st)
                rv = store_super0(st, di->fd);
 
                if (rv == 0 && (sb->state & (1<<MD_SB_BITMAP_PRESENT)))
-                       rv = st->ss->write_bitmap(st, di->fd);
+                       rv = st->ss->write_bitmap(st, di->fd, NoUpdate);
 
                if (rv)
                        pr_err("failed to write superblock to %s\n",
@@ -909,7 +925,6 @@ static int write_init_super0(struct supertype *st)
        }
        return rv;
 }
-#endif
 
 static int compare_super0(struct supertype *st, struct supertype *tst)
 {
@@ -930,7 +945,7 @@ static int compare_super0(struct supertype *st, struct supertype *tst)
                if (posix_memalign((void**)&first, 4096,
                             MD_SB_BYTES +
                             ROUND_UP(sizeof(struct bitmap_super_s), 4096)) != 0) {
-                       pr_err("%s could not allocate superblock\n", __func__);
+                       pr_err("could not allocate superblock\n");
                        return 1;
                }
                memcpy(first, second, MD_SB_BYTES + sizeof(struct bitmap_super_s));
@@ -998,7 +1013,7 @@ static int load_super0(struct supertype *st, int fd, char *devname)
        if (posix_memalign((void**)&super, 4096,
                           MD_SB_BYTES +
                           ROUND_UP(sizeof(bitmap_super_t), 4096)) != 0) {
-               pr_err("%s could not allocate superblock\n", __func__);
+               pr_err("could not allocate superblock\n");
                return 1;
        }
 
@@ -1044,8 +1059,8 @@ static int load_super0(struct supertype *st, int fd, char *devname)
         * valid.  If it doesn't clear the bit.  An --assemble --force
         * should get that written out.
         */
-       if (read(fd, super+1, ROUND_UP(sizeof(struct bitmap_super_s),4096))
-           != ROUND_UP(sizeof(struct bitmap_super_s),4096))
+       if (read(fd, super+1, ROUND_UP(sizeof(struct bitmap_super_s),4096)) !=
+           ROUND_UP(sizeof(struct bitmap_super_s), 4096))
                goto no_bitmap;
 
        uuid_from_super0(st, uuid);
@@ -1139,7 +1154,7 @@ static int add_internal_bitmap0(struct supertype *st, int *chunkp,
                if (chunk < 64*1024*1024)
                        chunk = 64*1024*1024;
        } else if ((unsigned long long)chunk < min_chunk)
-               return 0; /* chunk size too small */
+               return -EINVAL; /* chunk size too small */
 
        sb->state |= (1<<MD_SB_BITMAP_PRESENT);
 
@@ -1153,19 +1168,19 @@ static int add_internal_bitmap0(struct supertype *st, int *chunkp,
        bms->sync_size = __cpu_to_le64(size);
        bms->write_behind = __cpu_to_le32(write_behind);
        *chunkp = chunk;
-       return 1;
+       return 0;
 }
 
-static void locate_bitmap0(struct supertype *st, int fd)
+static int locate_bitmap0(struct supertype *st, int fd, int node_num)
 {
        unsigned long long dsize;
        unsigned long long offset;
 
        if (!get_dev_size(fd, NULL, &dsize))
-               return;
+               return -1;
 
        if (dsize < MD_RESERVED_SECTORS*512)
-               return;
+               return -1;
 
        offset = MD_NEW_SIZE_SECTORS(dsize>>9);
 
@@ -1174,9 +1189,10 @@ static void locate_bitmap0(struct supertype *st, int fd)
        offset += MD_SB_BYTES;
 
        lseek64(fd, offset, 0);
+       return 0;
 }
 
-static int write_bitmap0(struct supertype *st, int fd)
+static int write_bitmap0(struct supertype *st, int fd, enum bitmap_update update)
 {
        unsigned long long dsize;
        unsigned long long offset;
@@ -1239,13 +1255,12 @@ static void free_super0(struct supertype *st)
        st->sb = NULL;
 }
 
-#ifndef MDASSEMBLE
 static int validate_geometry0(struct supertype *st, int level,
                              int layout, int raiddisks,
                              int *chunk, unsigned long long size,
                              unsigned long long data_offset,
                              char *subdev, unsigned long long *freesize,
-                             int verbose)
+                             int consistency_policy, int verbose)
 {
        unsigned long long ldsize;
        int fd;
@@ -1270,13 +1285,18 @@ static int validate_geometry0(struct supertype *st, int level,
        }
        if (size >= tbmax * 2ULL*1024*1024*1024) {
                if (verbose)
-                       pr_err("0.90 metadata supports at most "
-                              "%d terabytes per device\n", tbmax);
+                       pr_err("0.90 metadata supports at most %d terabytes per device\n", tbmax);
                return 0;
        }
        if (*chunk == UnSet)
                *chunk = DEFAULT_CHUNK;
 
+       if (level == 0 && layout != UnSet) {
+               if (verbose)
+                       pr_err("0.90 metadata does not support layouts for RAID0\n");
+               return 0;
+       }
+
        if (!subdev)
                return 1;
 
@@ -1299,10 +1319,8 @@ static int validate_geometry0(struct supertype *st, int level,
        *freesize = MD_NEW_SIZE_SECTORS(ldsize >> 9);
        return 1;
 }
-#endif /* MDASSEMBLE */
 
 struct superswitch super0 = {
-#ifndef MDASSEMBLE
        .examine_super = examine_super0,
        .brief_examine_super = brief_examine_super0,
        .export_examine_super = export_examine_super0,
@@ -1312,7 +1330,6 @@ struct superswitch super0 = {
        .validate_geometry = validate_geometry0,
        .add_to_super = add_to_super0,
        .copy_metadata = copy_metadata0,
-#endif
        .match_home = match_home0,
        .uuid_from_super = uuid_from_super0,
        .getinfo_super = getinfo_super0,