]> git.ipfire.org Git - thirdparty/mdadm.git/blobdiff - mdadm.c
imsm: add support for checkpointing via 'curr_migr_unit'
[thirdparty/mdadm.git] / mdadm.c
diff --git a/mdadm.c b/mdadm.c
index e0f51b7fc87469365935fe7e6a95bb1de451fe86..74a39a88584b9bf4ac9885d9b21f534f8d8175f2 100644 (file)
--- a/mdadm.c
+++ b/mdadm.c
@@ -1,7 +1,7 @@
 /*
  * mdadm - manage Linux "md" devices aka RAID arrays.
  *
- * Copyright (C) 2001-2006 Neil Brown <neilb@suse.de>
+ * Copyright (C) 2001-2009 Neil Brown <neilb@suse.de>
  *
  *
  *    This program is free software; you can redistribute it and/or modify
  *    Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  *
  *    Author: Neil Brown
- *    Email: <neilb@cse.unsw.edu.au>
- *    Paper: Neil Brown
- *           School of Computer Science and Engineering
- *           The University of New South Wales
- *           Sydney, 2052
- *           Australia
+ *    Email: <neilb@suse.de>
  *
  *    Additions for bitmap and write-behind RAID options, Copyright (C) 2003-2004,
  *    Paul Clements, SteelEye Technology, Inc.
@@ -46,8 +41,10 @@ int main(int argc, char *argv[])
 
        int chunk = 0;
        long long size = -1;
+       long long array_size = -1;
        int level = UnSet;
        int layout = UnSet;
+       char *layout_str = NULL;
        int raiddisks = 0;
        int max_disks = MD_SB_DISKS; /* just a default */
        int sparedisks = 0;
@@ -91,8 +88,10 @@ int main(int argc, char *argv[])
 
        char *homehost = NULL;
        char sys_hostname[256];
+       int require_homehost = 1;
        char *mailaddr = NULL;
        char *program = NULL;
+       int increments = 20;
        int delay = 0;
        int daemonise = 0;
        char *pidfile = NULL;
@@ -105,7 +104,6 @@ int main(int argc, char *argv[])
        int rebuild_map = 0;
        int auto_update_home = 0;
 
-       int copies;
        int print_help = 0;
        FILE *outf;
 
@@ -132,7 +130,7 @@ int main(int argc, char *argv[])
                                shortopt, long_options,
                                &option_index)) != -1) {
                int newmode = mode;
-               /* firstly, some mode-independant options */
+               /* firstly, some mode-independent options */
                switch(opt) {
                case 'h':
                        if (option_index > 0 &&
@@ -166,7 +164,10 @@ int main(int argc, char *argv[])
                        continue;
 
                case HomeHost:
-                       homehost = optarg;
+                       if (strcasecmp(optarg, "<ignore>") == 0)
+                               require_homehost = 0;
+                       else
+                               homehost = optarg;
                        continue;
 
                case ':':
@@ -217,6 +218,7 @@ int main(int argc, char *argv[])
                case 'w':
                case 'W':
                case Waitclean:
+               case DetailPlatform:
                case 'K': if (!mode) newmode = MISC; break;
                }
                if (mode && newmode == mode) {
@@ -323,6 +325,7 @@ int main(int argc, char *argv[])
                 * could depend on the mode */
 #define O(a,b) ((a<<8)|b)
                switch (O(mode,opt)) {
+               case O(GROW,'c'):
                case O(CREATE,'c'):
                case O(BUILD,'c'): /* chunk or rounding */
                        if (chunk) {
@@ -338,9 +341,11 @@ int main(int argc, char *argv[])
                        }
                        continue;
 
+#if 0
                case O(ASSEMBLE,AutoHomeHost):
                        auto_update_home = 1;
                        continue;
+#endif
                case O(INCREMENTAL, 'e'):
                case O(CREATE,'e'):
                case O(ASSEMBLE,'e'):
@@ -373,7 +378,8 @@ int main(int argc, char *argv[])
 
 
                case O(GROW,'z'):
-               case O(CREATE,'z'): /* size */
+               case O(CREATE,'z'):
+               case O(BUILD,'z'): /* size */
                        if (size >= 0) {
                                fprintf(stderr, Name ": size may only be specified once. "
                                        "Second value is %s.\n", optarg);
@@ -382,16 +388,36 @@ int main(int argc, char *argv[])
                        if (strcmp(optarg, "max")==0)
                                size = 0;
                        else {
-                               size = strtoll(optarg, &c, 10);
-                               if (!optarg[0] || *c || size < 4) {
+                               size = parse_size(optarg);
+                               if (size < 8) {
                                        fprintf(stderr, Name ": invalid size: %s\n",
                                                optarg);
                                        exit(2);
                                }
+                               /* convert sectors to K */
+                               size /= 2;
                        }
                        continue;
 
-               case O(GROW,'l'): /* hack - needed to understand layout */
+               case O(GROW,'Z'): /* array size */
+                       if (array_size >= 0) {
+                               fprintf(stderr, Name ": array-size may only be specified once. "
+                                       "Second value is %s.\n", optarg);
+                               exit(2);
+                       }
+                       if (strcmp(optarg, "max") == 0)
+                               array_size = 0;
+                       else {
+                               array_size = parse_size(optarg);
+                               if (array_size <= 0) {
+                                       fprintf(stderr, Name ": invalid array size: %s\n",
+                                               optarg);
+                                       exit(2);
+                               }
+                       }
+                       continue;
+
+               case O(GROW,'l'):
                case O(CREATE,'l'):
                case O(BUILD,'l'): /* set raid level*/
                        if (level != UnSet) {
@@ -405,7 +431,10 @@ int main(int argc, char *argv[])
                                        optarg);
                                exit(2);
                        }
-                       if (level != 0 && level != -1 && level != 1 && level != -4 && level != -5 && mode == BUILD) {
+                       if (level != 0 && level != LEVEL_LINEAR && level != 1 &&
+                           level != LEVEL_MULTIPATH && level != LEVEL_FAULTY &&
+                           level != 10 &&
+                           mode == BUILD) {
                                fprintf(stderr, Name ": Raid level %s not permitted with --build.\n",
                                        optarg);
                                exit(2);
@@ -418,9 +447,18 @@ int main(int argc, char *argv[])
                        ident.level = level;
                        continue;
 
+               case O(GROW, 'p'): /* new layout */
+                       if (layout_str) {
+                               fprintf(stderr,Name ": layout may only be sent once.  "
+                                       "Second value was %s\n", optarg);
+                               exit(2);
+                       }
+                       layout_str = optarg;
+                       /* 'Grow' will parse the value */
+                       continue;
+
                case O(CREATE,'p'): /* raid5 layout */
                case O(BUILD,'p'): /* faulty layout */
-               case O(GROW, 'p'): /* faulty reconfig */
                        if (layout != UnSet) {
                                fprintf(stderr,Name ": layout may only be sent once.  "
                                        "Second value was %s\n", optarg);
@@ -436,7 +474,6 @@ int main(int argc, char *argv[])
                                exit(2);
 
                        case 5:
-                       case 6:
                                layout = map_name(r5layout, optarg);
                                if (layout==UnSet) {
                                        fprintf(stderr, Name ": layout %s not understood for raid5.\n",
@@ -444,40 +481,33 @@ int main(int argc, char *argv[])
                                        exit(2);
                                }
                                break;
+                       case 6:
+                               layout = map_name(r6layout, optarg);
+                               if (layout==UnSet) {
+                                       fprintf(stderr, Name ": layout %s not understood for raid6.\n",
+                                               optarg);
+                                       exit(2);
+                               }
+                               break;
 
                        case 10:
-                               /* 'f', 'o' or 'n' followed by a number <= raid_disks */
-                               if ((optarg[0] !=  'n' && optarg[0] != 'f' && optarg[0] != 'o') ||
-                                   (copies = strtoul(optarg+1, &cp, 10)) < 1 ||
-                                   copies > 200 ||
-                                   *cp) {
+                               layout = parse_layout_10(optarg);
+                               if (layout < 0) {
                                        fprintf(stderr, Name ": layout for raid10 must be 'nNN', 'oNN' or 'fNN' where NN is a number, not %s\n", optarg);
                                        exit(2);
                                }
-                               if (optarg[0] == 'n')
-                                       layout = 256 + copies;
-                               else if (optarg[0] == 'o')
-                                       layout = 0x10000 + (copies<<8) + 1;
-                               else
-                                       layout = 1 + (copies<<8);
                                break;
-                       case -5: /* Faulty
-                                 * modeNNN
-                                 */
-
-                       {
-                               int ln = strcspn(optarg, "0123456789");
-                               char *m = strdup(optarg);
-                               int mode;
-                               m[ln] = 0;
-                               mode = map_name(faultylayout, m);
-                               if (mode == UnSet) {
+                       case LEVEL_FAULTY:
+                               /* Faulty
+                                * modeNNN
+                                */
+                               layout = parse_layout_faulty(optarg);
+                               if (layout == -1) {
                                        fprintf(stderr, Name ": layout %s not understood for faulty.\n",
                                                optarg);
                                        exit(2);
                                }
-                               layout = mode | (atoi(optarg+ln)<< ModeShift);
-                       }
+                               break;
                        }
                        continue;
 
@@ -643,6 +673,7 @@ int main(int argc, char *argv[])
                "     'summaries', 'homehost', 'byteorder', 'devicesize'.\n");
                        exit(outf == stdout ? 0 : 2);
 
+               case O(INCREMENTAL,NoDegraded):
                case O(ASSEMBLE,NoDegraded): /* --no-degraded */
                        runstop = -1; /* --stop isn't allowed for --assemble,
                                       * so we overload slightly */
@@ -684,6 +715,14 @@ int main(int argc, char *argv[])
                                program = optarg;
                        continue;
 
+               case O(MONITOR,'r'): /* rebuild increments */
+                       increments = atoi(optarg);
+                       if (increments>99 || increments<1) {
+                               fprintf(stderr, Name ": please specify positive integer between 1 and 99 as rebuild increments.\n");
+                               exit(2);
+                       }
+                       continue;
+
                case O(MONITOR,'d'): /* delay in seconds */
                case O(GROW, 'd'):
                case O(BUILD,'d'): /* delay for bitmap updates */
@@ -769,6 +808,7 @@ int main(int argc, char *argv[])
                case O(MISC,'w'):
                case O(MISC,'W'):
                case O(MISC, Waitclean):
+               case O(MISC, DetailPlatform):
                        if (devmode && devmode != opt &&
                            (devmode == 'E' || (opt == 'E' && devmode != 'Q'))) {
                                fprintf(stderr, Name ": --examine/-E cannot be given with -%c\n",
@@ -999,7 +1039,7 @@ int main(int argc, char *argv[])
        }
 
        if (homehost == NULL)
-               homehost = conf_get_homehost();
+               homehost = conf_get_homehost(&require_homehost);
        if (homehost == NULL || strcmp(homehost, "<system>")==0) {
                if (gethostname(sys_hostname, sizeof(sys_hostname)) == 0) {
                        sys_hostname[sizeof(sys_hostname)-1] = 0;
@@ -1039,12 +1079,16 @@ int main(int argc, char *argv[])
                                        array_ident->autof = autof;
                                rv |= Assemble(ss, devlist->devname, array_ident,
                                               NULL, backup_file,
-                                              readonly, runstop, update, homehost, verbose-quiet, force);
+                                              readonly, runstop, update,
+                                              homehost, require_homehost,
+                                              verbose-quiet, force);
                        }
                } else if (!scan)
                        rv = Assemble(ss, devlist->devname, &ident,
                                      devlist->next, backup_file,
-                                     readonly, runstop, update, homehost, verbose-quiet, force);
+                                     readonly, runstop, update,
+                                     homehost, require_homehost,
+                                     verbose-quiet, force);
                else if (devs_found>0) {
                        if (update && devs_found > 1) {
                                fprintf(stderr, Name ": can only update a single array at a time\n");
@@ -1066,7 +1110,9 @@ int main(int argc, char *argv[])
                                        array_ident->autof = autof;
                                rv |= Assemble(ss, dv->devname, array_ident,
                                               NULL, backup_file,
-                                              readonly, runstop, update, homehost, verbose-quiet, force);
+                                              readonly, runstop, update,
+                                              homehost, require_homehost,
+                                              verbose-quiet, force);
                        }
                } else {
                        mddev_ident_t array_list =  conf_get_ident(NULL);
@@ -1085,13 +1131,18 @@ int main(int argc, char *argv[])
                                exit(1);
                        }
                        for (; array_list; array_list = array_list->next) {
+                               if (array_list->devname &&
+                                   strcasecmp(array_list->devname, "<ignore>") == 0)
+                                       continue;
                                if (array_list->autof == 0)
                                        array_list->autof = autof;
                                
                                rv |= Assemble(ss, array_list->devname,
                                               array_list,
                                               NULL, NULL,
-                                              readonly, runstop, NULL, homehost, verbose-quiet, force);
+                                              readonly, runstop, NULL,
+                                              homehost, require_homehost,
+                                              verbose-quiet, force);
                                cnt++;
                        }
                        if (homehost && cnt == 0) {
@@ -1109,7 +1160,9 @@ int main(int argc, char *argv[])
                                                rv2 = Assemble(ss, NULL,
                                                               &ident,
                                                               devlist, NULL,
-                                                              readonly, runstop, NULL, homehost, verbose-quiet, force);
+                                                              readonly, runstop, NULL,
+                                                              homehost, require_homehost,
+                                                              verbose-quiet, force);
                                                if (rv2==0) {
                                                        cnt++;
                                                        acnt++;
@@ -1122,6 +1175,7 @@ int main(int argc, char *argv[])
                                        } while (rv2!=2);
                                        /* Incase there are stacked devices, we need to go around again */
                                } while (acnt);
+#if 0
                                if (cnt == 0 && auto_update_home && homehost) {
                                        /* Nothing found, maybe we need to bootstrap homehost info */
                                        do {
@@ -1130,7 +1184,9 @@ int main(int argc, char *argv[])
                                                        rv2 = Assemble(ss, NULL,
                                                                       &ident,
                                                                       NULL, NULL,
-                                                                      readonly, runstop, "homehost", homehost, verbose-quiet, force);
+                                                                      readonly, runstop, "homehost",
+                                                                      homehost, require_homehost,
+                                                                      verbose-quiet, force);
                                                        if (rv2==0) {
                                                                cnt++;
                                                                acnt++;
@@ -1139,6 +1195,7 @@ int main(int argc, char *argv[])
                                                /* Incase there are stacked devices, we need to go around again */
                                        } while (acnt);
                                }
+#endif
                                if (cnt == 0 && rv == 0) {
                                        fprintf(stderr, Name ": No arrays found in config file or automatically\n");
                                        rv = 1;
@@ -1173,7 +1230,7 @@ int main(int argc, char *argv[])
                rv = Build(devlist->devname, chunk, level, layout,
                           raiddisks, devlist->next, assume_clean,
                           bitmap_file, bitmap_chunk, write_behind,
-                          delay, verbose-quiet, autof);
+                          delay, verbose-quiet, autof, size);
                break;
        case CREATE:
                if (delay == 0) delay = DEFAULT_BITMAP_DELAY;
@@ -1211,24 +1268,51 @@ int main(int argc, char *argv[])
                        rv = Examine(devlist, scan?(verbose>1?0:verbose+1):brief,
                                     export, scan,
                                     SparcAdjust, ss, homehost);
+               } else if (devmode == DetailPlatform) {
+                       rv = Detail_Platform(ss ? ss->ss : NULL, ss ? scan : 1, verbose);
                } else {
                        if (devlist == NULL) {
-                               if (devmode=='D' && scan) {
-                                       /* apply --detail to all devices in /proc/mdstat */
+                               if ((devmode=='D' || devmode == Waitclean) && scan) {
+                                       /* apply --detail or --wait-clean to
+                                        * all devices in /proc/mdstat
+                                        */
                                        struct mdstat_ent *ms = mdstat_read(0, 1);
                                        struct mdstat_ent *e;
+                                       struct map_ent *map = NULL;
+                                       int members;
+                                       int v = verbose>1?0:verbose+1;
+
+                                       for (members = 0; members <= 1; members++) {
                                        for (e=ms ; e ; e=e->next) {
-                                               char *name = get_md_name(e->devnum);
+                                               char *name;
+                                               struct map_ent *me;
+                                               int member = e->metadata_version &&
+                                                       strncmp(e->metadata_version,
+                                                               "external:/", 10) == 0;
+                                               if (members != member)
+                                                       continue;
+                                               me = map_by_devnum(&map, e->devnum);
+                                               if (me && me->path
+                                                   && strcmp(me->path, "/unknown") != 0)
+                                                       name = me->path;
+                                               else
+                                                       name = get_md_name(e->devnum);
 
                                                if (!name) {
                                                        fprintf(stderr, Name ": cannot find device file for %s\n",
                                                                e->dev);
                                                        continue;
                                                }
-                                               rv |= Detail(name, verbose>1?0:verbose+1,
-                                                            export, test, homehost);
+                                               if (devmode == 'D')
+                                                       rv |= Detail(name, v,
+                                                                    export, test,
+                                                                    homehost);
+                                               else
+                                                       rv |= WaitClean(name, -1, v);
                                                put_md_name(name);
                                        }
+                                       }
+                                       free_mdstat(ms);
                                } else  if (devmode == 'S' && scan) {
                                        /* apply --stop to all devices in /proc/mdstat */
                                        /* Due to possible stacking of devices, repeat until
@@ -1261,6 +1345,7 @@ int main(int argc, char *argv[])
 
                                                        put_md_name(name);
                                                }
+                                               free_mdstat(ms);
                                        } while (!last && err);
                                        if (err) rv |= 1;
                                } else {
@@ -1276,7 +1361,16 @@ int main(int argc, char *argv[])
                                                     export, test, homehost);
                                        continue;
                                case 'K': /* Zero superblock */
-                                       rv |= Kill(dv->devname, force, quiet,0);
+                                       if (ss)
+                                               rv |= Kill(dv->devname, ss, force, quiet,0);
+                                       else {
+                                               int q = quiet;
+                                               do {
+                                                       rv |= Kill(dv->devname, NULL, force, q, 0);
+                                                       q = 1;
+                                               } while ((rv & 2) == 0);
+                                               rv &= ~2;
+                                       }
                                        continue;
                                case 'Q':
                                        rv |= Query(dv->devname); continue;
@@ -1285,7 +1379,7 @@ int main(int argc, char *argv[])
                                case 'W':
                                        rv |= Wait(dv->devname); continue;
                                case Waitclean:
-                                       rv |= WaitClean(dv->devname, verbose-quiet); continue;
+                                       rv |= WaitClean(dv->devname, -1, verbose-quiet); continue;
                                }
                                mdfd = open_mddev(dv->devname, 1);
                                if (mdfd>=0) {
@@ -1316,17 +1410,55 @@ int main(int argc, char *argv[])
                        rv = 1;
                        break;
                }
+               if (delay == 0) {
+                       if (get_linux_version() > 20616)
+                               /* mdstat responds to poll */
+                               delay = 1000;
+                       else
+                               delay = 60;
+               }
                rv= Monitor(devlist, mailaddr, program,
                            delay?delay:60, daemonise, scan, oneshot,
-                           dosyslog, test, pidfile);
+                           dosyslog, test, pidfile, increments);
                break;
 
        case GROW:
+               if (array_size >= 0) {
+                       /* alway impose array size first, independent of
+                        * anything else
+                        * Do not allow level or raid_disks changes at the
+                        * same time as that can be irreversibly destructive.
+                        */
+                       struct mdinfo sra;
+                       int err;
+                       if (raiddisks || level != UnSet) {
+                               fprintf(stderr, Name ": cannot change array size in same operation "
+                                       "as changing raiddisks or level.\n"
+                                       "    Change size first, then check that data is still intact.\n");
+                               rv = 1;
+                               break;
+                       }
+                       sysfs_init(&sra, mdfd, 0);
+                       if (array_size == 0)
+                               err = sysfs_set_str(&sra, NULL, "array_size", "default");
+                       else
+                               err = sysfs_set_num(&sra, NULL, "array_size", array_size / 2);
+                       if (err < 0) {
+                               if (errno == E2BIG)
+                                       fprintf(stderr, Name ": --array-size setting"
+                                               " is too large.\n");
+                               else
+                                       fprintf(stderr, Name ": current kernel does"
+                                               " not support setting --array-size\n");
+                               rv = 1;
+                               break;
+                       }
+               }
                if (devs_found > 1) {
 
                        /* must be '-a'. */
-                       if (size >= 0 || raiddisks) {
-                               fprintf(stderr, Name ": --size, --raiddisks, and --add are exclusing in --grow mode\n");
+                       if (size >= 0 || raiddisks || chunk || layout_str != NULL || bitmap_file) {
+                               fprintf(stderr, Name ": --add cannot be used with other geometry changes in --grow mode\n");
                                rv = 1;
                                break;
                        }
@@ -1335,20 +1467,21 @@ int main(int argc, char *argv[])
                                if (rv)
                                        break;
                        }
-               } else if ((size >= 0) + (raiddisks != 0) +  (layout != UnSet) + (bitmap_file != NULL)> 1) {
-                       fprintf(stderr, Name ": can change at most one of size, raiddisks, bitmap, and layout\n");
-                       rv = 1;
-                       break;
-               } else if (layout != UnSet)
-                       rv = Manage_reconfig(devlist->devname, mdfd, layout);
-               else if (size >= 0 || raiddisks)
-                       rv = Grow_reshape(devlist->devname, mdfd, quiet, backup_file,
-                                         size, level, layout, chunk, raiddisks);
-               else if (bitmap_file) {
-                       if (delay == 0) delay = DEFAULT_BITMAP_DELAY;
+               } else if (bitmap_file) {
+                       if (size >= 0 || raiddisks || chunk || layout_str != NULL) {
+                               fprintf(stderr, Name ": --bitmap changes cannot be used with other geometry changes in --grow mode\n");
+                               rv = 1;
+                               break;
+                       }
+                       if (delay == 0)
+                               delay = DEFAULT_BITMAP_DELAY;
                        rv = Grow_addbitmap(devlist->devname, mdfd, bitmap_file,
                                            bitmap_chunk, delay, write_behind, force);
-               } else
+               } else if (size >= 0 || raiddisks != 0 || layout_str != NULL
+                          || chunk != 0 || level != UnSet) {
+                       rv = Grow_reshape(devlist->devname, mdfd, quiet, backup_file,
+                                         size, level, layout_str, chunk, raiddisks);
+               } else if (array_size < 0)
                        fprintf(stderr, Name ": no changes to --grow\n");
                break;
        case INCREMENTAL:
@@ -1378,7 +1511,7 @@ int main(int argc, char *argv[])
                        break;
                }
                rv = Incremental(devlist->devname, verbose-quiet, runstop,
-                                ss, homehost, autof);
+                                ss, homehost, require_homehost, autof);
                break;
        case AUTODETECT:
                autodetect();