]> git.ipfire.org Git - thirdparty/mdadm.git/blobdiff - Grow.c
Add mdstat_by_component
[thirdparty/mdadm.git] / Grow.c
diff --git a/Grow.c b/Grow.c
index d8d91cbca2aec9aae9f3fddae019e7310d0cbf90..28ed8d7390cdf3267ace5795ea96c8d99c384cb2 100644 (file)
--- a/Grow.c
+++ b/Grow.c
@@ -288,6 +288,11 @@ int Grow_addbitmap(char *devname, int fd, char *file, int chunk, int delay, int
                return 1;
        } else if (strcmp(file, "internal") == 0) {
                int d;
+               if (st->ss->add_internal_bitmap == NULL) {
+                       fprintf(stderr, Name ": Internal bitmaps not supported "
+                               "with %s metadata\n", st->ss->name);
+                       return 1;
+               }
                for (d=0; d< st->max_devs; d++) {
                        mdu_disk_info_t disk;
                        char *dv;
@@ -563,13 +568,17 @@ int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
                } else
                        rv = ioctl(fd, SET_ARRAY_INFO, &array);
                if (rv != 0) {
+                       int err = errno;
                        fprintf(stderr, Name ": Cannot set device size for %s: %s\n",
-                               devname, strerror(errno));
+                               devname, strerror(err));
+                       if (err == EBUSY && 
+                           (array.state & (1<<MD_SB_BITMAP_PRESENT)))
+                               fprintf(stderr, "       Bitmap must be removed before size can be changed\n");
                        rv = 1;
                        goto release;
                }
                ioctl(fd, GET_ARRAY_INFO, &array);
-               size = get_component_size(fd);
+               size = get_component_size(fd)/2;
                if (size == 0)
                        size = array.size;
                if (!quiet)
@@ -577,7 +586,7 @@ int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
                                devname, size);
                changed = 1;
        } else {
-               size = get_component_size(fd);
+               size = get_component_size(fd)/2;
                if (size == 0)
                        size = array.size;
        }
@@ -666,8 +675,12 @@ int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
                        }
                        err = sysfs_set_str(sra, NULL, "level", c);
                        if (err) {
+                               err = errno;
                                fprintf(stderr, Name ": %s: could not set level to %s\n",
                                        devname, c);
+                               if (err == EBUSY && 
+                                   (array.state & (1<<MD_SB_BITMAP_PRESENT)))
+                                       fprintf(stderr, "       Bitmap must be removed before level can be changed\n");
                                rv = 1;
                                goto release;
                        }
@@ -733,9 +746,14 @@ int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
                        c = map_num(pers, level);
                        if (c) {
                                rv = sysfs_set_str(sra, NULL, "level", c);
-                               if (rv)
+                               if (rv) {
+                                       int err = errno;
                                        fprintf(stderr, Name ": %s: could not set level to %s\n",
                                                devname, c);
+                                       if (err == EBUSY && 
+                                           (array.state & (1<<MD_SB_BITMAP_PRESENT)))
+                                               fprintf(stderr, "       Bitmap must be removed before level can be changed\n");
+                               }
                        }
                } else if (!changed && !quiet)
                        fprintf(stderr, Name ": %s: no change requested\n",
@@ -913,8 +931,8 @@ int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
                 * old stripes and a whole number of new stripes.
                 * So LCM for (chunksize*datadisks).
                 */
-               a = ochunk/512 * odata;
-               b = nchunk/512 * ndata;
+               a = (ochunk/512) * odata;
+               b = (nchunk/512) * ndata;
                /* Find GCD */
                while (a != b) {
                        if (a < b)
@@ -923,13 +941,20 @@ int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
                                a -= b;
                }
                /* LCM == product / GCD */
-               blocks = ochunk/512 * nchunk/512 * odata * ndata / a;
+               blocks = (ochunk/512) * (nchunk/512) * odata * ndata / a;
 
                sysfs_free(sra);
                sra = sysfs_read(fd, 0,
                                 GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|
                                 GET_CACHE);
 
+               if (!sra) {
+                       fprintf(stderr, Name ": %s: Cannot get array details from sysfs\n",
+                               devname);
+                       rv = 1;
+                       break;
+               }
+
                if (ndata == odata) {
                        /* Make 'blocks' bigger for better throughput, but
                         * not so big that we reject it below.
@@ -942,13 +967,6 @@ int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
                        fprintf(stderr, Name ": Need to backup %luK of critical "
                                "section..\n", blocks/2);
 
-               if (!sra) {
-                       fprintf(stderr, Name ": %s: Cannot get array details from sysfs\n",
-                               devname);
-                       rv = 1;
-                       break;
-               }
-
                if (blocks >= sra->component_size/2) {
                        fprintf(stderr, Name ": %s: Something wrong - reshape aborted\n",
                                devname);
@@ -1066,12 +1084,16 @@ int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
                if (ochunk == nchunk && olayout == nlayout) {
                        array.raid_disks = ndisks;
                        if (ioctl(fd, SET_ARRAY_INFO, &array) != 0) {
+                               int err = errno;
                                rv = 1;
                                fprintf(stderr, Name ": Cannot set device shape for %s: %s\n",
                                        devname, strerror(errno));
                                if (ndisks < odisks &&
                                    get_linux_version() < 2006030)
                                        fprintf(stderr, Name ": linux 2.6.30 or later required\n");
+                               if (err == EBUSY && 
+                                   (array.state & (1<<MD_SB_BITMAP_PRESENT)))
+                                       fprintf(stderr, "       Bitmap must be removed before shape can be changed\n");
 
                                break;
                        }
@@ -1079,17 +1101,24 @@ int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
                        /* set them all just in case some old 'new_*' value
                         * persists from some earlier problem
                         */
+                       int err = err; /* only used if rv==1, and always set if
+                                       * rv==1, so initialisation not needed,
+                                       * despite gcc warning
+                                       */
                        if (sysfs_set_num(sra, NULL, "chunk_size", nchunk) < 0)
-                               rv = 1;
-                       if (sysfs_set_num(sra, NULL, "layout", nlayout) < 0)
-                               rv = 1;
-                       if (sysfs_set_num(sra, NULL, "raid_disks", ndisks) < 0)
-                               rv = 1;
+                               rv = 1, err = errno;
+                       if (!rv && sysfs_set_num(sra, NULL, "layout", nlayout) < 0)
+                               rv = 1, err = errno;
+                       if (!rv && sysfs_set_num(sra, NULL, "raid_disks", ndisks) < 0)
+                               rv = 1, err = errno;
                        if (rv) {
                                fprintf(stderr, Name ": Cannot set device shape for %s\n",
                                        devname);
                                if (get_linux_version() < 2006030)
                                        fprintf(stderr, Name ": linux 2.6.30 or later required\n");
+                               if (err == EBUSY && 
+                                   (array.state & (1<<MD_SB_BITMAP_PRESENT)))
+                                       fprintf(stderr, "       Bitmap must be removed before shape can be changed\n");
                                break;
                        }
                }
@@ -1230,6 +1259,7 @@ int Grow_reshape(char *devname, int fd, int quiet, char *backup_file,
  * 
  */
 
+/* FIXME return status is never checked */
 int grow_backup(struct mdinfo *sra,
                unsigned long long offset, /* per device */
                unsigned long stripes, /* per device */
@@ -1254,7 +1284,7 @@ int grow_backup(struct mdinfo *sra,
                odata--;
        if (level == 6)
                odata--;
-       sysfs_set_num(sra, NULL, "suspend_hi", (offset + stripes * chunk/512) * odata);
+       sysfs_set_num(sra, NULL, "suspend_hi", (offset + stripes * (chunk/512)) * odata);
        /* Check that array hasn't become degraded, else we might backup the wrong data */
        sysfs_get_ll(sra, NULL, "degraded", &new_degraded);
        if (new_degraded != *degraded) {
@@ -1282,10 +1312,10 @@ int grow_backup(struct mdinfo *sra,
        }
        if (part) {
                bsb.arraystart2 = __cpu_to_le64(offset * odata);
-               bsb.length2 = __cpu_to_le64(stripes * chunk/512 * odata);
+               bsb.length2 = __cpu_to_le64(stripes * (chunk/512) * odata);
        } else {
                bsb.arraystart = __cpu_to_le64(offset * odata);
-               bsb.length = __cpu_to_le64(stripes * chunk/512 * odata);
+               bsb.length = __cpu_to_le64(stripes * (chunk/512) * odata);
        }
        if (part)
                bsb.magic[15] = '2';
@@ -1312,16 +1342,19 @@ int grow_backup(struct mdinfo *sra,
                        bsb.sb_csum2 = bsb_csum((char*)&bsb,
                                                ((char*)&bsb.sb_csum2)-((char*)&bsb));
 
-               lseek64(destfd[i], destoffsets[i] - 4096, 0);
-               write(destfd[i], &bsb, 512);
+               if (lseek64(destfd[i], destoffsets[i] - 4096, 0) != destoffsets[i] - 4096)
+                       rv = 1;
+               rv = rv ?: write(destfd[i], &bsb, 512);
                if (destoffsets[i] > 4096) {
-                       lseek64(destfd[i], destoffsets[i]+stripes*chunk*odata, 0);
-                       write(destfd[i], &bsb, 512);
+                       if (lseek64(destfd[i], destoffsets[i]+stripes*chunk*odata, 0) !=
+                           destoffsets[i]+stripes*chunk*odata)
+                               rv = 1;
+                       rv = rv ?: write(destfd[i], &bsb, 512);
                }
                fsync(destfd[i]);
        }
 
-       return 0;
+       return rv;
 }
 
 /* in 2.6.30, the value reported by sync_completed can be
@@ -1334,6 +1367,7 @@ int grow_backup(struct mdinfo *sra,
  * The various caller give appropriate values so that
  * every works.
  */
+/* FIXME return value is often ignored */
 int wait_backup(struct mdinfo *sra,
                unsigned long long offset, /* per device */
                unsigned long long blocks, /* per device */
@@ -1347,6 +1381,7 @@ int wait_backup(struct mdinfo *sra,
        int fd = sysfs_get_fd(sra, NULL, "sync_completed");
        unsigned long long completed;
        int i;
+       int rv;
 
        if (fd < 0)
                return -1;
@@ -1378,24 +1413,28 @@ int wait_backup(struct mdinfo *sra,
                bsb.length = __cpu_to_le64(0);
        }
        bsb.mtime = __cpu_to_le64(time(0));
+       rv = 0;
        for (i = 0; i < dests; i++) {
                bsb.devstart = __cpu_to_le64(destoffsets[i]/512);
                bsb.sb_csum = bsb_csum((char*)&bsb, ((char*)&bsb.sb_csum)-((char*)&bsb));
                if (memcmp(bsb.magic, "md_backup_data-2", 16) == 0)
                        bsb.sb_csum2 = bsb_csum((char*)&bsb,
                                                ((char*)&bsb.sb_csum2)-((char*)&bsb));
-               lseek64(destfd[i], destoffsets[i]-4096, 0);
-               write(destfd[i], &bsb, 512);
+               if (lseek64(destfd[i], destoffsets[i]-4096, 0) !=
+                   destoffsets[i]-4096)
+                       rv = 1;
+               rv = rv ?: write(destfd[i], &bsb, 512);
                fsync(destfd[i]);
        }
-       return 0;
+       return rv;
 }
 
 static void fail(char *msg)
 {
-       write(2, msg, strlen(msg));
-       write(2, "\n", 1);
-       exit(1);
+       int rv;
+       rv = write(2, msg, strlen(msg));
+       rv |= write(2, "\n", 1);
+       exit(rv ? 1 : 2);
 }
 
 static char *abuf, *bbuf;
@@ -1431,27 +1470,33 @@ static void validate(int afd, int bfd, unsigned long long offset)
                        free(abuf);
                        free(bbuf);
                        abuflen = len;
-                       posix_memalign((void**)&abuf, 4096, abuflen);
-                       posix_memalign((void**)&bbuf, 4096, abuflen);
+                       if (posix_memalign((void**)&abuf, 4096, abuflen) ||
+                           posix_memalign((void**)&bbuf, 4096, abuflen)) {
+                               abuflen = 0;
+                               /* just stop validating on mem-alloc failure */
+                               return;
+                       }
                }
 
                lseek64(bfd, offset, 0);
                if (read(bfd, bbuf, len) != len) {
-                       printf("len %llu\n", len);
+                       //printf("len %llu\n", len);
                        fail("read first backup failed");
                }
                lseek64(afd, __le64_to_cpu(bsb2.arraystart)*512, 0);
                if (read(afd, abuf, len) != len)
                        fail("read first from array failed");
                if (memcmp(bbuf, abuf, len) != 0) {
+                       #if 0
                        int i;
                        printf("offset=%llu len=%llu\n",
-                              __le64_to_cpu(bsb2.arraystart)*512, len);
+                              (unsigned long long)__le64_to_cpu(bsb2.arraystart)*512, len);
                        for (i=0; i<len; i++)
                                if (bbuf[i] != abuf[i]) {
                                        printf("first diff byte %d\n", i);
                                        break;
                                }
+                       #endif
                        fail("data1 compare failed");
                }
        }
@@ -1485,7 +1530,9 @@ static int child_grow(int afd, struct mdinfo *sra, unsigned long stripes,
        char *buf;
        int degraded = 0;
 
-       posix_memalign((void**)&buf, 4096, disks * chunk);
+       if (posix_memalign((void**)&buf, 4096, disks * chunk))
+               /* Don't start the 'reshape' */
+               return 0;
        sysfs_set_num(sra, NULL, "suspend_hi", 0);
        sysfs_set_num(sra, NULL, "suspend_lo", 0);
        grow_backup(sra, 0, stripes,
@@ -1493,10 +1540,10 @@ static int child_grow(int afd, struct mdinfo *sra, unsigned long stripes,
                    dests, destfd, destoffsets,
                    0, &degraded, buf);
        validate(afd, destfd[0], destoffsets[0]);
-       wait_backup(sra, 0, stripes * chunk / 512, stripes * chunk / 512,
+       wait_backup(sra, 0, stripes * (chunk / 512), stripes * (chunk / 512),
                    dests, destfd, destoffsets,
                    0);
-       sysfs_set_num(sra, NULL, "suspend_lo", (stripes * chunk/512) * data);
+       sysfs_set_num(sra, NULL, "suspend_lo", (stripes * (chunk/512)) * data);
        free(buf);
        /* FIXME this should probably be numeric */
        sysfs_set_str(sra, NULL, "sync_max", "max");
@@ -1513,13 +1560,14 @@ static int child_shrink(int afd, struct mdinfo *sra, unsigned long stripes,
        int rv;
        int degraded = 0;
 
-       posix_memalign((void**)&buf, 4096, disks * chunk);
-       start = sra->component_size - stripes * chunk/512;
+       if (posix_memalign((void**)&buf, 4096, disks * chunk))
+               return 0;
+       start = sra->component_size - stripes * (chunk/512);
        sysfs_set_num(sra, NULL, "sync_max", start);
        sysfs_set_str(sra, NULL, "sync_action", "reshape");
        sysfs_set_num(sra, NULL, "suspend_lo", 0);
        sysfs_set_num(sra, NULL, "suspend_hi", 0);
-       rv = wait_backup(sra, 0, start - stripes * chunk/512, stripes * chunk/512,
+       rv = wait_backup(sra, 0, start - stripes * (chunk/512), stripes * (chunk/512),
                         dests, destfd, destoffsets, 0);
        if (rv < 0)
                return 0;
@@ -1529,9 +1577,9 @@ static int child_shrink(int afd, struct mdinfo *sra, unsigned long stripes,
                    dests, destfd, destoffsets,
                    0, &degraded, buf);
        validate(afd, destfd[0], destoffsets[0]);
-       wait_backup(sra, start, stripes*chunk/512, 0,
+       wait_backup(sra, start, stripes*(chunk/512), 0,
                    dests, destfd, destoffsets, 0);
-       sysfs_set_num(sra, NULL, "suspend_lo", (stripes * chunk/512) * data);
+       sysfs_set_num(sra, NULL, "suspend_lo", (stripes * (chunk/512)) * data);
        free(buf);
        /* FIXME this should probably be numeric */
        sysfs_set_str(sra, NULL, "sync_max", "max");
@@ -1552,7 +1600,8 @@ static int child_same_size(int afd, struct mdinfo *sra, unsigned long stripes,
        int degraded = 0;
 
 
-       posix_memalign((void**)&buf, 4096, disks * chunk);
+       if (posix_memalign((void**)&buf, 4096, disks * chunk))
+               return 0;
 
        sysfs_set_num(sra, NULL, "suspend_lo", 0);
        sysfs_set_num(sra, NULL, "suspend_hi", 0);
@@ -1565,7 +1614,7 @@ static int child_same_size(int afd, struct mdinfo *sra, unsigned long stripes,
                    disks, chunk, level, layout,
                    dests, destfd, destoffsets,
                    0, &degraded, buf);
-       grow_backup(sra, (start + stripes) * chunk/512, stripes,
+       grow_backup(sra, (start + stripes) * (chunk/512), stripes,
                    fds, offsets,
                    disks, chunk, level, layout,
                    dests, destfd, destoffsets,
@@ -1575,16 +1624,16 @@ static int child_same_size(int afd, struct mdinfo *sra, unsigned long stripes,
        start += stripes * 2; /* where to read next */
        size = sra->component_size / (chunk/512);
        while (start < size) {
-               if (wait_backup(sra, (start-stripes*2)*chunk/512,
-                               stripes*chunk/512, 0,
+               if (wait_backup(sra, (start-stripes*2)*(chunk/512),
+                               stripes*(chunk/512), 0,
                                dests, destfd, destoffsets,
                                part) < 0)
                        return 0;
-               sysfs_set_num(sra, NULL, "suspend_lo", start*chunk/512 * data);
+               sysfs_set_num(sra, NULL, "suspend_lo", start*(chunk/512) * data);
                if (start + stripes > size)
                        tailstripes = (size - start);
 
-               grow_backup(sra, start*chunk/512, tailstripes,
+               grow_backup(sra, start*(chunk/512), tailstripes,
                            fds, offsets,
                            disks, chunk, level, layout,
                            dests, destfd, destoffsets,
@@ -1593,15 +1642,15 @@ static int child_same_size(int afd, struct mdinfo *sra, unsigned long stripes,
                part = 1 - part;
                validate(afd, destfd[0], destoffsets[0]);
        }
-       if (wait_backup(sra, (start-stripes*2) * chunk/512, stripes * chunk/512, 0,
+       if (wait_backup(sra, (start-stripes*2) * (chunk/512), stripes * (chunk/512), 0,
                        dests, destfd, destoffsets,
                        part) < 0)
                return 0;
-       sysfs_set_num(sra, NULL, "suspend_lo", ((start-stripes)*chunk/512) * data);
-       wait_backup(sra, (start-stripes) * chunk/512, tailstripes * chunk/512, 0,
+       sysfs_set_num(sra, NULL, "suspend_lo", ((start-stripes)*(chunk/512)) * data);
+       wait_backup(sra, (start-stripes) * (chunk/512), tailstripes * (chunk/512), 0,
                    dests, destfd, destoffsets,
                    1-part);
-       sysfs_set_num(sra, NULL, "suspend_lo", (size*chunk/512) * data);
+       sysfs_set_num(sra, NULL, "suspend_lo", (size*(chunk/512)) * data);
        sysfs_set_num(sra, NULL, "sync_speed_min", speed);
        free(buf);
        return 1;
@@ -1704,11 +1753,23 @@ int Grow_restart(struct supertype *st, struct mdinfo *info, int *fdlist, int cnt
                        continue; /* Wrong uuid */
                }
 
-               if (info->array.utime > __le64_to_cpu(bsb.mtime) + 10*60 ||
+               /* array utime and backup-mtime should be updated at much the same time, but it seems that
+                * sometimes they aren't... So allow considerable flexability in matching, and allow
+                * this test to be overridden by an environment variable.
+                */
+               if (info->array.utime > __le64_to_cpu(bsb.mtime) + 2*60*60 ||
                    info->array.utime < __le64_to_cpu(bsb.mtime) - 10*60) {
-                       if (verbose)
-                               fprintf(stderr, Name ": too-old timestamp on backup-metadata on %s\n", devname);
-                       continue; /* time stamp is too bad */
+                       if (check_env("MDADM_GROW_ALLOW_OLD")) {
+                               fprintf(stderr, Name ": accepting backup with timestamp %lu "
+                                       "for array with timestamp %lu\n",
+                                       (unsigned long)__le64_to_cpu(bsb.mtime),
+                                       (unsigned long)info->array.utime);
+                       } else {
+                               if (verbose)
+                                       fprintf(stderr, Name ": too-old timestamp on "
+                                               "backup-metadata on %s\n", devname);
+                               continue; /* time stamp is too bad */
+                       }
                }
 
                if (bsb.magic[15] == '1') {
@@ -1930,8 +1991,8 @@ int Grow_continue(int mdfd, struct supertype *st, struct mdinfo *info,
        nchunk = info->new_chunk;
 
 
-       a = ochunk/512 * odata;
-       b = nchunk/512 * ndata;
+       a = (ochunk/512) * odata;
+       b = (nchunk/512) * ndata;
        /* Find GCD */
        while (a != b) {
                if (a < b)
@@ -1940,7 +2001,7 @@ int Grow_continue(int mdfd, struct supertype *st, struct mdinfo *info,
                        a -= b;
        }
        /* LCM == product / GCD */
-       blocks = ochunk/512 * nchunk/512 * odata * ndata / a;
+       blocks = (ochunk/512) * (nchunk/512) * odata * ndata / a;
 
        sra = sysfs_read(-1, devname2devnum(info->sys_name),
                         GET_COMPONENT|GET_DEVS|GET_OFFSET|GET_STATE|