diff options
| -rw-r--r-- | Create.c | 565 | ||||
| -rw-r--r-- | ReadMe.c | 2 | ||||
| -rw-r--r-- | mdadm.8.in | 18 | ||||
| -rw-r--r-- | mdadm.c | 9 | ||||
| -rw-r--r-- | mdadm.h | 7 | ||||
| -rw-r--r-- | tests/00raid5-zero | 12 |
6 files changed, 437 insertions, 176 deletions
@@ -26,6 +26,10 @@ #include "md_u.h" #include "md_p.h" #include <ctype.h> +#include <fcntl.h> +#include <signal.h> +#include <sys/signalfd.h> +#include <sys/wait.h> static int round_size_and_verify(unsigned long long *size, int chunk) { @@ -91,6 +95,382 @@ int default_layout(struct supertype *st, int level, int verbose) return layout; } +static pid_t write_zeroes_fork(int fd, struct shape *s, struct supertype *st, + struct mddev_dev *dv) + +{ + const unsigned long long req_size = 1 << 30; + unsigned long long offset_bytes, size_bytes, sz; + sigset_t sigset; + int ret = 0; + pid_t pid; + + size_bytes = KIB_TO_BYTES(s->size); + + /* + * If size_bytes is zero, this is a zoned raid array where + * each disk is of a different size and uses its full + * disk. Thus zero the entire disk. + */ + if (!size_bytes && !get_dev_size(fd, dv->devname, &size_bytes)) + return -1; + + if (dv->data_offset != INVALID_SECTORS) + offset_bytes = SEC_TO_BYTES(dv->data_offset); + else + offset_bytes = SEC_TO_BYTES(st->data_offset); + + pr_info("zeroing data from %lld to %lld on: %s\n", + offset_bytes, size_bytes, dv->devname); + + pid = fork(); + if (pid < 0) { + pr_err("Could not fork to zero disks: %s\n", strerror(errno)); + return pid; + } else if (pid != 0) { + return pid; + } + + sigemptyset(&sigset); + sigaddset(&sigset, SIGINT); + sigprocmask(SIG_UNBLOCK, &sigset, NULL); + + while (size_bytes) { + /* + * Split requests to the kernel into 1GB chunks seeing the + * fallocate() call is not interruptible and blocking a + * ctrl-c for several minutes is not desirable. + * + * 1GB is chosen as a compromise: the user may still have + * to wait several seconds if they ctrl-c on devices that + * zero slowly, but will reduce the number of requests + * required and thus the overhead on devices that perform + * better. + */ + sz = size_bytes; + if (sz >= req_size) + sz = req_size; + + if (fallocate(fd, FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE, + offset_bytes, sz)) { + pr_err("zeroing %s failed: %s\n", dv->devname, + strerror(errno)); + ret = 1; + break; + } + + offset_bytes += sz; + size_bytes -= sz; + } + + exit(ret); +} + +static int wait_for_zero_forks(int *zero_pids, int count) +{ + int wstatus, ret = 0, i, sfd, wait_count = 0; + struct signalfd_siginfo fdsi; + bool interrupted = false; + sigset_t sigset; + ssize_t s; + + for (i = 0; i < count; i++) + if (zero_pids[i]) + wait_count++; + if (!wait_count) + return 0; + + sigemptyset(&sigset); + sigaddset(&sigset, SIGINT); + sigaddset(&sigset, SIGCHLD); + sigprocmask(SIG_BLOCK, &sigset, NULL); + + sfd = signalfd(-1, &sigset, 0); + if (sfd < 0) { + pr_err("Unable to create signalfd: %s\n", strerror(errno)); + return 1; + } + + while (1) { + s = read(sfd, &fdsi, sizeof(fdsi)); + if (s != sizeof(fdsi)) { + pr_err("Invalid signalfd read: %s\n", strerror(errno)); + close(sfd); + return 1; + } + + if (fdsi.ssi_signo == SIGINT) { + printf("\n"); + pr_info("Interrupting zeroing processes, please wait...\n"); + interrupted = true; + } else if (fdsi.ssi_signo == SIGCHLD) { + if (!--wait_count) + break; + } + } + + close(sfd); + + for (i = 0; i < count; i++) { + if (!zero_pids[i]) + continue; + + waitpid(zero_pids[i], &wstatus, 0); + zero_pids[i] = 0; + if (!WIFEXITED(wstatus) || WEXITSTATUS(wstatus)) + ret = 1; + } + + if (interrupted) { + pr_err("zeroing interrupted!\n"); + return 1; + } + + if (ret) + pr_err("zeroing failed!\n"); + else + pr_info("zeroing finished\n"); + + return ret; +} + +static int add_disk_to_super(int mdfd, struct shape *s, struct context *c, + struct supertype *st, struct mddev_dev *dv, + struct mdinfo *info, int have_container, int major_num, + int *zero_pid) +{ + dev_t rdev; + int fd; + + if (dv->disposition == 'j') { + info->disk.raid_disk = MD_DISK_ROLE_JOURNAL; + info->disk.state = (1<<MD_DISK_JOURNAL); + } else if (info->disk.raid_disk < s->raiddisks) { + info->disk.state = (1<<MD_DISK_ACTIVE) | + (1<<MD_DISK_SYNC); + } else { + info->disk.state = 0; + } + + if (dv->writemostly == FlagSet) { + if (major_num == BITMAP_MAJOR_CLUSTERED) { + pr_err("Can not set %s --write-mostly with a clustered bitmap\n",dv->devname); + return 1; + } else { + info->disk.state |= (1<<MD_DISK_WRITEMOSTLY); + } + + } + + if (dv->failfast == FlagSet) + info->disk.state |= (1<<MD_DISK_FAILFAST); + + if (have_container) { + fd = -1; + } else { + if (st->ss->external && st->container_devnm[0]) + fd = open(dv->devname, O_RDWR); + else + fd = open(dv->devname, O_RDWR|O_EXCL); + + if (fd < 0) { + pr_err("failed to open %s after earlier success - aborting\n", + dv->devname); + return 1; + } + if (!fstat_is_blkdev(fd, dv->devname, &rdev)) + return 1; + info->disk.major = major(rdev); + info->disk.minor = minor(rdev); + } + if (fd >= 0) + remove_partitions(fd); + if (st->ss->add_to_super(st, &info->disk, fd, dv->devname, + dv->data_offset)) { + ioctl(mdfd, STOP_ARRAY, NULL); + return 1; + } + st->ss->getinfo_super(st, info, NULL); + + if (fd >= 0 && s->write_zeroes) { + *zero_pid = write_zeroes_fork(fd, s, st, dv); + if (*zero_pid <= 0) { + ioctl(mdfd, STOP_ARRAY, NULL); + return 1; + } + } + + if (have_container && c->verbose > 0) + pr_err("Using %s for device %d\n", + map_dev(info->disk.major, info->disk.minor, 0), + info->disk.number); + + if (!have_container) { + /* getinfo_super might have lost these ... */ + info->disk.major = major(rdev); + info->disk.minor = minor(rdev); + } + + return 0; +} + +static int update_metadata(int mdfd, struct shape *s, struct supertype *st, + struct map_ent **map, struct mdinfo *info, + char *chosen_name) +{ + struct mdinfo info_new; + struct map_ent *me = NULL; + + /* check to see if the uuid has changed due to these + * metadata changes, and if so update the member array + * and container uuid. Note ->write_init_super clears + * the subarray cursor such that ->getinfo_super once + * again returns container info. + */ + st->ss->getinfo_super(st, &info_new, NULL); + if (st->ss->external && is_container(s->level) && + !same_uuid(info_new.uuid, info->uuid, 0)) { + map_update(map, fd2devnm(mdfd), + info_new.text_version, + info_new.uuid, chosen_name); + me = map_by_devnm(map, st->container_devnm); + } + + if (st->ss->write_init_super(st)) { + st->ss->free_super(st); + return 1; + } + + /* + * Before activating the array, perform extra steps + * required to configure the internal write-intent + * bitmap. + */ + if (info_new.consistency_policy == CONSISTENCY_POLICY_BITMAP && + st->ss->set_bitmap && st->ss->set_bitmap(st, info)) { + st->ss->free_super(st); + return 1; + } + + /* update parent container uuid */ + if (me) { + char *path = xstrdup(me->path); + + st->ss->getinfo_super(st, &info_new, NULL); + map_update(map, st->container_devnm, info_new.text_version, + info_new.uuid, path); + free(path); + } + + flush_metadata_updates(st); + st->ss->free_super(st); + + return 0; +} + +static int add_disks(int mdfd, struct mdinfo *info, struct shape *s, + struct context *c, struct supertype *st, + struct map_ent **map, struct mddev_dev *devlist, + int total_slots, int have_container, int insert_point, + int major_num, char *chosen_name) +{ + struct mddev_dev *moved_disk = NULL; + int pass, raid_disk_num, dnum; + int zero_pids[total_slots]; + struct mddev_dev *dv; + struct mdinfo *infos; + sigset_t sigset, orig_sigset; + int ret = 0; + + /* + * Block SIGINT so the main thread will always wait for the + * zeroing processes when being interrupted. Otherwise the + * zeroing processes will finish their work in the background + * keeping the disk busy. + */ + sigemptyset(&sigset); + sigaddset(&sigset, SIGINT); + sigprocmask(SIG_BLOCK, &sigset, &orig_sigset); + memset(zero_pids, 0, sizeof(zero_pids)); + + infos = xmalloc(sizeof(*infos) * total_slots); + enable_fds(total_slots); + for (pass = 1; pass <= 2; pass++) { + for (dnum = 0, raid_disk_num = 0, dv = devlist; dv; + dv = (dv->next) ? (dv->next) : moved_disk, dnum++) { + if (dnum >= total_slots) + abort(); + if (dnum == insert_point) { + raid_disk_num += 1; + moved_disk = dv; + continue; + } + if (strcasecmp(dv->devname, "missing") == 0) { + raid_disk_num += 1; + continue; + } + if (have_container) + moved_disk = NULL; + if (have_container && dnum < total_slots - 1) + /* repeatedly use the container */ + moved_disk = dv; + + switch(pass) { + case 1: + infos[dnum] = *info; + infos[dnum].disk.number = dnum; + infos[dnum].disk.raid_disk = raid_disk_num++; + + if (dv->disposition == 'j') + raid_disk_num--; + + ret = add_disk_to_super(mdfd, s, c, st, dv, + &infos[dnum], have_container, + major_num, &zero_pids[dnum]); + if (ret) + goto out; + + break; + case 2: + infos[dnum].errors = 0; + + ret = add_disk(mdfd, st, info, &infos[dnum]); + if (ret) { + pr_err("ADD_NEW_DISK for %s failed: %s\n", + dv->devname, strerror(errno)); + if (errno == EINVAL && + info->array.level == 0) { + pr_err("Possibly your kernel doesn't support RAID0 layouts.\n"); + pr_err("Either upgrade, or use --layout=dangerous\n"); + } + goto out; + } + break; + } + if (!have_container && + dv == moved_disk && dnum != insert_point) break; + } + + if (pass == 1) { + ret = wait_for_zero_forks(zero_pids, total_slots); + if (ret) + goto out; + + ret = update_metadata(mdfd, s, st, map, info, + chosen_name); + if (ret) + goto out; + } + } + +out: + if (ret) + wait_for_zero_forks(zero_pids, total_slots); + free(infos); + sigprocmask(SIG_SETMASK, &orig_sigset, NULL); + return ret; +} + int Create(struct supertype *st, char *mddev, char *name, int *uuid, int subdevs, struct mddev_dev *devlist, @@ -117,7 +497,7 @@ int Create(struct supertype *st, char *mddev, unsigned long long minsize = 0, maxsize = 0; char *mindisc = NULL; char *maxdisc = NULL; - int dnum, raid_disk_num; + int dnum; struct mddev_dev *dv; dev_t rdev; int fail = 0, warn = 0; @@ -126,18 +506,16 @@ int Create(struct supertype *st, char *mddev, int missing_disks = 0; int insert_point = subdevs * 2; /* where to insert a missing drive */ int total_slots; - int pass; int rv; int bitmap_fd; int have_container = 0; int container_fd = -1; int need_mdmon = 0; unsigned long long bitmapsize; - struct mdinfo info, *infos; + struct mdinfo info; int did_default = 0; int do_default_layout = 0; int do_default_chunk = 0; - unsigned long safe_mode_delay = 0; char chosen_name[1024]; struct map_ent *map = NULL; unsigned long long newsize; @@ -778,11 +1156,12 @@ int Create(struct supertype *st, char *mddev, mdi = sysfs_read(-1, devnm, GET_VERSION); - pr_err("Creating array inside %s container %s\n", + pr_info("Creating array inside %s container %s\n", mdi?mdi->text_version:"managed", devnm); sysfs_free(mdi); } else - pr_err("Defaulting to version %s metadata\n", info.text_version); + pr_info("Defaulting to version %s metadata\n", + info.text_version); } map_update(&map, fd2devnm(mdfd), info.text_version, @@ -870,175 +1249,11 @@ int Create(struct supertype *st, char *mddev, } } - infos = xmalloc(sizeof(*infos) * total_slots); - enable_fds(total_slots); - for (pass = 1; pass <= 2; pass++) { - struct mddev_dev *moved_disk = NULL; /* the disk that was moved out of the insert point */ - - for (dnum = 0, raid_disk_num = 0, dv = devlist; dv; - dv = (dv->next) ? (dv->next) : moved_disk, dnum++) { - int fd; - struct mdinfo *inf = &infos[dnum]; - - if (dnum >= total_slots) - abort(); - if (dnum == insert_point) { - raid_disk_num += 1; - moved_disk = dv; - continue; - } - if (strcasecmp(dv->devname, "missing") == 0) { - raid_disk_num += 1; - continue; - } - if (have_container) - moved_disk = NULL; - if (have_container && dnum < info.array.raid_disks - 1) - /* repeatedly use the container */ - moved_disk = dv; - - switch(pass) { - case 1: - *inf = info; - - inf->disk.number = dnum; - inf->disk.raid_disk = raid_disk_num++; - - if (dv->disposition == 'j') { - inf->disk.raid_disk = MD_DISK_ROLE_JOURNAL; - inf->disk.state = (1<<MD_DISK_JOURNAL); - raid_disk_num--; - } else if (inf->disk.raid_disk < s->raiddisks) - inf->disk.state = (1<<MD_DISK_ACTIVE) | - (1<<MD_DISK_SYNC); - else - inf->disk.state = 0; - - if (dv->writemostly == FlagSet) { - if (major_num == BITMAP_MAJOR_CLUSTERED) { - pr_err("Can not set %s --write-mostly with a clustered bitmap\n",dv->devname); - goto abort_locked; - } else - inf->disk.state |= (1<<MD_DISK_WRITEMOSTLY); - } - if (dv->failfast == FlagSet) - inf->disk.state |= (1<<MD_DISK_FAILFAST); - - if (have_container) - fd = -1; - else { - if (st->ss->external && - st->container_devnm[0]) - fd = open(dv->devname, O_RDWR); - else - fd = open(dv->devname, O_RDWR|O_EXCL); - - if (fd < 0) { - pr_err("failed to open %s after earlier success - aborting\n", - dv->devname); - goto abort_locked; - } - if (!fstat_is_blkdev(fd, dv->devname, &rdev)) - return 1; - inf->disk.major = major(rdev); - inf->disk.minor = minor(rdev); - } - if (fd >= 0) - remove_partitions(fd); - if (st->ss->add_to_super(st, &inf->disk, - fd, dv->devname, - dv->data_offset)) { - ioctl(mdfd, STOP_ARRAY, NULL); - goto abort_locked; - } - st->ss->getinfo_super(st, inf, NULL); - safe_mode_delay = inf->safe_mode_delay; - - if (have_container && c->verbose > 0) - pr_err("Using %s for device %d\n", - map_dev(inf->disk.major, - inf->disk.minor, - 0), dnum); - - if (!have_container) { - /* getinfo_super might have lost these ... */ - inf->disk.major = major(rdev); - inf->disk.minor = minor(rdev); - } - break; - case 2: - inf->errors = 0; - - rv = add_disk(mdfd, st, &info, inf); - - if (rv) { - pr_err("ADD_NEW_DISK for %s failed: %s\n", - dv->devname, strerror(errno)); - if (errno == EINVAL && - info.array.level == 0) { - pr_err("Possibly your kernel doesn't support RAID0 layouts.\n"); - pr_err("Either upgrade, or use --layout=dangerous\n"); - } - goto abort_locked; - } - break; - } - if (!have_container && - dv == moved_disk && dnum != insert_point) break; - } - if (pass == 1) { - struct mdinfo info_new; - struct map_ent *me = NULL; - - /* check to see if the uuid has changed due to these - * metadata changes, and if so update the member array - * and container uuid. Note ->write_init_super clears - * the subarray cursor such that ->getinfo_super once - * again returns container info. - */ - st->ss->getinfo_super(st, &info_new, NULL); - if (st->ss->external && !is_container(s->level) && - !same_uuid(info_new.uuid, info.uuid, 0)) { - map_update(&map, fd2devnm(mdfd), - info_new.text_version, - info_new.uuid, chosen_name); - me = map_by_devnm(&map, st->container_devnm); - } - - if (st->ss->write_init_super(st)) { - st->ss->free_super(st); - goto abort_locked; - } - /* - * Before activating the array, perform extra steps - * required to configure the internal write-intent - * bitmap. - */ - if (info_new.consistency_policy == - CONSISTENCY_POLICY_BITMAP && - st->ss->set_bitmap && - st->ss->set_bitmap(st, &info)) { - st->ss->free_super(st); - goto abort_locked; - } - - /* update parent container uuid */ - if (me) { - char *path = xstrdup(me->path); - - st->ss->getinfo_super(st, &info_new, NULL); - map_update(&map, st->container_devnm, - info_new.text_version, - info_new.uuid, path); - free(path); - } + if (add_disks(mdfd, &info, s, c, st, &map, devlist, total_slots, + have_container, insert_point, major_num, chosen_name)) + goto abort_locked; - flush_metadata_updates(st); - st->ss->free_super(st); - } - } map_unlock(&map); - free(infos); if (is_container(s->level)) { /* No need to start. But we should signal udev to @@ -1065,7 +1280,7 @@ int Create(struct supertype *st, char *mddev, "readonly"); break; } - sysfs_set_safemode(&info, safe_mode_delay); + sysfs_set_safemode(&info, info.safe_mode_delay); if (err) { pr_err("failed to activate array.\n"); ioctl(mdfd, STOP_ARRAY, NULL); @@ -1103,7 +1318,7 @@ int Create(struct supertype *st, char *mddev, ioctl(mdfd, RESTART_ARRAY_RW, NULL); } if (c->verbose >= 0) - pr_err("array %s started.\n", mddev); + pr_info("array %s started.\n", mddev); if (st->ss->external && st->container_devnm[0]) { if (need_mdmon) start_mdmon(st->container_devnm); @@ -138,6 +138,7 @@ struct option long_options[] = { {"size", 1, 0, 'z'}, {"auto", 1, 0, Auto}, /* also for --assemble */ {"assume-clean",0,0, AssumeClean }, + {"write-zeroes",0,0, WriteZeroes }, {"metadata", 1, 0, 'e'}, /* superblock format */ {"bitmap", 1, 0, Bitmap}, {"bitmap-chunk", 1, 0, BitmapChunk}, @@ -390,6 +391,7 @@ char Help_create[] = " --write-journal= : Specify journal device for RAID-4/5/6 array\n" " --consistency-policy= : Specify the policy that determines how the array\n" " -k : maintains consistency in case of unexpected shutdown.\n" +" --write-zeroes : Write zeroes to the disks before creating. This will bypass initial sync.\n" "\n" ; @@ -838,6 +838,22 @@ array is resynced at creation. From Linux version 3.0, can be used with that command to avoid the automatic resync. .TP +.BR \-\-write-zeroes +When creating an array, send write zeroes requests to all the block +devices. This should zero the data area on all disks such that the +initial sync is not necessary and, if successfull, will behave +as if +.B \-\-assume\-clean +was specified. +.IP +This is intended for use with devices that have hardware offload for +zeroing, but despite this zeroing can still take several minutes for +large disks. Thus a message is printed before and after zeroing and +each disk is zeroed in parallel with the others. +.IP +This is only meaningful with --create. + +.TP .BR \-\-backup\-file= This is needed when .B \-\-grow @@ -1370,7 +1386,7 @@ and .B layout\-alternate options are for RAID0 arrays with non-uniform devices size that were in use before Linux 5.4. If the array was being used with Linux 3.13 or -earlier, then to assemble the array on a new kernel, +earlier, then to assemble the array on a new kernel, .B \-\-update=layout\-original must be given. If the array was created and used with a kernel from Linux 3.14 to Linux 5.3, then @@ -590,6 +590,10 @@ int main(int argc, char *argv[]) s.assume_clean = 1; continue; + case O(CREATE, WriteZeroes): + s.write_zeroes = 1; + continue; + case O(GROW,'n'): case O(CREATE,'n'): case O(BUILD,'n'): /* number of raid disks */ @@ -1251,6 +1255,11 @@ int main(int argc, char *argv[]) } } + if (s.write_zeroes && !s.assume_clean) { + pr_info("Disk zeroing requested, setting --assume-clean to skip resync\n"); + s.assume_clean = 1; + } + if (!mode && devs_found) { mode = MISC; devmode = 'Q'; @@ -275,6 +275,9 @@ static inline void __put_unaligned32(__u32 val, void *p) #define ARRAY_SIZE(x) (sizeof(x)/sizeof(x[0])) +#define KIB_TO_BYTES(x) ((x) << 10) +#define SEC_TO_BYTES(x) ((x) << 9) + extern const char Name[]; struct md_bb_entry { @@ -435,6 +438,7 @@ extern char Version[], Usage[], Help[], OptionHelp[], */ enum special_options { AssumeClean = 300, + WriteZeroes, BitmapChunk, WriteBehind, ReAdd, @@ -640,6 +644,7 @@ struct shape { int bitmap_chunk; char *bitmap_file; int assume_clean; + bool write_zeroes; int write_behind; unsigned long long size; unsigned long long data_offset; @@ -1854,6 +1859,8 @@ static inline int xasprintf(char **strp, const char *fmt, ...) { #endif #define cont_err(fmt ...) fprintf(stderr, " " fmt) +#define pr_info(fmt, args...) printf("%s: "fmt, Name, ##args) + void *xmalloc(size_t len); void *xrealloc(void *ptr, size_t len); void *xcalloc(size_t num, size_t size); diff --git a/tests/00raid5-zero b/tests/00raid5-zero new file mode 100644 index 0000000..7d0f05a --- /dev/null +++ b/tests/00raid5-zero @@ -0,0 +1,12 @@ + +if mdadm -CfR $md0 -l 5 -n3 $dev0 $dev1 $dev2 --write-zeroes ; then + check nosync + echo check > /sys/block/md0/md/sync_action; + check wait +elif grep "zeroing [^ ]* failed: Operation not supported" \ + $targetdir/stderr; then + echo "write-zeros not supported, skipping" +else + echo >&2 "ERROR: mdadm return failure without not supported message" + exit 1 +fi |