* Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
*
- * Copyright (c) 2012, 2016, Intel Corporation.
+ * Copyright (c) 2012, 2017, Intel Corporation.
*/
/*
* This file is part of Lustre, http://www.lustre.org/
#include <sys/stat.h>
#include <sys/mount.h>
#include <sys/utsname.h>
+#include <sys/sysmacros.h>
#include <string.h>
#include <getopt.h>
extern char *progname;
-#define L_BLOCK_SIZE 4096
/* keep it less than LL_FID_NAMELEN */
#define DUMMY_FILE_NAME_LEN 25
#define EXT3_DIRENT_SIZE DUMMY_FILE_NAME_LEN
int ldiskfs_write_ldd(struct mkfs_opts *mop)
{
char mntpt[] = "/tmp/mntXXXXXX";
- char filepnm[128];
+ char filepnm[192];
char *dev;
FILE *filep;
int ret = 0;
E2FSPROGS, feature);
#if !(HAVE_LDISKFSPROGS)
fprintf(stderr, "Please install the latest version of e2fsprogs from\n"
- "https://downloads.hpdd.intel.com/public/e2fsprogs/latest/\n"
+ "https://downloads.whamcloud.com/public/e2fsprogs/latest/\n"
"to enable this feature.\n");
#endif
if (make_backfs)
static int enable_default_ext4_features(struct mkfs_opts *mop, char *anchor,
size_t maxbuflen, int user_spec)
{
+ int enable_64bit = 0;
+
+ /* Enable large block addresses if the LUN is over 2^32 blocks. */
+ if (mop->mo_device_kb / mop->mo_blocksize_kb > 0xffffffffULL &&
+ is_e2fsprogs_feature_supp("-O 64bit") == 0)
+ enable_64bit = 1;
+
if (IS_OST(&mop->mo_ldd)) {
append_unique(anchor, user_spec ? "," : " -O ",
"extents", NULL, maxbuflen);
append_unique(anchor, user_spec ? "," : " -O ",
"dirdata", NULL, maxbuflen);
append_unique(anchor, ",", "uninit_bg", NULL, maxbuflen);
- append_unique(anchor, ",", "^extents", NULL, maxbuflen);
+ if (enable_64bit)
+ append_unique(anchor, ",", "extents", NULL, maxbuflen);
+ else
+ append_unique(anchor, ",", "^extents", NULL, maxbuflen);
} else {
append_unique(anchor, user_spec ? "," : " -O ",
"uninit_bg", NULL, maxbuflen);
if (is_e2fsprogs_feature_supp("-O huge_file") == 0)
append_unique(anchor, ",", "huge_file", NULL, maxbuflen);
- /* Enable large block addresses if the LUN is over 2^32 blocks. */
- if (mop->mo_device_kb / (L_BLOCK_SIZE >> 10) >= 0x100002000ULL &&
- is_e2fsprogs_feature_supp("-O 64bit") == 0)
+ if (enable_64bit)
append_unique(anchor, ",", "64bit", NULL, maxbuflen);
/* Cluster inode/block bitmaps and inode table for more efficient IO.
if (IS_OST(&mop->mo_ldd) &&
strstr(mop->mo_mkfsopts, "-G") == NULL) {
snprintf(tmp_buf, sizeof(tmp_buf), " -G %u",
- (1 << 20) / L_BLOCK_SIZE);
+ 1024 / mop->mo_blocksize_kb);
strscat(anchor, tmp_buf, maxbuflen);
}
}
/* Build fs according to type */
int ldiskfs_make_lustre(struct mkfs_opts *mop)
{
- __u64 device_kb = mop->mo_device_kb, block_count = 0;
char mkfs_cmd[PATH_MAX];
char buf[64];
char *start;
char *dev;
int ret = 0, ext_opts = 0;
+ bool have_64bit = false;
size_t maxbuflen;
+ mop->mo_blocksize_kb = 4;
+
+ start = strstr(mop->mo_mkfsopts, "-b");
+ if (start) {
+ char *end = NULL;
+ long blocksize;
+
+ blocksize = strtol(start + 2, &end, 0);
+ if (end && (*end == 'k' || *end == 'K'))
+ blocksize *= 1024;
+ /* EXT4_MIN_BLOCK_SIZE || EXT4_MAX_BLOCK_SIZE */
+ if (blocksize < 1024 || blocksize > 65536) {
+ fprintf(stderr,
+ "%s: blocksize %lu not in 1024-65536 bytes, normally 4096 bytes\n",
+ progname, blocksize);
+ return EINVAL;
+ }
+
+ if ((blocksize & (blocksize - 1)) != 0) {
+ fprintf(stderr,
+ "%s: blocksize %lu not a power-of-two value\n",
+ progname, blocksize);
+ return EINVAL;
+ }
+ mop->mo_blocksize_kb = blocksize >> 10;
+ }
+
if (!(mop->mo_flags & MO_IS_LOOP)) {
- mop->mo_device_kb = get_device_size(mop->mo_device);
+ __u64 device_kb = get_device_size(mop->mo_device);
- if (mop->mo_device_kb == 0)
+ if (device_kb == 0)
return ENODEV;
/* Compare to real size */
- if (device_kb == 0 || device_kb > mop->mo_device_kb)
- device_kb = mop->mo_device_kb;
- else
+ if (mop->mo_device_kb == 0 || device_kb < mop->mo_device_kb)
mop->mo_device_kb = device_kb;
}
if (mop->mo_device_kb != 0) {
+ __u64 block_count;
+
if (mop->mo_device_kb < 32384) {
fprintf(stderr, "%s: size of filesystem must be larger "
"than 32MB, but is set to %lldKB\n",
progname, (long long)mop->mo_device_kb);
return EINVAL;
}
- block_count = mop->mo_device_kb / (L_BLOCK_SIZE >> 10);
- /* If the LUN size is just over 2^32 blocks, limit the
- * filesystem size to 2^32-1 blocks to avoid problems with
- * ldiskfs/mkfs not handling this size. Bug 22906 */
- if (block_count > 0xffffffffULL && block_count < 0x100002000ULL)
- block_count = 0xffffffffULL;
+ block_count = mop->mo_device_kb / mop->mo_blocksize_kb;
+ if (block_count > 0xffffffffULL) {
+ /* If the LUN size is just over 2^32 blocks, limit the
+ * filesystem size to 2^32-1 blocks to avoid problems
+ * with ldiskfs/mkfs not handling this well. b=22906
+ */
+ if (block_count < 0x100002000ULL)
+ mop->mo_device_kb =
+ 0xffffffffULL * mop->mo_blocksize_kb;
+ else
+ have_64bit = true;
+ }
}
+
if ((mop->mo_ldd.ldd_mount_type == LDD_MT_EXT3) ||
(mop->mo_ldd.ldd_mount_type == LDD_MT_LDISKFS) ||
(mop->mo_ldd.ldd_mount_type == LDD_MT_LDISKFS2)) {
/* Journal size in MB */
if (strstr(mop->mo_mkfsopts, "-J") == NULL &&
- device_kb > 1024 * 1024) {
+ mop->mo_device_kb > 1024 * 1024) {
/* Choose our own default journal size */
long journal_mb = 0, max_mb;
max_mb = 0;
/* Use at most 4% of device for journal */
- journal_mb = device_kb * 4 / (1024 * 100);
+ journal_mb = mop->mo_device_kb * 4 / (1024 * 100);
if (journal_mb > max_mb)
journal_mb = max_mb;
* (assuming all files are in composite layout and has
* 3 components):
*
- * ldiskfs inode size: 156
- * extended attributes size, including:
+ * ldiskfs inode size: 160
+ * MDT extended attributes size, including:
* ext4_xattr_header: 32
* LOV EA size: 32(lov_comp_md_v1) +
* 3 * 40(lov_comp_md_entry_v1) +
* 3 * 32(lov_mds_md) +
* stripes * 24(lov_ost_data) +
- * 16(xattr_entry) + 3(lov)
+ * 16(xattr_entry) + 4("lov")
* LMA EA size: 24(lustre_mdt_attrs) +
- * 16(xattr_entry) + 3(lma)
+ * 16(xattr_entry) + 4("lma")
+ * SOM EA size: 24(lustre_som_attrs) +
+ * 16(xattr_entry) + 4("som")
* link EA size: 24(link_ea_header) + 18(link_ea_entry) +
- * (filename) + 16(xattr_entry) + 4(link)
+ * 16(filename) + 16(xattr_entry) + 4("link")
* and some margin for 4-byte alignment, ACLs and other EAs.
*
* If we say the average filename length is about 32 bytes,
* the calculation looks like:
- * 156 + 32 + (32+3*(40 + 32)+24*N+19) + (24+19) +
- * (24+18+~32+20) + other <= 512*2^m, {m=0,1,2,3}
+ * 160 + 32 + (32+3*(40+32)+24*stripes+20) + (24+20) + (24+20) +
+ * (24+20) + (~42+16+20) + other <= 512*2^m, {m=0,1,2,3}
*/
if (strstr(mop->mo_mkfsopts, "-I") == NULL) {
if (IS_MDT(&mop->mo_ldd)) {
inode_size = 1024;
} else if (IS_OST(&mop->mo_ldd)) {
/* We store MDS FID and necessary composite
- * layout information in the OST object EA. */
+ * layout information in the OST object EA:
+ * ldiskfs inode size: 160
+ * OST extended attributes size, including:
+ * ext4_xattr_header: 32
+ * LMA EA size: 24(lustre_mdt_attrs) +
+ * 16(xattr_entry) + 4("lma")
+ * FID EA size: 52(filter_fid) +
+ * 16(xattr_entry) + 4("fid")
+ * 160 + 32 + (24+20) + (52+20) = 308
+ */
inode_size = 512;
}
* this, but it is impossible to know in advance. */
if (IS_OST(&mop->mo_ldd)) {
/* OST > 16TB assume average file size 1MB */
- if (device_kb > (16ULL << 30))
+ if (mop->mo_device_kb > (16ULL << 30))
bytes_per_inode = 1024 * 1024;
/* OST > 4TB assume average file size 512kB */
- else if (device_kb > (4ULL << 30))
+ else if (mop->mo_device_kb > (4ULL << 30))
bytes_per_inode = 512 * 1024;
/* OST > 1TB assume average file size 256kB */
- else if (device_kb > (1ULL << 30))
+ else if (mop->mo_device_kb > (1ULL << 30))
bytes_per_inode = 256 * 1024;
/* OST > 10GB assume average file size 64kB,
* plus a bit so that inodes will fit into a
* 256x flex_bg without overflowing */
- else if (device_kb > (10ULL << 20))
+ else if (mop->mo_device_kb > (10ULL << 20))
bytes_per_inode = 69905;
}
sprintf(buf, " -i %ld", bytes_per_inode);
strscat(mop->mo_mkfsopts, buf,
sizeof(mop->mo_mkfsopts));
+ mop->mo_inode_size = bytes_per_inode;
}
}
start = moveopts_to_end(start);
maxbuflen = sizeof(mop->mo_mkfsopts) -
(start - mop->mo_mkfsopts) - strlen(start);
- ret = enable_default_ext4_features(mop, start, maxbuflen, 1);
+ ret = enable_default_ext4_features(mop, start,
+ maxbuflen, 1);
} else {
start = mop->mo_mkfsopts + strlen(mop->mo_mkfsopts),
maxbuflen = sizeof(mop->mo_mkfsopts) -
strlen(mop->mo_mkfsopts);
- ret = enable_default_ext4_features(mop, start, maxbuflen, 0);
+ ret = enable_default_ext4_features(mop, start,
+ maxbuflen, 0);
}
if (ret)
return ret;
* limitations. */
if (strstr(mop->mo_mkfsopts, "meta_bg") == NULL &&
IS_OST(&mop->mo_ldd) && mop->mo_device_kb > 100 * 1024 &&
- mop->mo_device_kb * 1024 / L_BLOCK_SIZE <= 0xffffffffULL) {
- unsigned group_blocks = L_BLOCK_SIZE * 8;
- unsigned desc_per_block = L_BLOCK_SIZE / 32;
- unsigned resize_blks;
+ !have_64bit) {
+ unsigned int group_blocks = mop->mo_blocksize_kb * 8192;
+ unsigned int desc_per_block =
+ mop->mo_blocksize_kb * 1024 / 32;
+ unsigned int resize_blks;
resize_blks = (1ULL<<32) - desc_per_block*group_blocks;
snprintf(buf, sizeof(buf), "%u", resize_blks);
strscat(mop->mo_mkfsopts, " -F", sizeof(mop->mo_mkfsopts));
snprintf(mkfs_cmd, sizeof(mkfs_cmd),
- "%s -j -b %d -L %s ", MKE2FS, L_BLOCK_SIZE,
- mop->mo_ldd.ldd_svname);
+ "%s -j -b %d -L %s ", MKE2FS,
+ mop->mo_blocksize_kb * 1024, mop->mo_ldd.ldd_svname);
} else {
fprintf(stderr,"%s: unsupported fs type: %d (%s)\n",
progname, mop->mo_ldd.ldd_mount_type,
vprint("formatting backing filesystem %s on %s\n",
MT_STR(&mop->mo_ldd), dev);
vprint("\ttarget name %s\n", mop->mo_ldd.ldd_svname);
- vprint("\t4k blocks %ju\n", (uintmax_t)block_count);
+ vprint("\tkilobytes %llu\n", mop->mo_device_kb);
vprint("\toptions %s\n", mop->mo_mkfsopts);
/* mkfs_cmd's trailing space is important! */
strscat(mkfs_cmd, mop->mo_mkfsopts, sizeof(mkfs_cmd));
strscat(mkfs_cmd, " ", sizeof(mkfs_cmd));
strscat(mkfs_cmd, dev, sizeof(mkfs_cmd));
- if (block_count != 0) {
- snprintf(buf, sizeof(buf), " %ju",
- (uintmax_t)block_count);
+ if (mop->mo_device_kb != 0) {
+ snprintf(buf, sizeof(buf), " %lluk",
+ (unsigned long long)mop->mo_device_kb);
strscat(mkfs_cmd, buf, sizeof(mkfs_cmd));
}
char *real_sys_path = NULL;
int rc;
+ /*
+ * Don't apply block device tuning for MDT or MGT devices,
+ * since we don't need huge IO sizes to get good performance
+ */
+ if (!IS_OST(&mop->mo_ldd))
+ return 0;
+
if (src == NULL)
return EINVAL;
goto have_whole_dev;
if (verbose)
- fprintf(stderr,
- "warning: cannot access '%s': %s\n",
+ fprintf(stderr, "warning: cannot access '%s': %s\n",
partition_path, strerror(errno));
rc = errno;
goto out;
git://git.whamcloud.com / fs / lustre-release.git / blobdiff