exofs: RAID0 support

We now support striping over mirror devices. Including variable sized
stripe_unit.

Some limits:
* stripe_unit must be a multiple of PAGE_SIZE
* stripe_unit * stripe_count is maximum upto 32-bit (4Gb)

Tested RAID0 over mirrors, RAID0 only, mirrors only. All check.

Design notes:
* I'm not using a vectored raid-engine mechanism yet. Following the
  pnfs-objects-layout data-map structure, "Mirror" is just a private
  case of "group_width" == 1, and RAID0 is a private case of
  "Mirrors" == 1. The performance lose of the general case over the
  particular special case optimization is totally negligible, also
  considering the extra code size.

* In general I added a prepare_stripes() stage that divides the
  to-be-io pages to the participating devices, the previous
  exofs_ios_write/read, now becomes _write/read_mirrors and a new
  write/read upper layer loops on all devices calling
  _write/read_mirrors. Effectively the prepare_stripes stage is the all
  secret.
  Also truncate need fixing to accommodate for striping.

* In a RAID0 arrangement, in a regular usage scenario, if all inode
  layouts will start at the same device, the small files fill up the
  first device and the later devices stay empty, the farther the device
  the emptier it is.

  To fix that, each inode will start at a different stripe_unit,
  according to it's obj_id modulus number-of-stripe-units. And
  will then span all stripe-units in the same incrementing order
  wrapping back to the beginning of the device table. We call it
  a stripe-units moving window.

  Special consideration was taken to keep all devices in a mirror
  arrangement identical. So a broken osd-device could just be cloned
  from one of the mirrors and no FS scrubbing is needed. (We do that
  by rotating stripe-unit at a time and not a single device at a time.)

TODO:
 We no longer verify object_length == inode->i_size in exofs_iget.
 (since i_size is stripped on multiple objects now).
 I should introduce a multiple-device attribute reading, and use
 it in exofs_iget.

Signed-off-by: Boaz Harrosh <bharrosh@panasas.com>
This commit is contained in:
Boaz Harrosh 2010-02-01 13:35:51 +02:00
parent d9c740d225
commit 5d952b8391
4 changed files with 333 additions and 83 deletions

View file

@ -58,6 +58,14 @@
struct exofs_layout {
osd_id s_pid; /* partition ID of file system*/
/* Our way of looking at the data_map */
unsigned stripe_unit;
unsigned mirrors_p1;
unsigned group_width;
enum exofs_inode_layout_gen_functions lay_func;
unsigned s_numdevs; /* Num of devices in array */
struct osd_dev *s_ods[0]; /* Variable length */
};
@ -133,6 +141,9 @@ struct exofs_io_state {
struct exofs_per_dev_state {
struct osd_request *or;
struct bio *bio;
loff_t offset;
unsigned length;
unsigned dev;
} per_dev[];
};

View file

@ -869,18 +869,17 @@ static const struct osd_attr g_attr_inode_dir_layout = ATTR_DEF(
0);
/*
* Read an inode from the OSD, and return it as is. We also return the size
* attribute in the 'obj_size' argument.
* Read the Linux inode info from the OSD, and return it as is. In exofs the
* inode info is in an application specific page/attribute of the osd-object.
*/
static int exofs_get_inode(struct super_block *sb, struct exofs_i_info *oi,
struct exofs_fcb *inode, uint64_t *obj_size)
struct exofs_fcb *inode)
{
struct exofs_sb_info *sbi = sb->s_fs_info;
struct osd_attr attrs[] = {
[0] = g_attr_inode_data,
[1] = g_attr_inode_file_layout,
[2] = g_attr_inode_dir_layout,
[3] = g_attr_logical_length,
};
struct exofs_io_state *ios;
struct exofs_on_disk_inode_layout *layout;
@ -944,15 +943,6 @@ static int exofs_get_inode(struct super_block *sb, struct exofs_i_info *oi,
}
}
*obj_size = ~0;
ret = extract_attr_from_ios(ios, &attrs[3]);
if (ret) {
EXOFS_ERR("%s: extract_attr of logical_length failed\n",
__func__);
goto out;
}
*obj_size = get_unaligned_be64(attrs[3].val_ptr);
out:
exofs_put_io_state(ios);
return ret;
@ -971,7 +961,6 @@ struct inode *exofs_iget(struct super_block *sb, unsigned long ino)
struct exofs_i_info *oi;
struct exofs_fcb fcb;
struct inode *inode;
uint64_t obj_size;
int ret;
inode = iget_locked(sb, ino);
@ -983,7 +972,7 @@ struct inode *exofs_iget(struct super_block *sb, unsigned long ino)
__oi_init(oi);
/* read the inode from the osd */
ret = exofs_get_inode(sb, oi, &fcb, &obj_size);
ret = exofs_get_inode(sb, oi, &fcb);
if (ret)
goto bad_inode;
@ -1004,13 +993,6 @@ struct inode *exofs_iget(struct super_block *sb, unsigned long ino)
inode->i_blkbits = EXOFS_BLKSHIFT;
inode->i_generation = le32_to_cpu(fcb.i_generation);
if ((inode->i_size != obj_size) &&
(!exofs_inode_is_fast_symlink(inode))) {
EXOFS_ERR("WARNING: Size of inode=%llu != object=%llu\n",
inode->i_size, _LLU(obj_size));
/* FIXME: call exofs_inode_recovery() */
}
oi->i_dir_start_lookup = 0;
if ((inode->i_nlink == 0) && (inode->i_mode == 0)) {

View file

@ -23,6 +23,7 @@
*/
#include <scsi/scsi_device.h>
#include <asm/div64.h>
#include "exofs.h"
@ -110,7 +111,17 @@ void exofs_put_io_state(struct exofs_io_state *ios)
unsigned exofs_layout_od_id(struct exofs_layout *layout,
osd_id obj_no, unsigned layout_index)
{
return layout_index;
/* switch (layout->lay_func) {
case LAYOUT_MOVING_WINDOW:
{*/
unsigned dev_mod = obj_no;
return (layout_index + dev_mod * layout->mirrors_p1) %
layout->s_numdevs;
/* }
case LAYOUT_FUNC_IMPLICT:
return layout->devs[layout_index];
}*/
}
static inline struct osd_dev *exofs_ios_od(struct exofs_io_state *ios,
@ -225,8 +236,8 @@ int exofs_check_io(struct exofs_io_state *ios, u64 *resid)
_clear_bio(ios->per_dev[i].bio);
EXOFS_DBGMSG("start read offset passed end of file "
"offset=0x%llx, length=0x%llx\n",
_LLU(ios->offset),
_LLU(ios->length));
_LLU(ios->per_dev[i].offset),
_LLU(ios->per_dev[i].length));
continue; /* we recovered */
}
@ -248,6 +259,127 @@ int exofs_check_io(struct exofs_io_state *ios, u64 *resid)
return acumulated_lin_err;
}
/* REMOVEME: After review
Some quoteing from the standard
L = logical offset into the file
W = number of data components in a stripe
S = W * stripe_unit (S is Stripe length)
N = L / S (N is the stripe Number)
C = (L-(N*S)) / stripe_unit (C is the component)
O = (N*stripe_unit)+(L%stripe_unit) (O is the object's offset)
*/
static void _offset_dev_unit_off(struct exofs_io_state *ios, u64 file_offset,
u64 *obj_offset, unsigned *dev, unsigned *unit_off)
{
unsigned stripe_unit = ios->layout->stripe_unit;
unsigned stripe_length = stripe_unit * ios->layout->group_width;
u64 stripe_no = file_offset;
unsigned stripe_mod = do_div(stripe_no, stripe_length);
*unit_off = stripe_mod % stripe_unit;
*obj_offset = stripe_no * stripe_unit + *unit_off;
*dev = stripe_mod / stripe_unit * ios->layout->mirrors_p1;
}
static int _add_stripe_unit(struct exofs_io_state *ios, unsigned *cur_bvec,
struct exofs_per_dev_state *per_dev, int cur_len)
{
unsigned bv = *cur_bvec;
struct request_queue *q =
osd_request_queue(exofs_ios_od(ios, per_dev->dev));
per_dev->length += cur_len;
if (per_dev->bio == NULL) {
unsigned pages_in_stripe = ios->layout->group_width *
(ios->layout->stripe_unit / PAGE_SIZE);
unsigned bio_size = (ios->bio->bi_vcnt + pages_in_stripe) /
ios->layout->group_width;
per_dev->bio = bio_kmalloc(GFP_KERNEL, bio_size);
if (unlikely(!per_dev->bio)) {
EXOFS_DBGMSG("Faild to allocate BIO size=%u\n",
bio_size);
return -ENOMEM;
}
}
while (cur_len > 0) {
int added_len;
struct bio_vec *bvec = &ios->bio->bi_io_vec[bv];
BUG_ON(ios->bio->bi_vcnt <= bv);
cur_len -= bvec->bv_len;
added_len = bio_add_pc_page(q, per_dev->bio, bvec->bv_page,
bvec->bv_len, bvec->bv_offset);
if (unlikely(bvec->bv_len != added_len))
return -ENOMEM;
++bv;
}
BUG_ON(cur_len);
*cur_bvec = bv;
return 0;
}
static int _prepare_for_striping(struct exofs_io_state *ios)
{
u64 length = ios->length;
u64 offset = ios->offset;
unsigned stripe_unit = ios->layout->stripe_unit;
unsigned comp = 0;
unsigned stripes = 0;
unsigned cur_bvec = 0;
int ret;
if (!ios->bio) {
if (ios->kern_buff) {
struct exofs_per_dev_state *per_dev = &ios->per_dev[0];
unsigned unit_off;
_offset_dev_unit_off(ios, offset, &per_dev->offset,
&per_dev->dev, &unit_off);
/* no cross device without page array */
BUG_ON((ios->layout->group_width > 1) &&
(unit_off + length > stripe_unit));
}
ios->numdevs = ios->layout->mirrors_p1;
return 0;
}
while (length) {
struct exofs_per_dev_state *per_dev = &ios->per_dev[comp];
unsigned cur_len;
if (!per_dev->length) {
unsigned unit_off;
_offset_dev_unit_off(ios, offset, &per_dev->offset,
&per_dev->dev, &unit_off);
stripes++;
cur_len = min_t(u64, stripe_unit - unit_off, length);
offset += cur_len;
} else {
cur_len = min_t(u64, stripe_unit, length);
}
ret = _add_stripe_unit(ios, &cur_bvec, per_dev, cur_len);
if (unlikely(ret))
goto out;
comp += ios->layout->mirrors_p1;
comp %= ios->layout->s_numdevs;
length -= cur_len;
}
out:
ios->numdevs = stripes * ios->layout->mirrors_p1;
return ret;
}
int exofs_sbi_create(struct exofs_io_state *ios)
{
int i, ret;
@ -296,61 +428,71 @@ int exofs_sbi_remove(struct exofs_io_state *ios)
return ret;
}
int exofs_sbi_write(struct exofs_io_state *ios)
static int _sbi_write_mirror(struct exofs_io_state *ios, int cur_comp)
{
int i, ret;
struct exofs_per_dev_state *master_dev = &ios->per_dev[cur_comp];
unsigned dev = ios->per_dev[cur_comp].dev;
unsigned last_comp = cur_comp + ios->layout->mirrors_p1;
int ret = 0;
for (i = 0; i < ios->layout->s_numdevs; i++) {
for (; cur_comp < last_comp; ++cur_comp, ++dev) {
struct exofs_per_dev_state *per_dev = &ios->per_dev[cur_comp];
struct osd_request *or;
or = osd_start_request(exofs_ios_od(ios, i), GFP_KERNEL);
or = osd_start_request(exofs_ios_od(ios, dev), GFP_KERNEL);
if (unlikely(!or)) {
EXOFS_ERR("%s: osd_start_request failed\n", __func__);
ret = -ENOMEM;
goto out;
}
ios->per_dev[i].or = or;
ios->numdevs++;
per_dev->or = or;
per_dev->offset = master_dev->offset;
if (ios->bio) {
struct bio *bio;
if (i != 0) {
if (per_dev != master_dev) {
bio = bio_kmalloc(GFP_KERNEL,
ios->bio->bi_max_vecs);
master_dev->bio->bi_max_vecs);
if (unlikely(!bio)) {
EXOFS_DBGMSG(
"Faild to allocate BIO size=%u\n",
ios->bio->bi_max_vecs);
master_dev->bio->bi_max_vecs);
ret = -ENOMEM;
goto out;
}
__bio_clone(bio, ios->bio);
__bio_clone(bio, master_dev->bio);
bio->bi_bdev = NULL;
bio->bi_next = NULL;
ios->per_dev[i].bio = bio;
per_dev->length = master_dev->length;
per_dev->bio = bio;
per_dev->dev = dev;
} else {
bio = ios->bio;
bio = master_dev->bio;
/* FIXME: bio_set_dir() */
bio->bi_rw |= (1 << BIO_RW);
}
osd_req_write(or, &ios->obj, ios->offset, bio,
ios->length);
osd_req_write(or, &ios->obj, per_dev->offset, bio,
per_dev->length);
EXOFS_DBGMSG("write(0x%llx) offset=0x%llx "
"length=0x%llx dev=%d\n",
_LLU(ios->obj.id), _LLU(ios->offset),
_LLU(ios->length), i);
_LLU(ios->obj.id), _LLU(per_dev->offset),
_LLU(per_dev->length), dev);
} else if (ios->kern_buff) {
osd_req_write_kern(or, &ios->obj, ios->offset,
ret = osd_req_write_kern(or, &ios->obj, per_dev->offset,
ios->kern_buff, ios->length);
if (unlikely(ret))
goto out;
EXOFS_DBGMSG2("write_kern(0x%llx) offset=0x%llx "
"length=0x%llx dev=%d\n",
_LLU(ios->obj.id), _LLU(ios->offset),
_LLU(ios->length), i);
_LLU(ios->obj.id), _LLU(per_dev->offset),
_LLU(ios->length), dev);
} else {
osd_req_set_attributes(or, &ios->obj);
EXOFS_DBGMSG2("obj(0x%llx) set_attributes=%d dev=%d\n",
_LLU(ios->obj.id), ios->out_attr_len, i);
_LLU(ios->obj.id), ios->out_attr_len, dev);
}
if (ios->out_attr)
@ -361,40 +503,57 @@ int exofs_sbi_write(struct exofs_io_state *ios)
osd_req_add_get_attr_list(or, ios->in_attr,
ios->in_attr_len);
}
ret = exofs_io_execute(ios);
out:
return ret;
}
int exofs_sbi_read(struct exofs_io_state *ios)
int exofs_sbi_write(struct exofs_io_state *ios)
{
int i;
int ret;
ret = _prepare_for_striping(ios);
if (unlikely(ret))
return ret;
for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
ret = _sbi_write_mirror(ios, i);
if (unlikely(ret))
return ret;
}
ret = exofs_io_execute(ios);
return ret;
}
static int _sbi_read_mirror(struct exofs_io_state *ios, unsigned cur_comp)
{
struct osd_request *or;
struct exofs_per_dev_state *per_dev = &ios->per_dev[0];
struct exofs_per_dev_state *per_dev = &ios->per_dev[cur_comp];
unsigned first_dev = (unsigned)ios->obj.id;
first_dev %= ios->layout->s_numdevs;
first_dev = per_dev->dev + first_dev % ios->layout->mirrors_p1;
or = osd_start_request(exofs_ios_od(ios, first_dev), GFP_KERNEL);
if (unlikely(!or)) {
EXOFS_ERR("%s: osd_start_request failed\n", __func__);
return -ENOMEM;
}
per_dev->or = or;
ios->numdevs++;
if (ios->bio) {
osd_req_read(or, &ios->obj, ios->offset, ios->bio, ios->length);
osd_req_read(or, &ios->obj, per_dev->offset,
per_dev->bio, per_dev->length);
EXOFS_DBGMSG("read(0x%llx) offset=0x%llx length=0x%llx"
" dev=%d\n", _LLU(ios->obj.id),
_LLU(ios->offset), _LLU(ios->length),
_LLU(per_dev->offset), _LLU(per_dev->length),
first_dev);
} else if (ios->kern_buff) {
int ret = osd_req_read_kern(or, &ios->obj, ios->offset,
int ret = osd_req_read_kern(or, &ios->obj, per_dev->offset,
ios->kern_buff, ios->length);
EXOFS_DBGMSG2("read_kern(0x%llx) offset=0x%llx "
"length=0x%llx dev=%d ret=>%d\n",
_LLU(ios->obj.id), _LLU(ios->offset),
_LLU(ios->obj.id), _LLU(per_dev->offset),
_LLU(ios->length), first_dev, ret);
if (unlikely(ret))
return ret;
@ -403,14 +562,32 @@ int exofs_sbi_read(struct exofs_io_state *ios)
EXOFS_DBGMSG2("obj(0x%llx) get_attributes=%d dev=%d\n",
_LLU(ios->obj.id), ios->in_attr_len, first_dev);
}
if (ios->out_attr)
osd_req_add_set_attr_list(or, ios->out_attr, ios->out_attr_len);
if (ios->in_attr)
osd_req_add_get_attr_list(or, ios->in_attr, ios->in_attr_len);
return exofs_io_execute(ios);
return 0;
}
int exofs_sbi_read(struct exofs_io_state *ios)
{
int i;
int ret;
ret = _prepare_for_striping(ios);
if (unlikely(ret))
return ret;
for (i = 0; i < ios->numdevs; i += ios->layout->mirrors_p1) {
ret = _sbi_read_mirror(ios, i);
if (unlikely(ret))
return ret;
}
ret = exofs_io_execute(ios);
return ret;
}
int extract_attr_from_ios(struct exofs_io_state *ios, struct osd_attr *attr)
@ -434,42 +611,84 @@ int extract_attr_from_ios(struct exofs_io_state *ios, struct osd_attr *attr)
return -EIO;
}
static int _truncate_mirrors(struct exofs_io_state *ios, unsigned cur_comp,
struct osd_attr *attr)
{
int last_comp = cur_comp + ios->layout->mirrors_p1;
for (; cur_comp < last_comp; ++cur_comp) {
struct exofs_per_dev_state *per_dev = &ios->per_dev[cur_comp];
struct osd_request *or;
or = osd_start_request(exofs_ios_od(ios, cur_comp), GFP_KERNEL);
if (unlikely(!or)) {
EXOFS_ERR("%s: osd_start_request failed\n", __func__);
return -ENOMEM;
}
per_dev->or = or;
osd_req_set_attributes(or, &ios->obj);
osd_req_add_set_attr_list(or, attr, 1);
}
return 0;
}
int exofs_oi_truncate(struct exofs_i_info *oi, u64 size)
{
struct exofs_sb_info *sbi = oi->vfs_inode.i_sb->s_fs_info;
struct exofs_io_state *ios;
struct osd_attr attr;
__be64 newsize;
struct exofs_trunc_attr {
struct osd_attr attr;
__be64 newsize;
} *size_attrs;
u64 this_obj_size;
unsigned dev;
unsigned unit_off;
int i, ret;
if (exofs_get_io_state(&sbi->layout, &ios))
return -ENOMEM;
ret = exofs_get_io_state(&sbi->layout, &ios);
if (unlikely(ret))
return ret;
size_attrs = kcalloc(ios->layout->group_width, sizeof(*size_attrs),
GFP_KERNEL);
if (unlikely(!size_attrs)) {
ret = -ENOMEM;
goto out;
}
ios->obj.id = exofs_oi_objno(oi);
ios->cred = oi->i_cred;
newsize = cpu_to_be64(size);
attr = g_attr_logical_length;
attr.val_ptr = &newsize;
ios->numdevs = ios->layout->s_numdevs;
_offset_dev_unit_off(ios, size, &this_obj_size, &dev, &unit_off);
for (i = 0; i < sbi->layout.s_numdevs; i++) {
struct osd_request *or;
for (i = 0; i < ios->layout->group_width; ++i) {
struct exofs_trunc_attr *size_attr = &size_attrs[i];
u64 obj_size;
or = osd_start_request(exofs_ios_od(ios, i), GFP_KERNEL);
if (unlikely(!or)) {
EXOFS_ERR("%s: osd_start_request failed\n", __func__);
ret = -ENOMEM;
if (i < dev)
obj_size = this_obj_size +
ios->layout->stripe_unit - unit_off;
else if (i == dev)
obj_size = this_obj_size;
else /* i > dev */
obj_size = this_obj_size - unit_off;
size_attr->newsize = cpu_to_be64(obj_size);
size_attr->attr = g_attr_logical_length;
size_attr->attr.val_ptr = &size_attr->newsize;
ret = _truncate_mirrors(ios, i * ios->layout->mirrors_p1,
&size_attr->attr);
if (unlikely(ret))
goto out;
}
ios->per_dev[i].or = or;
ios->numdevs++;
osd_req_set_attributes(or, &ios->obj);
osd_req_add_set_attr_list(or, &attr, 1);
}
ret = exofs_io_execute(ios);
out:
kfree(size_attrs);
exofs_put_io_state(ios);
return ret;
}

View file

@ -308,6 +308,8 @@ static void exofs_put_super(struct super_block *sb)
static int _read_and_match_data_map(struct exofs_sb_info *sbi, unsigned numdevs,
struct exofs_device_table *dt)
{
u64 stripe_length;
sbi->data_map.odm_num_comps =
le32_to_cpu(dt->dt_data_map.cb_num_comps);
sbi->data_map.odm_stripe_unit =
@ -321,14 +323,47 @@ static int _read_and_match_data_map(struct exofs_sb_info *sbi, unsigned numdevs,
sbi->data_map.odm_raid_algorithm =
le32_to_cpu(dt->dt_data_map.cb_raid_algorithm);
/* FIXME: Hard coded mirror only for now. if not so do not mount */
if ((sbi->data_map.odm_num_comps != numdevs) ||
(sbi->data_map.odm_stripe_unit != EXOFS_BLKSIZE) ||
(sbi->data_map.odm_raid_algorithm != PNFS_OSD_RAID_0) ||
(sbi->data_map.odm_mirror_cnt != (numdevs - 1)))
/* FIXME: Only raid0 !group_width/depth for now. if not so, do not mount */
if (sbi->data_map.odm_group_width || sbi->data_map.odm_group_depth) {
EXOFS_ERR("Group width/depth not supported\n");
return -EINVAL;
else
return 0;
}
if (sbi->data_map.odm_num_comps != numdevs) {
EXOFS_ERR("odm_num_comps(%u) != numdevs(%u)\n",
sbi->data_map.odm_num_comps, numdevs);
return -EINVAL;
}
if (sbi->data_map.odm_raid_algorithm != PNFS_OSD_RAID_0) {
EXOFS_ERR("Only RAID_0 for now\n");
return -EINVAL;
}
if (0 != (numdevs % (sbi->data_map.odm_mirror_cnt + 1))) {
EXOFS_ERR("Data Map wrong, numdevs=%d mirrors=%d\n",
numdevs, sbi->data_map.odm_mirror_cnt);
return -EINVAL;
}
stripe_length = sbi->data_map.odm_stripe_unit *
(numdevs / (sbi->data_map.odm_mirror_cnt + 1));
if (stripe_length >= (1ULL << 32)) {
EXOFS_ERR("Total Stripe length(0x%llx)"
" >= 32bit is not supported\n", _LLU(stripe_length));
return -EINVAL;
}
if (0 != (sbi->data_map.odm_stripe_unit & ~PAGE_MASK)) {
EXOFS_ERR("Stripe Unit(0x%llx)"
" must be Multples of PAGE_SIZE(0x%lx)\n",
_LLU(sbi->data_map.odm_stripe_unit), PAGE_SIZE);
return -EINVAL;
}
sbi->layout.stripe_unit = sbi->data_map.odm_stripe_unit;
sbi->layout.mirrors_p1 = sbi->data_map.odm_mirror_cnt + 1;
sbi->layout.group_width = sbi->data_map.odm_num_comps /
sbi->layout.mirrors_p1;
return 0;
}
/* @odi is valid only as long as @fscb_dev is valid */
@ -502,6 +537,9 @@ static int exofs_fill_super(struct super_block *sb, void *data, int silent)
}
/* Default layout in case we do not have a device-table */
sbi->layout.stripe_unit = PAGE_SIZE;
sbi->layout.mirrors_p1 = 1;
sbi->layout.group_width = 1;
sbi->layout.s_ods[0] = od;
sbi->layout.s_numdevs = 1;
sbi->layout.s_pid = opts->pid;