linux/fs/gfs2/rgrp.c

1795 lines
45 KiB
C
Raw Normal View History

/*
* Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
* Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
*
* This copyrighted material is made available to anyone wishing to use,
* modify, copy, or redistribute it subject to the terms and conditions
* of the GNU General Public License version 2.
*/
#include <linux/slab.h>
#include <linux/spinlock.h>
#include <linux/completion.h>
#include <linux/buffer_head.h>
#include <linux/fs.h>
#include <linux/gfs2_ondisk.h>
#include <linux/prefetch.h>
#include <linux/blkdev.h>
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
#include <linux/rbtree.h>
#include "gfs2.h"
#include "incore.h"
#include "glock.h"
#include "glops.h"
#include "lops.h"
#include "meta_io.h"
#include "quota.h"
#include "rgrp.h"
#include "super.h"
#include "trans.h"
#include "util.h"
#include "log.h"
#include "inode.h"
#include "trace_gfs2.h"
#define BFITNOENT ((u32)~0)
#define NO_BLOCK ((u64)~0)
#if BITS_PER_LONG == 32
#define LBITMASK (0x55555555UL)
#define LBITSKIP55 (0x55555555UL)
#define LBITSKIP00 (0x00000000UL)
#else
#define LBITMASK (0x5555555555555555UL)
#define LBITSKIP55 (0x5555555555555555UL)
#define LBITSKIP00 (0x0000000000000000UL)
#endif
/*
* These routines are used by the resource group routines (rgrp.c)
* to keep track of block allocation. Each block is represented by two
* bits. So, each byte represents GFS2_NBBY (i.e. 4) blocks.
*
* 0 = Free
* 1 = Used (not metadata)
* 2 = Unlinked (still in use) inode
* 3 = Used (metadata)
*/
static const char valid_change[16] = {
/* current */
/* n */ 0, 1, 1, 1,
/* e */ 1, 0, 0, 0,
/* w */ 0, 0, 0, 1,
1, 0, 0, 0
};
static u32 rgblk_search(struct gfs2_rgrpd *rgd, u32 goal,
unsigned char old_state,
struct gfs2_bitmap **rbi);
/**
* gfs2_setbit - Set a bit in the bitmaps
* @buffer: the buffer that holds the bitmaps
* @buflen: the length (in bytes) of the buffer
* @block: the block to set
* @new_state: the new state of the block
*
*/
static inline void gfs2_setbit(struct gfs2_rgrpd *rgd, unsigned char *buf1,
unsigned char *buf2, unsigned int offset,
struct gfs2_bitmap *bi, u32 block,
unsigned char new_state)
{
unsigned char *byte1, *byte2, *end, cur_state;
unsigned int buflen = bi->bi_len;
const unsigned int bit = (block % GFS2_NBBY) * GFS2_BIT_SIZE;
byte1 = buf1 + offset + (block / GFS2_NBBY);
end = buf1 + offset + buflen;
BUG_ON(byte1 >= end);
cur_state = (*byte1 >> bit) & GFS2_BIT_MASK;
if (unlikely(!valid_change[new_state * 4 + cur_state])) {
printk(KERN_WARNING "GFS2: buf_blk = 0x%llx old_state=%d, "
"new_state=%d\n",
(unsigned long long)block, cur_state, new_state);
printk(KERN_WARNING "GFS2: rgrp=0x%llx bi_start=0x%lx\n",
(unsigned long long)rgd->rd_addr,
(unsigned long)bi->bi_start);
printk(KERN_WARNING "GFS2: bi_offset=0x%lx bi_len=0x%lx\n",
(unsigned long)bi->bi_offset,
(unsigned long)bi->bi_len);
dump_stack();
gfs2_consist_rgrpd(rgd);
return;
}
*byte1 ^= (cur_state ^ new_state) << bit;
if (buf2) {
byte2 = buf2 + offset + (block / GFS2_NBBY);
cur_state = (*byte2 >> bit) & GFS2_BIT_MASK;
*byte2 ^= (cur_state ^ new_state) << bit;
}
}
/**
* gfs2_testbit - test a bit in the bitmaps
* @buffer: the buffer that holds the bitmaps
* @buflen: the length (in bytes) of the buffer
* @block: the block to read
*
*/
static inline unsigned char gfs2_testbit(struct gfs2_rgrpd *rgd,
const unsigned char *buffer,
unsigned int buflen, u32 block)
{
const unsigned char *byte, *end;
unsigned char cur_state;
unsigned int bit;
byte = buffer + (block / GFS2_NBBY);
bit = (block % GFS2_NBBY) * GFS2_BIT_SIZE;
end = buffer + buflen;
gfs2_assert(rgd->rd_sbd, byte < end);
cur_state = (*byte >> bit) & GFS2_BIT_MASK;
return cur_state;
}
/**
* gfs2_bit_search
* @ptr: Pointer to bitmap data
* @mask: Mask to use (normally 0x55555.... but adjusted for search start)
* @state: The state we are searching for
*
* We xor the bitmap data with a patter which is the bitwise opposite
* of what we are looking for, this gives rise to a pattern of ones
* wherever there is a match. Since we have two bits per entry, we
* take this pattern, shift it down by one place and then and it with
* the original. All the even bit positions (0,2,4, etc) then represent
* successful matches, so we mask with 0x55555..... to remove the unwanted
* odd bit positions.
*
* This allows searching of a whole u64 at once (32 blocks) with a
* single test (on 64 bit arches).
*/
static inline u64 gfs2_bit_search(const __le64 *ptr, u64 mask, u8 state)
{
u64 tmp;
static const u64 search[] = {
[0] = 0xffffffffffffffffULL,
[1] = 0xaaaaaaaaaaaaaaaaULL,
[2] = 0x5555555555555555ULL,
[3] = 0x0000000000000000ULL,
};
tmp = le64_to_cpu(*ptr) ^ search[state];
tmp &= (tmp >> 1);
tmp &= mask;
return tmp;
}
/**
* gfs2_bitfit - Search an rgrp's bitmap buffer to find a bit-pair representing
* a block in a given allocation state.
* @buffer: the buffer that holds the bitmaps
* @len: the length (in bytes) of the buffer
* @goal: start search at this block's bit-pair (within @buffer)
* @state: GFS2_BLKST_XXX the state of the block we're looking for.
*
* Scope of @goal and returned block number is only within this bitmap buffer,
* not entire rgrp or filesystem. @buffer will be offset from the actual
* beginning of a bitmap block buffer, skipping any header structures, but
* headers are always a multiple of 64 bits long so that the buffer is
* always aligned to a 64 bit boundary.
*
* The size of the buffer is in bytes, but is it assumed that it is
* always ok to read a complete multiple of 64 bits at the end
* of the block in case the end is no aligned to a natural boundary.
*
* Return: the block number (bitmap buffer scope) that was found
*/
static u32 gfs2_bitfit(const u8 *buf, const unsigned int len,
u32 goal, u8 state)
{
u32 spoint = (goal << 1) & ((8*sizeof(u64)) - 1);
const __le64 *ptr = ((__le64 *)buf) + (goal >> 5);
const __le64 *end = (__le64 *)(buf + ALIGN(len, sizeof(u64)));
u64 tmp;
u64 mask = 0x5555555555555555ULL;
u32 bit;
BUG_ON(state > 3);
/* Mask off bits we don't care about at the start of the search */
mask <<= spoint;
tmp = gfs2_bit_search(ptr, mask, state);
ptr++;
while(tmp == 0 && ptr < end) {
tmp = gfs2_bit_search(ptr, 0x5555555555555555ULL, state);
ptr++;
}
/* Mask off any bits which are more than len bytes from the start */
if (ptr == end && (len & (sizeof(u64) - 1)))
tmp &= (((u64)~0) >> (64 - 8*(len & (sizeof(u64) - 1))));
/* Didn't find anything, so return */
if (tmp == 0)
return BFITNOENT;
ptr--;
bit = __ffs64(tmp);
bit /= 2; /* two bits per entry in the bitmap */
return (((const unsigned char *)ptr - buf) * GFS2_NBBY) + bit;
}
/**
* gfs2_bitcount - count the number of bits in a certain state
* @buffer: the buffer that holds the bitmaps
* @buflen: the length (in bytes) of the buffer
* @state: the state of the block we're looking for
*
* Returns: The number of bits
*/
static u32 gfs2_bitcount(struct gfs2_rgrpd *rgd, const u8 *buffer,
unsigned int buflen, u8 state)
{
const u8 *byte = buffer;
const u8 *end = buffer + buflen;
const u8 state1 = state << 2;
const u8 state2 = state << 4;
const u8 state3 = state << 6;
u32 count = 0;
for (; byte < end; byte++) {
if (((*byte) & 0x03) == state)
count++;
if (((*byte) & 0x0C) == state1)
count++;
if (((*byte) & 0x30) == state2)
count++;
if (((*byte) & 0xC0) == state3)
count++;
}
return count;
}
/**
* gfs2_rgrp_verify - Verify that a resource group is consistent
* @sdp: the filesystem
* @rgd: the rgrp
*
*/
void gfs2_rgrp_verify(struct gfs2_rgrpd *rgd)
{
struct gfs2_sbd *sdp = rgd->rd_sbd;
struct gfs2_bitmap *bi = NULL;
u32 length = rgd->rd_length;
u32 count[4], tmp;
int buf, x;
memset(count, 0, 4 * sizeof(u32));
/* Count # blocks in each of 4 possible allocation states */
for (buf = 0; buf < length; buf++) {
bi = rgd->rd_bits + buf;
for (x = 0; x < 4; x++)
count[x] += gfs2_bitcount(rgd,
bi->bi_bh->b_data +
bi->bi_offset,
bi->bi_len, x);
}
if (count[0] != rgd->rd_free) {
if (gfs2_consist_rgrpd(rgd))
fs_err(sdp, "free data mismatch: %u != %u\n",
count[0], rgd->rd_free);
return;
}
tmp = rgd->rd_data - rgd->rd_free - rgd->rd_dinodes;
if (count[1] != tmp) {
if (gfs2_consist_rgrpd(rgd))
fs_err(sdp, "used data mismatch: %u != %u\n",
count[1], tmp);
return;
}
if (count[2] + count[3] != rgd->rd_dinodes) {
if (gfs2_consist_rgrpd(rgd))
fs_err(sdp, "used metadata mismatch: %u != %u\n",
count[2] + count[3], rgd->rd_dinodes);
return;
}
}
static inline int rgrp_contains_block(struct gfs2_rgrpd *rgd, u64 block)
{
u64 first = rgd->rd_data0;
u64 last = first + rgd->rd_data;
return first <= block && block < last;
}
/**
* gfs2_blk2rgrpd - Find resource group for a given data/meta block number
* @sdp: The GFS2 superblock
* @n: The data block number
*
* Returns: The resource group, or NULL if not found
*/
struct gfs2_rgrpd *gfs2_blk2rgrpd(struct gfs2_sbd *sdp, u64 blk, bool exact)
{
struct rb_node *n, *next;
struct gfs2_rgrpd *cur;
spin_lock(&sdp->sd_rindex_spin);
n = sdp->sd_rindex_tree.rb_node;
while (n) {
cur = rb_entry(n, struct gfs2_rgrpd, rd_node);
next = NULL;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
if (blk < cur->rd_addr)
next = n->rb_left;
else if (blk >= cur->rd_data0 + cur->rd_data)
next = n->rb_right;
if (next == NULL) {
spin_unlock(&sdp->sd_rindex_spin);
if (exact) {
if (blk < cur->rd_addr)
return NULL;
if (blk >= cur->rd_data0 + cur->rd_data)
return NULL;
}
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
return cur;
}
n = next;
}
spin_unlock(&sdp->sd_rindex_spin);
return NULL;
}
/**
* gfs2_rgrpd_get_first - get the first Resource Group in the filesystem
* @sdp: The GFS2 superblock
*
* Returns: The first rgrp in the filesystem
*/
struct gfs2_rgrpd *gfs2_rgrpd_get_first(struct gfs2_sbd *sdp)
{
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
const struct rb_node *n;
struct gfs2_rgrpd *rgd;
spin_lock(&sdp->sd_rindex_spin);
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
n = rb_first(&sdp->sd_rindex_tree);
rgd = rb_entry(n, struct gfs2_rgrpd, rd_node);
spin_unlock(&sdp->sd_rindex_spin);
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
return rgd;
}
/**
* gfs2_rgrpd_get_next - get the next RG
* @rgd: A RG
*
* Returns: The next rgrp
*/
struct gfs2_rgrpd *gfs2_rgrpd_get_next(struct gfs2_rgrpd *rgd)
{
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
struct gfs2_sbd *sdp = rgd->rd_sbd;
const struct rb_node *n;
spin_lock(&sdp->sd_rindex_spin);
n = rb_next(&rgd->rd_node);
if (n == NULL)
n = rb_first(&sdp->sd_rindex_tree);
if (unlikely(&rgd->rd_node == n)) {
spin_unlock(&sdp->sd_rindex_spin);
return NULL;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
}
rgd = rb_entry(n, struct gfs2_rgrpd, rd_node);
spin_unlock(&sdp->sd_rindex_spin);
return rgd;
}
void gfs2_free_clones(struct gfs2_rgrpd *rgd)
{
int x;
for (x = 0; x < rgd->rd_length; x++) {
struct gfs2_bitmap *bi = rgd->rd_bits + x;
kfree(bi->bi_clone);
bi->bi_clone = NULL;
}
}
void gfs2_clear_rgrpd(struct gfs2_sbd *sdp)
{
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
struct rb_node *n;
struct gfs2_rgrpd *rgd;
struct gfs2_glock *gl;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
while ((n = rb_first(&sdp->sd_rindex_tree))) {
rgd = rb_entry(n, struct gfs2_rgrpd, rd_node);
gl = rgd->rd_gl;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
rb_erase(n, &sdp->sd_rindex_tree);
if (gl) {
spin_lock(&gl->gl_spin);
gl->gl_object = NULL;
spin_unlock(&gl->gl_spin);
gfs2_glock_add_to_lru(gl);
gfs2_glock_put(gl);
}
gfs2_free_clones(rgd);
kfree(rgd->rd_bits);
kmem_cache_free(gfs2_rgrpd_cachep, rgd);
}
}
static void gfs2_rindex_print(const struct gfs2_rgrpd *rgd)
{
printk(KERN_INFO " ri_addr = %llu\n", (unsigned long long)rgd->rd_addr);
printk(KERN_INFO " ri_length = %u\n", rgd->rd_length);
printk(KERN_INFO " ri_data0 = %llu\n", (unsigned long long)rgd->rd_data0);
printk(KERN_INFO " ri_data = %u\n", rgd->rd_data);
printk(KERN_INFO " ri_bitbytes = %u\n", rgd->rd_bitbytes);
}
/**
* gfs2_compute_bitstructs - Compute the bitmap sizes
* @rgd: The resource group descriptor
*
* Calculates bitmap descriptors, one for each block that contains bitmap data
*
* Returns: errno
*/
static int compute_bitstructs(struct gfs2_rgrpd *rgd)
{
struct gfs2_sbd *sdp = rgd->rd_sbd;
struct gfs2_bitmap *bi;
u32 length = rgd->rd_length; /* # blocks in hdr & bitmap */
u32 bytes_left, bytes;
int x;
if (!length)
return -EINVAL;
rgd->rd_bits = kcalloc(length, sizeof(struct gfs2_bitmap), GFP_NOFS);
if (!rgd->rd_bits)
return -ENOMEM;
bytes_left = rgd->rd_bitbytes;
for (x = 0; x < length; x++) {
bi = rgd->rd_bits + x;
bi->bi_flags = 0;
/* small rgrp; bitmap stored completely in header block */
if (length == 1) {
bytes = bytes_left;
bi->bi_offset = sizeof(struct gfs2_rgrp);
bi->bi_start = 0;
bi->bi_len = bytes;
/* header block */
} else if (x == 0) {
bytes = sdp->sd_sb.sb_bsize - sizeof(struct gfs2_rgrp);
bi->bi_offset = sizeof(struct gfs2_rgrp);
bi->bi_start = 0;
bi->bi_len = bytes;
/* last block */
} else if (x + 1 == length) {
bytes = bytes_left;
bi->bi_offset = sizeof(struct gfs2_meta_header);
bi->bi_start = rgd->rd_bitbytes - bytes_left;
bi->bi_len = bytes;
/* other blocks */
} else {
bytes = sdp->sd_sb.sb_bsize -
sizeof(struct gfs2_meta_header);
bi->bi_offset = sizeof(struct gfs2_meta_header);
bi->bi_start = rgd->rd_bitbytes - bytes_left;
bi->bi_len = bytes;
}
bytes_left -= bytes;
}
if (bytes_left) {
gfs2_consist_rgrpd(rgd);
return -EIO;
}
bi = rgd->rd_bits + (length - 1);
if ((bi->bi_start + bi->bi_len) * GFS2_NBBY != rgd->rd_data) {
if (gfs2_consist_rgrpd(rgd)) {
gfs2_rindex_print(rgd);
fs_err(sdp, "start=%u len=%u offset=%u\n",
bi->bi_start, bi->bi_len, bi->bi_offset);
}
return -EIO;
}
return 0;
}
[GFS2] kernel changes to support new gfs2_grow command This is another revision of my gfs2 kernel patch that allows gfs2_grow to function properly. Steve Whitehouse expressed some concerns about the previous patch and I restructured it based on his comments. The previous patch was doing the statfs_change at file close time, under its own transaction. The current patch does the statfs_change inside the gfs2_commit_write function, which keeps it under the umbrella of the inode transaction. I can't call ri_update to re-read the rindex file during the transaction because the transaction may have outstanding unwritten buffers attached to the rgrps that would be otherwise blown away. So instead, I created a new function, gfs2_ri_total, that will re-read the rindex file just to total the file system space for the sake of the statfs_change. The ri_update will happen later, when gfs2 realizes the version number has changed, as it happened before my patch. Since the statfs_change is happening at write_commit time and there may be multiple writes to the rindex file for one grow operation. So one consequence of this restructuring is that instead of getting one kernel message to indicate the change, you may see several. For example, before when you did a gfs2_grow, you'd get a single message like: GFS2: File system extended by 247876 blocks (968MB) Now you get something like: GFS2: File system extended by 207896 blocks (812MB) GFS2: File system extended by 39980 blocks (156MB) This version has also been successfully run against the hours-long "gfs2_fsck_hellfire" test that does several gfs2_grow and gfs2_fsck while interjecting file system damage. It does this repeatedly under a variety Resource Group conditions. Signed-off-By: Bob Peterson <rpeterso@redhat.com> Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-05-09 14:37:57 +00:00
/**
* gfs2_ri_total - Total up the file system space, according to the rindex.
*
*/
u64 gfs2_ri_total(struct gfs2_sbd *sdp)
{
u64 total_data = 0;
struct inode *inode = sdp->sd_rindex;
struct gfs2_inode *ip = GFS2_I(inode);
char buf[sizeof(struct gfs2_rindex)];
struct file_ra_state ra_state;
int error, rgrps;
file_ra_state_init(&ra_state, inode->i_mapping);
for (rgrps = 0;; rgrps++) {
loff_t pos = rgrps * sizeof(struct gfs2_rindex);
if (pos + sizeof(struct gfs2_rindex) > i_size_read(inode))
[GFS2] kernel changes to support new gfs2_grow command This is another revision of my gfs2 kernel patch that allows gfs2_grow to function properly. Steve Whitehouse expressed some concerns about the previous patch and I restructured it based on his comments. The previous patch was doing the statfs_change at file close time, under its own transaction. The current patch does the statfs_change inside the gfs2_commit_write function, which keeps it under the umbrella of the inode transaction. I can't call ri_update to re-read the rindex file during the transaction because the transaction may have outstanding unwritten buffers attached to the rgrps that would be otherwise blown away. So instead, I created a new function, gfs2_ri_total, that will re-read the rindex file just to total the file system space for the sake of the statfs_change. The ri_update will happen later, when gfs2 realizes the version number has changed, as it happened before my patch. Since the statfs_change is happening at write_commit time and there may be multiple writes to the rindex file for one grow operation. So one consequence of this restructuring is that instead of getting one kernel message to indicate the change, you may see several. For example, before when you did a gfs2_grow, you'd get a single message like: GFS2: File system extended by 247876 blocks (968MB) Now you get something like: GFS2: File system extended by 207896 blocks (812MB) GFS2: File system extended by 39980 blocks (156MB) This version has also been successfully run against the hours-long "gfs2_fsck_hellfire" test that does several gfs2_grow and gfs2_fsck while interjecting file system damage. It does this repeatedly under a variety Resource Group conditions. Signed-off-By: Bob Peterson <rpeterso@redhat.com> Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-05-09 14:37:57 +00:00
break;
error = gfs2_internal_read(ip, &ra_state, buf, &pos,
sizeof(struct gfs2_rindex));
if (error != sizeof(struct gfs2_rindex))
break;
total_data += be32_to_cpu(((struct gfs2_rindex *)buf)->ri_data);
[GFS2] kernel changes to support new gfs2_grow command This is another revision of my gfs2 kernel patch that allows gfs2_grow to function properly. Steve Whitehouse expressed some concerns about the previous patch and I restructured it based on his comments. The previous patch was doing the statfs_change at file close time, under its own transaction. The current patch does the statfs_change inside the gfs2_commit_write function, which keeps it under the umbrella of the inode transaction. I can't call ri_update to re-read the rindex file during the transaction because the transaction may have outstanding unwritten buffers attached to the rgrps that would be otherwise blown away. So instead, I created a new function, gfs2_ri_total, that will re-read the rindex file just to total the file system space for the sake of the statfs_change. The ri_update will happen later, when gfs2 realizes the version number has changed, as it happened before my patch. Since the statfs_change is happening at write_commit time and there may be multiple writes to the rindex file for one grow operation. So one consequence of this restructuring is that instead of getting one kernel message to indicate the change, you may see several. For example, before when you did a gfs2_grow, you'd get a single message like: GFS2: File system extended by 247876 blocks (968MB) Now you get something like: GFS2: File system extended by 207896 blocks (812MB) GFS2: File system extended by 39980 blocks (156MB) This version has also been successfully run against the hours-long "gfs2_fsck_hellfire" test that does several gfs2_grow and gfs2_fsck while interjecting file system damage. It does this repeatedly under a variety Resource Group conditions. Signed-off-By: Bob Peterson <rpeterso@redhat.com> Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-05-09 14:37:57 +00:00
}
return total_data;
}
static int rgd_insert(struct gfs2_rgrpd *rgd)
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
{
struct gfs2_sbd *sdp = rgd->rd_sbd;
struct rb_node **newn = &sdp->sd_rindex_tree.rb_node, *parent = NULL;
/* Figure out where to put new node */
while (*newn) {
struct gfs2_rgrpd *cur = rb_entry(*newn, struct gfs2_rgrpd,
rd_node);
parent = *newn;
if (rgd->rd_addr < cur->rd_addr)
newn = &((*newn)->rb_left);
else if (rgd->rd_addr > cur->rd_addr)
newn = &((*newn)->rb_right);
else
return -EEXIST;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
}
rb_link_node(&rgd->rd_node, parent, newn);
rb_insert_color(&rgd->rd_node, &sdp->sd_rindex_tree);
sdp->sd_rgrps++;
return 0;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
}
/**
* read_rindex_entry - Pull in a new resource index entry from the disk
* @gl: The glock covering the rindex inode
*
* Returns: 0 on success, > 0 on EOF, error code otherwise
*/
static int read_rindex_entry(struct gfs2_inode *ip,
struct file_ra_state *ra_state)
{
struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
loff_t pos = sdp->sd_rgrps * sizeof(struct gfs2_rindex);
struct gfs2_rindex buf;
int error;
struct gfs2_rgrpd *rgd;
if (pos >= i_size_read(&ip->i_inode))
return 1;
error = gfs2_internal_read(ip, ra_state, (char *)&buf, &pos,
sizeof(struct gfs2_rindex));
if (error != sizeof(struct gfs2_rindex))
return (error == 0) ? 1 : error;
rgd = kmem_cache_zalloc(gfs2_rgrpd_cachep, GFP_NOFS);
error = -ENOMEM;
if (!rgd)
return error;
rgd->rd_sbd = sdp;
rgd->rd_addr = be64_to_cpu(buf.ri_addr);
rgd->rd_length = be32_to_cpu(buf.ri_length);
rgd->rd_data0 = be64_to_cpu(buf.ri_data0);
rgd->rd_data = be32_to_cpu(buf.ri_data);
rgd->rd_bitbytes = be32_to_cpu(buf.ri_bitbytes);
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
error = compute_bitstructs(rgd);
if (error)
goto fail;
error = gfs2_glock_get(sdp, rgd->rd_addr,
&gfs2_rgrp_glops, CREATE, &rgd->rd_gl);
if (error)
goto fail;
rgd->rd_gl->gl_object = rgd;
rgd->rd_flags &= ~GFS2_RDF_UPTODATE;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
if (rgd->rd_data > sdp->sd_max_rg_data)
sdp->sd_max_rg_data = rgd->rd_data;
spin_lock(&sdp->sd_rindex_spin);
error = rgd_insert(rgd);
spin_unlock(&sdp->sd_rindex_spin);
if (!error)
return 0;
error = 0; /* someone else read in the rgrp; free it and ignore it */
gfs2_glock_put(rgd->rd_gl);
fail:
kfree(rgd->rd_bits);
kmem_cache_free(gfs2_rgrpd_cachep, rgd);
return error;
}
/**
* gfs2_ri_update - Pull in a new resource index from the disk
* @ip: pointer to the rindex inode
*
* Returns: 0 on successful update, error code otherwise
*/
static int gfs2_ri_update(struct gfs2_inode *ip)
{
struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
struct inode *inode = &ip->i_inode;
struct file_ra_state ra_state;
int error;
file_ra_state_init(&ra_state, inode->i_mapping);
do {
error = read_rindex_entry(ip, &ra_state);
} while (error == 0);
if (error < 0)
return error;
sdp->sd_rindex_uptodate = 1;
return 0;
}
/**
* gfs2_rindex_update - Update the rindex if required
* @sdp: The GFS2 superblock
*
* We grab a lock on the rindex inode to make sure that it doesn't
* change whilst we are performing an operation. We keep this lock
* for quite long periods of time compared to other locks. This
* doesn't matter, since it is shared and it is very, very rarely
* accessed in the exclusive mode (i.e. only when expanding the filesystem).
*
* This makes sure that we're using the latest copy of the resource index
* special file, which might have been updated if someone expanded the
* filesystem (via gfs2_grow utility), which adds new resource groups.
*
* Returns: 0 on succeess, error code otherwise
*/
int gfs2_rindex_update(struct gfs2_sbd *sdp)
{
struct gfs2_inode *ip = GFS2_I(sdp->sd_rindex);
struct gfs2_glock *gl = ip->i_gl;
struct gfs2_holder ri_gh;
int error = 0;
int unlock_required = 0;
/* Read new copy from disk if we don't have the latest */
if (!sdp->sd_rindex_uptodate) {
if (!gfs2_glock_is_locked_by_me(gl)) {
error = gfs2_glock_nq_init(gl, LM_ST_SHARED, 0, &ri_gh);
if (error)
return error;
unlock_required = 1;
}
if (!sdp->sd_rindex_uptodate)
error = gfs2_ri_update(ip);
if (unlock_required)
gfs2_glock_dq_uninit(&ri_gh);
}
return error;
}
static void gfs2_rgrp_in(struct gfs2_rgrpd *rgd, const void *buf)
{
const struct gfs2_rgrp *str = buf;
u32 rg_flags;
rg_flags = be32_to_cpu(str->rg_flags);
rg_flags &= ~GFS2_RDF_MASK;
rgd->rd_flags &= GFS2_RDF_MASK;
rgd->rd_flags |= rg_flags;
rgd->rd_free = be32_to_cpu(str->rg_free);
rgd->rd_dinodes = be32_to_cpu(str->rg_dinodes);
rgd->rd_igeneration = be64_to_cpu(str->rg_igeneration);
}
static void gfs2_rgrp_out(struct gfs2_rgrpd *rgd, void *buf)
{
struct gfs2_rgrp *str = buf;
str->rg_flags = cpu_to_be32(rgd->rd_flags & ~GFS2_RDF_MASK);
str->rg_free = cpu_to_be32(rgd->rd_free);
str->rg_dinodes = cpu_to_be32(rgd->rd_dinodes);
str->__pad = cpu_to_be32(0);
str->rg_igeneration = cpu_to_be64(rgd->rd_igeneration);
memset(&str->rg_reserved, 0, sizeof(str->rg_reserved));
}
/**
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
* gfs2_rgrp_go_lock - Read in a RG's header and bitmaps
* @rgd: the struct gfs2_rgrpd describing the RG to read in
*
* Read in all of a Resource Group's header and bitmap blocks.
* Caller must eventually call gfs2_rgrp_relse() to free the bitmaps.
*
* Returns: errno
*/
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
int gfs2_rgrp_go_lock(struct gfs2_holder *gh)
{
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
struct gfs2_rgrpd *rgd = gh->gh_gl->gl_object;
struct gfs2_sbd *sdp = rgd->rd_sbd;
struct gfs2_glock *gl = rgd->rd_gl;
unsigned int length = rgd->rd_length;
struct gfs2_bitmap *bi;
unsigned int x, y;
int error;
for (x = 0; x < length; x++) {
bi = rgd->rd_bits + x;
error = gfs2_meta_read(gl, rgd->rd_addr + x, 0, &bi->bi_bh);
if (error)
goto fail;
}
for (y = length; y--;) {
bi = rgd->rd_bits + y;
error = gfs2_meta_wait(sdp, bi->bi_bh);
if (error)
goto fail;
if (gfs2_metatype_check(sdp, bi->bi_bh, y ? GFS2_METATYPE_RB :
GFS2_METATYPE_RG)) {
error = -EIO;
goto fail;
}
}
if (!(rgd->rd_flags & GFS2_RDF_UPTODATE)) {
for (x = 0; x < length; x++)
clear_bit(GBF_FULL, &rgd->rd_bits[x].bi_flags);
gfs2_rgrp_in(rgd, (rgd->rd_bits[0].bi_bh)->b_data);
rgd->rd_flags |= (GFS2_RDF_UPTODATE | GFS2_RDF_CHECK);
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
rgd->rd_free_clone = rgd->rd_free;
}
return 0;
fail:
while (x--) {
bi = rgd->rd_bits + x;
brelse(bi->bi_bh);
bi->bi_bh = NULL;
gfs2_assert_warn(sdp, !bi->bi_clone);
}
return error;
}
/**
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
* gfs2_rgrp_go_unlock - Release RG bitmaps read in with gfs2_rgrp_bh_get()
* @rgd: the struct gfs2_rgrpd describing the RG to read in
*
*/
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
void gfs2_rgrp_go_unlock(struct gfs2_holder *gh)
{
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
struct gfs2_rgrpd *rgd = gh->gh_gl->gl_object;
int x, length = rgd->rd_length;
for (x = 0; x < length; x++) {
struct gfs2_bitmap *bi = rgd->rd_bits + x;
brelse(bi->bi_bh);
bi->bi_bh = NULL;
}
}
int gfs2_rgrp_send_discards(struct gfs2_sbd *sdp, u64 offset,
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
struct buffer_head *bh,
const struct gfs2_bitmap *bi, unsigned minlen, u64 *ptrimmed)
{
struct super_block *sb = sdp->sd_vfs;
struct block_device *bdev = sb->s_bdev;
const unsigned int sects_per_blk = sdp->sd_sb.sb_bsize /
bdev_logical_block_size(sb->s_bdev);
u64 blk;
sector_t start = 0;
sector_t nr_sects = 0;
int rv;
unsigned int x;
u32 trimmed = 0;
u8 diff;
for (x = 0; x < bi->bi_len; x++) {
const u8 *clone = bi->bi_clone ? bi->bi_clone : bi->bi_bh->b_data;
clone += bi->bi_offset;
clone += x;
if (bh) {
const u8 *orig = bh->b_data + bi->bi_offset + x;
diff = ~(*orig | (*orig >> 1)) & (*clone | (*clone >> 1));
} else {
diff = ~(*clone | (*clone >> 1));
}
diff &= 0x55;
if (diff == 0)
continue;
blk = offset + ((bi->bi_start + x) * GFS2_NBBY);
blk *= sects_per_blk; /* convert to sectors */
while(diff) {
if (diff & 1) {
if (nr_sects == 0)
goto start_new_extent;
if ((start + nr_sects) != blk) {
if (nr_sects >= minlen) {
rv = blkdev_issue_discard(bdev,
start, nr_sects,
GFP_NOFS, 0);
if (rv)
goto fail;
trimmed += nr_sects;
}
nr_sects = 0;
start_new_extent:
start = blk;
}
nr_sects += sects_per_blk;
}
diff >>= 2;
blk += sects_per_blk;
}
}
if (nr_sects >= minlen) {
rv = blkdev_issue_discard(bdev, start, nr_sects, GFP_NOFS, 0);
if (rv)
goto fail;
trimmed += nr_sects;
}
if (ptrimmed)
*ptrimmed = trimmed;
return 0;
fail:
if (sdp->sd_args.ar_discard)
fs_warn(sdp, "error %d on discard request, turning discards off for this filesystem", rv);
sdp->sd_args.ar_discard = 0;
return -EIO;
}
/**
* gfs2_fitrim - Generate discard requests for unused bits of the filesystem
* @filp: Any file on the filesystem
* @argp: Pointer to the arguments (also used to pass result)
*
* Returns: 0 on success, otherwise error code
*/
int gfs2_fitrim(struct file *filp, void __user *argp)
{
struct inode *inode = filp->f_dentry->d_inode;
struct gfs2_sbd *sdp = GFS2_SB(inode);
struct request_queue *q = bdev_get_queue(sdp->sd_vfs->s_bdev);
struct buffer_head *bh;
struct gfs2_rgrpd *rgd;
struct gfs2_rgrpd *rgd_end;
struct gfs2_holder gh;
struct fstrim_range r;
int ret = 0;
u64 amt;
u64 trimmed = 0;
unsigned int x;
if (!capable(CAP_SYS_ADMIN))
return -EPERM;
if (!blk_queue_discard(q))
return -EOPNOTSUPP;
if (argp == NULL) {
r.start = 0;
r.len = ULLONG_MAX;
r.minlen = 0;
} else if (copy_from_user(&r, argp, sizeof(r)))
return -EFAULT;
ret = gfs2_rindex_update(sdp);
if (ret)
return ret;
rgd = gfs2_blk2rgrpd(sdp, r.start, 0);
rgd_end = gfs2_blk2rgrpd(sdp, r.start + r.len, 0);
while (1) {
ret = gfs2_glock_nq_init(rgd->rd_gl, LM_ST_EXCLUSIVE, 0, &gh);
if (ret)
goto out;
if (!(rgd->rd_flags & GFS2_RGF_TRIMMED)) {
/* Trim each bitmap in the rgrp */
for (x = 0; x < rgd->rd_length; x++) {
struct gfs2_bitmap *bi = rgd->rd_bits + x;
ret = gfs2_rgrp_send_discards(sdp, rgd->rd_data0, NULL, bi, r.minlen, &amt);
if (ret) {
gfs2_glock_dq_uninit(&gh);
goto out;
}
trimmed += amt;
}
/* Mark rgrp as having been trimmed */
ret = gfs2_trans_begin(sdp, RES_RG_HDR, 0);
if (ret == 0) {
bh = rgd->rd_bits[0].bi_bh;
rgd->rd_flags |= GFS2_RGF_TRIMMED;
gfs2_trans_add_bh(rgd->rd_gl, bh, 1);
gfs2_rgrp_out(rgd, bh->b_data);
gfs2_trans_end(sdp);
}
}
gfs2_glock_dq_uninit(&gh);
if (rgd == rgd_end)
break;
rgd = gfs2_rgrpd_get_next(rgd);
}
out:
r.len = trimmed << 9;
if (argp && copy_to_user(argp, &r, sizeof(r)))
return -EFAULT;
return ret;
}
/**
* gfs2_qadata_get - get the struct gfs2_qadata structure for an inode
* @ip: the incore GFS2 inode structure
*
* Returns: the struct gfs2_qadata
*/
struct gfs2_qadata *gfs2_qadata_get(struct gfs2_inode *ip)
{
struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
int error;
BUG_ON(ip->i_qadata != NULL);
ip->i_qadata = kzalloc(sizeof(struct gfs2_qadata), GFP_NOFS);
error = gfs2_rindex_update(sdp);
if (error)
fs_warn(sdp, "rindex update returns %d\n", error);
return ip->i_qadata;
}
/**
* gfs2_blkrsv_get - get the struct gfs2_blkreserv structure for an inode
* @ip: the incore GFS2 inode structure
*
* Returns: the struct gfs2_qadata
*/
static int gfs2_blkrsv_get(struct gfs2_inode *ip)
{
BUG_ON(ip->i_res != NULL);
ip->i_res = kmem_cache_zalloc(gfs2_rsrv_cachep, GFP_NOFS);
if (!ip->i_res)
return -ENOMEM;
return 0;
}
/**
* try_rgrp_fit - See if a given reservation will fit in a given RG
* @rgd: the RG data
* @ip: the inode
*
* If there's room for the requested blocks to be allocated from the RG:
*
* Returns: 1 on success (it fits), 0 on failure (it doesn't fit)
*/
static int try_rgrp_fit(const struct gfs2_rgrpd *rgd, const struct gfs2_inode *ip)
{
const struct gfs2_blkreserv *rs = ip->i_res;
if (rgd->rd_flags & (GFS2_RGF_NOALLOC | GFS2_RDF_ERROR))
return 0;
if (rgd->rd_free_clone >= rs->rs_requested)
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
return 1;
return 0;
}
static inline u32 gfs2_bi2rgd_blk(struct gfs2_bitmap *bi, u32 blk)
{
return (bi->bi_start * GFS2_NBBY) + blk;
}
/**
* try_rgrp_unlink - Look for any unlinked, allocated, but unused inodes
* @rgd: The rgrp
*
* Returns: 0 if no error
* The inode, if one has been found, in inode.
*/
static void try_rgrp_unlink(struct gfs2_rgrpd *rgd, u64 *last_unlinked, u64 skip)
{
u32 goal = 0, block;
u64 no_addr;
struct gfs2_sbd *sdp = rgd->rd_sbd;
struct gfs2_glock *gl;
struct gfs2_inode *ip;
int error;
int found = 0;
struct gfs2_bitmap *bi;
while (goal < rgd->rd_data) {
down_write(&sdp->sd_log_flush_lock);
block = rgblk_search(rgd, goal, GFS2_BLKST_UNLINKED, &bi);
up_write(&sdp->sd_log_flush_lock);
if (block == BFITNOENT)
break;
block = gfs2_bi2rgd_blk(bi, block);
/* rgblk_search can return a block < goal, so we need to
keep it marching forward. */
no_addr = block + rgd->rd_data0;
goal = max(block + 1, goal + 1);
if (*last_unlinked != NO_BLOCK && no_addr <= *last_unlinked)
continue;
if (no_addr == skip)
continue;
*last_unlinked = no_addr;
error = gfs2_glock_get(sdp, no_addr, &gfs2_inode_glops, CREATE, &gl);
if (error)
continue;
/* If the inode is already in cache, we can ignore it here
* because the existing inode disposal code will deal with
* it when all refs have gone away. Accessing gl_object like
* this is not safe in general. Here it is ok because we do
* not dereference the pointer, and we only need an approx
* answer to whether it is NULL or not.
*/
ip = gl->gl_object;
if (ip || queue_work(gfs2_delete_workqueue, &gl->gl_delete) == 0)
gfs2_glock_put(gl);
else
found++;
/* Limit reclaim to sensible number of tasks */
if (found > NR_CPUS)
return;
}
rgd->rd_flags &= ~GFS2_RDF_CHECK;
return;
}
/**
* get_local_rgrp - Choose and lock a rgrp for allocation
* @ip: the inode to reserve space for
* @rgp: the chosen and locked rgrp
*
* Try to acquire rgrp in way which avoids contending with others.
*
* Returns: errno
*/
static int get_local_rgrp(struct gfs2_inode *ip, u64 *last_unlinked)
{
struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
struct gfs2_rgrpd *rgd, *begin = NULL;
struct gfs2_blkreserv *rs = ip->i_res;
int error, rg_locked, flags = LM_FLAG_TRY;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
int loops = 0;
if (ip->i_rgd && rgrp_contains_block(ip->i_rgd, ip->i_goal))
rgd = begin = ip->i_rgd;
else
rgd = begin = gfs2_blk2rgrpd(sdp, ip->i_goal, 1);
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
if (rgd == NULL)
return -EBADSLT;
while (loops < 3) {
rg_locked = 0;
if (gfs2_glock_is_locked_by_me(rgd->rd_gl)) {
rg_locked = 1;
error = 0;
} else {
error = gfs2_glock_nq_init(rgd->rd_gl, LM_ST_EXCLUSIVE,
flags, &rs->rs_rgd_gh);
}
switch (error) {
case 0:
if (try_rgrp_fit(rgd, ip)) {
ip->i_rgd = rgd;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
return 0;
}
if (rgd->rd_flags & GFS2_RDF_CHECK)
try_rgrp_unlink(rgd, last_unlinked, ip->i_no_addr);
if (!rg_locked)
gfs2_glock_dq_uninit(&rs->rs_rgd_gh);
/* fall through */
case GLR_TRYFAILED:
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
rgd = gfs2_rgrpd_get_next(rgd);
if (rgd == begin) {
flags = 0;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
loops++;
}
break;
default:
return error;
}
}
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
return -ENOSPC;
}
static void gfs2_blkrsv_put(struct gfs2_inode *ip)
{
BUG_ON(ip->i_res == NULL);
kmem_cache_free(gfs2_rsrv_cachep, ip->i_res);
ip->i_res = NULL;
}
/**
* gfs2_inplace_reserve - Reserve space in the filesystem
* @ip: the inode to reserve space for
*
* Returns: errno
*/
int gfs2_inplace_reserve(struct gfs2_inode *ip, u32 requested)
{
struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
struct gfs2_blkreserv *rs;
int error;
u64 last_unlinked = NO_BLOCK;
int tries = 0;
error = gfs2_blkrsv_get(ip);
if (error)
return error;
rs = ip->i_res;
rs->rs_requested = requested;
if (gfs2_assert_warn(sdp, requested)) {
error = -EINVAL;
goto out;
}
do {
error = get_local_rgrp(ip, &last_unlinked);
if (error != -ENOSPC)
break;
/* Check that fs hasn't grown if writing to rindex */
if (ip == GFS2_I(sdp->sd_rindex) && !sdp->sd_rindex_uptodate) {
error = gfs2_ri_update(ip);
if (error)
break;
continue;
}
/* Flushing the log may release space */
gfs2_log_flush(sdp, NULL);
} while (tries++ < 3);
out:
if (error)
gfs2_blkrsv_put(ip);
return error;
}
/**
* gfs2_inplace_release - release an inplace reservation
* @ip: the inode the reservation was taken out on
*
* Release a reservation made by gfs2_inplace_reserve().
*/
void gfs2_inplace_release(struct gfs2_inode *ip)
{
struct gfs2_blkreserv *rs = ip->i_res;
if (rs->rs_rgd_gh.gh_gl)
gfs2_glock_dq_uninit(&rs->rs_rgd_gh);
gfs2_blkrsv_put(ip);
}
/**
* gfs2_get_block_type - Check a block in a RG is of given type
* @rgd: the resource group holding the block
* @block: the block number
*
* Returns: The block type (GFS2_BLKST_*)
*/
static unsigned char gfs2_get_block_type(struct gfs2_rgrpd *rgd, u64 block)
{
struct gfs2_bitmap *bi = NULL;
u32 length, rgrp_block, buf_block;
unsigned int buf;
unsigned char type;
length = rgd->rd_length;
rgrp_block = block - rgd->rd_data0;
for (buf = 0; buf < length; buf++) {
bi = rgd->rd_bits + buf;
if (rgrp_block < (bi->bi_start + bi->bi_len) * GFS2_NBBY)
break;
}
gfs2_assert(rgd->rd_sbd, buf < length);
buf_block = rgrp_block - bi->bi_start * GFS2_NBBY;
type = gfs2_testbit(rgd, bi->bi_bh->b_data + bi->bi_offset,
bi->bi_len, buf_block);
return type;
}
/**
* rgblk_search - find a block in @state
* @rgd: the resource group descriptor
* @goal: the goal block within the RG (start here to search for avail block)
* @state: GFS2_BLKST_XXX the before-allocation state to find
* @dinode: TRUE if the first block we allocate is for a dinode
* @rbi: address of the pointer to the bitmap containing the block found
*
* Walk rgrp's bitmap to find bits that represent a block in @state.
*
* This function never fails, because we wouldn't call it unless we
* know (from reservation results, etc.) that a block is available.
*
* Scope of @goal is just within rgrp, not the whole filesystem.
* Scope of @returned block is just within bitmap, not the whole filesystem.
*
* Returns: the block number found relative to the bitmap rbi
*/
static u32 rgblk_search(struct gfs2_rgrpd *rgd, u32 goal,
unsigned char state,
struct gfs2_bitmap **rbi)
{
struct gfs2_bitmap *bi = NULL;
const u32 length = rgd->rd_length;
u32 blk = BFITNOENT;
unsigned int buf, x;
const u8 *buffer = NULL;
*rbi = NULL;
/* Find bitmap block that contains bits for goal block */
for (buf = 0; buf < length; buf++) {
bi = rgd->rd_bits + buf;
/* Convert scope of "goal" from rgrp-wide to within found bit block */
if (goal < (bi->bi_start + bi->bi_len) * GFS2_NBBY) {
goal -= bi->bi_start * GFS2_NBBY;
goto do_search;
}
}
buf = 0;
goal = 0;
do_search:
/* Search (up to entire) bitmap in this rgrp for allocatable block.
"x <= length", instead of "x < length", because we typically start
the search in the middle of a bit block, but if we can't find an
allocatable block anywhere else, we want to be able wrap around and
search in the first part of our first-searched bit block. */
for (x = 0; x <= length; x++) {
bi = rgd->rd_bits + buf;
if (test_bit(GBF_FULL, &bi->bi_flags) &&
(state == GFS2_BLKST_FREE))
goto skip;
/* The GFS2_BLKST_UNLINKED state doesn't apply to the clone
bitmaps, so we must search the originals for that. */
buffer = bi->bi_bh->b_data + bi->bi_offset;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
WARN_ON(!buffer_uptodate(bi->bi_bh));
if (state != GFS2_BLKST_UNLINKED && bi->bi_clone)
buffer = bi->bi_clone + bi->bi_offset;
blk = gfs2_bitfit(buffer, bi->bi_len, goal, state);
if (blk != BFITNOENT)
break;
if ((goal == 0) && (state == GFS2_BLKST_FREE))
set_bit(GBF_FULL, &bi->bi_flags);
/* Try next bitmap block (wrap back to rgrp header if at end) */
skip:
buf++;
buf %= length;
goal = 0;
}
if (blk != BFITNOENT)
*rbi = bi;
GFS2: Use rbtree for resource groups and clean up bitmap buffer ref count scheme Here is an update of Bob's original rbtree patch which, in addition, also resolves the rather strange ref counting that was being done relating to the bitmap blocks. Originally we had a dual system for journaling resource groups. The metadata blocks were journaled and also the rgrp itself was added to a list. The reason for adding the rgrp to the list in the journal was so that the "repolish clones" code could be run to update the free space, and potentially send any discard requests when the log was flushed. This was done by comparing the "cloned" bitmap with what had been written back on disk during the transaction commit. Due to this, there was a requirement to hang on to the rgrps' bitmap buffers until the journal had been flushed. For that reason, there was a rather complicated set up in the ->go_lock ->go_unlock functions for rgrps involving both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference count on the buffers. However, the journal maintains a reference count on the buffers anyway, since they are being journaled as metadata buffers. So by moving the code which deals with the post-journal accounting for bitmap blocks to the metadata journaling code, we can entirely dispense with the rather strange buffer ref counting scheme and also the requirement to journal the rgrps. The net result of all this is that the ->sd_rindex_spin is left to do exactly one job, and that is to look after the rbtree or rgrps. This patch is designed to be a stepping stone towards using RCU for the rbtree of resource groups, however the reduction in the number of uses of the ->sd_rindex_spin is likely to have benefits for multi-threaded workloads, anyway. The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also be removed in future in favour of calling the functions directly where required in the code. That will allow locking of resource groups without needing to actually read them in - something that could be useful in speeding up statfs. In the mean time though it is valid to dereference ->bi_bh only when the rgrp is locked. This is basically the same rule as before, modulo the references not being valid until the following journal flush. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com> Cc: Benjamin Marzinski <bmarzins@redhat.com>
2011-08-31 08:53:19 +00:00
return blk;
}
/**
* gfs2_alloc_extent - allocate an extent from a given bitmap
* @rgd: the resource group descriptor
* @bi: the bitmap within the rgrp
* @blk: the block within the bitmap
* @dinode: TRUE if the first block we allocate is for a dinode
* @n: The extent length
*
* Add the found bitmap buffer to the transaction.
* Set the found bits to @new_state to change block's allocation state.
* Returns: starting block number of the extent (fs scope)
*/
static u64 gfs2_alloc_extent(struct gfs2_rgrpd *rgd, struct gfs2_bitmap *bi,
u32 blk, bool dinode, unsigned int *n)
{
const unsigned int elen = *n;
u32 goal;
const u8 *buffer = NULL;
*n = 0;
buffer = bi->bi_bh->b_data + bi->bi_offset;
gfs2_trans_add_bh(rgd->rd_gl, bi->bi_bh, 1);
gfs2_setbit(rgd, bi->bi_bh->b_data, bi->bi_clone, bi->bi_offset,
bi, blk, dinode ? GFS2_BLKST_DINODE : GFS2_BLKST_USED);
(*n)++;
goal = blk;
while (*n < elen) {
goal++;
if (goal >= (bi->bi_len * GFS2_NBBY))
break;
if (gfs2_testbit(rgd, buffer, bi->bi_len, goal) !=
GFS2_BLKST_FREE)
break;
gfs2_setbit(rgd, bi->bi_bh->b_data, bi->bi_clone, bi->bi_offset,
bi, goal, GFS2_BLKST_USED);
(*n)++;
}
blk = gfs2_bi2rgd_blk(bi, blk);
rgd->rd_last_alloc = blk + *n - 1;
return rgd->rd_data0 + blk;
}
/**
* rgblk_free - Change alloc state of given block(s)
* @sdp: the filesystem
* @bstart: the start of a run of blocks to free
* @blen: the length of the block run (all must lie within ONE RG!)
* @new_state: GFS2_BLKST_XXX the after-allocation block state
*
* Returns: Resource group containing the block(s)
*/
static struct gfs2_rgrpd *rgblk_free(struct gfs2_sbd *sdp, u64 bstart,
u32 blen, unsigned char new_state)
{
struct gfs2_rgrpd *rgd;
struct gfs2_bitmap *bi = NULL;
u32 length, rgrp_blk, buf_blk;
unsigned int buf;
rgd = gfs2_blk2rgrpd(sdp, bstart, 1);
if (!rgd) {
if (gfs2_consist(sdp))
fs_err(sdp, "block = %llu\n", (unsigned long long)bstart);
return NULL;
}
length = rgd->rd_length;
rgrp_blk = bstart - rgd->rd_data0;
while (blen--) {
for (buf = 0; buf < length; buf++) {
bi = rgd->rd_bits + buf;
if (rgrp_blk < (bi->bi_start + bi->bi_len) * GFS2_NBBY)
break;
}
gfs2_assert(rgd->rd_sbd, buf < length);
buf_blk = rgrp_blk - bi->bi_start * GFS2_NBBY;
rgrp_blk++;
if (!bi->bi_clone) {
bi->bi_clone = kmalloc(bi->bi_bh->b_size,
GFP_NOFS | __GFP_NOFAIL);
memcpy(bi->bi_clone + bi->bi_offset,
bi->bi_bh->b_data + bi->bi_offset,
bi->bi_len);
}
gfs2_trans_add_bh(rgd->rd_gl, bi->bi_bh, 1);
gfs2_setbit(rgd, bi->bi_bh->b_data, NULL, bi->bi_offset,
bi, buf_blk, new_state);
}
return rgd;
}
/**
* gfs2_rgrp_dump - print out an rgrp
* @seq: The iterator
* @gl: The glock in question
*
*/
int gfs2_rgrp_dump(struct seq_file *seq, const struct gfs2_glock *gl)
{
const struct gfs2_rgrpd *rgd = gl->gl_object;
if (rgd == NULL)
return 0;
gfs2_print_dbg(seq, " R: n:%llu f:%02x b:%u/%u i:%u\n",
(unsigned long long)rgd->rd_addr, rgd->rd_flags,
rgd->rd_free, rgd->rd_free_clone, rgd->rd_dinodes);
return 0;
}
static void gfs2_rgrp_error(struct gfs2_rgrpd *rgd)
{
struct gfs2_sbd *sdp = rgd->rd_sbd;
fs_warn(sdp, "rgrp %llu has an error, marking it readonly until umount\n",
(unsigned long long)rgd->rd_addr);
fs_warn(sdp, "umount on all nodes and run fsck.gfs2 to fix the error\n");
gfs2_rgrp_dump(NULL, rgd->rd_gl);
rgd->rd_flags |= GFS2_RDF_ERROR;
}
/**
* gfs2_alloc_blocks - Allocate one or more blocks of data and/or a dinode
* @ip: the inode to allocate the block for
* @bn: Used to return the starting block number
* @ndata: requested number of blocks/extent length (value/result)
* @dinode: 1 if we're allocating a dinode block, else 0
* @generation: the generation number of the inode
*
* Returns: 0 or error
*/
int gfs2_alloc_blocks(struct gfs2_inode *ip, u64 *bn, unsigned int *nblocks,
bool dinode, u64 *generation)
{
struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
struct buffer_head *dibh;
struct gfs2_rgrpd *rgd;
unsigned int ndata;
u32 goal, blk; /* block, within the rgrp scope */
u64 block; /* block, within the file system scope */
int error;
struct gfs2_bitmap *bi;
/* Only happens if there is a bug in gfs2, return something distinctive
* to ensure that it is noticed.
*/
if (ip->i_res == NULL)
return -ECANCELED;
rgd = ip->i_rgd;
if (!dinode && rgrp_contains_block(rgd, ip->i_goal))
goal = ip->i_goal - rgd->rd_data0;
else
goal = rgd->rd_last_alloc;
blk = rgblk_search(rgd, goal, GFS2_BLKST_FREE, &bi);
/* Since all blocks are reserved in advance, this shouldn't happen */
if (blk == BFITNOENT)
goto rgrp_error;
block = gfs2_alloc_extent(rgd, bi, blk, dinode, nblocks);
ndata = *nblocks;
if (dinode)
ndata--;
if (!dinode) {
ip->i_goal = block + ndata - 1;
error = gfs2_meta_inode_buffer(ip, &dibh);
if (error == 0) {
struct gfs2_dinode *di =
(struct gfs2_dinode *)dibh->b_data;
gfs2_trans_add_bh(ip->i_gl, dibh, 1);
di->di_goal_meta = di->di_goal_data =
cpu_to_be64(ip->i_goal);
brelse(dibh);
}
}
if (rgd->rd_free < *nblocks)
goto rgrp_error;
rgd->rd_free -= *nblocks;
if (dinode) {
rgd->rd_dinodes++;
*generation = rgd->rd_igeneration++;
if (*generation == 0)
*generation = rgd->rd_igeneration++;
}
gfs2_trans_add_bh(rgd->rd_gl, rgd->rd_bits[0].bi_bh, 1);
gfs2_rgrp_out(rgd, rgd->rd_bits[0].bi_bh->b_data);
gfs2_statfs_change(sdp, 0, -(s64)*nblocks, dinode ? 1 : 0);
if (dinode)
gfs2_trans_add_unrevoke(sdp, block, 1);
/*
* This needs reviewing to see why we cannot do the quota change
* at this point in the dinode case.
*/
if (ndata)
gfs2_quota_change(ip, ndata, ip->i_inode.i_uid,
ip->i_inode.i_gid);
rgd->rd_free_clone -= *nblocks;
trace_gfs2_block_alloc(ip, block, *nblocks,
dinode ? GFS2_BLKST_DINODE : GFS2_BLKST_USED);
*bn = block;
return 0;
rgrp_error:
gfs2_rgrp_error(rgd);
return -EIO;
}
/**
* __gfs2_free_blocks - free a contiguous run of block(s)
* @ip: the inode these blocks are being freed from
* @bstart: first block of a run of contiguous blocks
* @blen: the length of the block run
* @meta: 1 if the blocks represent metadata
*
*/
void __gfs2_free_blocks(struct gfs2_inode *ip, u64 bstart, u32 blen, int meta)
{
struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
struct gfs2_rgrpd *rgd;
rgd = rgblk_free(sdp, bstart, blen, GFS2_BLKST_FREE);
if (!rgd)
return;
trace_gfs2_block_alloc(ip, bstart, blen, GFS2_BLKST_FREE);
rgd->rd_free += blen;
rgd->rd_flags &= ~GFS2_RGF_TRIMMED;
gfs2_trans_add_bh(rgd->rd_gl, rgd->rd_bits[0].bi_bh, 1);
gfs2_rgrp_out(rgd, rgd->rd_bits[0].bi_bh->b_data);
/* Directories keep their data in the metadata address space */
if (meta || ip->i_depth)
gfs2_meta_wipe(ip, bstart, blen);
GFS2: deallocation performance patch This patch is a performance improvement to GFS2's dealloc code. Rather than update the quota file and statfs file for every single block that's stripped off in unlink function do_strip, this patch keeps track and updates them once for every layer that's stripped. This is done entirely inside the existing transaction, so there should be no risk of corruption. The other functions that deallocate blocks will be unaffected because they are using wrapper functions that do the same thing that they do today. I tested this code on my roth cluster by creating 200 files in a directory, each of which is 100MB, then on four nodes, I simultaneously deleted the files, thus competing for GFS2 resources (but different files). The commands I used were: [root@roth-01]# time for i in `seq 1 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done [root@roth-02]# time for i in `seq 2 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done [root@roth-03]# time for i in `seq 3 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done [root@roth-05]# time for i in `seq 4 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done The performance increase was significant: roth-01 roth-02 roth-03 roth-05 --------- --------- --------- --------- old: real 0m34.027 0m25.021s 0m23.906s 0m35.646s new: real 0m22.379s 0m24.362s 0m24.133s 0m18.562s Total time spent deleting: old: 118.6s new: 89.4 For this particular case, this showed a 25% performance increase for GFS2 unlinks. Signed-off-by: Bob Peterson <rpeterso@redhat.com> Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2011-02-23 21:11:33 +00:00
}
GFS2: deallocation performance patch This patch is a performance improvement to GFS2's dealloc code. Rather than update the quota file and statfs file for every single block that's stripped off in unlink function do_strip, this patch keeps track and updates them once for every layer that's stripped. This is done entirely inside the existing transaction, so there should be no risk of corruption. The other functions that deallocate blocks will be unaffected because they are using wrapper functions that do the same thing that they do today. I tested this code on my roth cluster by creating 200 files in a directory, each of which is 100MB, then on four nodes, I simultaneously deleted the files, thus competing for GFS2 resources (but different files). The commands I used were: [root@roth-01]# time for i in `seq 1 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done [root@roth-02]# time for i in `seq 2 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done [root@roth-03]# time for i in `seq 3 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done [root@roth-05]# time for i in `seq 4 4 200` ; do rm /mnt/gfs2/bigdir/gfs2.$i; done The performance increase was significant: roth-01 roth-02 roth-03 roth-05 --------- --------- --------- --------- old: real 0m34.027 0m25.021s 0m23.906s 0m35.646s new: real 0m22.379s 0m24.362s 0m24.133s 0m18.562s Total time spent deleting: old: 118.6s new: 89.4 For this particular case, this showed a 25% performance increase for GFS2 unlinks. Signed-off-by: Bob Peterson <rpeterso@redhat.com> Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2011-02-23 21:11:33 +00:00
/**
* gfs2_free_meta - free a contiguous run of data block(s)
* @ip: the inode these blocks are being freed from
* @bstart: first block of a run of contiguous blocks
* @blen: the length of the block run
*
*/
void gfs2_free_meta(struct gfs2_inode *ip, u64 bstart, u32 blen)
{
struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
__gfs2_free_blocks(ip, bstart, blen, 1);
gfs2_statfs_change(sdp, 0, +blen, 0);
gfs2_quota_change(ip, -(s64)blen, ip->i_inode.i_uid, ip->i_inode.i_gid);
}
void gfs2_unlink_di(struct inode *inode)
{
struct gfs2_inode *ip = GFS2_I(inode);
struct gfs2_sbd *sdp = GFS2_SB(inode);
struct gfs2_rgrpd *rgd;
u64 blkno = ip->i_no_addr;
rgd = rgblk_free(sdp, blkno, 1, GFS2_BLKST_UNLINKED);
if (!rgd)
return;
trace_gfs2_block_alloc(ip, blkno, 1, GFS2_BLKST_UNLINKED);
gfs2_trans_add_bh(rgd->rd_gl, rgd->rd_bits[0].bi_bh, 1);
gfs2_rgrp_out(rgd, rgd->rd_bits[0].bi_bh->b_data);
}
static void gfs2_free_uninit_di(struct gfs2_rgrpd *rgd, u64 blkno)
{
struct gfs2_sbd *sdp = rgd->rd_sbd;
struct gfs2_rgrpd *tmp_rgd;
tmp_rgd = rgblk_free(sdp, blkno, 1, GFS2_BLKST_FREE);
if (!tmp_rgd)
return;
gfs2_assert_withdraw(sdp, rgd == tmp_rgd);
if (!rgd->rd_dinodes)
gfs2_consist_rgrpd(rgd);
rgd->rd_dinodes--;
rgd->rd_free++;
gfs2_trans_add_bh(rgd->rd_gl, rgd->rd_bits[0].bi_bh, 1);
gfs2_rgrp_out(rgd, rgd->rd_bits[0].bi_bh->b_data);
gfs2_statfs_change(sdp, 0, +1, -1);
}
void gfs2_free_di(struct gfs2_rgrpd *rgd, struct gfs2_inode *ip)
{
gfs2_free_uninit_di(rgd, ip->i_no_addr);
trace_gfs2_block_alloc(ip, ip->i_no_addr, 1, GFS2_BLKST_FREE);
gfs2_quota_change(ip, -1, ip->i_inode.i_uid, ip->i_inode.i_gid);
gfs2_meta_wipe(ip, ip->i_no_addr, 1);
}
/**
* gfs2_check_blk_type - Check the type of a block
* @sdp: The superblock
* @no_addr: The block number to check
* @type: The block type we are looking for
*
* Returns: 0 if the block type matches the expected type
* -ESTALE if it doesn't match
* or -ve errno if something went wrong while checking
*/
int gfs2_check_blk_type(struct gfs2_sbd *sdp, u64 no_addr, unsigned int type)
{
struct gfs2_rgrpd *rgd;
struct gfs2_holder rgd_gh;
int error = -EINVAL;
rgd = gfs2_blk2rgrpd(sdp, no_addr, 1);
if (!rgd)
goto fail;
error = gfs2_glock_nq_init(rgd->rd_gl, LM_ST_SHARED, 0, &rgd_gh);
if (error)
goto fail;
if (gfs2_get_block_type(rgd, no_addr) != type)
error = -ESTALE;
gfs2_glock_dq_uninit(&rgd_gh);
fail:
return error;
}
/**
* gfs2_rlist_add - add a RG to a list of RGs
* @ip: the inode
* @rlist: the list of resource groups
* @block: the block
*
* Figure out what RG a block belongs to and add that RG to the list
*
* FIXME: Don't use NOFAIL
*
*/
void gfs2_rlist_add(struct gfs2_inode *ip, struct gfs2_rgrp_list *rlist,
u64 block)
{
struct gfs2_sbd *sdp = GFS2_SB(&ip->i_inode);
struct gfs2_rgrpd *rgd;
struct gfs2_rgrpd **tmp;
unsigned int new_space;
unsigned int x;
if (gfs2_assert_warn(sdp, !rlist->rl_ghs))
return;
if (ip->i_rgd && rgrp_contains_block(ip->i_rgd, block))
rgd = ip->i_rgd;
else
rgd = gfs2_blk2rgrpd(sdp, block, 1);
if (!rgd) {
fs_err(sdp, "rlist_add: no rgrp for block %llu\n", (unsigned long long)block);
return;
}
ip->i_rgd = rgd;
for (x = 0; x < rlist->rl_rgrps; x++)
if (rlist->rl_rgd[x] == rgd)
return;
if (rlist->rl_rgrps == rlist->rl_space) {
new_space = rlist->rl_space + 10;
tmp = kcalloc(new_space, sizeof(struct gfs2_rgrpd *),
GFP_NOFS | __GFP_NOFAIL);
if (rlist->rl_rgd) {
memcpy(tmp, rlist->rl_rgd,
rlist->rl_space * sizeof(struct gfs2_rgrpd *));
kfree(rlist->rl_rgd);
}
rlist->rl_space = new_space;
rlist->rl_rgd = tmp;
}
rlist->rl_rgd[rlist->rl_rgrps++] = rgd;
}
/**
* gfs2_rlist_alloc - all RGs have been added to the rlist, now allocate
* and initialize an array of glock holders for them
* @rlist: the list of resource groups
* @state: the lock state to acquire the RG lock in
* @flags: the modifier flags for the holder structures
*
* FIXME: Don't use NOFAIL
*
*/
void gfs2_rlist_alloc(struct gfs2_rgrp_list *rlist, unsigned int state)
{
unsigned int x;
rlist->rl_ghs = kcalloc(rlist->rl_rgrps, sizeof(struct gfs2_holder),
GFP_NOFS | __GFP_NOFAIL);
for (x = 0; x < rlist->rl_rgrps; x++)
gfs2_holder_init(rlist->rl_rgd[x]->rd_gl,
state, 0,
&rlist->rl_ghs[x]);
}
/**
* gfs2_rlist_free - free a resource group list
* @list: the list of resource groups
*
*/
void gfs2_rlist_free(struct gfs2_rgrp_list *rlist)
{
unsigned int x;
kfree(rlist->rl_rgd);
if (rlist->rl_ghs) {
for (x = 0; x < rlist->rl_rgrps; x++)
gfs2_holder_uninit(&rlist->rl_ghs[x]);
kfree(rlist->rl_ghs);
}
}