linux/fs/gfs2/incore.h

866 lines
23 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.
*/
#ifndef __INCORE_DOT_H__
#define __INCORE_DOT_H__
#include <linux/fs.h>
#include <linux/kobject.h>
#include <linux/workqueue.h>
#include <linux/dlm.h>
#include <linux/buffer_head.h>
#include <linux/rcupdate.h>
#include <linux/rculist_bl.h>
#include <linux/completion.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>
GFS2: glock statistics gathering The stats are divided into two sets: those relating to the super block and those relating to an individual glock. The super block stats are done on a per cpu basis in order to try and reduce the overhead of gathering them. They are also further divided by glock type. In the case of both the super block and glock statistics, the same information is gathered in each case. The super block statistics are used to provide default values for most of the glock statistics, so that newly created glocks should have, as far as possible, a sensible starting point. The statistics are divided into three pairs of mean and variance, plus two counters. The mean/variance pairs are smoothed exponential estimates and the algorithm used is one which will be very familiar to those used to calculation of round trip times in network code. The three pairs of mean/variance measure the following things: 1. DLM lock time (non-blocking requests) 2. DLM lock time (blocking requests) 3. Inter-request time (again to the DLM) A non-blocking request is one which will complete right away, whatever the state of the DLM lock in question. That currently means any requests when (a) the current state of the lock is exclusive (b) the requested state is either null or unlocked or (c) the "try lock" flag is set. A blocking request covers all the other lock requests. There are two counters. The first is there primarily to show how many lock requests have been made, and thus how much data has gone into the mean/variance calculations. The other counter is counting queueing of holders at the top layer of the glock code. Hopefully that number will be a lot larger than the number of dlm lock requests issued. So why gather these statistics? There are several reasons we'd like to get a better idea of these timings: 1. To be able to better set the glock "min hold time" 2. To spot performance issues more easily 3. To improve the algorithm for selecting resource groups for allocation (to base it on lock wait time, rather than blindly using a "try lock") Due to the smoothing action of the updates, a step change in some input quantity being sampled will only fully be taken into account after 8 samples (or 4 for the variance) and this needs to be carefully considered when interpreting the results. Knowing both the time it takes a lock request to complete and the average time between lock requests for a glock means we can compute the total percentage of the time for which the node is able to use a glock vs. time that the rest of the cluster has its share. That will be very useful when setting the lock min hold time. The other point to remember is that all times are in nanoseconds. Great care has been taken to ensure that we measure exactly the quantities that we want, as accurately as possible. There are always inaccuracies in any measuring system, but I hope this is as accurate as we can reasonably make it. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-01-20 10:38:36 +00:00
#include <linux/ktime.h>
#include <linux/percpu.h>
#include <linux/lockref.h>
#include <linux/rhashtable.h>
#define DIO_WAIT 0x00000010
#define DIO_METADATA 0x00000020
struct gfs2_log_operations;
struct gfs2_bufdata;
struct gfs2_holder;
struct gfs2_glock;
struct gfs2_quota_data;
struct gfs2_trans;
struct gfs2_jdesc;
struct gfs2_sbd;
struct lm_lockops;
typedef void (*gfs2_glop_bh_t) (struct gfs2_glock *gl, unsigned int ret);
struct gfs2_log_header_host {
u64 lh_sequence; /* Sequence number of this transaction */
u32 lh_flags; /* GFS2_LOG_HEAD_... */
u32 lh_tail; /* Block number of log tail */
u32 lh_blkno;
};
/*
* Structure of operations that are associated with each
* type of element in the log.
*/
struct gfs2_log_operations {
void (*lo_before_commit) (struct gfs2_sbd *sdp, struct gfs2_trans *tr);
void (*lo_after_commit) (struct gfs2_sbd *sdp, struct gfs2_trans *tr);
void (*lo_before_scan) (struct gfs2_jdesc *jd,
struct gfs2_log_header_host *head, int pass);
int (*lo_scan_elements) (struct gfs2_jdesc *jd, unsigned int start,
struct gfs2_log_descriptor *ld, __be64 *ptr,
int pass);
void (*lo_after_scan) (struct gfs2_jdesc *jd, int error, int pass);
const char *lo_name;
};
#define GBF_FULL 1
struct gfs2_bitmap {
struct buffer_head *bi_bh;
char *bi_clone;
unsigned long bi_flags;
u32 bi_offset;
u32 bi_start;
u32 bi_len;
u32 bi_blocks;
};
struct gfs2_rgrpd {
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 rd_node; /* Link with superblock */
struct gfs2_glock *rd_gl; /* Glock for this rgrp */
u64 rd_addr; /* grp block disk address */
u64 rd_data0; /* first data location */
u32 rd_length; /* length of rgrp header in fs blocks */
u32 rd_data; /* num of data blocks in rgrp */
u32 rd_bitbytes; /* number of bytes in data bitmaps */
u32 rd_free;
u32 rd_reserved; /* number of blocks reserved */
u32 rd_free_clone;
u32 rd_dinodes;
u64 rd_igeneration;
struct gfs2_bitmap *rd_bits;
struct gfs2_sbd *rd_sbd;
struct gfs2_rgrp_lvb *rd_rgl;
u32 rd_last_alloc;
u32 rd_flags;
u32 rd_extfail_pt; /* extent failure point */
#define GFS2_RDF_CHECK 0x10000000 /* check for unlinked inodes */
#define GFS2_RDF_UPTODATE 0x20000000 /* rg is up to date */
#define GFS2_RDF_ERROR 0x40000000 /* error in rg */
#define GFS2_RDF_PREFERRED 0x80000000 /* This rgrp is preferred */
#define GFS2_RDF_MASK 0xf0000000 /* mask for internal flags */
spinlock_t rd_rsspin; /* protects reservation related vars */
struct rb_root rd_rstree; /* multi-block reservation tree */
};
struct gfs2_rbm {
struct gfs2_rgrpd *rgd;
u32 offset; /* The offset is bitmap relative */
int bii; /* Bitmap index */
};
static inline struct gfs2_bitmap *rbm_bi(const struct gfs2_rbm *rbm)
{
return rbm->rgd->rd_bits + rbm->bii;
}
static inline u64 gfs2_rbm_to_block(const struct gfs2_rbm *rbm)
{
BUG_ON(rbm->offset >= rbm->rgd->rd_data);
return rbm->rgd->rd_data0 + (rbm_bi(rbm)->bi_start * GFS2_NBBY) +
rbm->offset;
}
GFS2: Replace rgblk_search with gfs2_rbm_find This is part of a series of patches which are introducing the gfs2_rbm structure throughout the block allocation code. The main aim of this part is to create a search function which can deal directly with struct gfs2_rbm. In this case it specifies the initial position at which to start the search and also the point at which the search terminates. The net result of this is to clean up the search code and make it rather more readable, and the various possible exceptions which may occur during the search are partitioned into their own functions. There are some bug fixes too. We should not be checking the reservations while allocating extents - the time for that is when we are searching for where to put the extent, not when we've already made that decision. Also, rgblk_search had two uses, and in only one of those cases did it make sense to check for reservations. This is fixed in the new gfs2_rbm_find function, which has a cleaner interface. The reservation checking has been improved by always checking for contiguous reservations, and returning the first free block after all contiguous reservations. This is done under the spin lock to ensure consistancy of the tree. The allocation of extents is now in all cases done by the existing allocation code, and if there is an active reservation, that is updated after the fact. Again this is done under the spin lock, since it entails changing the lookup key for the reservation in question. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-08-01 19:35:05 +00:00
static inline bool gfs2_rbm_eq(const struct gfs2_rbm *rbm1,
const struct gfs2_rbm *rbm2)
{
return (rbm1->rgd == rbm2->rgd) && (rbm1->bii == rbm2->bii) &&
GFS2: Replace rgblk_search with gfs2_rbm_find This is part of a series of patches which are introducing the gfs2_rbm structure throughout the block allocation code. The main aim of this part is to create a search function which can deal directly with struct gfs2_rbm. In this case it specifies the initial position at which to start the search and also the point at which the search terminates. The net result of this is to clean up the search code and make it rather more readable, and the various possible exceptions which may occur during the search are partitioned into their own functions. There are some bug fixes too. We should not be checking the reservations while allocating extents - the time for that is when we are searching for where to put the extent, not when we've already made that decision. Also, rgblk_search had two uses, and in only one of those cases did it make sense to check for reservations. This is fixed in the new gfs2_rbm_find function, which has a cleaner interface. The reservation checking has been improved by always checking for contiguous reservations, and returning the first free block after all contiguous reservations. This is done under the spin lock to ensure consistancy of the tree. The allocation of extents is now in all cases done by the existing allocation code, and if there is an active reservation, that is updated after the fact. Again this is done under the spin lock, since it entails changing the lookup key for the reservation in question. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-08-01 19:35:05 +00:00
(rbm1->offset == rbm2->offset);
}
enum gfs2_state_bits {
BH_Pinned = BH_PrivateStart,
BH_Escaped = BH_PrivateStart + 1,
};
BUFFER_FNS(Pinned, pinned)
TAS_BUFFER_FNS(Pinned, pinned)
BUFFER_FNS(Escaped, escaped)
TAS_BUFFER_FNS(Escaped, escaped)
struct gfs2_bufdata {
struct buffer_head *bd_bh;
struct gfs2_glock *bd_gl;
u64 bd_blkno;
struct list_head bd_list;
const struct gfs2_log_operations *bd_ops;
struct gfs2_trans *bd_tr;
struct list_head bd_ail_st_list;
struct list_head bd_ail_gl_list;
};
/*
* Internally, we prefix things with gdlm_ and GDLM_ (for gfs-dlm) since a
* prefix of lock_dlm_ gets awkward.
*/
#define GDLM_STRNAME_BYTES 25
#define GDLM_LVB_SIZE 32
/*
* ls_recover_flags:
*
* DFL_BLOCK_LOCKS: dlm is in recovery and will grant locks that had been
* held by failed nodes whose journals need recovery. Those locks should
* only be used for journal recovery until the journal recovery is done.
* This is set by the dlm recover_prep callback and cleared by the
* gfs2_control thread when journal recovery is complete. To avoid
* races between recover_prep setting and gfs2_control clearing, recover_spin
* is held while changing this bit and reading/writing recover_block
* and recover_start.
*
* DFL_NO_DLM_OPS: dlm lockspace ops/callbacks are not being used.
*
* DFL_FIRST_MOUNT: this node is the first to mount this fs and is doing
* recovery of all journals before allowing other nodes to mount the fs.
* This is cleared when FIRST_MOUNT_DONE is set.
*
* DFL_FIRST_MOUNT_DONE: this node was the first mounter, and has finished
* recovery of all journals, and now allows other nodes to mount the fs.
*
* DFL_MOUNT_DONE: gdlm_mount has completed successfully and cleared
* BLOCK_LOCKS for the first time. The gfs2_control thread should now
* control clearing BLOCK_LOCKS for further recoveries.
*
* DFL_UNMOUNT: gdlm_unmount sets to keep sdp off gfs2_control_wq.
*
* DFL_DLM_RECOVERY: set while dlm is in recovery, between recover_prep()
* and recover_done(), i.e. set while recover_block == recover_start.
*/
enum {
DFL_BLOCK_LOCKS = 0,
DFL_NO_DLM_OPS = 1,
DFL_FIRST_MOUNT = 2,
DFL_FIRST_MOUNT_DONE = 3,
DFL_MOUNT_DONE = 4,
DFL_UNMOUNT = 5,
DFL_DLM_RECOVERY = 6,
};
/*
* We are using struct lm_lockname as an rhashtable key. Avoid holes within
* the struct; padding at the end is fine.
*/
struct lm_lockname {
u64 ln_number;
struct gfs2_sbd *ln_sbd;
unsigned int ln_type;
};
#define lm_name_equal(name1, name2) \
(((name1)->ln_number == (name2)->ln_number) && \
((name1)->ln_type == (name2)->ln_type) && \
((name1)->ln_sbd == (name2)->ln_sbd))
struct gfs2_glock_operations {
void (*go_sync) (struct gfs2_glock *gl);
int (*go_xmote_bh) (struct gfs2_glock *gl, struct gfs2_holder *gh);
void (*go_inval) (struct gfs2_glock *gl, int flags);
int (*go_demote_ok) (const struct gfs2_glock *gl);
int (*go_lock) (struct gfs2_holder *gh);
void (*go_unlock) (struct gfs2_holder *gh);
void (*go_dump)(struct seq_file *seq, const struct gfs2_glock *gl);
void (*go_callback)(struct gfs2_glock *gl, bool remote);
const int go_type;
const unsigned long go_flags;
#define GLOF_ASPACE 1
#define GLOF_LVB 2
#define GLOF_LRU 4
};
GFS2: glock statistics gathering The stats are divided into two sets: those relating to the super block and those relating to an individual glock. The super block stats are done on a per cpu basis in order to try and reduce the overhead of gathering them. They are also further divided by glock type. In the case of both the super block and glock statistics, the same information is gathered in each case. The super block statistics are used to provide default values for most of the glock statistics, so that newly created glocks should have, as far as possible, a sensible starting point. The statistics are divided into three pairs of mean and variance, plus two counters. The mean/variance pairs are smoothed exponential estimates and the algorithm used is one which will be very familiar to those used to calculation of round trip times in network code. The three pairs of mean/variance measure the following things: 1. DLM lock time (non-blocking requests) 2. DLM lock time (blocking requests) 3. Inter-request time (again to the DLM) A non-blocking request is one which will complete right away, whatever the state of the DLM lock in question. That currently means any requests when (a) the current state of the lock is exclusive (b) the requested state is either null or unlocked or (c) the "try lock" flag is set. A blocking request covers all the other lock requests. There are two counters. The first is there primarily to show how many lock requests have been made, and thus how much data has gone into the mean/variance calculations. The other counter is counting queueing of holders at the top layer of the glock code. Hopefully that number will be a lot larger than the number of dlm lock requests issued. So why gather these statistics? There are several reasons we'd like to get a better idea of these timings: 1. To be able to better set the glock "min hold time" 2. To spot performance issues more easily 3. To improve the algorithm for selecting resource groups for allocation (to base it on lock wait time, rather than blindly using a "try lock") Due to the smoothing action of the updates, a step change in some input quantity being sampled will only fully be taken into account after 8 samples (or 4 for the variance) and this needs to be carefully considered when interpreting the results. Knowing both the time it takes a lock request to complete and the average time between lock requests for a glock means we can compute the total percentage of the time for which the node is able to use a glock vs. time that the rest of the cluster has its share. That will be very useful when setting the lock min hold time. The other point to remember is that all times are in nanoseconds. Great care has been taken to ensure that we measure exactly the quantities that we want, as accurately as possible. There are always inaccuracies in any measuring system, but I hope this is as accurate as we can reasonably make it. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-01-20 10:38:36 +00:00
enum {
GFS2_LKS_SRTT = 0, /* Non blocking smoothed round trip time */
GFS2_LKS_SRTTVAR = 1, /* Non blocking smoothed variance */
GFS2_LKS_SRTTB = 2, /* Blocking smoothed round trip time */
GFS2_LKS_SRTTVARB = 3, /* Blocking smoothed variance */
GFS2_LKS_SIRT = 4, /* Smoothed Inter-request time */
GFS2_LKS_SIRTVAR = 5, /* Smoothed Inter-request variance */
GFS2_LKS_DCOUNT = 6, /* Count of dlm requests */
GFS2_LKS_QCOUNT = 7, /* Count of gfs2_holder queues */
GFS2_NR_LKSTATS
};
struct gfs2_lkstats {
u64 stats[GFS2_NR_LKSTATS];
GFS2: glock statistics gathering The stats are divided into two sets: those relating to the super block and those relating to an individual glock. The super block stats are done on a per cpu basis in order to try and reduce the overhead of gathering them. They are also further divided by glock type. In the case of both the super block and glock statistics, the same information is gathered in each case. The super block statistics are used to provide default values for most of the glock statistics, so that newly created glocks should have, as far as possible, a sensible starting point. The statistics are divided into three pairs of mean and variance, plus two counters. The mean/variance pairs are smoothed exponential estimates and the algorithm used is one which will be very familiar to those used to calculation of round trip times in network code. The three pairs of mean/variance measure the following things: 1. DLM lock time (non-blocking requests) 2. DLM lock time (blocking requests) 3. Inter-request time (again to the DLM) A non-blocking request is one which will complete right away, whatever the state of the DLM lock in question. That currently means any requests when (a) the current state of the lock is exclusive (b) the requested state is either null or unlocked or (c) the "try lock" flag is set. A blocking request covers all the other lock requests. There are two counters. The first is there primarily to show how many lock requests have been made, and thus how much data has gone into the mean/variance calculations. The other counter is counting queueing of holders at the top layer of the glock code. Hopefully that number will be a lot larger than the number of dlm lock requests issued. So why gather these statistics? There are several reasons we'd like to get a better idea of these timings: 1. To be able to better set the glock "min hold time" 2. To spot performance issues more easily 3. To improve the algorithm for selecting resource groups for allocation (to base it on lock wait time, rather than blindly using a "try lock") Due to the smoothing action of the updates, a step change in some input quantity being sampled will only fully be taken into account after 8 samples (or 4 for the variance) and this needs to be carefully considered when interpreting the results. Knowing both the time it takes a lock request to complete and the average time between lock requests for a glock means we can compute the total percentage of the time for which the node is able to use a glock vs. time that the rest of the cluster has its share. That will be very useful when setting the lock min hold time. The other point to remember is that all times are in nanoseconds. Great care has been taken to ensure that we measure exactly the quantities that we want, as accurately as possible. There are always inaccuracies in any measuring system, but I hope this is as accurate as we can reasonably make it. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-01-20 10:38:36 +00:00
};
enum {
/* States */
HIF_HOLDER = 6, /* Set for gh that "holds" the glock */
HIF_FIRST = 7,
HIF_WAIT = 10,
};
struct gfs2_holder {
struct list_head gh_list;
struct gfs2_glock *gh_gl;
struct pid *gh_owner_pid;
u16 gh_flags;
u16 gh_state;
int gh_error;
unsigned long gh_iflags; /* HIF_... */
unsigned long gh_ip;
};
/* Number of quota types we support */
#define GFS2_MAXQUOTAS 2
struct gfs2_qadata { /* quota allocation data */
/* Quota stuff */
struct gfs2_quota_data *qa_qd[2 * GFS2_MAXQUOTAS];
struct gfs2_holder qa_qd_ghs[2 * GFS2_MAXQUOTAS];
unsigned int qa_qd_num;
};
/* Resource group multi-block reservation, in order of appearance:
Step 1. Function prepares to write, allocates a mb, sets the size hint.
Step 2. User calls inplace_reserve to target an rgrp, sets the rgrp info
Step 3. Function get_local_rgrp locks the rgrp, determines which bits to use
Step 4. Bits are assigned from the rgrp based on either the reservation
or wherever it can.
*/
struct gfs2_blkreserv {
/* components used during write (step 1): */
atomic_t rs_sizehint; /* hint of the write size */
struct gfs2_holder rs_rgd_gh; /* Filled in by get_local_rgrp */
struct rb_node rs_node; /* link to other block reservations */
struct gfs2_rbm rs_rbm; /* Start of reservation */
u32 rs_free; /* how many blocks are still free */
u64 rs_inum; /* Inode number for reservation */
};
/*
* Allocation parameters
* @target: The number of blocks we'd ideally like to allocate
* @aflags: The flags (e.g. Orlov flag)
*
* The intent is to gradually expand this structure over time in
* order to give more information, e.g. alignment, min extent size
* to the allocation code.
*/
struct gfs2_alloc_parms {
u64 target;
u32 min_target;
u32 aflags;
u64 allowed;
};
enum {
GLF_LOCK = 1,
GLF_DEMOTE = 3,
GLF_PENDING_DEMOTE = 4,
GLF_DEMOTE_IN_PROGRESS = 5,
GLF_DIRTY = 6,
GLF_LFLUSH = 7,
GLF_INVALIDATE_IN_PROGRESS = 8,
GLF_REPLY_PENDING = 9,
GLF_INITIAL = 10,
GLF_FROZEN = 11,
GLF_QUEUED = 12,
GLF_LRU = 13,
GLF_OBJECT = 14, /* Used only for tracing */
GFS2: glock statistics gathering The stats are divided into two sets: those relating to the super block and those relating to an individual glock. The super block stats are done on a per cpu basis in order to try and reduce the overhead of gathering them. They are also further divided by glock type. In the case of both the super block and glock statistics, the same information is gathered in each case. The super block statistics are used to provide default values for most of the glock statistics, so that newly created glocks should have, as far as possible, a sensible starting point. The statistics are divided into three pairs of mean and variance, plus two counters. The mean/variance pairs are smoothed exponential estimates and the algorithm used is one which will be very familiar to those used to calculation of round trip times in network code. The three pairs of mean/variance measure the following things: 1. DLM lock time (non-blocking requests) 2. DLM lock time (blocking requests) 3. Inter-request time (again to the DLM) A non-blocking request is one which will complete right away, whatever the state of the DLM lock in question. That currently means any requests when (a) the current state of the lock is exclusive (b) the requested state is either null or unlocked or (c) the "try lock" flag is set. A blocking request covers all the other lock requests. There are two counters. The first is there primarily to show how many lock requests have been made, and thus how much data has gone into the mean/variance calculations. The other counter is counting queueing of holders at the top layer of the glock code. Hopefully that number will be a lot larger than the number of dlm lock requests issued. So why gather these statistics? There are several reasons we'd like to get a better idea of these timings: 1. To be able to better set the glock "min hold time" 2. To spot performance issues more easily 3. To improve the algorithm for selecting resource groups for allocation (to base it on lock wait time, rather than blindly using a "try lock") Due to the smoothing action of the updates, a step change in some input quantity being sampled will only fully be taken into account after 8 samples (or 4 for the variance) and this needs to be carefully considered when interpreting the results. Knowing both the time it takes a lock request to complete and the average time between lock requests for a glock means we can compute the total percentage of the time for which the node is able to use a glock vs. time that the rest of the cluster has its share. That will be very useful when setting the lock min hold time. The other point to remember is that all times are in nanoseconds. Great care has been taken to ensure that we measure exactly the quantities that we want, as accurately as possible. There are always inaccuracies in any measuring system, but I hope this is as accurate as we can reasonably make it. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-01-20 10:38:36 +00:00
GLF_BLOCKING = 15,
GLF_INODE_CREATING = 16, /* Inode creation occurring */
};
struct gfs2_glock {
unsigned long gl_flags; /* GLF_... */
struct lm_lockname gl_name;
struct lockref gl_lockref;
/* State fields protected by gl_lockref.lock */
unsigned int gl_state:2, /* Current state */
gl_target:2, /* Target state */
gl_demote_state:2, /* State requested by remote node */
gl_req:2, /* State in last dlm request */
gl_reply:8; /* Last reply from the dlm */
unsigned long gl_demote_time; /* time of first demote request */
long gl_hold_time;
struct list_head gl_holders;
const struct gfs2_glock_operations *gl_ops;
GFS2: glock statistics gathering The stats are divided into two sets: those relating to the super block and those relating to an individual glock. The super block stats are done on a per cpu basis in order to try and reduce the overhead of gathering them. They are also further divided by glock type. In the case of both the super block and glock statistics, the same information is gathered in each case. The super block statistics are used to provide default values for most of the glock statistics, so that newly created glocks should have, as far as possible, a sensible starting point. The statistics are divided into three pairs of mean and variance, plus two counters. The mean/variance pairs are smoothed exponential estimates and the algorithm used is one which will be very familiar to those used to calculation of round trip times in network code. The three pairs of mean/variance measure the following things: 1. DLM lock time (non-blocking requests) 2. DLM lock time (blocking requests) 3. Inter-request time (again to the DLM) A non-blocking request is one which will complete right away, whatever the state of the DLM lock in question. That currently means any requests when (a) the current state of the lock is exclusive (b) the requested state is either null or unlocked or (c) the "try lock" flag is set. A blocking request covers all the other lock requests. There are two counters. The first is there primarily to show how many lock requests have been made, and thus how much data has gone into the mean/variance calculations. The other counter is counting queueing of holders at the top layer of the glock code. Hopefully that number will be a lot larger than the number of dlm lock requests issued. So why gather these statistics? There are several reasons we'd like to get a better idea of these timings: 1. To be able to better set the glock "min hold time" 2. To spot performance issues more easily 3. To improve the algorithm for selecting resource groups for allocation (to base it on lock wait time, rather than blindly using a "try lock") Due to the smoothing action of the updates, a step change in some input quantity being sampled will only fully be taken into account after 8 samples (or 4 for the variance) and this needs to be carefully considered when interpreting the results. Knowing both the time it takes a lock request to complete and the average time between lock requests for a glock means we can compute the total percentage of the time for which the node is able to use a glock vs. time that the rest of the cluster has its share. That will be very useful when setting the lock min hold time. The other point to remember is that all times are in nanoseconds. Great care has been taken to ensure that we measure exactly the quantities that we want, as accurately as possible. There are always inaccuracies in any measuring system, but I hope this is as accurate as we can reasonably make it. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-01-20 10:38:36 +00:00
ktime_t gl_dstamp;
struct gfs2_lkstats gl_stats;
struct dlm_lksb gl_lksb;
unsigned long gl_tchange;
void *gl_object;
struct list_head gl_lru;
struct list_head gl_ail_list;
atomic_t gl_ail_count;
atomic_t gl_revokes;
struct delayed_work gl_work;
union {
/* For inode and iopen glocks only */
struct work_struct gl_delete;
/* For rgrp glocks only */
struct {
loff_t start;
loff_t end;
} gl_vm;
};
struct rcu_head gl_rcu;
struct rhash_head gl_node;
};
#define GFS2_MIN_LVB_SIZE 32 /* Min size of LVB that gfs2 supports */
enum {
GIF_INVALID = 0,
GIF_QD_LOCKED = 1,
GIF_ALLOC_FAILED = 2,
GIF_SW_PAGED = 3,
GIF_ORDERED = 4,
GIF_FREE_VFS_INODE = 5,
GIF_GLOP_PENDING = 6,
};
struct gfs2_inode {
struct inode i_inode;
u64 i_no_addr;
u64 i_no_formal_ino;
u64 i_generation;
u64 i_eattr;
unsigned long i_flags; /* GIF_... */
struct gfs2_glock *i_gl; /* Move into i_gh? */
struct gfs2_holder i_iopen_gh;
struct gfs2_holder i_gh; /* for prepare/commit_write only */
struct gfs2_qadata *i_qadata; /* quota allocation data */
struct gfs2_blkreserv i_res; /* rgrp multi-block reservation */
struct gfs2_rgrpd *i_rgd;
u64 i_goal; /* goal block for allocations */
struct rw_semaphore i_rw_mutex;
struct list_head i_ordered;
struct list_head i_trunc_list;
__be64 *i_hash_cache;
u32 i_entries;
u32 i_diskflags;
u8 i_height;
u8 i_depth;
u16 i_rahead;
};
/*
* Since i_inode is the first element of struct gfs2_inode,
* this is effectively a cast.
*/
static inline struct gfs2_inode *GFS2_I(struct inode *inode)
{
return container_of(inode, struct gfs2_inode, i_inode);
}
static inline struct gfs2_sbd *GFS2_SB(const struct inode *inode)
{
return inode->i_sb->s_fs_info;
}
struct gfs2_file {
struct mutex f_fl_mutex;
struct gfs2_holder f_fl_gh;
};
struct gfs2_revoke_replay {
struct list_head rr_list;
u64 rr_blkno;
unsigned int rr_where;
};
enum {
QDF_CHANGE = 1,
QDF_LOCKED = 2,
QDF_REFRESH = 3,
QDF_QMSG_QUIET = 4,
};
struct gfs2_quota_data {
struct hlist_bl_node qd_hlist;
struct list_head qd_list;
struct kqid qd_id;
struct gfs2_sbd *qd_sbd;
struct lockref qd_lockref;
struct list_head qd_lru;
unsigned qd_hash;
unsigned long qd_flags; /* QDF_... */
s64 qd_change;
s64 qd_change_sync;
unsigned int qd_slot;
unsigned int qd_slot_count;
struct buffer_head *qd_bh;
struct gfs2_quota_change *qd_bh_qc;
unsigned int qd_bh_count;
struct gfs2_glock *qd_gl;
struct gfs2_quota_lvb qd_qb;
u64 qd_sync_gen;
unsigned long qd_last_warn;
struct rcu_head qd_rcu;
};
enum {
TR_TOUCHED = 1,
TR_ATTACHED = 2,
TR_ALLOCED = 3,
};
struct gfs2_trans {
unsigned long tr_ip;
unsigned int tr_blocks;
unsigned int tr_revokes;
unsigned int tr_reserved;
unsigned long tr_flags;
unsigned int tr_num_buf_new;
[GFS2] assertion failure after writing to journaled file, umount This patch passes all my nasty tests that were causing the code to fail under one circumstance or another. Here is a complete summary of all changes from today's git tree, in order of appearance: 1. There are now separate variables for metadata buffer accounting. 2. Variable sd_log_num_hdrs is no longer needed, since the header accounting is taken care of by the reserve/refund sequence. 3. Fixed a tiny grammatical problem in a comment. 4. Added a new function "calc_reserved" to calculate the reserved log space. This isn't entirely necessary, but it has two benefits: First, it simplifies the gfs2_log_refund function greatly. Second, it allows for easier debugging because I could sprinkle the code with calls to this function to make sure the accounting is proper (by adding asserts and printks) at strategic point of the code. 5. In log_pull_tail there apparently was a kludge to fix up the accounting based on a "pull" parameter. The buffer accounting is now done properly, so the kludge was removed. 6. File sync operations were making a call to gfs2_log_flush that writes another journal header. Since that header was unplanned for (reserved) by the reserve/refund sequence, the free space had to be decremented so that when log_pull_tail gets called, the free space is be adjusted properly. (Did I hear you call that a kludge? well, maybe, but a lot more justifiable than the one I removed). 7. In the gfs2_log_shutdown code, it optionally syncs the log by specifying the PULL parameter to log_write_header. I'm not sure this is necessary anymore. It just seems to me there could be cases where shutdown is called while there are outstanding log buffers. 8. In the (data)buf_lo_before_commit functions, I changed some offset values from being calculated on the fly to being constants. That simplified some code and we might as well let the compiler do the calculation once rather than redoing those cycles at run time. 9. This version has my rewritten databuf_lo_add function. This version is much more like its predecessor, buf_lo_add, which makes it easier to understand. Again, this might not be necessary, but it seems as if this one works as well as the previous one, maybe even better, so I decided to leave it in. 10. In databuf_lo_before_commit, a previous data corruption problem was caused by going off the end of the buffer. The proper solution is to have the proper limit in place, rather than stopping earlier. (Thus my previous attempt to fix it is wrong). If you don't wrap the buffer, you're stopping too early and that causes more log buffer accounting problems. 11. In lops.h there are two new (previously mentioned) constants for figuring out the data offset for the journal buffers. 12. There are also two new functions, buf_limit and databuf_limit to calculate how many entries will fit in the buffer. 13. In function gfs2_meta_wipe, it needs to distinguish between pinned metadata buffers and journaled data buffers for proper journal buffer accounting. It can't use the JDATA gfs2_inode flag because it's sometimes passed the "real" inode and sometimes the "metadata inode" and the inode flags will be random bits in a metadata gfs2_inode. It needs to base its decision on which was passed in. Signed-off-by: Bob Peterson <rpeterso@redhat.com> Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-06-18 19:50:20 +00:00
unsigned int tr_num_databuf_new;
unsigned int tr_num_buf_rm;
[GFS2] assertion failure after writing to journaled file, umount This patch passes all my nasty tests that were causing the code to fail under one circumstance or another. Here is a complete summary of all changes from today's git tree, in order of appearance: 1. There are now separate variables for metadata buffer accounting. 2. Variable sd_log_num_hdrs is no longer needed, since the header accounting is taken care of by the reserve/refund sequence. 3. Fixed a tiny grammatical problem in a comment. 4. Added a new function "calc_reserved" to calculate the reserved log space. This isn't entirely necessary, but it has two benefits: First, it simplifies the gfs2_log_refund function greatly. Second, it allows for easier debugging because I could sprinkle the code with calls to this function to make sure the accounting is proper (by adding asserts and printks) at strategic point of the code. 5. In log_pull_tail there apparently was a kludge to fix up the accounting based on a "pull" parameter. The buffer accounting is now done properly, so the kludge was removed. 6. File sync operations were making a call to gfs2_log_flush that writes another journal header. Since that header was unplanned for (reserved) by the reserve/refund sequence, the free space had to be decremented so that when log_pull_tail gets called, the free space is be adjusted properly. (Did I hear you call that a kludge? well, maybe, but a lot more justifiable than the one I removed). 7. In the gfs2_log_shutdown code, it optionally syncs the log by specifying the PULL parameter to log_write_header. I'm not sure this is necessary anymore. It just seems to me there could be cases where shutdown is called while there are outstanding log buffers. 8. In the (data)buf_lo_before_commit functions, I changed some offset values from being calculated on the fly to being constants. That simplified some code and we might as well let the compiler do the calculation once rather than redoing those cycles at run time. 9. This version has my rewritten databuf_lo_add function. This version is much more like its predecessor, buf_lo_add, which makes it easier to understand. Again, this might not be necessary, but it seems as if this one works as well as the previous one, maybe even better, so I decided to leave it in. 10. In databuf_lo_before_commit, a previous data corruption problem was caused by going off the end of the buffer. The proper solution is to have the proper limit in place, rather than stopping earlier. (Thus my previous attempt to fix it is wrong). If you don't wrap the buffer, you're stopping too early and that causes more log buffer accounting problems. 11. In lops.h there are two new (previously mentioned) constants for figuring out the data offset for the journal buffers. 12. There are also two new functions, buf_limit and databuf_limit to calculate how many entries will fit in the buffer. 13. In function gfs2_meta_wipe, it needs to distinguish between pinned metadata buffers and journaled data buffers for proper journal buffer accounting. It can't use the JDATA gfs2_inode flag because it's sometimes passed the "real" inode and sometimes the "metadata inode" and the inode flags will be random bits in a metadata gfs2_inode. It needs to base its decision on which was passed in. Signed-off-by: Bob Peterson <rpeterso@redhat.com> Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2007-06-18 19:50:20 +00:00
unsigned int tr_num_databuf_rm;
unsigned int tr_num_revoke;
unsigned int tr_num_revoke_rm;
struct list_head tr_list;
struct list_head tr_databuf;
struct list_head tr_buf;
unsigned int tr_first;
struct list_head tr_ail1_list;
struct list_head tr_ail2_list;
};
struct gfs2_journal_extent {
struct list_head list;
unsigned int lblock; /* First logical block */
u64 dblock; /* First disk block */
u64 blocks;
};
struct gfs2_jdesc {
struct list_head jd_list;
struct list_head extent_list;
unsigned int nr_extents;
struct work_struct jd_work;
struct inode *jd_inode;
unsigned long jd_flags;
#define JDF_RECOVERY 1
unsigned int jd_jid;
unsigned int jd_blocks;
int jd_recover_error;
/* Replay stuff */
unsigned int jd_found_blocks;
unsigned int jd_found_revokes;
unsigned int jd_replayed_blocks;
struct list_head jd_revoke_list;
unsigned int jd_replay_tail;
};
struct gfs2_statfs_change_host {
s64 sc_total;
s64 sc_free;
s64 sc_dinodes;
};
#define GFS2_QUOTA_DEFAULT GFS2_QUOTA_OFF
#define GFS2_QUOTA_OFF 0
#define GFS2_QUOTA_ACCOUNT 1
#define GFS2_QUOTA_ON 2
#define GFS2_DATA_DEFAULT GFS2_DATA_ORDERED
#define GFS2_DATA_WRITEBACK 1
#define GFS2_DATA_ORDERED 2
#define GFS2_ERRORS_DEFAULT GFS2_ERRORS_WITHDRAW
#define GFS2_ERRORS_WITHDRAW 0
#define GFS2_ERRORS_CONTINUE 1 /* place holder for future feature */
#define GFS2_ERRORS_RO 2 /* place holder for future feature */
#define GFS2_ERRORS_PANIC 3
struct gfs2_args {
char ar_lockproto[GFS2_LOCKNAME_LEN]; /* Name of the Lock Protocol */
char ar_locktable[GFS2_LOCKNAME_LEN]; /* Name of the Lock Table */
char ar_hostdata[GFS2_LOCKNAME_LEN]; /* Host specific data */
unsigned int ar_spectator:1; /* Don't get a journal */
unsigned int ar_localflocks:1; /* Let the VFS do flock|fcntl */
unsigned int ar_debug:1; /* Oops on errors */
unsigned int ar_posix_acl:1; /* Enable posix acls */
unsigned int ar_quota:2; /* off/account/on */
unsigned int ar_suiddir:1; /* suiddir support */
unsigned int ar_data:2; /* ordered/writeback */
unsigned int ar_meta:1; /* mount metafs */
unsigned int ar_discard:1; /* discard requests */
unsigned int ar_errors:2; /* errors=withdraw | panic */
unsigned int ar_nobarrier:1; /* do not send barriers */
unsigned int ar_rgrplvb:1; /* use lvbs for rgrp info */
gfs2: change gfs2 readdir cookie gfs2 currently returns 31 bits of filename hash as a cookie that readdir uses for an offset into the directory. When there are a large number of directory entries, the likelihood of a collision goes up way too quickly. GFS2 will now return cookies that are guaranteed unique for a while, and then fail back to using 30 bits of filename hash. Specifically, the directory leaf blocks are divided up into chunks based on the minimum size of a gfs2 directory entry (48 bytes). Each entry's cookie is based off the chunk where it starts, in the linked list of leaf blocks that it hashes to (there are 131072 hash buckets). Directory entries will have unique names until they take reach chunk 8192. Assuming the largest filenames possible, and the least efficient spacing possible, this new method will still be able to return unique names when the previous method has statistically more than a 99% chance of a collision. The non-unique names it fails back to are guaranteed to not collide with the unique names. unique cookies will be in this format: - 1 bit "0" to make sure the the returned cookie is positive - 17 bits for the hash table index - 1 bit for the mode "0" - 13 bits for the offset non-unique cookies will be in this format: - 1 bit "0" to make sure the the returned cookie is positive - 17 bits for the hash table index - 1 bit for the mode "1" - 13 more bits of the name hash Another benefit of location based cookies, is that once a directory's exhash table is fully extended (so that multiple hash table indexs do not use the same leaf blocks), gfs2 can skip sorting the directory entries until it reaches the non-unique ones, and then it only needs to sort these. This provides a significant speed up for directory reads of very large directories. The only issue is that for these cookies to continue to point to the correct entry as files are added and removed from the directory, gfs2 must keep the entries at the same offset in the leaf block when they are split (see my previous patch). This means that until all the nodes in a cluster are running with code that will split the directory leaf blocks this way, none of the nodes can use the new cookie code. To deal with this, gfs2 now has the mount option loccookie, which, if set, will make it return these new location based cookies. This option must not be set until all nodes in the cluster are at least running this version of the kernel code, and you have guaranteed that there are no outstanding cookies required by other software, such as NFS. gfs2 uses some of the extra space at the end of the gfs2_dirent structure to store the calculated readdir cookies. This keeps us from needing to allocate a seperate array to hold these values. gfs2 recomputes the cookie stored in de_cookie for every readdir call. The time it takes to do so is small, and if gfs2 expected this value to be saved on disk, the new code wouldn't work correctly on filesystems created with an earlier version of gfs2. One issue with adding de_cookie to the union in the gfs2_dirent structure is that it caused the union to align itself to a 4 byte boundary, instead of its previous 2 byte boundary. This changed the offset of de_rahead. To solve that, I pulled de_rahead out of the union, since it does not need to be there. Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com>
2015-12-01 14:46:55 +00:00
unsigned int ar_loccookie:1; /* use location based readdir
cookies */
int ar_commit; /* Commit interval */
int ar_statfs_quantum; /* The fast statfs interval */
int ar_quota_quantum; /* The quota interval */
int ar_statfs_percent; /* The % change to force sync */
};
struct gfs2_tune {
spinlock_t gt_spin;
unsigned int gt_logd_secs;
unsigned int gt_quota_warn_period; /* Secs between quota warn msgs */
unsigned int gt_quota_scale_num; /* Numerator */
unsigned int gt_quota_scale_den; /* Denominator */
unsigned int gt_quota_quantum; /* Secs between syncs to quota file */
unsigned int gt_new_files_jdata;
unsigned int gt_max_readahead; /* Max bytes to read-ahead from disk */
unsigned int gt_complain_secs;
unsigned int gt_statfs_quantum;
unsigned int gt_statfs_slow;
};
enum {
SDF_JOURNAL_CHECKED = 0,
SDF_JOURNAL_LIVE = 1,
SDF_SHUTDOWN = 2,
SDF_NOBARRIERS = 3,
SDF_NORECOVERY = 4,
SDF_DEMOTE = 5,
SDF_NOJOURNALID = 6,
SDF_RORECOVERY = 7, /* read only recovery */
SDF_SKIP_DLM_UNLOCK = 8,
SDF_FORCE_AIL_FLUSH = 9,
};
enum gfs2_freeze_state {
SFS_UNFROZEN = 0,
SFS_STARTING_FREEZE = 1,
SFS_FROZEN = 2,
};
#define GFS2_FSNAME_LEN 256
struct gfs2_inum_host {
u64 no_formal_ino;
u64 no_addr;
};
struct gfs2_sb_host {
u32 sb_magic;
u32 sb_type;
u32 sb_format;
u32 sb_fs_format;
u32 sb_multihost_format;
u32 sb_bsize;
u32 sb_bsize_shift;
struct gfs2_inum_host sb_master_dir;
struct gfs2_inum_host sb_root_dir;
char sb_lockproto[GFS2_LOCKNAME_LEN];
char sb_locktable[GFS2_LOCKNAME_LEN];
};
/*
* lm_mount() return values
*
* ls_jid - the journal ID this node should use
* ls_first - this node is the first to mount the file system
* ls_lockspace - lock module's context for this file system
* ls_ops - lock module's functions
*/
struct lm_lockstruct {
int ls_jid;
unsigned int ls_first;
const struct lm_lockops *ls_ops;
dlm_lockspace_t *ls_dlm;
int ls_recover_jid_done; /* These two are deprecated, */
int ls_recover_jid_status; /* used previously by gfs_controld */
struct dlm_lksb ls_mounted_lksb; /* mounted_lock */
struct dlm_lksb ls_control_lksb; /* control_lock */
char ls_control_lvb[GDLM_LVB_SIZE]; /* control_lock lvb */
struct completion ls_sync_wait; /* {control,mounted}_{lock,unlock} */
char *ls_lvb_bits;
spinlock_t ls_recover_spin; /* protects following fields */
unsigned long ls_recover_flags; /* DFL_ */
uint32_t ls_recover_mount; /* gen in first recover_done cb */
uint32_t ls_recover_start; /* gen in last recover_done cb */
uint32_t ls_recover_block; /* copy recover_start in last recover_prep */
uint32_t ls_recover_size; /* size of recover_submit, recover_result */
uint32_t *ls_recover_submit; /* gen in last recover_slot cb per jid */
uint32_t *ls_recover_result; /* result of last jid recovery */
};
GFS2: glock statistics gathering The stats are divided into two sets: those relating to the super block and those relating to an individual glock. The super block stats are done on a per cpu basis in order to try and reduce the overhead of gathering them. They are also further divided by glock type. In the case of both the super block and glock statistics, the same information is gathered in each case. The super block statistics are used to provide default values for most of the glock statistics, so that newly created glocks should have, as far as possible, a sensible starting point. The statistics are divided into three pairs of mean and variance, plus two counters. The mean/variance pairs are smoothed exponential estimates and the algorithm used is one which will be very familiar to those used to calculation of round trip times in network code. The three pairs of mean/variance measure the following things: 1. DLM lock time (non-blocking requests) 2. DLM lock time (blocking requests) 3. Inter-request time (again to the DLM) A non-blocking request is one which will complete right away, whatever the state of the DLM lock in question. That currently means any requests when (a) the current state of the lock is exclusive (b) the requested state is either null or unlocked or (c) the "try lock" flag is set. A blocking request covers all the other lock requests. There are two counters. The first is there primarily to show how many lock requests have been made, and thus how much data has gone into the mean/variance calculations. The other counter is counting queueing of holders at the top layer of the glock code. Hopefully that number will be a lot larger than the number of dlm lock requests issued. So why gather these statistics? There are several reasons we'd like to get a better idea of these timings: 1. To be able to better set the glock "min hold time" 2. To spot performance issues more easily 3. To improve the algorithm for selecting resource groups for allocation (to base it on lock wait time, rather than blindly using a "try lock") Due to the smoothing action of the updates, a step change in some input quantity being sampled will only fully be taken into account after 8 samples (or 4 for the variance) and this needs to be carefully considered when interpreting the results. Knowing both the time it takes a lock request to complete and the average time between lock requests for a glock means we can compute the total percentage of the time for which the node is able to use a glock vs. time that the rest of the cluster has its share. That will be very useful when setting the lock min hold time. The other point to remember is that all times are in nanoseconds. Great care has been taken to ensure that we measure exactly the quantities that we want, as accurately as possible. There are always inaccuracies in any measuring system, but I hope this is as accurate as we can reasonably make it. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-01-20 10:38:36 +00:00
struct gfs2_pcpu_lkstats {
/* One struct for each glock type */
struct gfs2_lkstats lkstats[10];
};
struct gfs2_sbd {
struct super_block *sd_vfs;
GFS2: glock statistics gathering The stats are divided into two sets: those relating to the super block and those relating to an individual glock. The super block stats are done on a per cpu basis in order to try and reduce the overhead of gathering them. They are also further divided by glock type. In the case of both the super block and glock statistics, the same information is gathered in each case. The super block statistics are used to provide default values for most of the glock statistics, so that newly created glocks should have, as far as possible, a sensible starting point. The statistics are divided into three pairs of mean and variance, plus two counters. The mean/variance pairs are smoothed exponential estimates and the algorithm used is one which will be very familiar to those used to calculation of round trip times in network code. The three pairs of mean/variance measure the following things: 1. DLM lock time (non-blocking requests) 2. DLM lock time (blocking requests) 3. Inter-request time (again to the DLM) A non-blocking request is one which will complete right away, whatever the state of the DLM lock in question. That currently means any requests when (a) the current state of the lock is exclusive (b) the requested state is either null or unlocked or (c) the "try lock" flag is set. A blocking request covers all the other lock requests. There are two counters. The first is there primarily to show how many lock requests have been made, and thus how much data has gone into the mean/variance calculations. The other counter is counting queueing of holders at the top layer of the glock code. Hopefully that number will be a lot larger than the number of dlm lock requests issued. So why gather these statistics? There are several reasons we'd like to get a better idea of these timings: 1. To be able to better set the glock "min hold time" 2. To spot performance issues more easily 3. To improve the algorithm for selecting resource groups for allocation (to base it on lock wait time, rather than blindly using a "try lock") Due to the smoothing action of the updates, a step change in some input quantity being sampled will only fully be taken into account after 8 samples (or 4 for the variance) and this needs to be carefully considered when interpreting the results. Knowing both the time it takes a lock request to complete and the average time between lock requests for a glock means we can compute the total percentage of the time for which the node is able to use a glock vs. time that the rest of the cluster has its share. That will be very useful when setting the lock min hold time. The other point to remember is that all times are in nanoseconds. Great care has been taken to ensure that we measure exactly the quantities that we want, as accurately as possible. There are always inaccuracies in any measuring system, but I hope this is as accurate as we can reasonably make it. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-01-20 10:38:36 +00:00
struct gfs2_pcpu_lkstats __percpu *sd_lkstats;
struct kobject sd_kobj;
unsigned long sd_flags; /* SDF_... */
struct gfs2_sb_host sd_sb;
/* Constants computed on mount */
u32 sd_fsb2bb;
u32 sd_fsb2bb_shift;
u32 sd_diptrs; /* Number of pointers in a dinode */
u32 sd_inptrs; /* Number of pointers in a indirect block */
u32 sd_jbsize; /* Size of a journaled data block */
u32 sd_hash_bsize; /* sizeof(exhash block) */
u32 sd_hash_bsize_shift;
u32 sd_hash_ptrs; /* Number of pointers in a hash block */
u32 sd_qc_per_block;
GFS2: Speed up gfs2_rbm_from_block This patch is a rewrite of function gfs2_rbm_from_block. Rather than looping to find the right bitmap, the code now does a few simple math calculations. I compared the performance of both algorithms side by side and the new algorithm is noticeably faster. Sample instrumentation output from a "fast" machine: 5 million calls: millisec spent: Orig: 166 New: 113 5 million calls: millisec spent: Orig: 189 New: 114 In addition, I ran postmark (on a somewhat slowr CPU) before the after the new algorithm was put in place and postmark showed a decent improvement: Before the new algorithm: ------------------------- Time: 645 seconds total 584 seconds of transactions (171 per second) Files: 150087 created (232 per second) Creation alone: 100000 files (2083 per second) Mixed with transactions: 50087 files (85 per second) 49995 read (85 per second) 49991 appended (85 per second) 150087 deleted (232 per second) Deletion alone: 100174 files (7705 per second) Mixed with transactions: 49913 files (85 per second) Data: 273.42 megabytes read (434.08 kilobytes per second) 852.13 megabytes written (1.32 megabytes per second) With the new algorithm: ----------------------- Time: 599 seconds total 530 seconds of transactions (188 per second) Files: 150087 created (250 per second) Creation alone: 100000 files (1886 per second) Mixed with transactions: 50087 files (94 per second) 49995 read (94 per second) 49991 appended (94 per second) 150087 deleted (250 per second) Deletion alone: 100174 files (6260 per second) Mixed with transactions: 49913 files (94 per second) Data: 273.42 megabytes read (467.42 kilobytes per second) 852.13 megabytes written (1.42 megabytes per second) Signed-off-by: Bob Peterson <rpeterso@redhat.com> Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-10-19 12:32:51 +00:00
u32 sd_blocks_per_bitmap;
u32 sd_max_dirres; /* Max blocks needed to add a directory entry */
u32 sd_max_height; /* Max height of a file's metadata tree */
u64 sd_heightsize[GFS2_MAX_META_HEIGHT + 1];
gfs2: change gfs2 readdir cookie gfs2 currently returns 31 bits of filename hash as a cookie that readdir uses for an offset into the directory. When there are a large number of directory entries, the likelihood of a collision goes up way too quickly. GFS2 will now return cookies that are guaranteed unique for a while, and then fail back to using 30 bits of filename hash. Specifically, the directory leaf blocks are divided up into chunks based on the minimum size of a gfs2 directory entry (48 bytes). Each entry's cookie is based off the chunk where it starts, in the linked list of leaf blocks that it hashes to (there are 131072 hash buckets). Directory entries will have unique names until they take reach chunk 8192. Assuming the largest filenames possible, and the least efficient spacing possible, this new method will still be able to return unique names when the previous method has statistically more than a 99% chance of a collision. The non-unique names it fails back to are guaranteed to not collide with the unique names. unique cookies will be in this format: - 1 bit "0" to make sure the the returned cookie is positive - 17 bits for the hash table index - 1 bit for the mode "0" - 13 bits for the offset non-unique cookies will be in this format: - 1 bit "0" to make sure the the returned cookie is positive - 17 bits for the hash table index - 1 bit for the mode "1" - 13 more bits of the name hash Another benefit of location based cookies, is that once a directory's exhash table is fully extended (so that multiple hash table indexs do not use the same leaf blocks), gfs2 can skip sorting the directory entries until it reaches the non-unique ones, and then it only needs to sort these. This provides a significant speed up for directory reads of very large directories. The only issue is that for these cookies to continue to point to the correct entry as files are added and removed from the directory, gfs2 must keep the entries at the same offset in the leaf block when they are split (see my previous patch). This means that until all the nodes in a cluster are running with code that will split the directory leaf blocks this way, none of the nodes can use the new cookie code. To deal with this, gfs2 now has the mount option loccookie, which, if set, will make it return these new location based cookies. This option must not be set until all nodes in the cluster are at least running this version of the kernel code, and you have guaranteed that there are no outstanding cookies required by other software, such as NFS. gfs2 uses some of the extra space at the end of the gfs2_dirent structure to store the calculated readdir cookies. This keeps us from needing to allocate a seperate array to hold these values. gfs2 recomputes the cookie stored in de_cookie for every readdir call. The time it takes to do so is small, and if gfs2 expected this value to be saved on disk, the new code wouldn't work correctly on filesystems created with an earlier version of gfs2. One issue with adding de_cookie to the union in the gfs2_dirent structure is that it caused the union to align itself to a 4 byte boundary, instead of its previous 2 byte boundary. This changed the offset of de_rahead. To solve that, I pulled de_rahead out of the union, since it does not need to be there. Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com> Signed-off-by: Bob Peterson <rpeterso@redhat.com>
2015-12-01 14:46:55 +00:00
u32 sd_max_dents_per_leaf; /* Max number of dirents in a leaf block */
struct gfs2_args sd_args; /* Mount arguments */
struct gfs2_tune sd_tune; /* Filesystem tuning structure */
/* Lock Stuff */
struct lm_lockstruct sd_lockstruct;
struct gfs2_holder sd_live_gh;
struct gfs2_glock *sd_rename_gl;
GFS2: remove transaction glock GFS2 has a transaction glock, which must be grabbed for every transaction, whose purpose is to deal with freezing the filesystem. Aside from this involving a large amount of locking, it is very easy to make the current fsfreeze code hang on unfreezing. This patch rewrites how gfs2 handles freezing the filesystem. The transaction glock is removed. In it's place is a freeze glock, which is cached (but not held) in a shared state by every node in the cluster when the filesystem is mounted. This lock only needs to be grabbed on freezing, and actions which need to be safe from freezing, like recovery. When a node wants to freeze the filesystem, it grabs this glock exclusively. When the freeze glock state changes on the nodes (either from shared to unlocked, or shared to exclusive), the filesystem does a special log flush. gfs2_log_flush() does all the work for flushing out the and shutting down the incore log, and then it tries to grab the freeze glock in a shared state again. Since the filesystem is stuck in gfs2_log_flush, no new transaction can start, and nothing can be written to disk. Unfreezing the filesytem simply involes dropping the freeze glock, allowing gfs2_log_flush() to grab and then release the shared lock, so it is cached for next time. However, in order for the unfreezing ioctl to occur, gfs2 needs to get a shared lock on the filesystem root directory inode to check permissions. If that glock has already been grabbed exclusively, fsfreeze will be unable to get the shared lock and unfreeze the filesystem. In order to allow the unfreeze, this patch makes gfs2 grab a shared lock on the filesystem root directory during the freeze, and hold it until it unfreezes the filesystem. The functions which need to grab a shared lock in order to allow the unfreeze ioctl to be issued now use the lock grabbed by the freeze code instead. The freeze and unfreeze code take care to make sure that this shared lock will not be dropped while another process is using it. Signed-off-by: Benjamin Marzinski <bmarzins@redhat.com> Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2014-05-02 03:26:55 +00:00
struct gfs2_glock *sd_freeze_gl;
struct work_struct sd_freeze_work;
wait_queue_head_t sd_glock_wait;
atomic_t sd_glock_disposal;
struct completion sd_locking_init;
struct completion sd_wdack;
struct delayed_work sd_control_work;
/* Inode Stuff */
struct dentry *sd_master_dir;
struct dentry *sd_root_dir;
struct inode *sd_jindex;
struct inode *sd_statfs_inode;
struct inode *sd_sc_inode;
struct inode *sd_qc_inode;
struct inode *sd_rindex;
struct inode *sd_quota_inode;
/* StatFS stuff */
spinlock_t sd_statfs_spin;
struct gfs2_statfs_change_host sd_statfs_master;
struct gfs2_statfs_change_host sd_statfs_local;
int sd_statfs_force_sync;
/* Resource group stuff */
int sd_rindex_uptodate;
spinlock_t 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
struct rb_root sd_rindex_tree;
unsigned int sd_rgrps;
unsigned int sd_max_rg_data;
/* Journal index stuff */
struct list_head sd_jindex_list;
spinlock_t sd_jindex_spin;
struct mutex sd_jindex_mutex;
unsigned int sd_journals;
struct gfs2_jdesc *sd_jdesc;
struct gfs2_holder sd_journal_gh;
struct gfs2_holder sd_jinode_gh;
struct gfs2_holder sd_sc_gh;
struct gfs2_holder sd_qc_gh;
struct completion sd_journal_ready;
/* Daemon stuff */
struct task_struct *sd_logd_process;
struct task_struct *sd_quotad_process;
/* Quota stuff */
struct list_head sd_quota_list;
atomic_t sd_quota_count;
struct mutex sd_quota_mutex;
struct mutex sd_quota_sync_mutex;
wait_queue_head_t sd_quota_wait;
struct list_head sd_trunc_list;
spinlock_t sd_trunc_lock;
unsigned int sd_quota_slots;
unsigned long *sd_quota_bitmap;
spinlock_t sd_bitmap_lock;
u64 sd_quota_sync_gen;
/* Log stuff */
struct address_space sd_aspace;
spinlock_t sd_log_lock;
struct gfs2_trans *sd_log_tr;
unsigned int sd_log_blks_reserved;
int sd_log_commited_revoke;
atomic_t sd_log_pinned;
unsigned int sd_log_num_revoke;
struct list_head sd_log_le_revoke;
struct list_head sd_log_le_ordered;
spinlock_t sd_ordered_lock;
atomic_t sd_log_thresh1;
atomic_t sd_log_thresh2;
atomic_t sd_log_blks_free;
atomic_t sd_log_blks_needed;
wait_queue_head_t sd_log_waitq;
wait_queue_head_t sd_logd_waitq;
u64 sd_log_sequence;
unsigned int sd_log_head;
unsigned int sd_log_tail;
int sd_log_idle;
struct rw_semaphore sd_log_flush_lock;
atomic_t sd_log_in_flight;
struct bio *sd_log_bio;
wait_queue_head_t sd_log_flush_wait;
int sd_log_error;
atomic_t sd_reserving_log;
wait_queue_head_t sd_reserving_log_wait;
unsigned int sd_log_flush_head;
spinlock_t sd_ail_lock;
struct list_head sd_ail1_list;
struct list_head sd_ail2_list;
/* For quiescing the filesystem */
struct gfs2_holder sd_freeze_gh;
atomic_t sd_freeze_state;
struct mutex sd_freeze_mutex;
char sd_fsname[GFS2_FSNAME_LEN + 3 * sizeof(int) + 2];
char sd_table_name[GFS2_FSNAME_LEN];
char sd_proto_name[GFS2_FSNAME_LEN];
/* Debugging crud */
unsigned long sd_last_warning;
struct dentry *debugfs_dir; /* debugfs directory */
GFS2: glock statistics gathering The stats are divided into two sets: those relating to the super block and those relating to an individual glock. The super block stats are done on a per cpu basis in order to try and reduce the overhead of gathering them. They are also further divided by glock type. In the case of both the super block and glock statistics, the same information is gathered in each case. The super block statistics are used to provide default values for most of the glock statistics, so that newly created glocks should have, as far as possible, a sensible starting point. The statistics are divided into three pairs of mean and variance, plus two counters. The mean/variance pairs are smoothed exponential estimates and the algorithm used is one which will be very familiar to those used to calculation of round trip times in network code. The three pairs of mean/variance measure the following things: 1. DLM lock time (non-blocking requests) 2. DLM lock time (blocking requests) 3. Inter-request time (again to the DLM) A non-blocking request is one which will complete right away, whatever the state of the DLM lock in question. That currently means any requests when (a) the current state of the lock is exclusive (b) the requested state is either null or unlocked or (c) the "try lock" flag is set. A blocking request covers all the other lock requests. There are two counters. The first is there primarily to show how many lock requests have been made, and thus how much data has gone into the mean/variance calculations. The other counter is counting queueing of holders at the top layer of the glock code. Hopefully that number will be a lot larger than the number of dlm lock requests issued. So why gather these statistics? There are several reasons we'd like to get a better idea of these timings: 1. To be able to better set the glock "min hold time" 2. To spot performance issues more easily 3. To improve the algorithm for selecting resource groups for allocation (to base it on lock wait time, rather than blindly using a "try lock") Due to the smoothing action of the updates, a step change in some input quantity being sampled will only fully be taken into account after 8 samples (or 4 for the variance) and this needs to be carefully considered when interpreting the results. Knowing both the time it takes a lock request to complete and the average time between lock requests for a glock means we can compute the total percentage of the time for which the node is able to use a glock vs. time that the rest of the cluster has its share. That will be very useful when setting the lock min hold time. The other point to remember is that all times are in nanoseconds. Great care has been taken to ensure that we measure exactly the quantities that we want, as accurately as possible. There are always inaccuracies in any measuring system, but I hope this is as accurate as we can reasonably make it. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-01-20 10:38:36 +00:00
struct dentry *debugfs_dentry_glocks;
struct dentry *debugfs_dentry_glstats;
struct dentry *debugfs_dentry_sbstats;
};
GFS2: glock statistics gathering The stats are divided into two sets: those relating to the super block and those relating to an individual glock. The super block stats are done on a per cpu basis in order to try and reduce the overhead of gathering them. They are also further divided by glock type. In the case of both the super block and glock statistics, the same information is gathered in each case. The super block statistics are used to provide default values for most of the glock statistics, so that newly created glocks should have, as far as possible, a sensible starting point. The statistics are divided into three pairs of mean and variance, plus two counters. The mean/variance pairs are smoothed exponential estimates and the algorithm used is one which will be very familiar to those used to calculation of round trip times in network code. The three pairs of mean/variance measure the following things: 1. DLM lock time (non-blocking requests) 2. DLM lock time (blocking requests) 3. Inter-request time (again to the DLM) A non-blocking request is one which will complete right away, whatever the state of the DLM lock in question. That currently means any requests when (a) the current state of the lock is exclusive (b) the requested state is either null or unlocked or (c) the "try lock" flag is set. A blocking request covers all the other lock requests. There are two counters. The first is there primarily to show how many lock requests have been made, and thus how much data has gone into the mean/variance calculations. The other counter is counting queueing of holders at the top layer of the glock code. Hopefully that number will be a lot larger than the number of dlm lock requests issued. So why gather these statistics? There are several reasons we'd like to get a better idea of these timings: 1. To be able to better set the glock "min hold time" 2. To spot performance issues more easily 3. To improve the algorithm for selecting resource groups for allocation (to base it on lock wait time, rather than blindly using a "try lock") Due to the smoothing action of the updates, a step change in some input quantity being sampled will only fully be taken into account after 8 samples (or 4 for the variance) and this needs to be carefully considered when interpreting the results. Knowing both the time it takes a lock request to complete and the average time between lock requests for a glock means we can compute the total percentage of the time for which the node is able to use a glock vs. time that the rest of the cluster has its share. That will be very useful when setting the lock min hold time. The other point to remember is that all times are in nanoseconds. Great care has been taken to ensure that we measure exactly the quantities that we want, as accurately as possible. There are always inaccuracies in any measuring system, but I hope this is as accurate as we can reasonably make it. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-01-20 10:38:36 +00:00
static inline void gfs2_glstats_inc(struct gfs2_glock *gl, int which)
{
gl->gl_stats.stats[which]++;
}
static inline void gfs2_sbstats_inc(const struct gfs2_glock *gl, int which)
{
const struct gfs2_sbd *sdp = gl->gl_name.ln_sbd;
GFS2: glock statistics gathering The stats are divided into two sets: those relating to the super block and those relating to an individual glock. The super block stats are done on a per cpu basis in order to try and reduce the overhead of gathering them. They are also further divided by glock type. In the case of both the super block and glock statistics, the same information is gathered in each case. The super block statistics are used to provide default values for most of the glock statistics, so that newly created glocks should have, as far as possible, a sensible starting point. The statistics are divided into three pairs of mean and variance, plus two counters. The mean/variance pairs are smoothed exponential estimates and the algorithm used is one which will be very familiar to those used to calculation of round trip times in network code. The three pairs of mean/variance measure the following things: 1. DLM lock time (non-blocking requests) 2. DLM lock time (blocking requests) 3. Inter-request time (again to the DLM) A non-blocking request is one which will complete right away, whatever the state of the DLM lock in question. That currently means any requests when (a) the current state of the lock is exclusive (b) the requested state is either null or unlocked or (c) the "try lock" flag is set. A blocking request covers all the other lock requests. There are two counters. The first is there primarily to show how many lock requests have been made, and thus how much data has gone into the mean/variance calculations. The other counter is counting queueing of holders at the top layer of the glock code. Hopefully that number will be a lot larger than the number of dlm lock requests issued. So why gather these statistics? There are several reasons we'd like to get a better idea of these timings: 1. To be able to better set the glock "min hold time" 2. To spot performance issues more easily 3. To improve the algorithm for selecting resource groups for allocation (to base it on lock wait time, rather than blindly using a "try lock") Due to the smoothing action of the updates, a step change in some input quantity being sampled will only fully be taken into account after 8 samples (or 4 for the variance) and this needs to be carefully considered when interpreting the results. Knowing both the time it takes a lock request to complete and the average time between lock requests for a glock means we can compute the total percentage of the time for which the node is able to use a glock vs. time that the rest of the cluster has its share. That will be very useful when setting the lock min hold time. The other point to remember is that all times are in nanoseconds. Great care has been taken to ensure that we measure exactly the quantities that we want, as accurately as possible. There are always inaccuracies in any measuring system, but I hope this is as accurate as we can reasonably make it. Signed-off-by: Steven Whitehouse <swhiteho@redhat.com>
2012-01-20 10:38:36 +00:00
preempt_disable();
this_cpu_ptr(sdp->sd_lkstats)->lkstats[gl->gl_name.ln_type].stats[which]++;
preempt_enable();
}
extern struct gfs2_rgrpd *gfs2_glock2rgrp(struct gfs2_glock *gl);
static inline unsigned gfs2_max_stuffed_size(const struct gfs2_inode *ip)
{
return GFS2_SB(&ip->i_inode)->sd_sb.sb_bsize - sizeof(struct gfs2_dinode);
}
#endif /* __INCORE_DOT_H__ */