blob: fcc79957be63ef6c7110f78e83c50b00d24dcecd [file] [log] [blame]
Steven Whitehouse9f1585c2008-06-26 08:25:57 +01001 Glock internal locking rules
2 ------------------------------
3
4This documents the basic principles of the glock state machine
5internals. Each glock (struct gfs2_glock in fs/gfs2/incore.h)
6has two main (internal) locks:
7
8 1. A spinlock (gl_spin) which protects the internal state such
9 as gl_state, gl_target and the list of holders (gl_holders)
10 2. A non-blocking bit lock, GLF_LOCK, which is used to prevent other
11 threads from making calls to the DLM, etc. at the same time. If a
12 thread takes this lock, it must then call run_queue (usually via the
13 workqueue) when it releases it in order to ensure any pending tasks
14 are completed.
15
16The gl_holders list contains all the queued lock requests (not
17just the holders) associated with the glock. If there are any
18held locks, then they will be contiguous entries at the head
19of the list. Locks are granted in strictly the order that they
20are queued, except for those marked LM_FLAG_PRIORITY which are
21used only during recovery, and even then only for journal locks.
22
23There are three lock states that users of the glock layer can request,
24namely shared (SH), deferred (DF) and exclusive (EX). Those translate
25to the following DLM lock modes:
26
27Glock mode | DLM lock mode
28------------------------------
29 UN | IV/NL Unlocked (no DLM lock associated with glock) or NL
30 SH | PR (Protected read)
31 DF | CW (Concurrent write)
32 EX | EX (Exclusive)
33
34Thus DF is basically a shared mode which is incompatible with the "normal"
35shared lock mode, SH. In GFS2 the DF mode is used exclusively for direct I/O
36operations. The glocks are basically a lock plus some routines which deal
37with cache management. The following rules apply for the cache:
38
39Glock mode | Cache data | Cache Metadata | Dirty Data | Dirty Metadata
40--------------------------------------------------------------------------
41 UN | No | No | No | No
42 SH | Yes | Yes | No | No
43 DF | No | Yes | No | No
44 EX | Yes | Yes | Yes | Yes
45
46These rules are implemented using the various glock operations which
47are defined for each type of glock. Not all types of glocks use
48all the modes. Only inode glocks use the DF mode for example.
49
50Table of glock operations and per type constants:
51
52Field | Purpose
53----------------------------------------------------------------------------
54go_xmote_th | Called before remote state change (e.g. to sync dirty data)
55go_xmote_bh | Called after remote state change (e.g. to refill cache)
56go_inval | Called if remote state change requires invalidating the cache
57go_demote_ok | Returns boolean value of whether its ok to demote a glock
58 | (e.g. checks timeout, and that there is no cached data)
59go_lock | Called for the first local holder of a lock
60go_unlock | Called on the final local unlock of a lock
61go_dump | Called to print content of object for debugfs file, or on
62 | error to dump glock to the log.
Steven Whitehousee9ccb732009-05-19 10:23:23 +010063go_type | The type of the glock, LM_TYPE_.....
Steven Whitehouse2ebc3f82012-05-10 12:41:40 +010064go_callback | Called if the DLM sends a callback to drop this lock
65go_flags | GLOF_ASPACE is set, if the glock has an address space
66 | associated with it
Steven Whitehouse9f1585c2008-06-26 08:25:57 +010067
68The minimum hold time for each lock is the time after a remote lock
69grant for which we ignore remote demote requests. This is in order to
70prevent a situation where locks are being bounced around the cluster
71from node to node with none of the nodes making any progress. This
72tends to show up most with shared mmaped files which are being written
73to by multiple nodes. By delaying the demotion in response to a
74remote callback, that gives the userspace program time to make
75some progress before the pages are unmapped.
76
77There is a plan to try and remove the go_lock and go_unlock callbacks
78if possible, in order to try and speed up the fast path though the locking.
79Also, eventually we hope to make the glock "EX" mode locally shared
80such that any local locking will be done with the i_mutex as required
81rather than via the glock.
82
83Locking rules for glock operations:
84
85Operation | GLF_LOCK bit lock held | gl_spin spinlock held
86-----------------------------------------------------------------
87go_xmote_th | Yes | No
88go_xmote_bh | Yes | No
89go_inval | Yes | No
90go_demote_ok | Sometimes | Yes
91go_lock | Yes | No
92go_unlock | Yes | No
93go_dump | Sometimes | Yes
Steven Whitehouse2ebc3f82012-05-10 12:41:40 +010094go_callback | Sometimes (N/A) | Yes
Steven Whitehouse9f1585c2008-06-26 08:25:57 +010095
96N.B. Operations must not drop either the bit lock or the spinlock
97if its held on entry. go_dump and do_demote_ok must never block.
98Note that go_dump will only be called if the glock's state
99indicates that it is caching uptodate data.
100
101Glock locking order within GFS2:
102
103 1. i_mutex (if required)
104 2. Rename glock (for rename only)
105 3. Inode glock(s)
106 (Parents before children, inodes at "same level" with same parent in
107 lock number order)
108 4. Rgrp glock(s) (for (de)allocation operations)
109 5. Transaction glock (via gfs2_trans_begin) for non-read operations
110 6. Page lock (always last, very important!)
111
112There are two glocks per inode. One deals with access to the inode
113itself (locking order as above), and the other, known as the iopen
114glock is used in conjunction with the i_nlink field in the inode to
115determine the lifetime of the inode in question. Locking of inodes
116is on a per-inode basis. Locking of rgrps is on a per rgrp basis.
Steven Whitehouse2ebc3f82012-05-10 12:41:40 +0100117In general we prefer to lock local locks prior to cluster locks.
118
119 Glock Statistics
120 ------------------
121
122The stats are divided into two sets: those relating to the
123super block and those relating to an individual glock. The
124super block stats are done on a per cpu basis in order to
125try and reduce the overhead of gathering them. They are also
126further divided by glock type. All timings are in nanoseconds.
127
128In the case of both the super block and glock statistics,
129the same information is gathered in each case. The super
130block timing statistics are used to provide default values for
131the glock timing statistics, so that newly created glocks
132should have, as far as possible, a sensible starting point.
133The per-glock counters are initialised to zero when the
134glock is created. The per-glock statistics are lost when
135the glock is ejected from memory.
136
137The statistics are divided into three pairs of mean and
138variance, plus two counters. The mean/variance pairs are
139smoothed exponential estimates and the algorithm used is
140one which will be very familiar to those used to calculation
141of round trip times in network code. See "TCP/IP Illustrated,
142Volume 1", W. Richard Stevens, sect 21.3, "Round-Trip Time Measurement",
143p. 299 and onwards. Also, Volume 2, Sect. 25.10, p. 838 and onwards.
144Unlike the TCP/IP Illustrated case, the mean and variance are
145not scaled, but are in units of integer nanoseconds.
146
147The three pairs of mean/variance measure the following
148things:
149
150 1. DLM lock time (non-blocking requests)
151 2. DLM lock time (blocking requests)
152 3. Inter-request time (again to the DLM)
153
154A non-blocking request is one which will complete right
155away, whatever the state of the DLM lock in question. That
156currently means any requests when (a) the current state of
157the lock is exclusive, i.e. a lock demotion (b) the requested
158state is either null or unlocked (again, a demotion) or (c) the
159"try lock" flag is set. A blocking request covers all the other
160lock requests.
161
162There are two counters. The first is there primarily to show
163how many lock requests have been made, and thus how much data
164has gone into the mean/variance calculations. The other counter
165is counting queuing of holders at the top layer of the glock
166code. Hopefully that number will be a lot larger than the number
167of dlm lock requests issued.
168
169So why gather these statistics? There are several reasons
170we'd like to get a better idea of these timings:
171
1721. To be able to better set the glock "min hold time"
1732. To spot performance issues more easily
1743. To improve the algorithm for selecting resource groups for
175allocation (to base it on lock wait time, rather than blindly
176using a "try lock")
177
178Due to the smoothing action of the updates, a step change in
179some input quantity being sampled will only fully be taken
180into account after 8 samples (or 4 for the variance) and this
181needs to be carefully considered when interpreting the
182results.
183
184Knowing both the time it takes a lock request to complete and
185the average time between lock requests for a glock means we
186can compute the total percentage of the time for which the
187node is able to use a glock vs. time that the rest of the
188cluster has its share. That will be very useful when setting
189the lock min hold time.
190
191Great care has been taken to ensure that we
192measure exactly the quantities that we want, as accurately
193as possible. There are always inaccuracies in any
194measuring system, but I hope this is as accurate as we
195can reasonably make it.
196
197Per sb stats can be found here:
198/sys/kernel/debug/gfs2/<fsname>/sbstats
199Per glock stats can be found here:
200/sys/kernel/debug/gfs2/<fsname>/glstats
201
202Assuming that debugfs is mounted on /sys/kernel/debug and also
203that <fsname> is replaced with the name of the gfs2 filesystem
204in question.
205
206The abbreviations used in the output as are follows:
207
208srtt - Smoothed round trip time for non-blocking dlm requests
209srttvar - Variance estimate for srtt
210srttb - Smoothed round trip time for (potentially) blocking dlm requests
211srttvarb - Variance estimate for srttb
212sirt - Smoothed inter-request time (for dlm requests)
213sirtvar - Variance estimate for sirt
214dlm - Number of dlm requests made (dcnt in glstats file)
215queue - Number of glock requests queued (qcnt in glstats file)
216
217The sbstats file contains a set of these stats for each glock type (so 8 lines
218for each type) and for each cpu (one column per cpu). The glstats file contains
219a set of these stats for each glock in a similar format to the glocks file, but
220using the format mean/variance for each of the timing stats.
221
222The gfs2_glock_lock_time tracepoint prints out the current values of the stats
223for the glock in question, along with some addition information on each dlm
224reply that is received:
225
226status - The status of the dlm request
227flags - The dlm request flags
228tdiff - The time taken by this specific request
229(remaining fields as per above list)
230
Steven Whitehouse9f1585c2008-06-26 08:25:57 +0100231