Commit | Line | Data |
---|---|---|
80aafd50 JL |
1 | The errseq_t datatype |
2 | ===================== | |
3 | An errseq_t is a way of recording errors in one place, and allowing any | |
4 | number of "subscribers" to tell whether it has changed since a previous | |
5 | point where it was sampled. | |
6 | ||
7 | The initial use case for this is tracking errors for file | |
8 | synchronization syscalls (fsync, fdatasync, msync and sync_file_range), | |
9 | but it may be usable in other situations. | |
10 | ||
11 | It's implemented as an unsigned 32-bit value. The low order bits are | |
12 | designated to hold an error code (between 1 and MAX_ERRNO). The upper bits | |
13 | are used as a counter. This is done with atomics instead of locking so that | |
14 | these functions can be called from any context. | |
15 | ||
16 | Note that there is a risk of collisions if new errors are being recorded | |
17 | frequently, since we have so few bits to use as a counter. | |
18 | ||
19 | To mitigate this, the bit between the error value and counter is used as | |
20 | a flag to tell whether the value has been sampled since a new value was | |
21 | recorded. That allows us to avoid bumping the counter if no one has | |
22 | sampled it since the last time an error was recorded. | |
23 | ||
24 | Thus we end up with a value that looks something like this:: | |
25 | ||
26 | bit: 31..13 12 11..0 | |
27 | +-----------------+----+----------------+ | |
28 | | counter | SF | errno | | |
29 | +-----------------+----+----------------+ | |
30 | ||
31 | The general idea is for "watchers" to sample an errseq_t value and keep | |
32 | it as a running cursor. That value can later be used to tell whether | |
33 | any new errors have occurred since that sampling was done, and atomically | |
34 | record the state at the time that it was checked. This allows us to | |
35 | record errors in one place, and then have a number of "watchers" that | |
36 | can tell whether the value has changed since they last checked it. | |
37 | ||
38 | A new errseq_t should always be zeroed out. An errseq_t value of all zeroes | |
39 | is the special (but common) case where there has never been an error. An all | |
40 | zero value thus serves as the "epoch" if one wishes to know whether there | |
41 | has ever been an error set since it was first initialized. | |
42 | ||
43 | API usage | |
44 | ========= | |
45 | Let me tell you a story about a worker drone. Now, he's a good worker | |
46 | overall, but the company is a little...management heavy. He has to | |
47 | report to 77 supervisors today, and tomorrow the "big boss" is coming in | |
48 | from out of town and he's sure to test the poor fellow too. | |
49 | ||
50 | They're all handing him work to do -- so much he can't keep track of who | |
51 | handed him what, but that's not really a big problem. The supervisors | |
52 | just want to know when he's finished all of the work they've handed him so | |
53 | far and whether he made any mistakes since they last asked. | |
54 | ||
55 | He might have made the mistake on work they didn't actually hand him, | |
56 | but he can't keep track of things at that level of detail, all he can | |
57 | remember is the most recent mistake that he made. | |
58 | ||
59 | Here's our worker_drone representation:: | |
60 | ||
61 | struct worker_drone { | |
62 | errseq_t wd_err; /* for recording errors */ | |
63 | }; | |
64 | ||
65 | Every day, the worker_drone starts out with a blank slate:: | |
66 | ||
67 | struct worker_drone wd; | |
68 | ||
69 | wd.wd_err = (errseq_t)0; | |
70 | ||
71 | The supervisors come in and get an initial read for the day. They | |
72 | don't care about anything that happened before their watch begins:: | |
73 | ||
74 | struct supervisor { | |
75 | errseq_t s_wd_err; /* private "cursor" for wd_err */ | |
76 | spinlock_t s_wd_err_lock; /* protects s_wd_err */ | |
77 | } | |
78 | ||
79 | struct supervisor su; | |
80 | ||
81 | su.s_wd_err = errseq_sample(&wd.wd_err); | |
82 | spin_lock_init(&su.s_wd_err_lock); | |
83 | ||
84 | Now they start handing him tasks to do. Every few minutes they ask him to | |
85 | finish up all of the work they've handed him so far. Then they ask him | |
86 | whether he made any mistakes on any of it:: | |
87 | ||
88 | spin_lock(&su.su_wd_err_lock); | |
89 | err = errseq_check_and_advance(&wd.wd_err, &su.s_wd_err); | |
90 | spin_unlock(&su.su_wd_err_lock); | |
91 | ||
92 | Up to this point, that just keeps returning 0. | |
93 | ||
94 | Now, the owners of this company are quite miserly and have given him | |
95 | substandard equipment with which to do his job. Occasionally it | |
96 | glitches and he makes a mistake. He sighs a heavy sigh, and marks it | |
97 | down:: | |
98 | ||
99 | errseq_set(&wd.wd_err, -EIO); | |
100 | ||
101 | ...and then gets back to work. The supervisors eventually poll again | |
102 | and they each get the error when they next check. Subsequent calls will | |
103 | return 0, until another error is recorded, at which point it's reported | |
104 | to each of them once. | |
105 | ||
106 | Note that the supervisors can't tell how many mistakes he made, only | |
107 | whether one was made since they last checked, and the latest value | |
108 | recorded. | |
109 | ||
110 | Occasionally the big boss comes in for a spot check and asks the worker | |
111 | to do a one-off job for him. He's not really watching the worker | |
112 | full-time like the supervisors, but he does need to know whether a | |
113 | mistake occurred while his job was processing. | |
114 | ||
115 | He can just sample the current errseq_t in the worker, and then use that | |
116 | to tell whether an error has occurred later:: | |
117 | ||
118 | errseq_t since = errseq_sample(&wd.wd_err); | |
119 | /* submit some work and wait for it to complete */ | |
120 | err = errseq_check(&wd.wd_err, since); | |
121 | ||
122 | Since he's just going to discard "since" after that point, he doesn't | |
123 | need to advance it here. He also doesn't need any locking since it's | |
124 | not usable by anyone else. | |
125 | ||
126 | Serializing errseq_t cursor updates | |
127 | =================================== | |
128 | Note that the errseq_t API does not protect the errseq_t cursor during a | |
129 | check_and_advance_operation. Only the canonical error code is handled | |
130 | atomically. In a situation where more than one task might be using the | |
131 | same errseq_t cursor at the same time, it's important to serialize | |
132 | updates to that cursor. | |
133 | ||
134 | If that's not done, then it's possible for the cursor to go backward | |
135 | in which case the same error could be reported more than once. | |
136 | ||
137 | Because of this, it's often advantageous to first do an errseq_check to | |
138 | see if anything has changed, and only later do an | |
139 | errseq_check_and_advance after taking the lock. e.g.:: | |
140 | ||
141 | if (errseq_check(&wd.wd_err, READ_ONCE(su.s_wd_err)) { | |
142 | /* su.s_wd_err is protected by s_wd_err_lock */ | |
143 | spin_lock(&su.s_wd_err_lock); | |
144 | err = errseq_check_and_advance(&wd.wd_err, &su.s_wd_err); | |
145 | spin_unlock(&su.s_wd_err_lock); | |
146 | } | |
147 | ||
148 | That avoids the spinlock in the common case where nothing has changed | |
149 | since the last time it was checked. |