summaryrefslogtreecommitdiff
path: root/open_issues/alarm_setitimer.mdwn
blob: 5999808ceeaf88198a000a7ee7db8cb8f30f3b31 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
[[!meta copyright="Copyright © 2012, 2013 Free Software Foundation, Inc."]]

[[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable
id="license" text="Permission is granted to copy, distribute and/or modify this
document under the terms of the GNU Free Documentation License, Version 1.2 or
any later version published by the Free Software Foundation; with no Invariant
Sections, no Front-Cover Texts, and no Back-Cover Texts.  A copy of the license
is included in the section entitled [[GNU Free Documentation
License|/fdl]]."]]"""]]

[[!meta title="alarm/setitimer"]]

[[!tag open_issue_glibc open_issue_hurd]]

`setitimer()`, called by `alarm()` when setting a new alarm, it is not able
to disable on its own the timer when the alarm is fired the first time.
On the other hand, manually invoking `alarm(0)` can cancel the running timer
for `SIGALRM`.

See also the attached file: on other OSes (e.g. Linux) it blocks waiting
for a signal, while on GNU/Hurd it gets a new alarm and exits.

[[alrm.c]]


# IRC, freenode, #hurd, 2012-07-29

    <braunr> our setitimer is bugged
    <braunr> it seems doesn't seem to leave a timer disarmed when the interval
      is set to 0
    <braunr> (which means a one shot timer is actually periodic ..)


## IRC, freenode, #hurd, 2012-12-26

    <braunr> youpi: tschwinge: the setitimer issue
      http://www.gnu.org/software/hurd/open_issues/alarm_setitimer.html) is
      because of the global preemptor installed by setitimer not being run when
      sigalrm is catched
    <braunr> if anyone has a good definition for a preemptor, let us know (mine
      is currently "something that is scanned on signal delivery and can alter
      this delivery")
    <youpi> I don't have any better definition
    <pinotree> braunr: ah, that explains indeed
    <pinotree> thanks
    <braunr> i think i found the problem :)
    <braunr> seems to be a minor overlook from drepper
    <braunr> (or the real author if he was only the committer)
    <braunr> hurd_preempt_signals augments _hurdsig_preempted_set with the
      signals from the installed preemptor
    <braunr> but the inline version in setitimer doesn't
    <braunr> and post_signal actually checks that
    <braunr> the preemptor itself looks wrong, since its sigcode range is 0, 0
      whereas SI_TIMER is used when raising SIGALRM ...
    <braunr> ah but that's a recent change, right
    <braunr> it came with "implement SA_SIGINFO signal handlers"
      (e19a2fad70b187e5efe79768f86a1f05cb5e0390, Tue Feb 21 02:41:18 2012)
    <braunr> yes, fixed :)
    <braunr> patch committed at
      http://git.savannah.gnu.org/cgit/hurd/glibc.git/log/?h=rbraun/setitimer_fix
    <youpi> and pushed to the debian package


## IRC, freenode, #hurd, 2012-12-27

    <braunr> do we know any application that was broken because of setitimer ?
    <pinotree> braunr: bits in the python and perl test suites
    <braunr> ok


## IRC, freenode, #hurd, 2012-12-28

    <pinotree> braunr: ah, also libglib-perl's testsuite is affected by the
      alarm/setitimer issue
    <braunr> pinotree: only tests ? :(
    <pinotree> braunr: yeah
    <braunr> ok, we don't win that much on this fix, but anyway, still good to
      have
    <pinotree> but that source is pretty quick to compile and check
    <pinotree> braunr: eh, so far that's what i found myself


## IRC, freenode, #hurd, 2013-01-04

See also [[select]].

    <youpi> bummer, we have broken ghc completely with the latest glibc patches
    <pinotree> youpi: what do you mean?
    <youpi> pinotree: it just hangs on installation


## IRC, freenode, #hurd, 2013-01-05

    <youpi> pinotree: it seems ghc was disturbed by the setitimer patch
    <youpi> pinotree: http://paste.debian.net/221807/
    <youpi> pinotree: it seems to be simply due to nested locking of
      _hurd_siglock :/
    <youpi> pinotree: I wonder whether this code has ever been really tested
    <youpi> oops
    <youpi> braunr: my comments above were for you actually :)
    <youpi> braunr: see the update I've just commited to the debian patch
    <youpi> I've added a parameter to setitimer_locked, to know whether the
      lock is already taken or not
    <youpi> that does fix ghc
    <youpi> as well as the gdb ntpdate hang, apparently
    <youpi> I can confirm that the single-select patch breaks ntpdate for some
      reason
    <youpi> I wonder whether it could be due to port set behavior being
      different from single reply port
    <youpi> I believe I understand what happens

[[select_vs_signals]].

    <youpi> I'll rebuild ntpdate with a 1s timeout
    <youpi> that'll at least fix that
    <youpi> rah, no, doesn't work, it insists on getting its alarm
    <youpi> Mmm, no, the __mach_msg call doesn't even return
    <youpi> even though MACH_RCV_TIMEOUT is set, and to is 1000
    <braunr> youpi: i see
    <braunr> gnu_srs: and you, see how youpi analysed and understood the
      problem, instead of just guessing :p
    <braunr> youpi: it doesn't return ?
    <braunr> iirc, the __mach_msg wrapper deals with the interruptible flag
    <youpi> braunr: yes, __mach_msg deals with the interruptible flag by
      looping !
    <youpi> and the info page says it: if it's interrupted too often, it may
      just never return
    <youpi> that's what actually happens here
    <youpi> (ntpdate sets an itimer more often than every 1s)
    <braunr> youpi: ew :)
    <youpi> I'll test a bit more, and submit a patch
    <pinotree> youpi: otoh a _locker function usually means it expects a locked
      mutex ;)
    <pinotree> i also i wondered whether there could be a race in the settimer
      mini-thread, between its mach_msg and its reading of the interval
    <youpi> pinotree: right, we could as well just lock anyway
    <youpi> there could be indeed
    <pinotree> youpi: i don't know much about the internals of signal
      dispatching, but could it happen the following:
    <pinotree> in timer thread, mach_msg expires → sig_post_request → before
      the main thread receives/processes the signal, the timer thread iterates
      again on its while(1), using the same interval previously used
    <pinotree> ?
    <youpi> did you check the comment above __msg_sig_post_request?
    <pinotree> ah ok
    <youpi> I'm not sure how that works, but it's supposed to :)
    <pinotree> just wonder: wouldn't it be simplier if the logic to change the
      timeout would be in the timer thread, instead of relying on the main
      thread adjusting it?
    <youpi> maybe there are some semantic details that wouldn't be right with
      such approach
    <pinotree> i see
    <pinotree> i guess so if the new interval is 0, the thread can be properly
      suspened (or killed, if the former fails)
    <youpi> could be something like this, yes
    <pinotree> youpi: ah, wrt your comments of tonight: at least with the
      current setitimer patch (in -38), a simple alarm() test app works, and i
      saw few python tests can be reenabled now
    <youpi> ok
    <pinotree> so even if not totally correct, at least it had some positive
      effects
    <pinotree> youpi: wrt the double lock issue of _hurd_siglock, what about
      using the "crit" parameter of setitimer_locked?
    <youpi> it may have various values
    <youpi> depending whether we're already in the critical section etc.
    <pinotree> restart_itimer does not take that lock, so we could check
      whether crit is null, and in that case not even bothering to check the
      signal preemptors, since it was called as a result of own setitimer
      thread?
    <youpi> I'd rather avoid binding whether the mutex is held to whether the
      call is coming from the actual premptor
    <youpi> again, crit may be null if we're already in the critical section
      when setitimer is called
    <braunr> setitimer already does unclean things with preemptors
    <youpi> not a good thing to add more :)
    <pinotree> fair enough, so a simple bool should do the job
    <braunr> i mean, the whole thing is "cheezoid" :)
    <braunr> it probably needs a rewrite some day
    <braunr> so "in the meantime" (of years, i know)
    <pinotree> braunr: and temporary, too
    <braunr> but a bool is fine too, sure :)