Interrupted system call messages / read errors during peak network traffic

[kallisti5]

Downloading a large file at a high rate with wget results in "Interrupted system call" messages.

Example:

wget ​https://www.kernel.org/pub/linux/kernel/v4.x/linux-4.1.tar.xz

(screenshot attached)

Test: hrev49332, x86_64

[kallisti5]

I also see the following messages in syslog with network traffic:

KERN: thread XXX tried to acquire kernel semaphore 0.
KERN: thread XXX tried to release kernel semaphore.

[waddlesplash]

Semaphore issues should've been fixed by hrev49333. Possibly related to #10747?

[pulkomandy]

"Interrupted system call" is not a bug, but a feature. When a thread gets a signal, whatever syscall it is waiting on returns immediately with error EINTR and must be restarted. This makes it possible to ctrl+C a thread while it is blocked on a read(), for example.

It's possible to configure the OS to automatically restart the syscall in that case (SA_RESTART), so the apps don't have to bother with managing EINTR.

It looks like wget is not properly configured on Haiku to set this as it wants. Maybe there is some legacy BeOS code still enabled in the wget port? I don't think there are any problems on Haiku side, however.

[kallisti5]

it seems like pkgman is also giving "Interrupted system call" messages... doesn't it use libcurl?

[pulkomandy]

Yes, it does.

strace output would help to know which syscalls are interrupted and why?

[kallisti5]

Interestingly I see this same issue with one core (SMP disabled due to other issues) every time I "resize the terminal window"

----------------------------------------------------------------------
media-video::ffmpeg-3.1.1
        /Code/haikuports/media-video/ffmpeg/ffmpeg-3.1.1.recipe
----------------------------------------------------------------------

Downloading: https://ffmpeg.org/releases/ffmpeg-3.1.1.tar.bz2 ...
--2016-07-04 21:57:44--  https://ffmpeg.org/releases/ffmpeg-3.1.1.tar.bz2
Resolving ffmpeg.org... 178.63.43.86
Connecting to ffmpeg.org|178.63.43.86|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 9327512 (8.9M) [application/x-bzip2]
Saving to: ‘/Code/haikuports/media-video/ffmpeg/download/ffmpeg-3.1.1.tar.bz2’

/Code/haikuports/media-video/ffmp  29%[=================>                                            ]   2.62M   449KB/s    in 6.2s    

2016-07-04 21:57:50 (432 KB/s) - Read error at byte 2752512/9327512 (Interrupted system call). Retrying.

--2016-07-04 21:57:51--  (try: 2)  https://ffmpeg.org/releases/ffmpeg-3.1.1.tar.bz2
Connecting to ffmpeg.org|178.63.43.86|:443... connected.
HTTP request sent, awaiting response... 206 Partial Content
Length: 9327512 (8.9M), 6575000 (6.3M) remaining [application/x-bzip2]
Saving to: ‘/Code/haikuports/media-video/ffmpeg/download/ffmpeg-3.1.1.tar.bz2’

/Code/haikuports/media-video/ffmpeg/downloa  45%[++++++++++++++++++++++++++=============>                                                 ]   4.03M   421KB/s    in 3.4s    

2016-07-04 21:57:55 (421 KB/s) - Read error at byte 4227072/9327512 (Interrupted system call). Retrying.

--2016-07-04 21:57:57--  (try: 3)  https://ffmpeg.org/releases/ffmpeg-3.1.1.tar.bz2
Connecting to ffmpeg.org|178.63.43.86|:443... connected.
HTTP request sent, awaiting response... 206 Partial Content
Length: 9327512 (8.9M), 5100440 (4.9M) remaining [application/x-bzip2]
Saving to: ‘/Code/haikuports/media-video/ffmpeg/download/ffmpeg-3.1.1.tar.bz2’

/Code/haikuports/media-video/ffmpeg/download 100%[+++++++++++++++++++++++++++++++++++++++++==================================================>]   8.89M   426KB/s    in 12s     

2016-07-04 21:58:10 (421 KB/s) - ‘/Code/haikuports/media-video/ffmpeg/download/ffmpeg-3.1.1.tar.bz2’ saved [9327512/9327512]
"

Every time that error is seen, i've resized the terminal window :-|

[pulkomandy]

This is a feature: resizing a terminal window (also in xterm) sends the SIGWINCH signal ("window changed"). This allows ncurses apps to adjust their size to the terminal when it is resized.

wget uses SIGWINCH to resize its progress bar (see ​https://github.com/juaristi/wget/blob/master/src/progress.c).

We should be using ​https://github.com/cooljeanius/wget/blob/master/gnulib/lib/safe-read.c which protects against this by using IS_EINTR (not sure where that comes from).

Sounds like a bug in how we configure and build wget, and nothing wrong on Haiku side?

[axeld]

It could be a bug in how we build wget, but it may also be a bug on the Haiku side -- that's something someone should look into :-)

BTW IS_EINTR comes from the same source file, just a couple of lines above (line 35ff).

[pulkomandy]

I see that this only happens with SSL (https) downloads. So I think the problem is in openssl, which is used by both wget and curl.

wget does check for EINTR from OpenSSL: ​https://github.com/jnothman/wget/blob/master/src/openssl.c#L269 (OpenSSL just forwards the error and does nothing special).

[pulkomandy]

I'm still tracking this down...

wget properly manages this in ​https://github.com/greearb/ct-wget/blob/master/src/openssl.c#L360

However, our OpenSSL does not behave as expected, it gives an SSL_WANT_READ instead of an SSL_ERROR_SYSCALL and we jump out of the loop. Digging into OpenSSL sources next...

[pulkomandy]

Fixed in ​https://github.com/haikuports/haikuports/commit/8af9d2525297b6c201a61f339d7b97bcf7982b45