View Issue Details
| ID | Project | Category | View Status | Date Submitted | Last Update |
|---|---|---|---|---|---|
| 0001536 | libmicrohttpd | documentation | public | 2010-01-19 21:24 | 2010-07-26 23:33 |
| Reporter | squishyent | Assigned To | Christian Grothoff | ||
| Priority | normal | Severity | text | Reproducibility | N/A |
| Status | closed | Resolution | fixed | ||
| Product Version | 0.4.4 | ||||
| Summary | 0001536: Occasionally crashes with SIGPIPE due to client-terminated connections in Solaris 8 | ||||
| Description | I am currently using libmicrohttpd in a C++ server which is using HTTP as its primary request protocol. This server is running both on Solaris 8 on a SPARC processor and Red Hat Linux on an x86 processor. Some of the requests can take a couple of minutes before posting responses, and if the requesting client terminates the connection before the work is complete the program will occasionally crash. I believe I have tracked the problem down to receiving SIGPIPE due to an attempt to write to a closed SOCK_STREAM socket connection. I have managed to reliably reproduce this problem in Solaris (but not Linux!) using the following steps: 1. Start with the hellobrowser.c tutorial example. 2. Add a significant delay in the request by adding a call to sleep( 3 ) to the beginning of answer_to_connection(). 3. Start the server and use a web browser to load a page from it. This should work. 4. Hold down F5 to cause the browser to refresh the page many times. On one of the calls, after answer_to_connection() returns, the program receives SIGPIPE and dies (with the message "Broken pipe" sent to STDERR). Some notes: * Using Google Chrome on Windows XP, I was able to make this happen every time when returning from the fifth call to answer_to_connection() after six or more refreshes. This may be useless information, as I don't know when Chrome closes its connections, but fewer than six refreshes will not cause the problem to occur. * If the program is modified to report errors, it sometimes sends the following message to the error handler immediately before dying: "Failed to send data: Broken pipe". I never see this message when the program doesn't die. It certainly seems odd that we can get both a message in the error handler and an uncaught SIGPIPE which kills the program. This is wild speculation, but perhaps Solaris 8 is not re-registering the signal handler on the first call, and a subsequent attempt to write to the same socket causes an unhandled SIGPIPE to be sent? | ||||
| Tags | No tags attached. | ||||
|
|
Note that if I call sigignore( SIGPIPE ) before MHD_start_daemon() is called, the issue seems to be resolved. I can refresh the page as much as I want. I will get occasional "Failed to send data: Broken pipe" errors reported, and eventually I start to see "Server reached connection limit (closing inbound connection)" errors, and life is wonderful. I am going to add this call to my program as a workaround for the problem. |
|
|
This is not really "just" a workaround but the (strongly) suggested solution. MHD does not install a handler for SIGPIPE since, as a library, messing with signal handlers might cause problems for an app that has (conflicting) handlers installed already. So on platforms where SIGPIPE cannot be elimianted by MSG_NOSIGNAL, MHD requires developers to install a handler. Now, the examples don't contain this -- largely to keep them simple. Still, the documentation might want to stress this point a bit more, so I'll file this bug report under "improve documentation". |
|
|
I can't reproduce this problem on Windows Vista with Chrome and the fileserver_example. |
|
|
Issue documented in MHD manual for 0.9 (introduction, subsection 1.1). |
| Date Modified | Username | Field | Change |
|---|---|---|---|
| 2010-01-19 21:24 | squishyent | New Issue | |
| 2010-01-19 22:11 | squishyent | Note Added: 0003991 | |
| 2010-01-20 20:21 | Christian Grothoff | Note Added: 0003992 | |
| 2010-01-20 20:21 | Christian Grothoff | Severity | crash => text |
| 2010-01-20 20:21 | Christian Grothoff | Reproducibility | have not tried => N/A |
| 2010-01-20 20:21 | Christian Grothoff | Status | new => confirmed |
| 2010-01-20 20:21 | Christian Grothoff | Category | portability => documentation |
| 2010-01-20 22:26 | NDurner | Note Added: 0003993 | |
| 2010-07-26 20:12 | Christian Grothoff | Status | confirmed => assigned |
| 2010-07-26 20:12 | Christian Grothoff | Assigned To | => Christian Grothoff |
| 2010-07-26 20:12 | Christian Grothoff | Note Added: 0004078 | |
| 2010-07-26 20:12 | Christian Grothoff | Status | assigned => resolved |
| 2010-07-26 20:12 | Christian Grothoff | Resolution | open => fixed |
| 2010-07-26 23:33 | Christian Grothoff | Status | resolved => closed |
