summaryrefslogtreecommitdiffstats
path: root/README
diff options
context:
space:
mode:
authorgraham.gower <graham.gower@e8e0d7a0-c8d9-11dd-a880-a1081c7ac358>2009-11-19 17:40:55 (EST)
committer graham.gower <graham.gower@e8e0d7a0-c8d9-11dd-a880-a1081c7ac358>2009-11-19 17:40:55 (EST)
commitb6c404659768c0be9d4d7b93a8d475c76ee26380 (patch)
tree876956fca419e79ccd2ad4b630ec270e112661f2 /README
parentc03e55b08e74a0d6ddb8cfd3d691444dd911a90e (diff)
Fix hang in waitpid, exposed by r310. Patch from Enrico Scholz
<enrico.scholz@informatik.tu-chemnitz.de>, his analysis follows. libbb/unarchive.c: prevent opkg hang when subprocess is stuck in a write() call on a filled pipe and main process (which assumes that end of data from pipe has been reached) waits for this died subprocess. This patch swaps the original wait(pid) + close(pipe) sequencse so that pipe is closed first. The 'ar' code path has been fixed too by breaking the loop when requested data have been found. Previously, the loop continued at the (wrongly calculated) next position in the stream. The patch moves the stream cleanup at a better place. Variable declarations were moved to inner scopes too to ease detection of broken deallocation. NOTE: the | f = fdopen(...); | while (... /* do fread(f) */ ...) { /* ==1== */ | /* this is done in gz_open() */ | pid = fork(); | if (pid == 0) { | fread(f); /* ==2== */ | _exit(0); | } | } code looks problematic because '==2==' might update f's internal buffer. As ==2== is done in an own process, these changes are not seen by ==1==. It works only because gz_open() is called not more than one time and the loops break (after the patch). git-svn-id: http://opkg.googlecode.com/svn/trunk@340 e8e0d7a0-c8d9-11dd-a880-a1081c7ac358
Diffstat (limited to 'README')
0 files changed, 0 insertions, 0 deletions