While working on bug 439144, by code inspection, Wan-Teh found a nasty bug in a routine common to NSPR sendfile and transmit file on HPUX only. This may well have been causing crashes or data corruption in sent files. The patch he wrote for it is: >@@ -1012,17 +1012,17 @@ static PRBool pt_hpux_sendfile_cont(pt_C >- hdtrl[0].iov_base = ((char *) hdtrl[0].iov_len) + count; >+ hdtrl[0].iov_base = ((char *) hdtrl[0].iov_base) + count; > hdtrl[0].iov_len -= count;

This bug was introduced in NSPR 3.5 in 1999, and has been present in all version of NSPR for HPUX produced since then.

I guess it's extremely rare for a non-blocking sendfile call to be unable to send the entire header before blocking, which is why we haven't seen reports of this crash.

Comment on attachment 376582 [details] [diff] [review] Proposed patch r=nelson

I checked in the patch on the NSPR trunk (NSPR 4.8). Checking in ptio.c; /cvsroot/mozilla/nsprpub/pr/src/pthreads/ptio.c,v <-- ptio.c new revision: 3.113; previous revision: 3.112 done

I backported the patch (attachment 376582 [details] [diff] [review]) to the NSPR_4_7_BRANCH for NSPR 4.7.5. Checking in ptio.c; /cvsroot/mozilla/nsprpub/pr/src/pthreads/ptio.c,v <-- ptio.c new revision: 3.110.4.1; previous revision: 3.110 done