Messages in this thread | | | Date | Tue, 2 Sep 2008 15:14:26 -0700 | From | Andrew Morton <> | Subject | Re: [PATCH] Provide control over core name even for multithreaded processes |
| |
On Fri, 29 Aug 2008 15:22:15 -0400 "La Monte H.P. Yarroll" <piggy@laurelnetworks.com> wrote:
> Our system explicitly sets core_pattern to force a relatively small > limit on the number of different core dump names. > > Unfortunately, the kernel unconditionally appends .<tid> for > multithreaded processes. > > The attached patch introduces "%T" to the core_pattern language, which > expands to ".<tid>" for multithreaded processes. It then changes the > default core_pattern to "core%T" replicating the current default > behavior. > > Signed-off-by: La Monte H.P. Yarroll <piggy@laurelnetworks.com> 74-256740 > > diff -Naur linux-2.6.26.3/Documentation/sysctl/kernel.txt linux-2.6.26.3-new/Documentation/sysctl/kernel.txt > --- linux-2.6.26.3/Documentation/sysctl/kernel.txt 2008-08-20 14:11:37.000000000 -0400 > +++ linux-2.6.26.3-new/Documentation/sysctl/kernel.txt 2008-08-28 13:27:50.239325000 -0400 > @@ -88,7 +88,7 @@ > core_pattern: > > core_pattern is used to specify a core dumpfile pattern name. > -. max length 128 characters; default value is "core" > +. max length 128 characters; default value is "core%T" > . core_pattern is used as a pattern template for the output filename; > certain string patterns (beginning with '%') are substituted with > their actual values. > @@ -104,6 +104,7 @@ > %g gid > %s signal number > %t UNIX time of dump > + %T .tgid or nothing if process is not threaded > %h hostname > %e executable filename > %<OTHER> both are dropped > diff -Naur linux-2.6.26.3/fs/exec.c linux-2.6.26.3-new/fs/exec.c > --- linux-2.6.26.3/fs/exec.c 2008-08-20 14:11:37.000000000 -0400 > +++ linux-2.6.26.3-new/fs/exec.c 2008-08-28 13:46:57.919768000 -0400 > @@ -66,7 +66,7 @@ > #endif > > int core_uses_pid; > -char core_pattern[CORENAME_MAX_SIZE] = "core"; > +char core_pattern[CORENAME_MAX_SIZE] = "core%T"; > int suid_dumpable = 0; > > /* The maximal length of core_pattern is also specified in sysctl.c */ > @@ -1415,7 +1415,7 @@ > case 'p': > pid_in_pattern = 1; > rc = snprintf(out_ptr, out_end - out_ptr, > - "%d", task_tgid_vnr(current)); > + "%d", task_pid_vnr(current)); > if (rc > out_end - out_ptr) > goto out; > out_ptr += rc; > @@ -1455,6 +1455,17 @@ > out_ptr += rc; > break; > } > + /* Conditional tgid */ > + case 'T': { > + if (atomic_read(¤t->mm->mm_users) != 1) { > + rc = snprintf(out_ptr, out_end - out_ptr, > + ".%d", task_tgid_vnr(current)); > + if (rc > out_end - out_ptr) > + goto out; > + out_ptr += rc; > + } > + break; > + } > /* hostname */ > case 'h': > down_read(&uts_sem); > @@ -1492,10 +1503,9 @@ > * If core_pattern does not include a %p (as is the default) > * and core_uses_pid is set, then .%pid will be appended to > * the filename. Do not do this for piped commands. */ > - if (!ispipe && !pid_in_pattern > - && (core_uses_pid || atomic_read(¤t->mm->mm_users) != 1)) { > + if (!ispipe && !pid_in_pattern && core_uses_pid) { > rc = snprintf(out_ptr, out_end - out_ptr, > - ".%d", task_tgid_vnr(current)); > + ".%d", current->pid); > if (rc > out_end - out_ptr) > goto out; > out_ptr += rc;
Please review Oleg's http://userweb.kernel.org/~akpm/mmotm/broken-out/coredump-format_corename-dont-append-%25pid-if-multi-threaded.patch, which I have queued for 2.6.28.
| |