Re: [pushed] Block signals during the initial part of dlopen

[Florian Weimer <fw at deneb dot enyo dot de>]

* Florian Weimer:

> * Joseph Myers:
>
>> On Wed, 27 Nov 2019, Sourceware to Gerrit sync (Code Review) wrote:
>>
>>> Block signals during the initial part of dlopen
>>> 
>>> Lazy binding in a signal handler that interrupts a dlopen sees
>>> intermediate dynamic linker state.  This has likely been always
>>> unsafe, but with the new pending NODELETE state, this is clearly
>>> incorrect.  Other threads are excluded via the loader lock, but the
>>> current thread is not.  Blocking signals until right before ELF
>>> constructors run is the safe thing to do.
>>
>> This change (commit a2e8aa0d9ea648068d8be52dd7b15f1b6a008e23) breaks the 
>> build for i686-gnu (unfortunately this was after a syntax error was 
>> introduced in build-many-glibcs.py so my bots quietly died on re-execing 
>> with the syntax error, and so didn't get to find this build failure).
>>
>> There are a series of errors linking ld.so starting with:
>>
>> /scratch/jmyers/glibc/many9/install/compilers/i686-gnu/lib/gcc/i686-glibc-gnu/9.2.1/../../../../i686-glibc-gnu/bin/ld:
>> /scratch/jmyers/glibc/many9/build/glibcs/i686-gnu/glibc/libc_pic.a(libc_fatal.os):
>> in function `__GI___libc_fatal':
>> /scratch/jmyers/glibc/many9/src/glibc/libio/../sysdeps/posix/libc_fatal.c:161:
>> multiple definition of `__libc_fatal';
>> /scratch/jmyers/glibc/many9/build/glibcs/i686-gnu/glibc/elf/dl-allobjs.os:/scratch/jmyers/glibc/many9/src/glibc/elf/dl-minimal.c:188:
>> first defined here
>>
>> and leading up to:
>>
>> make[3]: ***
>> [/scratch/jmyers/glibc/many9/build/glibcs/i686-gnu/glibc/elf/librtld.map]
>> Error 1
>
> I failed to test the later commits using build-many-glibcs.py because
> I deemed them more straightforward.
>
> The patch below should fix it.

Posted too soon.  The localplt update was missing.

8<------------------------------------------------------------------8<
hurd: Add stub implementation of __sigprocmask to the dynamic loader

The actual __sigprocmask cannot be built for the loader since it
depends on per-thread state.

Fixes commit a2e8aa0d9ea648068d8be52dd7b15f1b6a008e23 ("Block signals
during the initial part of dlopen").

diff --git a/sysdeps/mach/hurd/Versions b/sysdeps/mach/hurd/Versions
index f69d5fef67..57d45e6d0d 100644
--- a/sysdeps/mach/hurd/Versions
+++ b/sysdeps/mach/hurd/Versions
@@ -41,6 +41,6 @@ ld {
 
     # functions that must be shared with libc
     __access_noerrno; __libc_read; __libc_write; __libc_lseek64;
-    __libc_lock_self0;
+    __libc_lock_self0; __sigprocmask;
   }
 }
diff --git a/sysdeps/mach/hurd/dl-sysdep.c b/sysdeps/mach/hurd/dl-sysdep.c
index 719d603f44..e35ce69c4a 100644
--- a/sysdeps/mach/hurd/dl-sysdep.c
+++ b/sysdeps/mach/hurd/dl-sysdep.c
@@ -693,6 +693,15 @@ _hurd_intr_rpc_mach_msg (mach_msg_header_t *msg,
 		     timeout, notify);
 }
 
+/* This function does nothing and will be interposed with the actual
+   implementation once is libc is loaded.  Before that, manipulating
+   the signal mask does not make sense because no signal handlers have
+   been installed.  */
+int weak_function
+__sigprocmask (int how, const sigset_t *set, sigset_t *oset)
+{
+  return 0;
+}
 
 void
 _dl_show_auxv (void)
diff --git a/sysdeps/mach/hurd/i386/localplt.data b/sysdeps/mach/hurd/i386/localplt.data
index a5b5241b84..b3f5cefeea 100644
--- a/sysdeps/mach/hurd/i386/localplt.data
+++ b/sysdeps/mach/hurd/i386/localplt.data
@@ -41,6 +41,7 @@ ld.so: __strtoul_internal
 #ld.so: _exit
 ld.so: abort
 ld.so: _hurd_intr_rpc_mach_msg
+ld.so: __sigprocmask
 ld.so: __errno_location
 # rtld_hidden is currently disabled to avoid having to special-case the
 # functions above which do need a PLT.  These are thus currently expected.