The problem: On IA64, the implementation of umount(2) in glibc and of sys_umount in the kernel results in the kernel performing umounts with essentially random flag words. $ dpkg -l | grep libc6 ii libc6.1 2.3.1-17 GNU C Library: Shared libraries and Timezone ii libc6.1-dev 2.3.1-17 GNU C Library: Development Libraries and Hea In libc, umount and umount2 are the same code. They both use syscall number 1044 without introducing a call layer. $ gdb /lib/libc.so.6.1 ... (gdb) disas umount Dump of assembler code for function umount: 0x1a63a0 <umount>: [MII] mov r15=1044 0x1a63a1 <umount+1>: break.i 0x100000;; 0x1a63a2 <umount+2>: cmp.eq p6,p0=-1,r10;; 0x1a63b0 <umount+16>: [BBB] (p06) br.cond.spnt.few 0x2d820 <__libc_start_main+640> 0x1a63b1 <umount+17>: br.ret.sptk.few b0 0x1a63b2 <umount+18>: nop.b 0x0;; End of assembler dump. (gdb) disas umount2 Dump of assembler code for function umount2: 0x1a63c0 <umount2>: [MII] mov r15=1044 0x1a63c1 <umount2+1>: break.i 0x100000;; 0x1a63c2 <umount2+2>: cmp.eq p6,p0=-1,r10;; 0x1a63d0 <umount2+16>: [BBB] (p06) br.cond.spnt.few 0x2d820 <__libc_start_main+640> 0x1a63d1 <umount2+17>: br.ret.sptk.few b0 0x1a63d2 <umount2+18>: nop.b 0x0;; End of assembler dump. And from fs/namespace.c, the interface for sys_umount is: asmlinkage long sys_umount(char * name, int flags) The problem is that sys_umount expects two parameters, the sys/umount.h header only allows the program to specify one parameter for umount(), and the kernel picks up whatever is in the next output register of the caller as the flags word. So umount calls are executed with essentially random flag words. The problem could be solved in libc by implementing a wrapper for umount that looks something like: extern int umount2 (__const char *__special_file, int __flags) ; int umount (const char *target) { umount2(target, 0); } In this case, umount2 could use the same lightweight method that it uses now. Bob Montgomery, HPReceived on Tue May 13 13:06:54 2003
This archive was generated by hypermail 2.1.8 : 2005-08-02 09:20:14 EST