Eric B Munson wrote:
> On Tue, 28 Dec 2010, Andrew Hastings wrote:
>
>> libhugetlbfs redefines shmget() so that it can transparently optionally
>> add the hugepage flag. Unfortunately, it uses dlsym() to to find the
>> next shmget() to call in the dynamic library call chain. This does not
>> work on statically-linked executables.
>>
>> Work around this by making all references to libdl weak, and testing
>> for its presence. If libdl symbols are not not present, directly
>> make the shmget syscall.
>>
>> Making the dl* symbols weak does not affect dynamic executables as
>> libdl.so is:
>> * Automatically added to the library dependency list when liked with
>> -lhugetlbfs
>> * Mapped into the programs address space when used with LD_PRELOAD.
>>
>> Based on a patch originally authored by Dean Luick <lu...@cray.com>.
>>
>> Signed-off-by: Andrew Hastings <a...@cray.com> on behalf of Cray Inc.
>> ---
>>
>> -Andrew Hastings
>> Cray Inc.
>>
>> diff -ruNp libhugetlbfs-2.11/shm.c libhugetlbfs-2.11-shm/shm.c
>> --- libhugetlbfs-2.11/shm.c 2010-12-16 11:38:22.000000000 -0600
>> +++ libhugetlbfs-2.11-shm/shm.c 2010-12-28 17:26:00.725063000 -0600
>> @@ -29,6 +29,37 @@
>> #include "libhugetlbfs_internal.h"
>> #include "hugetlbfs.h"
>>
>> +#if defined(__x86_64)
>
> Why is this patch x86_64 specific, does it not work for x86, or ppc64, or any
> other arch we support? I am reluctant to take functionality like this for
> a single arch unless there is no way to make it work elsewhere.
ia32 doesn't have a SYS_shmget system call; instead shmget gets funneled
through SYS_ipc and the definitions needed to use it are not available outside
glibc.
We no longer have a way to test on ppc or ppc64, unfortunately.
-Andrew Hastings
Cray Inc.
>
>> +/*
>> + * The calls to dlsym() and dlerror() in the shmget() wrapper below force
>> + * a dependency on libdl.so. This does not work for static executables
>> + * as the glibc dynamic library implementation does not automatically
>> + * have static dl* function stubs linked into static executables.
>> + *
>> + * Work around this problem by adding a weak attribute to the declarations
>> + * of dlsym() and dlerror(). (The declaration is otherwise the same as in
>> + * <dlfcn.h>). This allows a static executable to be linked without -ldl.
>> + * If &dlsym is NULL then this is a static executable and a call to the
>> + * system shmget() may be performed without worry as there is no dynamic
>> + * call chain.
>> + *
>> + * Although this problem appears to be platform-independent, the solution
>> + * unfortunately depends on the platform, thus the __x86_64 ifdef.
>> + */
>> +#include <sys/syscall.h> /* for SYS_xxx definitions */
>> +
>> +extern void *dlsym (void *__restrict __handle, __const char *__restrict
>> __name)
>> + __attribute__((weak)) __THROW __nonnull ((2));
>> +extern char *dlerror (void) __attribute__((weak)) __THROW;
>> +
>> +
>> +/* call syscall shmget through the generic syscall mechanism */
>> +static int syscall_shmget(key_t key, size_t size, int shmflg)
>> +{
>> + return syscall(SYS_shmget, key, size, shmflg);
>> +}
>> +#endif /* __x86_64 */
>> +
>> int shmget(key_t key, size_t size, int shmflg)
>> {
>> static int (*real_shmget)(key_t key, size_t size, int shmflg) = NULL;
>> @@ -40,10 +71,18 @@ int shmget(key_t key, size_t size, int s
>>
>> /* Get a handle to the "real" shmget system call */
>> if (!real_shmget) {
>> - real_shmget = dlsym(RTLD_NEXT, "shmget");
>> - if ((error = dlerror()) != NULL) {
>> - ERROR("%s", error);
>> - return -1;
>> +#if defined(__x86_64)
>> + if (&dlsym == NULL) {
>> + /* in a static executable, call shmget directly */
>> + real_shmget = syscall_shmget;
>> + } else
>> +#endif /* __x86_64 */
>> + {
>> + real_shmget = dlsym(RTLD_NEXT, "shmget");
>> + if ((error = dlerror()) != NULL) {
>> + ERROR("%s", error);
>> + return -1;
>> + }
>> }
>> }
>>
>>
>> ------------------------------------------------------------------------------
>> Learn how Oracle Real Application Clusters (RAC) One Node allows customers
>> to consolidate database storage, standardize their database environment,
>> and,
>> should the need arise, upgrade to a full multi-node Oracle RAC database
>> without downtime or disruption
>> http://p.sf.net/sfu/oracle-sfdevnl
>> _______________________________________________
>> Libhugetlbfs-devel mailing list
>> Libhugetlbfs-devel@lists.sourceforge.net
>> https://lists.sourceforge.net/lists/listinfo/libhugetlbfs-devel
>>
------------------------------------------------------------------------------
Learn how Oracle Real Application Clusters (RAC) One Node allows customers
to consolidate database storage, standardize their database environment, and,
should the need arise, upgrade to a full multi-node Oracle RAC database
without downtime or disruption
http://p.sf.net/sfu/oracle-sfdevnl
_______________________________________________
Libhugetlbfs-devel mailing list
Libhugetlbfs-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libhugetlbfs-devel
