Message ID | e661c73438b37a7c931089e6ba0421a5711d468d.1605283657.git.fweimer@redhat.com |
---|---|
State | New |
Headers | show |
Series | [1/2] Replace __libc_multiple_libcs with tri-state __libc_type | expand |
On 13/11/2020 13:14, Florian Weimer via Libc-alpha wrote: > The inner libc must not use sbrk. _dl_addr occasionally shows up > in profiles, but had to be used before because __libc_multiple_libs > was unreliable. LGTM. Is the performance the only advantage of using sbrk on the inner libc? > --- > malloc/arena.c | 7 +------ > malloc/malloc.c | 2 ++ > 2 files changed, 3 insertions(+), 6 deletions(-) > > diff --git a/malloc/arena.c b/malloc/arena.c > index 202daf15b0..865eaa1a84 100644 > --- a/malloc/arena.c > +++ b/malloc/arena.c > @@ -296,12 +296,7 @@ ptmalloc_init (void) > #ifdef SHARED > /* In case this libc copy is in a non-default namespace, never use brk. > Likewise if dlopened from statically linked program. */ > - Dl_info di; > - struct link_map *l; > - > - if (_dl_open_hook != NULL > - || (_dl_addr (ptmalloc_init, &di, &l, NULL) != 0 > - && l->l_ns != LM_ID_BASE)) > + if (__libc_type == libc_type_secondary) > __morecore = __failing_morecore; > #endif > > diff --git a/malloc/malloc.c b/malloc/malloc.c > index 5b87bdb081..0834e4181a 100644 > --- a/malloc/malloc.c > +++ b/malloc/malloc.c > @@ -247,6 +247,8 @@ > /* For SINGLE_THREAD_P. */ > #include <sysdep-cancel.h> > > +#include <libc-internal.h> > + > /* > Debugging: > >
* Adhemerval Zanella via Libc-alpha: > On 13/11/2020 13:14, Florian Weimer via Libc-alpha wrote: >> The inner libc must not use sbrk. _dl_addr occasionally shows up >> in profiles, but had to be used before because __libc_multiple_libs >> was unreliable. > > LGTM. Is the performance the only advantage of using sbrk on > the inner libc? Avoiding sbrk in the inner libc is required for correctness because sbrk is not thread-safe, and there is no external synchronization. The startup performance improvement comes from not calling _dl_addr. On POWER, it saves around 150,000 instructions, according to a quick test. Thanks, Florian
On 13/11/2020 16:05, Florian Weimer wrote: > * Adhemerval Zanella via Libc-alpha: > >> On 13/11/2020 13:14, Florian Weimer via Libc-alpha wrote: >>> The inner libc must not use sbrk. _dl_addr occasionally shows up >>> in profiles, but had to be used before because __libc_multiple_libs >>> was unreliable. >> >> LGTM. Is the performance the only advantage of using sbrk on >> the inner libc? > > Avoiding sbrk in the inner libc is required for correctness because sbrk > is not thread-safe, and there is no external synchronization. > > The startup performance improvement comes from not calling _dl_addr. On > POWER, it saves around 150,000 instructions, according to a quick test. In fact I meant what is the gain of using sbrk on outer libc. I am asking because maybe an option would be to use the same strategy regardless whether it is a inner or outer libc.
* Adhemerval Zanella: > On 13/11/2020 16:05, Florian Weimer wrote: >> * Adhemerval Zanella via Libc-alpha: >> >>> On 13/11/2020 13:14, Florian Weimer via Libc-alpha wrote: >>>> The inner libc must not use sbrk. _dl_addr occasionally shows up >>>> in profiles, but had to be used before because __libc_multiple_libs >>>> was unreliable. >>> >>> LGTM. Is the performance the only advantage of using sbrk on >>> the inner libc? >> >> Avoiding sbrk in the inner libc is required for correctness because sbrk >> is not thread-safe, and there is no external synchronization. >> >> The startup performance improvement comes from not calling _dl_addr. On >> POWER, it saves around 150,000 instructions, according to a quick test. > > In fact I meant what is the gain of using sbrk on outer libc. I am asking > because maybe an option would be to use the same strategy regardless whether > it is a inner or outer libc. Oh, that's totally unclear. The main arena has a very different layout from the other arenas. The __morecore hook exposes the inner sbrk-based workings of our malloc as part of the ABI. I'm not sure if we can restructure the code so that __morecore is never called, or if that would break existing users (such as libhugetlbfs). Thanks, Florian
diff --git a/malloc/arena.c b/malloc/arena.c index 202daf15b0..865eaa1a84 100644 --- a/malloc/arena.c +++ b/malloc/arena.c @@ -296,12 +296,7 @@ ptmalloc_init (void) #ifdef SHARED /* In case this libc copy is in a non-default namespace, never use brk. Likewise if dlopened from statically linked program. */ - Dl_info di; - struct link_map *l; - - if (_dl_open_hook != NULL - || (_dl_addr (ptmalloc_init, &di, &l, NULL) != 0 - && l->l_ns != LM_ID_BASE)) + if (__libc_type == libc_type_secondary) __morecore = __failing_morecore; #endif diff --git a/malloc/malloc.c b/malloc/malloc.c index 5b87bdb081..0834e4181a 100644 --- a/malloc/malloc.c +++ b/malloc/malloc.c @@ -247,6 +247,8 @@ /* For SINGLE_THREAD_P. */ #include <sysdep-cancel.h> +#include <libc-internal.h> + /* Debugging: