Add a slab corruption tracepoint
Subsystems (2)
PATCH 1/1 Add a slab corruption tracepoint
Add a tracepoint to log slab corruption messages to the trace log also so that it's easier to correlate with other trace messages that are being used to track refcounting.
--- include/trace/events/kmem.h | 23 +++++++++++++++++++++++ mm/slab.c | 2 ++ 2 files changed, 25 insertions(+)
include/trace/events/kmem.h +23 -0
@@ -315,6 +315,29 @@ TRACE_EVENT(mm_page_alloc_extfrag, __entry->change_ownership) ); +TRACE_EVENT (slab_corruption, + TP_PROTO(const char *slab, void *object, unsigned int size, unsigned int offset), + +TP_ARGS (slab, object, size, offset), + +TP_STRUCT__entry ( + __field( void *, object ) + __field( unsigned int, size ) + __field( unsigned int, offset ) +__array ( char, slab,16 ) + ), + +TP_fast_assign ( + strlcpy(__entry->slab, slab, sizeof(__entry->slab)); + __entry->object = object; + __entry->size = size; + __entry->offset = offset; + ), + +TP_printk ("slab=%s obj=%px size=%x off=%x" , + __entry->slab, __entry->object, __entry->size, __entry->offset) +); +#endif /* _TRACE_KMEM_H */ /* This part must be outside protection */
mm/slab.c +2 -0
@@ -1527,6 +1527,8 @@ static void check_poison_obj(struct kmem_cache *cachep, void *objp)print_tainted (), cachep->name, realobj, size);print_objinfo (cachep, objp,0 ); +trace_slab_corruption (cachep->name, realobj, + size, i); }/* Hexdump the affected line */ i = (i /16 ) *16 ;
Quoted 6 lines · David Howells
Add a tracepoint to log slab corruption messages to the trace log also so that it's easier to correlate with other trace messages that are being used to track refcounting.
Signed-off-by David Howells <[email protected]>
Shouldn't that include SLUB? I'm surprised to see SLAB used for debugging refcounting these days, as the SLUB debugging features are vastly superior, while SLAB ones are being sometimes found to be broken for years and removed.
Quoted 53 lines
--- include/trace/events/kmem.h | 23 +++++++++++++++++++++++ mm/slab.c | 2 ++ 2 files changed, 25 insertions(+)include/trace/events/kmem.h +23
-0@@ -315,6 +315,29 @@ TRACE_EVENT(mm_page_alloc_extfrag, __entry->change_ownership) ); +TRACE_EVENT (slab_corruption, + TP_PROTO(const char *slab, void *object, unsigned int size, unsigned int offset), + +TP_ARGS (slab, object, size, offset), + +TP_STRUCT__entry ( + __field( void *, object ) + __field( unsigned int, size ) + __field( unsigned int, offset ) +__array ( char, slab,16 ) + ), + +TP_fast_assign ( + strlcpy(__entry->slab, slab, sizeof(__entry->slab)); + __entry->object = object; + __entry->size = size; + __entry->offset = offset; + ), + +TP_printk ("slab=%s obj=%px size=%x off=%x" , + __entry->slab, __entry->object, __entry->size, __entry->offset) +); +#endif /* _TRACE_KMEM_H */ /* This part must be outside protection */ mm/slab.c +2
-0@@ -1527,6 +1527,8 @@ static void check_poison_obj(struct kmem_cache *cachep, void *objp)print_tainted (), cachep->name, realobj, size);print_objinfo (cachep, objp,0 ); +trace_slab_corruption (cachep->name, realobj, + size, i); }/* Hexdump the affected line */ i = (i /16 ) *16 ;
Quoted 1 line · Vlastimil Babka
Shouldn't that include SLUB?
I've been using SLAB. Looking in SLUB it's much less obvious where to insert the tracepoint. check_bytes_and_report() maybe?
Quoted 1 line
I'm surprised to see SLAB used for debugging refcounting these days,
The refcount debugging in question is not in SLAB, but rather in rxrpc; it's just SLAB detected the resulting memory corruption. rxrpc has tracepoints that track the refcounting, but SLAB printks a message to indicate the corruption and it's a bit tricky to work out where the printk happened with respect to the trace.
Quoted 2 lines
as the SLUB debugging features are vastly superior, while SLAB ones are being sometimes found to be broken for years and removed.
If SLUB is better than SLAB, shouldn't SLAB be removed?
David
Quoted 6 lines · David Howells
Quoted 1 line · Vlastimil Babka
Shouldn't that include SLUB?
I've been using SLAB. Looking in SLUB it's much less obvious where to insert the tracepoint.
check_bytes_and_report()maybe?
Perhaps object_err() and slab_err().
Quoted 4 lines
Quoted 1 line
I'm surprised to see SLAB used for debugging refcounting these days,
The refcount debugging in question is not in SLAB, but rather in rxrpc; it's just SLAB detected the resulting memory corruption.
I understand, and it's what I meant - SLUB is better suited to debug wrong uses (use-after-free, double free) that may be e.g. result of a refcounting bug. You can configure it to perform the extra checks only on the single cache you care about, and when it detects inconsistency it will also print stacktraces of who last allocated and freed the object.
Quoted 4 lines
rxrpc has tracepoints
that track the refcounting, but SLAB printks a message to indicate the corruption and it's a bit tricky to work out where the printk happened with respect to the trace.
Could you correlate by timestamps?
Quoted 4 lines
Quoted 2 lines
as the SLUB debugging features are vastly superior, while SLAB ones are being sometimes found to be broken for years and removed.
If SLUB is better than SLAB, shouldn't SLAB be removed?
It has been proposed before, but SLAB still has its users, who don't care about the debugging scenarios that much.
Quoted 2 lines
David