On 3/1/22 09:51, Hyeonggon Yoo wrote:
> commit ba10d4b46655 ("mm/slub: use stackdepot to save stack trace in
> objects") initializes stack depot in cache creation if SLAB_STORE_USER
> flag is set.
As pointed out, this is not a stable commit, the series was just posted for
review and there will be v2. So instead of "this fixes the commit..." I
suggest writing the patch assuming it's a preparation for the patch
"mm/slub: use stackdepot"... and I can then make it part of the series.
So it should instead explain that for slub_debug we will need a way to
trigger stack_depot_early_init() based on boot options and so this patch
introduces it...
> This can make kernel crash because a cache can be crashed in various
> contexts. For example if user sets slub_debug=U, kernel crashes
> because create_boot_cache() calls stack_depot_init(), which tries to
> allocate hash table using memblock_alloc() if slab is not available.
> But memblock is also not available at that time.
>
> This patch solves the problem by initializing stack depot early
> in boot process if SLAB_STORE_USER debug flag is set globally
> or the flag is set for at least one cache.
>
> [ elver@xxxxxxxxxx: initialize stack depot depending on slub_debug
> parameter instead of allowing stack_depot_init() to be called
> during kmem_cache_init() for simplicity. ]
>
> [ vbabka@xxxxxxx: parse slub_debug parameter in setup_slub_debug()
> and initialize stack depot in stack_depot_early_init(). ]
>
> [ lkp@xxxxxxxxx: Fix build error. ]
>
> Link: https://lore.kernel.org/all/YhyeaP8lrzKgKm5A@xxxxxxxxxxxxxxxxxxx-northeast-1.compute.internal/
> Fixes: ba10d4b46655 ("mm/slub: use stackdepot to save stack trace in objects")
> Signed-off-by: Hyeonggon Yoo <42.hyeyoo@xxxxxxxxx>
> ---
> include/linux/slab.h | 1 +
> include/linux/stackdepot.h | 3 ++-
> mm/slab.c | 5 +++++
> mm/slob.c | 5 +++++
> mm/slub.c | 19 ++++++++++++++++---
> 5 files changed, 29 insertions(+), 4 deletions(-)
>
> diff --git a/include/linux/slab.h b/include/linux/slab.h
> index 37bde99b74af..d2b0f8f9e5e6 100644
> --- a/include/linux/slab.h
> +++ b/include/linux/slab.h
> @@ -762,6 +762,7 @@ extern void kvfree_sensitive(const void *addr, size_t len);
>
> unsigned int kmem_cache_size(struct kmem_cache *s);
> void __init kmem_cache_init_late(void);
> +int __init slab_stack_depot_init(void);
>
> #if defined(CONFIG_SMP) && defined(CONFIG_SLAB)
> int slab_prepare_cpu(unsigned int cpu);
> diff --git a/include/linux/stackdepot.h b/include/linux/stackdepot.h
> index 17f992fe6355..a813a2673c48 100644
> --- a/include/linux/stackdepot.h
> +++ b/include/linux/stackdepot.h
> @@ -12,6 +12,7 @@
> #define _LINUX_STACKDEPOT_H
>
> #include <linux/gfp.h>
> +#include <linux/slab.h>
>
> typedef u32 depot_stack_handle_t;
>
> @@ -32,7 +33,7 @@ int stack_depot_init(void);
> #ifdef CONFIG_STACKDEPOT_ALWAYS_INIT
> static inline int stack_depot_early_init(void) { return stack_depot_init(); }
> #else
> -static inline int stack_depot_early_init(void) { return 0; }
> +static inline int stack_depot_early_init(void) { return slab_stack_depot_init(); }
> #endif
I think the approach should be generic for stackdepot, not tied to a
function that belongs to slab with 3 different implementations.
E.g. in stackdepot.h declare a variable e.g. "stack_depot_want_early_init"
that is checked in stack_depot_early_init() above to call stack_depot_init().
> depot_stack_handle_t stack_depot_save(unsigned long *entries,
> diff --git a/mm/slab.c b/mm/slab.c
> index ddf5737c63d9..c7f929665fbe 100644
> --- a/mm/slab.c
> +++ b/mm/slab.c
> @@ -1196,6 +1196,11 @@ static void __init set_up_node(struct kmem_cache *cachep, int index)
> }
> }
>
> +int __init slab_stack_depot_init(void)
> +{
> + return 0;
> +}
> +
> /*
> * Initialisation. Called after the page allocator have been initialised and
> * before smp_init().
> diff --git a/mm/slob.c b/mm/slob.c
> index 60c5842215f1..7597c219f061 100644
> --- a/mm/slob.c
> +++ b/mm/slob.c
> @@ -725,3 +725,8 @@ void __init kmem_cache_init_late(void)
> {
> slab_state = FULL;
> }
> +
> +int __init slab_stack_depot_init(void)
> +{
> + return 0;
> +}
> diff --git a/mm/slub.c b/mm/slub.c
> index a74afe59a403..8f130f917977 100644
> --- a/mm/slub.c
> +++ b/mm/slub.c
> @@ -646,6 +646,16 @@ static slab_flags_t slub_debug;
>
> static char *slub_debug_string;
> static int disable_higher_order_debug;
> +static bool __initdata init_stack_depot;
> +
> +int __init slab_stack_depot_init(void)
> +{
> +#ifdef CONFIG_STACKDEPOT
> + if (init_stack_depot)
> + stack_depot_init();
> +#endif
> + return 0;
> +}
> /*
> * slub is about to manipulate internal object metadata. This memory lies
> @@ -1531,6 +1541,8 @@ static int __init setup_slub_debug(char *str)
> global_slub_debug_changed = true;
> } else {
> slab_list_specified = true;
> + if (flags & SLAB_STORE_USER)
> + init_stack_depot = true;
> }
> }
>
> @@ -1546,6 +1558,10 @@ static int __init setup_slub_debug(char *str)
> global_flags = slub_debug;
> slub_debug_string = saved_str;
> }
> +
> + if (global_flags & SLAB_STORE_USER)
> + init_stack_depot = true;
This looks good, it would just set the "stack_depot_want_early_init"
variable instead. But logically should be part of "mm/slub: use
stackdepot...", not part of patch that introduces the variable. That so if
you don't mind I would move it there with credit.
> out:
> slub_debug = global_flags;
> if (slub_debug != 0 || slub_debug_string)
> @@ -4221,9 +4237,6 @@ static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
> s->remote_node_defrag_ratio = 1000;
> #endif
>
> - if (s->flags & SLAB_STORE_USER && IS_ENABLED(CONFIG_STACKDEPOT))
> - stack_depot_init();
> -
> /* Initialize the pre-computed randomized freelist if slab is up */
> if (slab_state >= UP) {
> if (init_cache_random_seq(s))
