Re: [patch 1/2] mm, mempool: poison elements backed by slab allocator
On 03/19/2015 06:20 PM, David Rientjes wrote: > On Mon, 16 Mar 2015, Rasmus Villemoes wrote: > >>> Mempools keep elements in a reserved pool for contexts in which >>> allocation may not be possible. When an element is allocated from the >>> reserved pool, its memory contents is the same as when it was added to >>> the reserved pool. >>> >>> Because of this, elements lack any free poisoning to detect >>> use-after-free errors. >>> >>> This patch adds free poisoning for elements backed by the slab allocator. >>> This is possible because the mempool layer knows the object size of each >>> element. >>> >>> When an element is added to the reserved pool, it is poisoned with >>> POISON_FREE. When it is removed from the reserved pool, the contents are >>> checked for POISON_FREE. If there is a mismatch, a warning is emitted to >>> the kernel log. >>> >>> + >>> +static void poison_slab_element(mempool_t *pool, void *element) >>> +{ >>> + if (pool->alloc == mempool_alloc_slab || >>> + pool->alloc == mempool_kmalloc) { >>> + size_t size = ksize(element); >>> + u8 *obj = element; >>> + >>> + memset(obj, POISON_FREE, size - 1); >>> + obj[size - 1] = POISON_END; >>> + } >>> +} >> >> Maybe a stupid question, but what happens if the underlying slab >> allocator has non-trivial ->ctor? >> > > Not a stupid question at all, it's very legitimate, thanks for thinking > about it! > > Using slab constructors with mempools is inherently risky because you > don't know where the element is coming from when returned by > mempool_alloc(): was it able to be allocated in GFP_NOFS context from the > slab allocator, or is it coming from mempool's reserved pool of elements? > > In the former, the constructor is properly called and in the latter it's > not called: we simply pop the element from the reserved pool and return it > to the caller. > > For that reason, without mempool element poisoning, we need to take care > that objects are properly deconstructed when freed to the reserved pool so > that they are in a state we expect when returned from mempool_alloc(). > Thus, at least part of the slab constructor must be duplicated before > calling mempool_free(). > > There's one user in the tree that actually does this: it's the > mempool_create_slab_pool() in fs/jfs/jfs_metapage.c, and it does exactly > what is described above: it clears necessary fields before doing > mempool_free() that are duplicated in the slab object constructor. > > We'd like to be able to do this: > > diff --git a/mm/mempool.c b/mm/mempool.c > --- a/mm/mempool.c > +++ b/mm/mempool.c > @@ -15,6 +15,7 @@ > #include > #include > #include > +#include "slab.h" > > static void add_element(mempool_t *pool, void *element) > { > @@ -332,6 +333,7 @@ EXPORT_SYMBOL(mempool_free); > void *mempool_alloc_slab(gfp_t gfp_mask, void *pool_data) > { > struct kmem_cache *mem = pool_data; > + BUG_ON(mem->ctor); > return kmem_cache_alloc(mem, gfp_mask); > } > EXPORT_SYMBOL(mempool_alloc_slab); > > Since it would be difficult to reproduce an error with an improperly > decosntructed mempool element when used with a mempool based on a slab > cache that has a constructor: normally, slab objects are allocatable even > with GFP_NOFS since there are free objects available and there's less of a > liklihood that we'll need to use the mempool reserved pool. > > But we obviously can't do that if jfs is actively using mempools based on > slab caches with object constructors. I think it would be much better to > simply initialize objects as they are allocated, regardless of whether > they come from the slab allocator or mempool reserved pool, and avoid > trying to set the state up properly before mempool_free(). > > This patch properly initializes all fields that are currently done by the > object constructor (with the exception of mp->flags since it is > immediately overwritten by the caller anyway). It also removes META_free > since nothing is checking for it. > > Jfs folks, would this be acceptable to you? Sure. I have no objection. > --- > diff --git a/fs/jfs/jfs_metapage.c b/fs/jfs/jfs_metapage.c > --- a/fs/jfs/jfs_metapage.c > +++ b/fs/jfs/jfs_metapage.c > @@ -183,30 +183,23 @@ static inline void remove_metapage(struct page *page, > struct metapage *mp) > > #endif > > -static void init_once(void *foo) > -{ > - struct metapage *mp = (struct metapage *)foo; > - > - mp->lid = 0; > - mp->lsn = 0; > - mp->flag = 0; > - mp->data = NULL; > - mp->clsn = 0; > - mp->log = NULL; > - set_bit(META_free, &mp->flag); > - init_waitqueue_head(&mp->wait); > -} > - > static inline struct metapage *alloc_metapage(gfp_t gfp_mask) > { > - return mempool_alloc(metapage_mempool, gfp_mask); > + struct metapage *mp = mempool_alloc(metapage_mempool, gfp_mask); > + > + if (mp) { > + mp->lid = 0; > + mp->lsn = 0; > +
Re: [patch 1/2] mm, mempool: poison elements backed by slab allocator
On Mon, 16 Mar 2015, Rasmus Villemoes wrote: > > Mempools keep elements in a reserved pool for contexts in which > > allocation may not be possible. When an element is allocated from the > > reserved pool, its memory contents is the same as when it was added to > > the reserved pool. > > > > Because of this, elements lack any free poisoning to detect > > use-after-free errors. > > > > This patch adds free poisoning for elements backed by the slab allocator. > > This is possible because the mempool layer knows the object size of each > > element. > > > > When an element is added to the reserved pool, it is poisoned with > > POISON_FREE. When it is removed from the reserved pool, the contents are > > checked for POISON_FREE. If there is a mismatch, a warning is emitted to > > the kernel log. > > > > + > > +static void poison_slab_element(mempool_t *pool, void *element) > > +{ > > + if (pool->alloc == mempool_alloc_slab || > > + pool->alloc == mempool_kmalloc) { > > + size_t size = ksize(element); > > + u8 *obj = element; > > + > > + memset(obj, POISON_FREE, size - 1); > > + obj[size - 1] = POISON_END; > > + } > > +} > > Maybe a stupid question, but what happens if the underlying slab > allocator has non-trivial ->ctor? > Not a stupid question at all, it's very legitimate, thanks for thinking about it! Using slab constructors with mempools is inherently risky because you don't know where the element is coming from when returned by mempool_alloc(): was it able to be allocated in GFP_NOFS context from the slab allocator, or is it coming from mempool's reserved pool of elements? In the former, the constructor is properly called and in the latter it's not called: we simply pop the element from the reserved pool and return it to the caller. For that reason, without mempool element poisoning, we need to take care that objects are properly deconstructed when freed to the reserved pool so that they are in a state we expect when returned from mempool_alloc(). Thus, at least part of the slab constructor must be duplicated before calling mempool_free(). There's one user in the tree that actually does this: it's the mempool_create_slab_pool() in fs/jfs/jfs_metapage.c, and it does exactly what is described above: it clears necessary fields before doing mempool_free() that are duplicated in the slab object constructor. We'd like to be able to do this: diff --git a/mm/mempool.c b/mm/mempool.c --- a/mm/mempool.c +++ b/mm/mempool.c @@ -15,6 +15,7 @@ #include #include #include +#include "slab.h" static void add_element(mempool_t *pool, void *element) { @@ -332,6 +333,7 @@ EXPORT_SYMBOL(mempool_free); void *mempool_alloc_slab(gfp_t gfp_mask, void *pool_data) { struct kmem_cache *mem = pool_data; + BUG_ON(mem->ctor); return kmem_cache_alloc(mem, gfp_mask); } EXPORT_SYMBOL(mempool_alloc_slab); Since it would be difficult to reproduce an error with an improperly decosntructed mempool element when used with a mempool based on a slab cache that has a constructor: normally, slab objects are allocatable even with GFP_NOFS since there are free objects available and there's less of a liklihood that we'll need to use the mempool reserved pool. But we obviously can't do that if jfs is actively using mempools based on slab caches with object constructors. I think it would be much better to simply initialize objects as they are allocated, regardless of whether they come from the slab allocator or mempool reserved pool, and avoid trying to set the state up properly before mempool_free(). This patch properly initializes all fields that are currently done by the object constructor (with the exception of mp->flags since it is immediately overwritten by the caller anyway). It also removes META_free since nothing is checking for it. Jfs folks, would this be acceptable to you? --- diff --git a/fs/jfs/jfs_metapage.c b/fs/jfs/jfs_metapage.c --- a/fs/jfs/jfs_metapage.c +++ b/fs/jfs/jfs_metapage.c @@ -183,30 +183,23 @@ static inline void remove_metapage(struct page *page, struct metapage *mp) #endif -static void init_once(void *foo) -{ - struct metapage *mp = (struct metapage *)foo; - - mp->lid = 0; - mp->lsn = 0; - mp->flag = 0; - mp->data = NULL; - mp->clsn = 0; - mp->log = NULL; - set_bit(META_free, &mp->flag); - init_waitqueue_head(&mp->wait); -} - static inline struct metapage *alloc_metapage(gfp_t gfp_mask) { - return mempool_alloc(metapage_mempool, gfp_mask); + struct metapage *mp = mempool_alloc(metapage_mempool, gfp_mask); + + if (mp) { + mp->lid = 0; + mp->lsn = 0; + mp->data = NULL; + mp->clsn = 0; + mp->log = NULL; + init_waitqueue_head(&mp->wait); + } + return mp; } static inline void free_metapage(struct metapage *mp) { - mp->
Re: [patch 1/2] mm, mempool: poison elements backed by slab allocator
On Mon, Mar 09 2015, David Rientjes wrote: > Mempools keep elements in a reserved pool for contexts in which > allocation may not be possible. When an element is allocated from the > reserved pool, its memory contents is the same as when it was added to > the reserved pool. > > Because of this, elements lack any free poisoning to detect > use-after-free errors. > > This patch adds free poisoning for elements backed by the slab allocator. > This is possible because the mempool layer knows the object size of each > element. > > When an element is added to the reserved pool, it is poisoned with > POISON_FREE. When it is removed from the reserved pool, the contents are > checked for POISON_FREE. If there is a mismatch, a warning is emitted to > the kernel log. > > + > +static void poison_slab_element(mempool_t *pool, void *element) > +{ > + if (pool->alloc == mempool_alloc_slab || > + pool->alloc == mempool_kmalloc) { > + size_t size = ksize(element); > + u8 *obj = element; > + > + memset(obj, POISON_FREE, size - 1); > + obj[size - 1] = POISON_END; > + } > +} Maybe a stupid question, but what happens if the underlying slab allocator has non-trivial ->ctor? Rasmus -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
Re: [patch 1/2] mm, mempool: poison elements backed by slab allocator
On Thu, 12 Mar 2015, Andrew Morton wrote: > > Mempools keep elements in a reserved pool for contexts in which > > allocation may not be possible. When an element is allocated from the > > reserved pool, its memory contents is the same as when it was added to > > the reserved pool. > > > > Because of this, elements lack any free poisoning to detect > > use-after-free errors. > > > > This patch adds free poisoning for elements backed by the slab allocator. > > This is possible because the mempool layer knows the object size of each > > element. > > > > When an element is added to the reserved pool, it is poisoned with > > POISON_FREE. When it is removed from the reserved pool, the contents are > > checked for POISON_FREE. If there is a mismatch, a warning is emitted to > > the kernel log. > > > > This is only effective for configs with CONFIG_DEBUG_VM. > > At present CONFIG_DEBUG_VM is pretty lightweight (I hope) and using it > for mempool poisoning might be inappropriately costly. Would it be > better to tie this to something else? Either standalone or reuse some > slab debug option, perhaps. > Ok, I agree. I'll use CONFIG_DEBUG_SLAB and CONFIG_SLUB_DEBUG_ON and allow it to be enabled by slub debugging when that is enabled. It probably doesn't make a lot of sense to do mempool poisoning without slab poisoning. > Did you measure the overhead btw? It might be significant with fast > devices. > It's certainly costly: with a new 128-byte slab cache, allocating 64 objects took about 480 cycles longer per object to do the poison checking and in-use poisoning on one of my 2.2GHz machines (~90 cycles/object without CONFIG_DEBUG_VM). To do the free poisoning, it was about ~130 cycles longer per object (~140 cycles/object without CONFIG_DEBUG_VM). For cache cold pages from the page allocator, it's more expensive, allocating and freeing 64 pages, it's ~620 cycles longer per page and freeing is an additional ~60 cycles/page. Keep in mind that overhead is only incurred when the mempool alloc function fails to allocate memory directly from the slab allocator or page allocator in the given context and on mempool_create() to create the new mempool. I didn't benchmark high-order page poisoning, but that's only used by bcache and I'm looking at that separately: allocating high-order pages from a mempool sucks. > > --- a/mm/mempool.c > > +++ b/mm/mempool.c > > @@ -16,16 +16,77 @@ > > #include > > #include > > > > +#ifdef CONFIG_DEBUG_VM > > +static void poison_error(mempool_t *pool, void *element, size_t size, > > +size_t byte) > > +{ > > + const int nr = pool->curr_nr; > > + const int start = max_t(int, byte - (BITS_PER_LONG / 8), 0); > > + const int end = min_t(int, byte + (BITS_PER_LONG / 8), size); > > + int i; > > + > > + pr_err("BUG: mempool element poison mismatch\n"); > > + pr_err("Mempool %p size %ld\n", pool, size); > > + pr_err(" nr=%d @ %p: %s0x", nr, element, start > 0 ? "... " : ""); > > + for (i = start; i < end; i++) > > + pr_cont("%x ", *(u8 *)(element + i)); > > + pr_cont("%s\n", end < size ? "..." : ""); > > + dump_stack(); > > +} > > "byte" wasn't a very useful identifier, and it's called "i" in > check_slab_element(). Rename it to "offset" in both places? > > > +static void check_slab_element(mempool_t *pool, void *element) > > +{ > > + if (pool->free == mempool_free_slab || pool->free == mempool_kfree) { > > + size_t size = ksize(element); > > + u8 *obj = element; > > + size_t i; > > + > > + for (i = 0; i < size; i++) { > > + u8 exp = (i < size - 1) ? POISON_FREE : POISON_END; > > + > > + if (obj[i] != exp) { > > + poison_error(pool, element, size, i); > > + return; > > + } > > + } > > + memset(obj, POISON_INUSE, size); > > + } > > +} > > I question the reuse of POISON_FREE/POISON_INUSE. If this thing > triggers, it may be hard to tell if it was due to a slab thing or to a > mempool thing. Using a distinct poison pattern for mempool would clear > that up? > Hmm, I think it would actually make it more confusing: mempools only allocate from the reserved pool (those poisoned by this patchset) when doing kmalloc() or kmem_cache_free() in context fails. Normally, the reserved pool isn't used because there are free objects sitting on slab free or partial slabs and the context is irrelevant. If slab poisoning is enabled, they are already POISON_FREE as anticipated. We only fallback to the reserved pool when new slab needs to be allocated and fails in the given context, so the poison value would differ depending on where the objects came from. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.htm
Re: [patch 1/2] mm, mempool: poison elements backed by slab allocator
On Mon, 9 Mar 2015 00:21:56 -0700 (PDT) David Rientjes wrote: > Mempools keep elements in a reserved pool for contexts in which > allocation may not be possible. When an element is allocated from the > reserved pool, its memory contents is the same as when it was added to > the reserved pool. > > Because of this, elements lack any free poisoning to detect > use-after-free errors. > > This patch adds free poisoning for elements backed by the slab allocator. > This is possible because the mempool layer knows the object size of each > element. > > When an element is added to the reserved pool, it is poisoned with > POISON_FREE. When it is removed from the reserved pool, the contents are > checked for POISON_FREE. If there is a mismatch, a warning is emitted to > the kernel log. > > This is only effective for configs with CONFIG_DEBUG_VM. At present CONFIG_DEBUG_VM is pretty lightweight (I hope) and using it for mempool poisoning might be inappropriately costly. Would it be better to tie this to something else? Either standalone or reuse some slab debug option, perhaps. Did you measure the overhead btw? It might be significant with fast devices. > --- a/mm/mempool.c > +++ b/mm/mempool.c > @@ -16,16 +16,77 @@ > #include > #include > > +#ifdef CONFIG_DEBUG_VM > +static void poison_error(mempool_t *pool, void *element, size_t size, > + size_t byte) > +{ > + const int nr = pool->curr_nr; > + const int start = max_t(int, byte - (BITS_PER_LONG / 8), 0); > + const int end = min_t(int, byte + (BITS_PER_LONG / 8), size); > + int i; > + > + pr_err("BUG: mempool element poison mismatch\n"); > + pr_err("Mempool %p size %ld\n", pool, size); > + pr_err(" nr=%d @ %p: %s0x", nr, element, start > 0 ? "... " : ""); > + for (i = start; i < end; i++) > + pr_cont("%x ", *(u8 *)(element + i)); > + pr_cont("%s\n", end < size ? "..." : ""); > + dump_stack(); > +} "byte" wasn't a very useful identifier, and it's called "i" in check_slab_element(). Rename it to "offset" in both places? > +static void check_slab_element(mempool_t *pool, void *element) > +{ > + if (pool->free == mempool_free_slab || pool->free == mempool_kfree) { > + size_t size = ksize(element); > + u8 *obj = element; > + size_t i; > + > + for (i = 0; i < size; i++) { > + u8 exp = (i < size - 1) ? POISON_FREE : POISON_END; > + > + if (obj[i] != exp) { > + poison_error(pool, element, size, i); > + return; > + } > + } > + memset(obj, POISON_INUSE, size); > + } > +} I question the reuse of POISON_FREE/POISON_INUSE. If this thing triggers, it may be hard to tell if it was due to a slab thing or to a mempool thing. Using a distinct poison pattern for mempool would clear that up? > ... -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
[patch 1/2] mm, mempool: poison elements backed by slab allocator
Mempools keep elements in a reserved pool for contexts in which allocation may not be possible. When an element is allocated from the reserved pool, its memory contents is the same as when it was added to the reserved pool. Because of this, elements lack any free poisoning to detect use-after-free errors. This patch adds free poisoning for elements backed by the slab allocator. This is possible because the mempool layer knows the object size of each element. When an element is added to the reserved pool, it is poisoned with POISON_FREE. When it is removed from the reserved pool, the contents are checked for POISON_FREE. If there is a mismatch, a warning is emitted to the kernel log. This is only effective for configs with CONFIG_DEBUG_VM. Signed-off-by: David Rientjes --- mm/mempool.c | 65 ++-- 1 file changed, 63 insertions(+), 2 deletions(-) diff --git a/mm/mempool.c b/mm/mempool.c --- a/mm/mempool.c +++ b/mm/mempool.c @@ -16,16 +16,77 @@ #include #include +#ifdef CONFIG_DEBUG_VM +static void poison_error(mempool_t *pool, void *element, size_t size, +size_t byte) +{ + const int nr = pool->curr_nr; + const int start = max_t(int, byte - (BITS_PER_LONG / 8), 0); + const int end = min_t(int, byte + (BITS_PER_LONG / 8), size); + int i; + + pr_err("BUG: mempool element poison mismatch\n"); + pr_err("Mempool %p size %ld\n", pool, size); + pr_err(" nr=%d @ %p: %s0x", nr, element, start > 0 ? "... " : ""); + for (i = start; i < end; i++) + pr_cont("%x ", *(u8 *)(element + i)); + pr_cont("%s\n", end < size ? "..." : ""); + dump_stack(); +} + +static void check_slab_element(mempool_t *pool, void *element) +{ + if (pool->free == mempool_free_slab || pool->free == mempool_kfree) { + size_t size = ksize(element); + u8 *obj = element; + size_t i; + + for (i = 0; i < size; i++) { + u8 exp = (i < size - 1) ? POISON_FREE : POISON_END; + + if (obj[i] != exp) { + poison_error(pool, element, size, i); + return; + } + } + memset(obj, POISON_INUSE, size); + } +} + +static void poison_slab_element(mempool_t *pool, void *element) +{ + if (pool->alloc == mempool_alloc_slab || + pool->alloc == mempool_kmalloc) { + size_t size = ksize(element); + u8 *obj = element; + + memset(obj, POISON_FREE, size - 1); + obj[size - 1] = POISON_END; + } +} +#else /* CONFIG_DEBUG_VM */ +static inline void check_slab_element(mempool_t *pool, void *element) +{ +} +static inline void poison_slab_element(mempool_t *pool, void *element) +{ +} +#endif /* CONFIG_DEBUG_VM */ + static void add_element(mempool_t *pool, void *element) { BUG_ON(pool->curr_nr >= pool->min_nr); + poison_slab_element(pool, element); pool->elements[pool->curr_nr++] = element; } static void *remove_element(mempool_t *pool) { - BUG_ON(pool->curr_nr <= 0); - return pool->elements[--pool->curr_nr]; + void *element = pool->elements[--pool->curr_nr]; + + BUG_ON(pool->curr_nr < 0); + check_slab_element(pool, element); + return element; } /** -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/