On 2019/6/8 6:22 下午, Coly Li wrote:
> Recently people report bcache code compiled with gcc9 is broken, one of
> the buggy behavior I observe is that two adjacent 4KB I/Os should merge
> into one but they don't. Finally it turns out to be a stack corruption
> caused by macro PRECEDING_KEY().
>
> See how PRECEDING_KEY() is defined in bset.h,
> 437 #define PRECEDING_KEY(_k) \
> 438 ({ \
> 439 struct bkey *_ret = NULL; \
> 440 \
> 441 if (KEY_INODE(_k) || KEY_OFFSET(_k)) { \
> 442 _ret = &KEY(KEY_INODE(_k), KEY_OFFSET(_k), 0); \
> 443 \
> 444 if (!_ret->low) \
> 445 _ret->high--; \
> 446 _ret->low--; \
> 447 } \
> 448 \
> 449 _ret; \
> 450 })
>
> At line 442, _ret points to address of a on-stack variable combined by
> KEY(), the life range of this on-stack variable is in line 442-446,
> once _ret is returned to bch_btree_insert_key(), the returned address
> points to an invalid stack address and this adress is overwritten in
> the following called bch_btree_iter_init(). Then argument 'search' of
> bch_btree_iter_init() points to some address inside stackframe of
> bch_btree_iter_init(), exact address depends on how the compiler
> allocates stack space. Now the stack is corrupted.
>
> The fix is to avoid to allocate and return an on-stack variable only
> in range of PRECEDING_KEY(). This patch changes macro PRECEDING_KEY()
> to an inline function, and allocate another on-stack variable from
> function bch_btree_insert_key(), then the allocated memory address
> will be always valid in life range of bch_btree_insert_key().
>
> NOTE: This is only a RFC patch for more people to test. During my
> test I find bcache code does not complain out-of-order bkeys in btree
> node anymore, but the adjacent keys still don't totally merge as
> expected (e.g. they should be merged into one single key). So now I
> still continue to check what needs to be fixed more.
>
Hi folks,
After more testing, I realize the cached bkeys are not always merged,
this is the bkey dump information from
/sys/kernel/debug/bcache/bcache-87adfbc4-0b11-45b9-9a11-a11cfe5df2eb,
0:16 len 120 -> [0:377856 gen 1] dirty
0:136 len 8 -> [0:377976 gen 1] dirty
0:144 len 896 -> [0:721000 gen 1]
0:4112 len 8 -> [0:393136 gen 1]
So the patched bcache code is behaving correctly, IMHO no more fix
necessary.
I see Shenghui tested and verified the fix, more testing or review
comments are welcome.
Thanks.
Coly Li
> Signed-off-by: Coly Li <colyli@xxxxxxx>
> Cc: Kent Overstreet <kent.overstreet@xxxxxxxxx>
> Cc: Rolf Fokkens <rolf@xxxxxxxxxxxxxx>
> Cc: Nix <nix@xxxxxxxxxxxxx>
> Cc: Pierre JUHEN <pierre.juhen@xxxxxxxxx>
> Cc: linux-bcache@xxxxxxxxxxxxxxx
> Cc: linux-block@xxxxxxxxxxxxxxx
> ---
> drivers/md/bcache/bset.c | 16 +++++++++++++---
> drivers/md/bcache/bset.h | 34 ++++++++++++++++++++--------------
> 2 files changed, 33 insertions(+), 17 deletions(-)
>
> diff --git a/drivers/md/bcache/bset.c b/drivers/md/bcache/bset.c
> index 8f07fa6e1739..9422f3f1c682 100644
> --- a/drivers/md/bcache/bset.c
> +++ b/drivers/md/bcache/bset.c
> @@ -887,12 +887,22 @@ unsigned int bch_btree_insert_key(struct btree_keys *b, struct bkey *k,
> struct bset *i = bset_tree_last(b)->data;
> struct bkey *m, *prev = NULL;
> struct btree_iter iter;
> + struct bkey preceding_key_on_stack = ZERO_KEY;
> + struct bkey *preceding_key_p = &preceding_key_on_stack;
>
> BUG_ON(b->ops->is_extents && !KEY_SIZE(k));
>
> - m = bch_btree_iter_init(b, &iter, b->ops->is_extents
> - ? PRECEDING_KEY(&START_KEY(k))
> - : PRECEDING_KEY(k));
> + /*
> + * If k has preceding key, preceding_key_p will be set to address
> + * of k's preceding key; otherwise preceding_key_p will be set
> + * to NULL inside preceding_key().
> + */
> + if (b->ops->is_extents)
> + preceding_key(&START_KEY(k), preceding_key_p);
> + else
> + preceding_key(k, preceding_key_p);
> +
> + m = bch_btree_iter_init(b, &iter, preceding_key_p);
>
> if (b->ops->insert_fixup(b, k, &iter, replace_key))
> return status;
> diff --git a/drivers/md/bcache/bset.h b/drivers/md/bcache/bset.h
> index bac76aabca6d..6ab165dcb717 100644
> --- a/drivers/md/bcache/bset.h
> +++ b/drivers/md/bcache/bset.h
> @@ -434,20 +434,26 @@ static inline bool bch_cut_back(const struct bkey *where, struct bkey *k)
> return __bch_cut_back(where, k);
> }
>
> -#define PRECEDING_KEY(_k) \
> -({ \
> - struct bkey *_ret = NULL; \
> - \
> - if (KEY_INODE(_k) || KEY_OFFSET(_k)) { \
> - _ret = &KEY(KEY_INODE(_k), KEY_OFFSET(_k), 0); \
> - \
> - if (!_ret->low) \
> - _ret->high--; \
> - _ret->low--; \
> - } \
> - \
> - _ret; \
> -})
> +/*
> + * Pointer preceding_key_p points to a memory object to store preceding
> + * key of k. If the preceding key does not exist, set preceding_key_p to
> + * NULL. So the caller of preceding_key() needs to take care of memory
> + * which preceding_key_p pointed to before calling preceding_key().
> + * Currently the only caller of preceding_key() is bch_btree_insert_key(),
> + * and preceding_key_p points to an on-stack variable, so the memory
> + * release is handled by stackframe itself.
> + */
> +static inline void preceding_key(struct bkey *k, struct bkey *preceding_key_p)
> +{
> + if (KEY_INODE(k) || KEY_OFFSET(k)) {
> + *preceding_key_p = KEY(KEY_INODE(k), KEY_OFFSET(k), 0);
> + if (!preceding_key_p->low)
> + preceding_key_p->high--;
> + preceding_key_p->low--;
> + } else {
> + preceding_key_p = NULL;
> + }
> +}
>
> static inline bool bch_ptr_invalid(struct btree_keys *b, const struct bkey *k)
> {
>
