: Fix ALIGN() macro
David Miller
maintainerPATCH 1/1 : Fix ALIGN() macro
A regression or two were added by the crypto-2.6 tree merge.
I've tracked down one of them, and it's caused by the ALIGN() macro truncating things down to "int". Some of Herbert's new code is aligning pointers using ALIGN() and this thus explodes on 64-bit since the top 32-bits get chopped off.
It is arguable that perhaps we should make a special macro for pointer aligning, but even in that case ALIGN() as a general purpose macro should use the largest natural integer size in order to not cause surprises for people.
I'm still trying to track down the other regression added by the crypto merge. I have it git bisected down to a single changeset, but I haven't determined what's really wrong yet. I should be able to kill that over the weekend. I want to fix this before merging my networking tree so I can be absolutely sure that IPSEC doesn't break because of something in my tree :)
[KERNEL]: Do not truncate to 'int' in ALIGN() macro.
include/linux/kernel.h +1 -1
@@ -31,7 +31,7 @@ #define ULLONG_MAX (~0ULL)#define STACK_MAGIC 0xdeadbeef#define ARRAY_SIZE (x) (sizeof(x) / sizeof((x)[0])) -#define ALIGN (x,a) (((x)+(a)-1)&~((a)-1)) +#define ALIGN (x,a) (((x)+(a)-1UL)&~((a)-1UL))#define FIELD_SIZEOF (t, f) (sizeof(((t*)0)->f))#define roundup (x, y) ((((x) + ((y) - 1)) / (y)) * (y))
Herbert Xu
maintainerQuoted 7 lines · David Miller
I'm still trying to track down the other regression added by the crypto merge. I have it git bisected down to a single changeset, but I haven't determined what's really wrong yet. I should be able to kill that over the weekend. I want to fix this before merging my networking tree so I can be absolutely sure that IPSEC doesn't break because of something in my tree :)
Thanks for fixing this Dave. I recall being bitten earlier by the same thing as well. I really need to start testing on 64-bit.
BTW could you describe the other regression?
Thanks,
Signature · 4 lines
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txtHerbert Xu
maintainerQuoted 6 lines · herbert
Thanks for fixing this Dave. I recall being bitten earlier by the same thing as well. I really need to start testing on 64-bit.
BTW could you describe the other regression?
Nevermind, I saw your note on IRC. If you haven't found the problem yet, could you pleas try modprobe tcrypt mode=100?
That should better pin-point the problem.
Cheers,
Signature · 4 lines
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txtHerbert Xu
maintainerHi:
OK I think I've found the problem.
[CRYPTO] hmac: Fix hmac_init update call
The crypto_hash_update call in hmac_init gave the number 1 instead of the length of the sg list in bytes. This is a missed conversion from the digest => hash change.
As tcrypt only tests crypto_hash_digest it didn't catch this.
Cheers,
Signature · 5 lines
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
--crypto/hmac.c +1 -1
@@ -98,7 +98,7 @@ static int hmac_init(struct hash_desc *psg_set_buf (&tmp, ipad, bs);return unlikely (crypto_hash_init (&desc)) ?: -crypto_hash_update (&desc, &tmp,1 ); +crypto_hash_update (&desc, &tmp, bs); } static int hmac_update(struct hash_desc *pdesc,
David Miller
maintainerFrom: Herbert Xu <[email protected]>
Date: Sun, 24 Sep 2006 00:40:41 +1000
Quoted 11 lines
OK I think I've found the problem.
[CRYPTO] hmac: Fix
hmac_initupdate callThe
crypto_hash_updatecall inhmac_initgave the number 1 instead of the length of the sg list in bytes. This is a missed conversion from the digest => hash change.As tcrypt only tests
crypto_hash_digestit didn't catch this.Signed-off-by Herbert Xu <[email protected]>
This, along with the ALIGN() patch, fixes all the regressions for me. Good spotting Herbert!
Linus, please apply.
crypto/hmac.c static int hmac_init(struct hash_desc *p · line 104
Quoted 13 lines · Herbert Xu
crypto/hmac.c +1
-1@@ -98,7 +98,7 @@ static int hmac_init(struct hash_desc *psg_set_buf (&tmp, ipad, bs);return unlikely (crypto_hash_init (&desc)) ?: -crypto_hash_update (&desc, &tmp,1 ); +crypto_hash_update (&desc, &tmp, bs); } static int hmac_update(struct hash_desc *pdesc,
Quick question: does "crypto_hash_init()" ever return anything other than 0 or 1? If so this is a subtle bug, as "unlikely()" is implemented like this:
# define unlikely(x) __builtin_expect(!!(x), 0)
IMO any usage of likely/unlikely other than if(unlikely()), if(likely()) is probably a bug.
Cheers,
Kyle Moffett
Quoted 13 lines · Kyle Moffett
Quoted 12 lines · Herbert Xu
crypto/hmac.c +1
-1@@ -98,7 +98,7 @@ static int hmac_init(struct hash_desc *psg_set_buf (&tmp, ipad, bs);return unlikely (crypto_hash_init (&desc)) ?: -crypto_hash_update (&desc, &tmp,1 ); +crypto_hash_update (&desc, &tmp, bs); }
Quoted 7 lines
crypto/hmac.c static int hmac_init(struct hash_desc *p · line 104
Quoted 1 line
crypto/hmac.c +0
-0static int hmac_update(struct hash_desc *pdesc,Quick question: does "
crypto_hash_init()" ever return anything other than 0 or 1? If so this is a subtle bug, as "unlikely()" is implemented like this:# define unlikely(x) __builtin_expect(!!(x), 0)IMO any usage of likely/unlikely other than if(unlikely()), if(likely()) is probably a bug.
With a bit more contemplation, I think this is one place where (if the compiler or sparse are cooperative) we should really look at the _Bool type or similar. If we could cast likely()/unlikely() to return (_Bool), then these sorts of problems would be caught at compile time, and likewise for other functions which return boolean values.
Cheers,
Kyle Moffett
Herbert Xu
maintainerQuoted 4 lines · Kyle Moffett
Quick question: does "
crypto_hash_init()" ever return anything other than 0 or 1? If so this is a subtle bug, as "unlikely()" is implemented like this:
Good point. It's meant to be an errno value so this is a bug.
[CRYPTO] hmac: Fix error truncation by unlikely()
The error return values are truncated by unlikely so we need to save it first. Thanks to Kyle Moffett for spotting this.
Signature · 5 lines
Visit Openswan at http://www.openswan.org/
Email: Herbert Xu ~{PmV>HI~} <[email protected]>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
--crypto/hmac.c +18 -6
@@ -92,13 +92,17 @@ static int hmac_init(struct hash_desc *pstruct hmac_ctx *ctx =align_ptr (ipad + bs *2 + ds,sizeof (void *));struct hash_desc desc;struct scatterlist tmp; +int err; desc.tfm = ctx->child; desc.flags = pdesc->flags &CRYPTO_TFM_REQ_MAY_SLEEP ;sg_set_buf (&tmp, ipad, bs); -return unlikely (crypto_hash_init (&desc)) ?: -crypto_hash_update (&desc, &tmp, bs); + err =crypto_hash_init (&desc); +if (unlikely (err)) +return err; + +return crypto_hash_update (&desc, &tmp, bs); } static int hmac_update(struct hash_desc *pdesc,
@@ -123,13 +127,17 @@ static int hmac_final(struct hash_desc *struct hmac_ctx *ctx =align_ptr (digest + ds,sizeof (void *));struct hash_desc desc;struct scatterlist tmp; +int err; desc.tfm = ctx->child; desc.flags = pdesc->flags &CRYPTO_TFM_REQ_MAY_SLEEP ;sg_set_buf (&tmp, opad, bs + ds); -return unlikely (crypto_hash_final (&desc, digest)) ?: -crypto_hash_digest (&desc, &tmp, bs + ds, out); + err =crypto_hash_final (&desc, digest); +if (unlikely (err)) +return err; + +return crypto_hash_digest (&desc, &tmp, bs + ds, out); } static int hmac_digest(struct hash_desc *pdesc, struct scatterlist *sg,
@@ -145,6 +153,7 @@ static int hmac_digest(struct hash_descstruct hash_desc desc;struct scatterlist sg1[2 ];struct scatterlist sg2[1 ]; +int err; desc.tfm = ctx->child; desc.flags = pdesc->flags &CRYPTO_TFM_REQ_MAY_SLEEP ;
@@ -154,8 +163,11 @@ static int hmac_digest(struct hash_desc sg1[1 ].length =0 ;sg_set_buf (sg2, opad, bs + ds); -return unlikely (crypto_hash_digest (&desc, sg1, nbytes + bs, digest)) ?: -crypto_hash_digest (&desc, sg2, bs + ds, out); + err =crypto_hash_digest (&desc, sg1, nbytes + bs, digest); +if (unlikely (err)) +return err; + +return crypto_hash_digest (&desc, sg2, bs + ds, out); }static int hmac_init_tfm (struct crypto_tfm *tfm)
David Miller
maintainerFrom: Herbert Xu <[email protected]>
Date: Sat, 23 Sep 2006 22:54:59 +1000
Quoted 6 lines · herbert
Quoted 5 lines
Thanks for fixing this Dave. I recall being bitten earlier by the same thing as well. I really need to start testing on 64-bit.
Quoted 6 lines
Quoted 1 line
BTW could you describe the other regression?
Nevermind, I saw your note on IRC. If you haven't found the problem yet, could you pleas try modprobe tcrypt mode=100?
That should better pin-point the problem.
I'll do that, but currently I believe that the error has something to do with the "tail" skb handling in the ESP conversions within that changeset.
I should have a good handle on this bug by the end of today.
Thanks.