| Message ID | 20170912093553.2580986-2-arnd@arndb.de |
|---|---|
| State | Accepted |
| Commit | 532f419cde077ffe9616c97902af177fbb868b17 |
| Headers | show |
| Series | [1/2] crypto: stm32 - Fix uninitialized data usage | expand |
On Tue, Sep 12, 2017 at 11:35:39AM +0200, Arnd Bergmann wrote: > gcc warns that the length for the extra unaligned data in the hash > function may be used unaligned. In theory this could happen if > we pass a zero-length sg_list, or if sg_is_last() was never true: > > In file included from drivers/crypto/stm32/stm32-hash.c:23: > drivers/crypto/stm32/stm32-hash.c: In function 'stm32_hash_one_request': > include/uapi/linux/kernel.h:12:49: error: 'ncp' may be used uninitialized in this function [-Werror=maybe-uninitialized] > #define __KERNEL_DIV_ROUND_UP(n, d) (((n) + (d) - 1) / (d)) > > Neither of these can happen in practice, so the warning is harmless. > > However while trying to suppress the warning, I noticed multiple > problems with that code: > > - On big-endian kernels, we byte-swap the data like we do for > register accesses, however this is a data stream and almost > certainly needs to use a single writesl() instead of series > of writel() to give the correct hash. > > - If the length is not a multiple of four bytes, we skip the > last word entirely, since we write the truncated length > using stm32_hash_set_nblw(). > > - If we change the code to round the length up rather than > down, the last bytes contain stale data, so it needs some > form of padding. > > This tries to address all four problems, by correctly > initializing the length to zero, using endian-safe copy > functions, adding zero-padding and passing the padded length. > > I have done no testing on this patch, so please review > carefully and if possible test with an unaligned length > and big-endian kernel builds. > > Fixes: 8a1012d3f2ab ("crypto: stm32 - Support for STM32 HASH module") > Signed-off-by: Arnd Bergmann <arnd@arndb.de> Patch applied. Thanks. -- Email: Herbert Xu <herbert@gondor.apana.org.au> Home Page: http://gondor.apana.org.au/~herbert/ PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
diff --git a/drivers/crypto/stm32/stm32-hash.c b/drivers/crypto/stm32/stm32-hash.c index 3c23a23e9ee5..b2ed2bccdb66 100644 --- a/drivers/crypto/stm32/stm32-hash.c +++ b/drivers/crypto/stm32/stm32-hash.c @@ -553,9 +553,9 @@ static int stm32_hash_dma_send(struct stm32_hash_dev *hdev) { struct stm32_hash_request_ctx *rctx = ahash_request_ctx(hdev->req); struct scatterlist sg[1], *tsg; - int err = 0, len = 0, reg, ncp; + int err = 0, len = 0, reg, ncp = 0; unsigned int i; - const u32 *buffer = (const u32 *)rctx->buffer; + u32 *buffer = (void *)rctx->buffer; rctx->sg = hdev->req->src; rctx->total = hdev->req->nbytes; @@ -620,10 +620,13 @@ static int stm32_hash_dma_send(struct stm32_hash_dev *hdev) reg |= HASH_CR_DMAA; stm32_hash_write(hdev, HASH_CR, reg); - for (i = 0; i < DIV_ROUND_UP(ncp, sizeof(u32)); i++) - stm32_hash_write(hdev, HASH_DIN, buffer[i]); - - stm32_hash_set_nblw(hdev, ncp); + if (ncp) { + memset(buffer + ncp, 0, + DIV_ROUND_UP(ncp, sizeof(u32)) - ncp); + writesl(hdev->io_base + HASH_DIN, buffer, + DIV_ROUND_UP(ncp, sizeof(u32))); + } + stm32_hash_set_nblw(hdev, DIV_ROUND_UP(ncp, sizeof(u32))); reg = stm32_hash_read(hdev, HASH_STR); reg |= HASH_STR_DCAL; stm32_hash_write(hdev, HASH_STR, reg);
gcc warns that the length for the extra unaligned data in the hash function may be used unaligned. In theory this could happen if we pass a zero-length sg_list, or if sg_is_last() was never true: In file included from drivers/crypto/stm32/stm32-hash.c:23: drivers/crypto/stm32/stm32-hash.c: In function 'stm32_hash_one_request': include/uapi/linux/kernel.h:12:49: error: 'ncp' may be used uninitialized in this function [-Werror=maybe-uninitialized] #define __KERNEL_DIV_ROUND_UP(n, d) (((n) + (d) - 1) / (d)) Neither of these can happen in practice, so the warning is harmless. However while trying to suppress the warning, I noticed multiple problems with that code: - On big-endian kernels, we byte-swap the data like we do for register accesses, however this is a data stream and almost certainly needs to use a single writesl() instead of series of writel() to give the correct hash. - If the length is not a multiple of four bytes, we skip the last word entirely, since we write the truncated length using stm32_hash_set_nblw(). - If we change the code to round the length up rather than down, the last bytes contain stale data, so it needs some form of padding. This tries to address all four problems, by correctly initializing the length to zero, using endian-safe copy functions, adding zero-padding and passing the padded length. I have done no testing on this patch, so please review carefully and if possible test with an unaligned length and big-endian kernel builds. Fixes: 8a1012d3f2ab ("crypto: stm32 - Support for STM32 HASH module") Signed-off-by: Arnd Bergmann <arnd@arndb.de> --- drivers/crypto/stm32/stm32-hash.c | 15 +++++++++------ 1 file changed, 9 insertions(+), 6 deletions(-) -- 2.9.0