| Message ID | 20150526180635.6527.5580.stgit@tlendack-t1.amdoffice.net (mailing list archive) |
|---|---|
| State | Changes Requested |
| Delegated to: | Herbert Xu |
| Headers | show |
Tom Lendacky <thomas.lendacky@amd.com> wrote: > Scatter gather lists can be created with more available entries than are > actually used (e.g. using sg_init_table() to reserve a specific number > of sg entries, but in actuality using something less than that based on > the data length). The caller sometimes fails to mark the last entry > with sg_mark_end(). In these cases, sg_nents() will return the original > size of the sg list as opposed to the actual number of sg entries that > contain valid data. > > On arm64, if the sg_nents() value is used in a call to dma_map_sg() in > this situation, then it causes a BUG_ON in lib/swiotlb.c because an > "empty" sg list entry results in dma_capable() returning false and > swiotlb trying to create a bounce buffer of size 0. This occurred in > the userspace crypto interface before being fixed by > > 0f477b655a52 ("crypto: algif - Mark sgl end at the end of data") > > Protect against this in the future by counting the number of sg entries > needed to meet the length requirement and supplying that value to > dma_map_sg(). Is this needed for any reason other than this bug that's already been fixed? The reason I'm asking is because while this patch fixes your driver everybody else will still crash and burn should something like this happen again. Cheers,
On Wed, May 27, 2015 at 05:43:05PM +0800, Herbert Xu wrote: > Tom Lendacky <thomas.lendacky@amd.com> wrote: > > Scatter gather lists can be created with more available entries than are > > actually used (e.g. using sg_init_table() to reserve a specific number > > of sg entries, but in actuality using something less than that based on > > the data length). The caller sometimes fails to mark the last entry > > with sg_mark_end(). In these cases, sg_nents() will return the original > > size of the sg list as opposed to the actual number of sg entries that > > contain valid data. > > > > On arm64, if the sg_nents() value is used in a call to dma_map_sg() in > > this situation, then it causes a BUG_ON in lib/swiotlb.c because an > > "empty" sg list entry results in dma_capable() returning false and > > swiotlb trying to create a bounce buffer of size 0. This occurred in > > the userspace crypto interface before being fixed by > > > > 0f477b655a52 ("crypto: algif - Mark sgl end at the end of data") > > > > Protect against this in the future by counting the number of sg entries > > needed to meet the length requirement and supplying that value to > > dma_map_sg(). > > Is this needed for any reason other than this bug that's already > been fixed? Could this be needed if you have a properly marked SG list say of 100 bytes but len is only 10 bytes? Cheers,
On 05/27/2015 04:43 AM, Herbert Xu wrote: > Tom Lendacky <thomas.lendacky@amd.com> wrote: >> Scatter gather lists can be created with more available entries than are >> actually used (e.g. using sg_init_table() to reserve a specific number >> of sg entries, but in actuality using something less than that based on >> the data length). The caller sometimes fails to mark the last entry >> with sg_mark_end(). In these cases, sg_nents() will return the original >> size of the sg list as opposed to the actual number of sg entries that >> contain valid data. >> >> On arm64, if the sg_nents() value is used in a call to dma_map_sg() in >> this situation, then it causes a BUG_ON in lib/swiotlb.c because an >> "empty" sg list entry results in dma_capable() returning false and >> swiotlb trying to create a bounce buffer of size 0. This occurred in >> the userspace crypto interface before being fixed by >> >> 0f477b655a52 ("crypto: algif - Mark sgl end at the end of data") >> >> Protect against this in the future by counting the number of sg entries >> needed to meet the length requirement and supplying that value to >> dma_map_sg(). > > Is this needed for any reason other than this bug that's already > been fixed? > I added this just to protect against any other users of the API that may do something similar in the future (or if the user should re-use an sg list and leave leftover sg entries in it). Since software crypto implementations walk the sg list based on length and do not use DMA mappings it is possible for this bug to pop up again in another location since it is likely that the testing won't be done with hardware crypto devices. > The reason I'm asking is because while this patch fixes your driver > everybody else will still crash and burn should something like this > happen again. A number of other drivers already have similar sg-count functions in them. I'm ok if you decide that this patch shouldn't be applied. It's just that this is typically an issue that won't be found until after the release of a kernel rather than during the development stages. Thanks, Tom > > Cheers, > -- To unsubscribe from this list: send the line "unsubscribe linux-crypto" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On 05/27/2015 04:45 AM, Herbert Xu wrote: > On Wed, May 27, 2015 at 05:43:05PM +0800, Herbert Xu wrote: >> Tom Lendacky <thomas.lendacky@amd.com> wrote: >>> Scatter gather lists can be created with more available entries than are >>> actually used (e.g. using sg_init_table() to reserve a specific number >>> of sg entries, but in actuality using something less than that based on >>> the data length). The caller sometimes fails to mark the last entry >>> with sg_mark_end(). In these cases, sg_nents() will return the original >>> size of the sg list as opposed to the actual number of sg entries that >>> contain valid data. >>> >>> On arm64, if the sg_nents() value is used in a call to dma_map_sg() in >>> this situation, then it causes a BUG_ON in lib/swiotlb.c because an >>> "empty" sg list entry results in dma_capable() returning false and >>> swiotlb trying to create a bounce buffer of size 0. This occurred in >>> the userspace crypto interface before being fixed by >>> >>> 0f477b655a52 ("crypto: algif - Mark sgl end at the end of data") >>> >>> Protect against this in the future by counting the number of sg entries >>> needed to meet the length requirement and supplying that value to >>> dma_map_sg(). >> >> Is this needed for any reason other than this bug that's already >> been fixed? > > Could this be needed if you have a properly marked SG list say of > 100 bytes but len is only 10 bytes? I don't think that situation matters because the DMA mapping should succeed just fine at 100 bytes even if only needing/using 10 bytes. Thanks, Tom > > Cheers, > -- To unsubscribe from this list: send the line "unsubscribe linux-crypto" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
On Wed, May 27, 2015 at 09:12:02AM -0500, Tom Lendacky wrote: > > >The reason I'm asking is because while this patch fixes your driver > >everybody else will still crash and burn should something like this > >happen again. > > A number of other drivers already have similar sg-count functions in > them. Perhaps you can help abstract this into a helper that everybody can call? Cheers,
On 05/27/2015 07:36 PM, Herbert Xu wrote: > On Wed, May 27, 2015 at 09:12:02AM -0500, Tom Lendacky wrote: >> >>> The reason I'm asking is because while this patch fixes your driver >>> everybody else will still crash and burn should something like this >>> happen again. >> >> A number of other drivers already have similar sg-count functions in >> them. > > Perhaps you can help abstract this into a helper that everybody can > call? I can do that. Something like an sg_nents_for_len() function that takes an sg pointer and a u64 length as arguments. The function should also return an error if the length requirement isn't satisfied. Thanks, Tom > > Cheers, > -- To unsubscribe from this list: send the line "unsubscribe linux-crypto" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
diff --git a/drivers/crypto/ccp/ccp-ops.c b/drivers/crypto/ccp/ccp-ops.c index 542453c..8377ed6 100644 --- a/drivers/crypto/ccp/ccp-ops.c +++ b/drivers/crypto/ccp/ccp-ops.c @@ -477,6 +477,22 @@ static u32 ccp_gen_jobid(struct ccp_device *ccp) return atomic_inc_return(&ccp->current_id) & CCP_JOBID_MASK; } +static int ccp_sg_nents(struct scatterlist *sg, u64 len) +{ + int nents = 0; + + while (sg && len) { + nents++; + if (sg->length > len) + break; + + len -= sg->length; + sg = sg_next(sg); + } + + return nents; +} + static void ccp_sg_free(struct ccp_sg_workarea *wa) { if (wa->dma_count) @@ -495,7 +511,7 @@ static int ccp_init_sg_workarea(struct ccp_sg_workarea *wa, struct device *dev, if (!sg) return 0; - wa->nents = sg_nents(sg); + wa->nents = ccp_sg_nents(sg, len); wa->bytes_left = len; wa->sg_used = 0;
Scatter gather lists can be created with more available entries than are actually used (e.g. using sg_init_table() to reserve a specific number of sg entries, but in actuality using something less than that based on the data length). The caller sometimes fails to mark the last entry with sg_mark_end(). In these cases, sg_nents() will return the original size of the sg list as opposed to the actual number of sg entries that contain valid data. On arm64, if the sg_nents() value is used in a call to dma_map_sg() in this situation, then it causes a BUG_ON in lib/swiotlb.c because an "empty" sg list entry results in dma_capable() returning false and swiotlb trying to create a bounce buffer of size 0. This occurred in the userspace crypto interface before being fixed by 0f477b655a52 ("crypto: algif - Mark sgl end at the end of data") Protect against this in the future by counting the number of sg entries needed to meet the length requirement and supplying that value to dma_map_sg(). Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com> --- drivers/crypto/ccp/ccp-ops.c | 18 +++++++++++++++++- 1 file changed, 17 insertions(+), 1 deletion(-) -- To unsubscribe from this list: send the line "unsubscribe linux-crypto" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html