crypto: aesni - Use zero-copy for gcm(aes) buffers that are partially contiguous
Subsystems (2)
v2 cover crypto: aesni - Use zero-copy for gcm(aes) buffers that are partially contiguous
Changes in v2:
- Integrated https://patchwork.kernel.org/patch/10173981
Currently, the AESNI gcm(aes) implementation uses zero-copy only when the entire src and dest request buffers, including the AAD, the data and the Auth Tag are contiguous. This series enables the use of zero-copy even if the AAD and/or Auth Tag are in different buffers than the actual data, as long as each of them individually satisfies the zero-copy conditions (i.e. the entire buffer is either in low-mem or within a single high-mem page). Furthermore, it also enables the use of zero-copy even if only one of src and dest satisfies these conditions rather than only when both of them do.
Junaid Shahid (4):
crypto: aesni - Fix out-of-bounds access of the AAD buffer in AVX
gcm-aesni
crypto: aesni - Enable one-sided zero copy for gcm(aes) request
buffers
crypto: aesni - Directly use kmap_atomic instead of scatter_walk
object in gcm(aes)
crypto: aesni - Use zero-copy for gcm(aes) even if the
AAD/Data/AuthTag are separate
arch/x86/crypto/aesni-intel_avx-x86_64.S | 154 +++++----------- arch/x86/crypto/aesni-intel_glue.c | 307 +++++++++++++++++++------------ 2 files changed, 227 insertions(+), 234 deletions(-)
Signature · 1 line
2.16.0.rc1.238.g530d649a79-goog
PATCH v2 1/4 crypto: aesni - Fix out-of-bounds access of the AAD buffer in AVX gcm-aesni
The AVX/AVX2 versions of gcm-aes encryption/decryption functions can access memory after the end of the AAD buffer if the AAD length is not a multiple of 4 bytes. It didn't matter as long as the AAD and data buffers were always contiguous, since the AVX version are not used for small data sizes and hence enough data bytes were always present to cover the over-run. However, now that we have support for non-contiguous AAD and data buffers, that is no longer the case. This can potentially result in accessing a page that is not mapped and thus causing the machine to crash. This patch fixes that by reading the last <16 byte block of the AAD byte-by-byte and optionally via an 8-byte load if the block was at least 8 bytes.
--- arch/x86/crypto/aesni-intel_avx-x86_64.S | 154 +++++++++---------------------- 1 file changed, 42 insertions(+), 112 deletions(-)
arch/x86/crypto/aesni-intel_avx-x86_64.S +42 -112
@@ -106,14 +106,6 @@ ## ## AAD Format with 64-bit Extended Sequence Number ## -## -## aadLen: -## from the definition of the spec, aadLen can only be 8 or 12 bytes. -## The code additionally supports aadLen of length 16 bytes. -## -## TLen: -## from the definition of the spec, TLen can only be 8, 12 or 16 bytes. -## ## poly = x^128 + x^127 + x^126 + x^121 + 1 ## throughout the code, one tab and two tab indentations are used. one tab is ## for GHASH part, two tabs is for AES part.
@@ -155,30 +147,6 @@ SHIFT_MASK: .octa 0x0f0e0d0c0b0a09080706050403020100 ALL_F: .octa 0xffffffffffffffffffffffffffffffff .octa 0x00000000000000000000000000000000 -.section .rodata -.align 16 -.type aad_shift_arr, @object -.size aad_shift_arr, 272 -aad_shift_arr: - .octa 0xffffffffffffffffffffffffffffffff - .octa 0xffffffffffffffffffffffffffffff0C - .octa 0xffffffffffffffffffffffffffff0D0C - .octa 0xffffffffffffffffffffffffff0E0D0C - .octa 0xffffffffffffffffffffffff0F0E0D0C - .octa 0xffffffffffffffffffffff0C0B0A0908 - .octa 0xffffffffffffffffffff0D0C0B0A0908 - .octa 0xffffffffffffffffff0E0D0C0B0A0908 - .octa 0xffffffffffffffff0F0E0D0C0B0A0908 - .octa 0xffffffffffffff0C0B0A090807060504 - .octa 0xffffffffffff0D0C0B0A090807060504 - .octa 0xffffffffff0E0D0C0B0A090807060504 - .octa 0xffffffff0F0E0D0C0B0A090807060504 - .octa 0xffffff0C0B0A09080706050403020100 - .octa 0xffff0D0C0B0A09080706050403020100 - .octa 0xff0E0D0C0B0A09080706050403020100 - .octa 0x0F0E0D0C0B0A09080706050403020100 - - .text
@@ -280,6 +248,36 @@ VARIABLE_OFFSET = 16*8 vaesenclast 16*10(arg1), \XMM0, \XMM0 .endm +# Reads DLEN bytes starting at DPTR and stores in XMMDst +# where 0 < DLEN < 16 +# Clobbers %rax, DLEN and XMM1 +.macro READ_PARTIAL_BLOCK DPTR DLEN XMM1 XMMDst + cmp $8, \DLEN + jl _read_lt8_\@ + movq (\DPTR), \XMMDst + sub $8, \DLEN + jz _done_read_partial_block_\@ + xor %eax, %eax +_read_next_byte_\@: + shl $8, %rax + mov 7(\DPTR, \DLEN, 1), %al + dec \DLEN + jnz _read_next_byte_\@ + movq %rax, \XMM1 + pslldq $8, \XMM1 + por \XMM1, \XMMDst + jmp _done_read_partial_block_\@ +_read_lt8_\@: + xor %eax, %eax +_read_next_byte_lt8_\@: + shl $8, %rax + mov -1(\DPTR, \DLEN, 1), %al + dec \DLEN + jnz _read_next_byte_lt8_\@ + movq %rax, \XMMDst +_done_read_partial_block_\@: +.endm + #ifdef CONFIG_AS_AVX ############################################################################### # GHASH_MUL MACRO to implement: Data*HashKey mod (128,127,126,121,0)
@@ -400,63 +398,29 @@ VARIABLE_OFFSET = 16*8 setreg mov arg6, %r10 # r10 = AAD - mov arg7, %r12 # r12 = aadLen - - - mov %r12, %r11 + mov arg7, %r11 # r11 = aadLen vpxor reg_j, reg_j, reg_j vpxor reg_i, reg_i, reg_i cmp $16, %r11 - jl _get_AAD_rest8\@ + jl _get_AAD_rest\@ _get_AAD_blocks\@: vmovdqu (%r10), reg_i vpshufb SHUF_MASK(%rip), reg_i, reg_i vpxor reg_i, reg_j, reg_j GHASH_MUL_AVX reg_j, \T2, \T1, \T3, \T4, \T5, \T6 add $16, %r10 - sub $16, %r12 sub $16, %r11 cmp $16, %r11 jge _get_AAD_blocks\@ vmovdqu reg_j, reg_i + + /* read the last <16B of AAD. */ +_get_AAD_rest\@: cmp $0, %r11 je _get_AAD_done\@ - vpxor reg_i, reg_i, reg_i - - /* read the last <16B of AAD. since we have at least 4B of - data right after the AAD (the ICV, and maybe some CT), we can - read 4B/8B blocks safely, and then get rid of the extra stuff */ -_get_AAD_rest8\@: - cmp $4, %r11 - jle _get_AAD_rest4\@ - movq (%r10), \T1 - add $8, %r10 - sub $8, %r11 - vpslldq $8, \T1, \T1 - vpsrldq $8, reg_i, reg_i - vpxor \T1, reg_i, reg_i - jmp _get_AAD_rest8\@ -_get_AAD_rest4\@: - cmp $0, %r11 - jle _get_AAD_rest0\@ - mov (%r10), %eax - movq %rax, \T1 - add $4, %r10 - sub $4, %r11 - vpslldq $12, \T1, \T1 - vpsrldq $4, reg_i, reg_i - vpxor \T1, reg_i, reg_i -_get_AAD_rest0\@: - /* finalize: shift out the extra bytes we read, and align - left. since pslldq can only shift by an immediate, we use - vpshufb and an array of shuffle masks */ - movq %r12, %r11 - salq $4, %r11 - movdqu aad_shift_arr(%r11), \T1 - vpshufb \T1, reg_i, reg_i -_get_AAD_rest_final\@: + READ_PARTIAL_BLOCK %r10, %r11, \T1, reg_i vpshufb SHUF_MASK(%rip), reg_i, reg_i vpxor reg_j, reg_i, reg_i GHASH_MUL_AVX reg_i, \T2, \T1, \T3, \T4, \T5, \T6
@@ -1706,64 +1670,30 @@ ENDPROC(aesni_gcm_dec_avx_gen2) setreg mov arg6, %r10 # r10 = AAD - mov arg7, %r12 # r12 = aadLen - - - mov %r12, %r11 + mov arg7, %r11 # r11 = aadLen vpxor reg_j, reg_j, reg_j vpxor reg_i, reg_i, reg_i cmp $16, %r11 - jl _get_AAD_rest8\@ + jl _get_AAD_rest\@ _get_AAD_blocks\@: vmovdqu (%r10), reg_i vpshufb SHUF_MASK(%rip), reg_i, reg_i vpxor reg_i, reg_j, reg_j GHASH_MUL_AVX2 reg_j, \T2, \T1, \T3, \T4, \T5, \T6 add $16, %r10 - sub $16, %r12 sub $16, %r11 cmp $16, %r11 jge _get_AAD_blocks\@ vmovdqu reg_j, reg_i + + /* read the last <16B of AAD. */ +_get_AAD_rest\@: cmp $0, %r11 je _get_AAD_done\@ - vpxor reg_i, reg_i, reg_i - - /* read the last <16B of AAD. since we have at least 4B of - data right after the AAD (the ICV, and maybe some CT), we can - read 4B/8B blocks safely, and then get rid of the extra stuff */ -_get_AAD_rest8\@: - cmp $4, %r11 - jle _get_AAD_rest4\@ - movq (%r10), \T1 - add $8, %r10 - sub $8, %r11 - vpslldq $8, \T1, \T1 - vpsrldq $8, reg_i, reg_i - vpxor \T1, reg_i, reg_i - jmp _get_AAD_rest8\@ -_get_AAD_rest4\@: - cmp $0, %r11 - jle _get_AAD_rest0\@ - mov (%r10), %eax - movq %rax, \T1 - add $4, %r10 - sub $4, %r11 - vpslldq $12, \T1, \T1 - vpsrldq $4, reg_i, reg_i - vpxor \T1, reg_i, reg_i -_get_AAD_rest0\@: - /* finalize: shift out the extra bytes we read, and align - left. since pslldq can only shift by an immediate, we use - vpshufb and an array of shuffle masks */ - movq %r12, %r11 - salq $4, %r11 - movdqu aad_shift_arr(%r11), \T1 - vpshufb \T1, reg_i, reg_i -_get_AAD_rest_final\@: + READ_PARTIAL_BLOCK %r10, %r11, \T1, reg_i vpshufb SHUF_MASK(%rip), reg_i, reg_i vpxor reg_j, reg_i, reg_i GHASH_MUL_AVX2 reg_i, \T2, \T1, \T3, \T4, \T5, \T6
Signature · 1 line
2.16.0.rc1.238.g530d649a79-goog
PATCH v2 2/4 crypto: aesni - Enable one-sided zero copy for gcm(aes) request buffers
In reply to Junaid Shahidgcmaes_encrypt/decrypt perform zero-copy crypto if both the source and destination satisfy certain conditions (single sglist entry located in low-mem or within a single high-mem page). But two copies are done otherwise, even if one of source or destination still satisfies the zero-copy conditions. This optimization is now extended to avoid the copy on the side that does satisfy the zero-copy conditions.
--- arch/x86/crypto/aesni-intel_glue.c | 256 +++++++++++++++++++------------------ 1 file changed, 134 insertions(+), 122 deletions(-)
arch/x86/crypto/aesni-intel_glue.c +134 -122
@@ -744,136 +744,148 @@ static int generic_gcmaes_set_authsize(struct crypto_aead *tfm,return 0 ; } +static bool is_mappable (struct scatterlist *sgl,unsigned long len) +{ +return sgl->length >0 && len <= sgl->length && + (!PageHighMem (sg_page (sgl)) || sgl->offset + len <=PAGE_SIZE ); +} + +/* +* Maps the sglist buffer and returns a pointer to the mapped buffer in +* data_buf. +* +* If direct mapping is not feasible, then allocates a bounce buffer if one +* isn't already available in bounce_buf, and returns a pointer to the bounce +* buffer in data_buf. +* +* When the buffer is no longer needed, put_request_buffer() should be called on +* the data_buf and the bounce_buf should be freed using kfree(). +*/ +static int get_request_buffer (struct scatterlist *sgl, +struct scatter_walk *sg_walk, +unsigned long bounce_buf_size, +u8 **data_buf,u8 **bounce_buf,bool *mapped) +{ +if (sg_is_last (sgl) &&is_mappable (sgl, sgl->length)) { + *mapped = true; +scatterwalk_start (sg_walk, sgl); + *data_buf =scatterwalk_map (sg_walk); +return 0 ; + } + + *mapped = false; + +if (*bounce_buf ==NULL ) { + *bounce_buf =kmalloc (bounce_buf_size,GFP_ATOMIC ); +if (unlikely (*bounce_buf ==NULL )) +return -ENOMEM ; + } + + *data_buf = *bounce_buf; +return 0 ; +} + +static void put_request_buffer (u8 *data_buf,unsigned long len,bool mapped, +struct scatter_walk *sg_walk,bool output) +{ +if (mapped) { +scatterwalk_unmap (data_buf); +scatterwalk_advance (sg_walk, len); +scatterwalk_done (sg_walk, output,0 ); + } +} + +/* +* Performs the encryption/decryption operation for the given request. The src +* and dst sglists in the request are directly mapped if possible. Otherwise, a +* bounce buffer is allocated and used to copy the data from the src or to the +* dst, or both. +*/ +static int gcmaes_crypt (struct aead_request *req,unsigned int assoclen, +u8 *hash_subkey,u8 *iv,void *aes_ctx,bool decrypt) +{ +u8 *src, *dst, *assoc, *bounce_buf =NULL ; +bool src_mapped = false, dst_mapped = false; +struct crypto_aead *tfm =crypto_aead_reqtfm (req); +unsigned long auth_tag_len =crypto_aead_authsize (tfm); +unsigned long data_len = req->cryptlen - (decrypt ? auth_tag_len :0 ); +struct scatter_walk src_sg_walk; +struct scatter_walk dst_sg_walk = {}; +int retval =0 ; +unsigned long bounce_buf_size = data_len + auth_tag_len + req->assoclen; + +if (auth_tag_len >16 ) +return -EINVAL ; + + retval =get_request_buffer (req->src, &src_sg_walk, bounce_buf_size, + &assoc, &bounce_buf, &src_mapped); +if (retval) + goto exit; + + src = assoc + req->assoclen; + +if (req->src == req->dst) { + dst = src; + dst_mapped = src_mapped; + }else { + retval =get_request_buffer (req->dst, &dst_sg_walk, + bounce_buf_size, &dst, &bounce_buf, + &dst_mapped); +if (retval) + goto exit; + + dst += req->assoclen; + } + +if (!src_mapped) +scatterwalk_map_and_copy (bounce_buf, req->src,0 , + req->assoclen + req->cryptlen,0 ); + +kernel_fpu_begin (); + +if (decrypt) { +u8 gen_auth_tag[16 ]; + +aesni_gcm_dec_tfm (aes_ctx, dst, src, data_len, iv, + hash_subkey, assoc, assoclen, + gen_auth_tag, auth_tag_len); +/* Compare generated tag with passed in tag. */ +if (crypto_memneq (src + data_len, gen_auth_tag, auth_tag_len)) + retval = -EBADMSG ; + + }else +aesni_gcm_enc_tfm (aes_ctx, dst, src, data_len, iv, + hash_subkey, assoc, assoclen, + dst + data_len, auth_tag_len); + +kernel_fpu_end (); + +if (!dst_mapped) +scatterwalk_map_and_copy (dst, req->dst, req->assoclen, + data_len + (decrypt ?0 : auth_tag_len), +1 ); +exit: +if (req->dst != req->src) +put_request_buffer (dst - req->assoclen, req->dst->length, + dst_mapped, &dst_sg_walk, true); + +put_request_buffer (assoc, req->src->length, src_mapped, &src_sg_walk, + false); + +kfree (bounce_buf); +return retval; +} +static int gcmaes_encrypt (struct aead_request *req,unsigned int assoclen,u8 *hash_subkey,u8 *iv,void *aes_ctx) { -u8 one_entry_in_sg =0 ; -u8 *src, *dst, *assoc; -struct crypto_aead *tfm =crypto_aead_reqtfm (req); -unsigned long auth_tag_len =crypto_aead_authsize (tfm); -struct scatter_walk src_sg_walk; -struct scatter_walk dst_sg_walk = {}; - -if (sg_is_last (req->src) && - (!PageHighMem (sg_page (req->src)) || - req->src->offset + req->src->length <=PAGE_SIZE ) && -sg_is_last (req->dst) && - (!PageHighMem (sg_page (req->dst)) || - req->dst->offset + req->dst->length <=PAGE_SIZE )) { - one_entry_in_sg =1 ; -scatterwalk_start (&src_sg_walk, req->src); - assoc =scatterwalk_map (&src_sg_walk); - src = assoc + req->assoclen; - dst = src; -if (unlikely (req->src != req->dst)) { -scatterwalk_start (&dst_sg_walk, req->dst); - dst =scatterwalk_map (&dst_sg_walk) + req->assoclen; - } - }else { -/* Allocate memory for src, dst, assoc */ - assoc =kmalloc (req->cryptlen + auth_tag_len + req->assoclen, -GFP_ATOMIC ); -if (unlikely (!assoc)) -return -ENOMEM ; -scatterwalk_map_and_copy (assoc, req->src,0 , - req->assoclen + req->cryptlen,0 ); - src = assoc + req->assoclen; - dst = src; - } - -kernel_fpu_begin (); -aesni_gcm_enc_tfm (aes_ctx, dst, src, req->cryptlen, iv, - hash_subkey, assoc, assoclen, - dst + req->cryptlen, auth_tag_len); -kernel_fpu_end (); - -/* The authTag (aka the Integrity Check Value) needs to be written -* back to the packet. */ -if (one_entry_in_sg) { -if (unlikely (req->src != req->dst)) { -scatterwalk_unmap (dst - req->assoclen); -scatterwalk_advance (&dst_sg_walk, req->dst->length); -scatterwalk_done (&dst_sg_walk,1 ,0 ); - } -scatterwalk_unmap (assoc); -scatterwalk_advance (&src_sg_walk, req->src->length); -scatterwalk_done (&src_sg_walk, req->src == req->dst,0 ); - }else { -scatterwalk_map_and_copy (dst, req->dst, req->assoclen, - req->cryptlen + auth_tag_len,1 ); -kfree (assoc); - } -return 0 ; +return gcmaes_crypt (req, assoclen, hash_subkey, iv, aes_ctx, false); }static int gcmaes_decrypt (struct aead_request *req,unsigned int assoclen,u8 *hash_subkey,u8 *iv,void *aes_ctx) { -u8 one_entry_in_sg =0 ; -u8 *src, *dst, *assoc; -unsigned long tempCipherLen =0 ; -struct crypto_aead *tfm =crypto_aead_reqtfm (req); -unsigned long auth_tag_len =crypto_aead_authsize (tfm); -u8 authTag[16 ]; -struct scatter_walk src_sg_walk; -struct scatter_walk dst_sg_walk = {}; -int retval =0 ; - - tempCipherLen = (unsigned long )(req->cryptlen - auth_tag_len); - -if (sg_is_last (req->src) && - (!PageHighMem (sg_page (req->src)) || - req->src->offset + req->src->length <=PAGE_SIZE ) && -sg_is_last (req->dst) && - (!PageHighMem (sg_page (req->dst)) || - req->dst->offset + req->dst->length <=PAGE_SIZE )) { - one_entry_in_sg =1 ; -scatterwalk_start (&src_sg_walk, req->src); - assoc =scatterwalk_map (&src_sg_walk); - src = assoc + req->assoclen; - dst = src; -if (unlikely (req->src != req->dst)) { -scatterwalk_start (&dst_sg_walk, req->dst); - dst =scatterwalk_map (&dst_sg_walk) + req->assoclen; - } - }else { -/* Allocate memory for src, dst, assoc */ - assoc =kmalloc (req->cryptlen + req->assoclen,GFP_ATOMIC ); -if (!assoc) -return -ENOMEM ; -scatterwalk_map_and_copy (assoc, req->src,0 , - req->assoclen + req->cryptlen,0 ); - src = assoc + req->assoclen; - dst = src; - } - - -kernel_fpu_begin (); -aesni_gcm_dec_tfm (aes_ctx, dst, src, tempCipherLen, iv, - hash_subkey, assoc, assoclen, - authTag, auth_tag_len); -kernel_fpu_end (); - -/* Compare generated tag with passed in tag. */ - retval =crypto_memneq (src + tempCipherLen, authTag, auth_tag_len) ? - -EBADMSG :0 ; - -if (one_entry_in_sg) { -if (unlikely (req->src != req->dst)) { -scatterwalk_unmap (dst - req->assoclen); -scatterwalk_advance (&dst_sg_walk, req->dst->length); -scatterwalk_done (&dst_sg_walk,1 ,0 ); - } -scatterwalk_unmap (assoc); -scatterwalk_advance (&src_sg_walk, req->src->length); -scatterwalk_done (&src_sg_walk, req->src == req->dst,0 ); - }else { -scatterwalk_map_and_copy (dst, req->dst, req->assoclen, - tempCipherLen,1 ); -kfree (assoc); - } -return retval; - +return gcmaes_crypt (req, assoclen, hash_subkey, iv, aes_ctx, true); }static int helper_rfc4106_encrypt (struct aead_request *req)
Signature · 1 line
2.16.0.rc1.238.g530d649a79-goog
Re: [PATCH v2 2/4] crypto: aesni - Enable one-sided zero copy for gcm(aes) request buffers
Am Dienstag, 23. Januar 2018, 21:19:14 CET schrieb Junaid Shahid:
Hi Junaid,
Quoted 6 lines
gcmaes_encrypt/decrypt perform zero-copy crypto if both the source and destination satisfy certain conditions (single sglist entry located in low-mem or within a single high-mem page). But two copies are done otherwise, even if one of source or destination still satisfies the zero-copy conditions. This optimization is now extended to avoid the copy on the side that does satisfy the zero-copy conditions.
The patch I pointed you and you nicely integrated was accepted by Herbert into the current tree. This implies that your patch will not apply cleanly any more.
May I suggest to rebase it to the current code?
Ciao
Stephan
Re: [PATCH v2 2/4] crypto: aesni - Enable one-sided zero copy for gcm(aes) request buffers
Hi Stephan,
Sure, I'll rebase and send another revision.
Thanks,
Junaid
PATCH v2 3/4 crypto: aesni - Directly use kmap_atomic instead of scatter_walk object in gcm(aes)
In reply to Junaid Shahidgcmaes_crypt uses a scatter_walk object to map and unmap the crypto request sglists. But the only purpose that appears to serve here is to allow the D-Cache to be flushed at the end for pages that were used as output. However, that is not applicable on x86, so we can avoid using the scatter_walk object for simplicity.
--- arch/x86/crypto/aesni-intel_glue.c | 36 +++++++++++++++--------------------- 1 file changed, 15 insertions(+), 21 deletions(-)
arch/x86/crypto/aesni-intel_glue.c +15 -21
@@ -750,6 +750,11 @@ static bool is_mappable(struct scatterlist *sgl, unsigned long len) (!PageHighMem (sg_page (sgl)) || sgl->offset + len <=PAGE_SIZE ); } +static u8 *map_buffer (struct scatterlist *sgl) +{ +return kmap_atomic (sg_page (sgl)) + sgl->offset; +} +/* * Maps the sglist buffer and returns a pointer to the mapped buffer in * data_buf.
@@ -762,14 +767,12 @@ static bool is_mappable(struct scatterlist *sgl, unsigned long len)* the data_buf and the bounce_buf should be freed using kfree(). */ static int get_request_buffer (struct scatterlist *sgl, -struct scatter_walk *sg_walk,unsigned long bounce_buf_size,u8 **data_buf,u8 **bounce_buf,bool *mapped) {if (sg_is_last (sgl) &&is_mappable (sgl, sgl->length)) { *mapped = true; -scatterwalk_start (sg_walk, sgl); - *data_buf =scatterwalk_map (sg_walk); + *data_buf =map_buffer (sgl);return 0 ; }
@@ -785,14 +788,10 @@ static int get_request_buffer(struct scatterlist *sgl,return 0 ; } -static void put_request_buffer (u8 *data_buf,unsigned long len,bool mapped, -struct scatter_walk *sg_walk,bool output) +static void put_request_buffer (u8 *data_buf,bool mapped) { -if (mapped) { -scatterwalk_unmap (data_buf); -scatterwalk_advance (sg_walk, len); -scatterwalk_done (sg_walk, output,0 ); - } +if (mapped) +kunmap_atomic (data_buf); }/*
@@ -809,16 +808,14 @@ static int gcmaes_crypt(struct aead_request *req, unsigned int assoclen,struct crypto_aead *tfm =crypto_aead_reqtfm (req);unsigned long auth_tag_len =crypto_aead_authsize (tfm);unsigned long data_len = req->cryptlen - (decrypt ? auth_tag_len :0 ); -struct scatter_walk src_sg_walk; -struct scatter_walk dst_sg_walk = {};int retval =0 ;unsigned long bounce_buf_size = data_len + auth_tag_len + req->assoclen;if (auth_tag_len >16 )return -EINVAL ; - retval =get_request_buffer (req->src, &src_sg_walk, bounce_buf_size, - &assoc, &bounce_buf, &src_mapped); + retval =get_request_buffer (req->src, bounce_buf_size, &assoc, + &bounce_buf, &src_mapped);if (retval) goto exit;
@@ -828,9 +825,8 @@ static int gcmaes_crypt(struct aead_request *req, unsigned int assoclen, dst = src; dst_mapped = src_mapped; }else { - retval =get_request_buffer (req->dst, &dst_sg_walk, - bounce_buf_size, &dst, &bounce_buf, - &dst_mapped); + retval =get_request_buffer (req->dst, bounce_buf_size, &dst, + &bounce_buf, &dst_mapped);if (retval) goto exit;
@@ -866,11 +862,9 @@ static int gcmaes_crypt(struct aead_request *req, unsigned int assoclen, 1); exit:if (req->dst != req->src) -put_request_buffer (dst - req->assoclen, req->dst->length, - dst_mapped, &dst_sg_walk, true); +put_request_buffer (dst - req->assoclen, dst_mapped); -put_request_buffer (assoc, req->src->length, src_mapped, &src_sg_walk, - false); +put_request_buffer (assoc, src_mapped);kfree (bounce_buf);return retval;
Signature · 1 line
2.16.0.rc1.238.g530d649a79-goog
PATCH v2 4/4 crypto: aesni - Use zero-copy for gcm(aes) even if the AAD/Data/AuthTag are separate
In reply to Junaid ShahidEnable the use of zero-copy even if the AAD and/or Auth Tag are in different buffers than the actual data, as long as each of them individually satisfies the zero-copy conditions (i.e. the entire buffer is either in low-mem or within a single high-mem page).
--- arch/x86/crypto/aesni-intel_glue.c | 121 +++++++++++++++++++++++++++---------- 1 file changed, 89 insertions(+), 32 deletions(-)
arch/x86/crypto/aesni-intel_glue.c +89 -32
@@ -756,42 +756,91 @@ static u8 *map_buffer(struct scatterlist *sgl) }/* -* Maps the sglist buffer and returns a pointer to the mapped buffer in -* data_buf. +* Maps the sglist buffer and returns pointers to the mapped buffers in assoc, +* data and (optionally) auth_tag. * * If direct mapping is not feasible, then allocates a bounce buffer if one -* isn't already available in bounce_buf, and returns a pointer to the bounce -* buffer in data_buf. +* isn't already available in bounce_buf, and returns pointers within the bounce +* buffer in assoc, data and auth_tag. * -* When the buffer is no longer needed, put_request_buffer() should be called on -* the data_buf and the bounce_buf should be freed using kfree(). +* When the buffers are no longer needed, put_request_buffers() should be called +* and the bounce_buf should be freed using kfree(). */ -static int get_request_buffer (struct scatterlist *sgl, -unsigned long bounce_buf_size, -u8 **data_buf,u8 **bounce_buf,bool *mapped) +static int get_request_buffers (struct scatterlist *sgl, +unsigned long assoc_len,unsigned long data_len, +unsigned long auth_tag_len, +u8 **assoc,u8 **data,u8 **auth_tag, +u8 **bounce_buf,bool *mapped) { -if (sg_is_last (sgl) &&is_mappable (sgl, sgl->length)) { +struct scatterlist sgl_data_chain[2 ], sgl_auth_tag_chain[2 ]; +struct scatterlist *sgl_data, *sgl_auth_tag; + + sgl_data =scatterwalk_ffwd (sgl_data_chain, sgl, assoc_len); + sgl_auth_tag =scatterwalk_ffwd (sgl_auth_tag_chain, sgl, + assoc_len + data_len); + +if (is_mappable (sgl, assoc_len) &&is_mappable (sgl_data, data_len) && + (auth_tag ==NULL ||is_mappable (sgl_auth_tag, auth_tag_len))) { *mapped = true; - *data_buf =map_buffer (sgl); + + *assoc =map_buffer (sgl); + +if (sgl->length >= assoc_len + data_len) + *data = *assoc + assoc_len; +else + *data =map_buffer (sgl_data); + +if (auth_tag !=NULL ) { +if (sgl_data->length >= data_len + auth_tag_len) + *auth_tag = *data + data_len; +else + *auth_tag =map_buffer (sgl_auth_tag); + } +return 0 ; } *mapped = false;if (*bounce_buf ==NULL ) { - *bounce_buf =kmalloc (bounce_buf_size,GFP_ATOMIC ); + *bounce_buf =kmalloc (assoc_len + data_len + auth_tag_len, +GFP_ATOMIC );if (unlikely (*bounce_buf ==NULL ))return -ENOMEM ; } - *data_buf = *bounce_buf; + *assoc = *bounce_buf; + *data = *assoc + assoc_len; + +if (auth_tag !=NULL ) + *auth_tag = *data + data_len; +return 0 ; } -static void put_request_buffer (u8 *data_buf,bool mapped) +static void put_request_buffers (struct scatterlist *sgl,bool mapped, +u8 *assoc,u8 *data,u8 *auth_tag, +unsigned long assoc_len, +unsigned long data_len, +unsigned long auth_tag_len) { -if (mapped) -kunmap_atomic (data_buf); +struct scatterlist sgl_data_chain[2 ]; +struct scatterlist *sgl_data; + +if (!mapped) +return ; + + sgl_data =scatterwalk_ffwd (sgl_data_chain, sgl, assoc_len); + +/* The unmaps need to be done in reverse order of the maps. */ + +if (auth_tag !=NULL && sgl_data->length < data_len + auth_tag_len) +kunmap_atomic (auth_tag); + +if (sgl->length < assoc_len + data_len) +kunmap_atomic (data); + +kunmap_atomic (assoc); }/*
@@ -803,34 +852,38 @@ static void put_request_buffer(u8 *data_buf, bool mapped)static int gcmaes_crypt (struct aead_request *req,unsigned int assoclen,u8 *hash_subkey,u8 *iv,void *aes_ctx,bool decrypt) { -u8 *src, *dst, *assoc, *bounce_buf =NULL ; +u8 *src, *src_assoc; +u8 *dst, *dst_assoc; +u8 *auth_tag; +u8 *bounce_buf =NULL ;bool src_mapped = false, dst_mapped = false;struct crypto_aead *tfm =crypto_aead_reqtfm (req);unsigned long auth_tag_len =crypto_aead_authsize (tfm);unsigned long data_len = req->cryptlen - (decrypt ? auth_tag_len :0 );int retval =0 ; -unsigned long bounce_buf_size = data_len + auth_tag_len + req->assoclen;if (auth_tag_len >16 )return -EINVAL ; - retval =get_request_buffer (req->src, bounce_buf_size, &assoc, - &bounce_buf, &src_mapped); + retval =get_request_buffers (req->src, req->assoclen, data_len, + auth_tag_len, &src_assoc, &src, + (decrypt || req->src == req->dst) + ? &auth_tag :NULL , + &bounce_buf, &src_mapped);if (retval) goto exit; - src = assoc + req->assoclen; -if (req->src == req->dst) { + dst_assoc = src_assoc; dst = src; dst_mapped = src_mapped; }else { - retval =get_request_buffer (req->dst, bounce_buf_size, &dst, - &bounce_buf, &dst_mapped); + retval =get_request_buffers (req->dst, req->assoclen, data_len, + auth_tag_len, &dst_assoc, &dst, + decrypt ?NULL : &auth_tag, + &bounce_buf, &dst_mapped);if (retval) goto exit; - - dst += req->assoclen; }if (!src_mapped)
@@ -843,16 +896,16 @@ static int gcmaes_crypt(struct aead_request *req, unsigned int assoclen,u8 gen_auth_tag[16 ];aesni_gcm_dec_tfm (aes_ctx, dst, src, data_len, iv, - hash_subkey, assoc, assoclen, + hash_subkey, src_assoc, assoclen, gen_auth_tag, auth_tag_len);/* Compare generated tag with passed in tag. */ -if (crypto_memneq (src + data_len, gen_auth_tag, auth_tag_len)) +if (crypto_memneq (auth_tag, gen_auth_tag, auth_tag_len)) retval = -EBADMSG ; }else aesni_gcm_enc_tfm (aes_ctx, dst, src, data_len, iv, - hash_subkey, assoc, assoclen, - dst + data_len, auth_tag_len); + hash_subkey, src_assoc, assoclen, + auth_tag, auth_tag_len);kernel_fpu_end ();
@@ -862,9 +915,13 @@ static int gcmaes_crypt(struct aead_request *req, unsigned int assoclen, 1); exit:if (req->dst != req->src) -put_request_buffer (dst - req->assoclen, dst_mapped); +put_request_buffers (req->dst, dst_mapped, dst_assoc, dst, + decrypt ?NULL : auth_tag, + req->assoclen, data_len, auth_tag_len); -put_request_buffer (assoc, src_mapped); +put_request_buffers (req->src, src_mapped, src_assoc, src, + (decrypt || req->src == req->dst) ? auth_tag :NULL , + req->assoclen, data_len, auth_tag_len);kfree (bounce_buf);return retval;
Signature · 1 line
2.16.0.rc1.238.g530d649a79-goog
Hi Junaid.
Quoted 11 lines · Junaid Shahid
Changes in v2:
- Integrated https://patchwork.kernel.org/patch/10173981Currently, the AESNI gcm(aes) implementation uses zero-copy only when the entire src and dest request buffers, including the AAD, the data and the Auth Tag are contiguous. This series enables the use of zero-copy even if the AAD and/or Auth Tag are in different buffers than the actual data, as long as each of them individually satisfies the zero-copy conditions (i.e. the entire buffer is either in low-mem or within a single high-mem page). Furthermore, it also enables the use of zero-copy even if only one of src and dest satisfies these conditions rather than only when both of them do.
I wonder which special usecase you have in mind that will be improved by your patches.
For IPsec, the crypto layer either gets a contiguous buffer because networking already linearized the packet data, or the packet data and IPsec trailer are scattered over multiple page fragments. So in the first case, it is OK already without your patches. In the second case, your zero-copy conditions are not satisfied.
This looks a bit like a workaround of the real problem, that is that the gcm-aesni implementation does not support SG operations.
Maybe it would be better to investigate in the direction to support SG operations with gcm-aesni instead of trying to find all corner cases where the existing implementation can do zero-copy.
Hi Steffen,
Quoted 4 lines · Steffen Klassert
I wonder which special usecase you have in mind that will be improved by your patches.
This is not related to IPsec. We have an internal use case where the data buffer itself is a single memory page but the authentication tag needs to be in a separate buffer. This patch set allows us to have zero copy in that case.
Quoted 6 lines
Maybe it would be better to investigate in the direction to support SG operations with gcm-aesni instead of trying to find all corner cases where the existing implementation can do zero-copy.
Yes, if gcm-aesni could be implemented to handle arbitrarily fragmented buffers, that would be better, but it would likely be a much bigger and more complicated change. Just handling the case where the AAD/data/authtag are separate contiguous buffers is relatively simpler but still useful.
Thanks,
Junaid
Quoted 11 lines · Junaid Shahid
Quoted 4 lines · Steffen Klassert
I wonder which special usecase you have in mind that will be improved by your patches.
This is not related to IPsec. We have an internal use case where the data buffer itself is a single memory page but the authentication tag needs to be in a separate buffer. This patch set allows us to have zero copy in that case.
We usually don't add code for 'internal use cases' to the mainline kernel. If you need a SG version of gcm-aesni, then please implement the generic version so that everybody can benefit from it.
Think about how this will look like if everybody would add a workaround for its very special use case.