* g10/cipher-aead.c (do_hash): Add faster path for encrypting directly
from input buffer instead of memcpying then encrypting inplace.
--
When encrypting with AES256.OCB on AMD Ryzen 5800X, memcpy shows as
taking largest portion of computation time. Patch adds fast path for
AEAD encryption which largely eliminates need for memcpying when
doing AEAD encryption. AES256.OCB throughput increases from 2.2GB/s
to 4.2GB/s on said system (1.9x faster).
GnuPG-bug-id: T5828
Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>
* g10/cipher-aead.c (do_flush): Use PRIu64.
* g10/decrypt-data.c (aead_underflow): Likewise.
--
Even among LP64 data model machines, uint64_t type may differ;
unsigned long or unsigned long long.
Only portable way is use of PRIu64.
Signed-off-by: NIIBE Yutaka <gniibe@fsij.org>
* g10/cipher-aead.c (do_flush): No cast is correct.
* g10/decrypt-data.c (aead_underflow): No cast needed.
Use "%j" for uint64_t for chunklen.
Signed-off-by: NIIBE Yutaka <gniibe@fsij.org>
* g10/cipher-aead.c (write_header): Do not call set_nonce_and_ad.
(write_final_chunk): Do not increase chunkindex.
(do_flush): Call set_nonce_and_ad immediately before the first
encryption of a chunk. Bump up the chunkindex after writing the tag.
(do_free): Do not insert a zero length last chunk.
* g10/decrypt-data.c (aead_underflow): Fix the corresponding bug.
--
This fixes a bug in writing a zero length last chunk right before the
final chunk (which has by design a zero length). We also need to
adjust the decryption part because that assumed this zero length last
chunk.
Note that we use the term "last chunk" for the chunk which directly
precedes the "final chunk" which ends the entire encryption.
GnuPG-bug-id: 3774
Signed-off-by: Werner Koch <wk@gnupg.org>
* g10/cipher-aead.c (set_nonce, set_additional_data): Merge into ...
(set_nonce_and_ad): new function.
(write_auth_tag): Print error message here.
(do_flush): Rename var newchunk to finalize.
--
There is no need to have separate functions here. We should also
print a error message for writing the final tag.
Signed-off-by: Werner Koch <wk@gnupg.org>
* g10/cipher-aead.c (write_final_chunk): Do not bump up the chunk
index if the previous chunk was empty.
* g10/decrypt-data.c (aead_underflow): Likewise. Also handle a other
corner cases. Add more debug output.
--
GnuPG-bug-id: 3774
This fixes the reported case when the encrypted data is a multiple of
the chunk size. Then the chunk index for the final chunk was wrongly
incremented by 2. The actual fix makes use of the fact that the
current dfx->CHUNKLEN is 0 in this case. There is also some other
reorganizing to help with debugging. The thing seems to work now but
the code is not very clean - should be reworked. Creating test files
can be done with this script:
--8<---------------cut here---------------start------------->8---
csize=6
for len in 0 55 56 57; do
awk </dev/null -v i=$len 'BEGIN{while(i){i--;printf"~"}}' \
| gpg --no-options -v --rfc4880bis --batch --passphrase "abc" \
--s2k-count 1025 --s2k-digest-algo sha256 -z0 \
--force-aead --aead-algo eax --cipher aes -a \
--chunk-size $csize -c >symenc-aead-eax-c$csize-$len.asc
done
--8<---------------cut here---------------end--------------->8---
A LEN of 56 triggered the bug which can be seen by looking at the
"authdata:" line in the --debug=crypt,filter output.
Signed-off-by: Werner Koch <wk@gnupg.org>
* g10/cipher-aead.c (do_flush): Init ERR. Fix remaining chunklen
computation.
(do_free): Add dummy encryption. Close the cipher handle.
* g10/decrypt-data.c (aead_underflow): Rewrite.
--
Until we have integrated test into the test suite extensive tests can
also be done with a script like this:
--8<---------------cut here---------------start------------->8---
#!/bin/sh
set -e
GPG="../g10/gpg --rfc4880bis --pinentry-mode=loopback"
GPG="$GPG --passphrase abc --batch"
MKTDATA="$HOME/b/gnupg-2.0/tools/mk-tdata"
for chunksize in 6 7 12 13 14 30; do
for count in $(seq 1 200) $(seq 8100 8200) \
$(seq 16350 16400) $(seq 20000 20100); do
if [ ! -f "testfile-$count" ]; then
$MKTDATA $count >"testfile-$count"
fi
echo "testing chunk size 2^$chunksize with $count bytes"
$GPG --force-aead --aead-algo ocb --s2k-mode 0 --cipher AES -v -z 0 \
-c --chunk-size $chunksize \
<"testfile-$count" >"testfile-$count.gpg" 2>/dev/null
$GPG -vd <"testfile-$count.gpg" >"testfile-$count.out" 2>/dev/null
if ! cmp "testfile-$count" "testfile-$count.out"; then
echo "FAILED comparing count $count" >&2
exit 1
fi
done
done
echo All good
--8<---------------cut here---------------end--------------->8---
Signed-off-by: Werner Koch <wk@gnupg.org>
* g10/pkclist.c (select_aead_from_pklist): Return the AEAD_algo.
* g10/encrypt.c (use_aead): Return the AEAD algo.
(encrypt_simple): Adjust for this change.
(encrypt_crypt): Ditto.
(encrypt_filter): Ditto.
* g10/sign.c (sign_symencrypt_file): Ditto.
* g10/misc.c (MY_GCRY_CIPHER_MODE_EAX): New.
(openpgp_aead_algo_info): New.
* g10/cipher-aead.c (MY_GCRY_CIPHER_MODE_EAX): Remove.
(write_header): Use new fucntion.
* g10/decrypt-data.c (MY_GCRY_CIPHER_MODE_EAX): Remove.
(decrypt_data): Use new function. Also allow for chunkbytes other
than 10.
--
Note that other chunk bytes than 10 and in particular 0 (64 byte
chunks) have not yet been tested.
Signed-off-by: Werner Koch <wk@gnupg.org>
* g10/cipher-aead.c (MY_GCRY_CIPHER_MODE_EAX): New.
(write_header): Use it.
* g10/decrypt-data.c (MY_GCRY_CIPHER_MODE_EAX): New.
(decrypt_data): Use it.
* g10/misc.c (openpgp_aead_test_algo): Allow EAX.
--
This allows the use of EAX when the latest Libgcrypt master is used.
Signed-off-by: Werner Koch <wk@gnupg.org>
* common/openpgpdefs.h (PKT_ENCRYPTED_AEAD): New const.
* g10/dek.h (DEK): Increase size of use_aead to 4 bits.
* g10/filter.h (cipher_filter_context_t): Add new fields for AEAD.
* g10/packet.h (PKT_encrypted): Add fields aead_algo, cipher_algo, and
chunkbyte.
* g10/build-packet.c (do_encrypted_aead): New.
(build_packet): Call it.
* g10/parse-packet.c (dump_sig_subpkt): Handle SIGSUBPKT_PREF_AEAD.
(parse_one_sig_subpkt, can_handle_critical): Ditto.
(parse_encrypted): Clear new PKT_ENCRYPTED fields.
(parse_encrypted_aead): New.
(parse): Call it.
* g10/gpg.c (main): Take care of --rfc4880bis option when checking
compliance.
* g10/cipher-aead.c: Replace the stub by real code.
* g10/decrypt-data.c (decode_filter_ctx_t): Add fields for use with
AEAD.
(aead_set_nonce): New.
(aead_set_ad): New.
(decrypt_data): Support AEAD.
(aead_underflow): New.
(aead_decode_filter): New.
* g10/encrypt.c (use_aead): Make that new fucntion work.
(encrypt_simple): Use default_aead_algo() instead of EAX.
* g10/mainproc.c (proc_encrypted): Support AEAD.
(do_proc_packets): Support PKT_ENCRYPTED_AEAD.
--
This code has seen only a very few manual tests. Encrypting always
uses a 64k chunks and decryption has not been tested with larger
chunks. Those small chunks make debugging much faster.
Tests can be done using:
gpg --rfc4880bis --pinentry-mode=loopback --passphrase abc \
--force-aead --aead-algo ocb --s2k-mode 0 --cipher AES \
-v -z 0 --status-fd 2 -c <INFILE >OUTFILE
and
gpg --rfc4880bis --pinentry-mode=loopback --passphrase=abc \
--status-fd 2 -v -d <INFILE >OUTFILE
Signed-off-by: Werner Koch <wk@gnupg.org>
* g10/cipher.c (cipher_filter): Rename to cipher_filter_cfb.
* g10/cipher-aead.c: New. Right now only with a stub function.
* g10/Makefile.am (gpg_sources): Add file.
* g10/encrypt.c (encrypt_simple): Push either cipher_filter_cfb or
cipher_filter_aead.
(encrypt_crypt): Ditto.
(encrypt_filter): Ditto.
* g10/sign.c (sign_symencrypt_file): Ditto.
Signed-off-by: Werner Koch <wk@gnupg.org>
