Improve performance of re-used symmetric one shots #93124

vcsjones · 2023-10-06T14:22:56Z

Today, when you use the symmetric one-shots to decrypt or decrypt, it creates a new lite cipher, uses it, then clean it up every time.

(Context: "lite cipher" is the internal facade over the platform's native encryption components)

Creating a lite cipher is expensive from various things like creating handles, validating keys and inputs, and simply the natural of the algorithm itself from activities like key expansion.

This change allows caching the lite cipher for the duration of the key, so long as the key doesn't change. If the key changes, then the lite cipher is invalidated.

The scenario this improves is this:

using Aes aes = aes.Create();
aes.Key = GetTheKey();
byte[] iv1 = RandomNumberGenerator.GetBytes(16);
byte[] iv2 = RandomNumberGenerator.GetBytes(16);
byte[] ciphertext1 = aes.EncryptCbc(plaintext1, iv1);
byte[] ciphertext2 = aes.EncryptCbc(plaintext2, iv2);
// N or more calls to EncryptCbc / DecryptCbc, etc.

Because the one-shots were previously more-or-less fully contained, they were thread safe - such that calling the one shots from a single instance of SymmetricAlgorithm would work and produce correct results. This is trait I would like to keep and I am not comfortable breaking. Because of this then, the caching uses interlocked exchanges of handles. Since we are only caching one instance at a given time, using an instance from multiple threads may not benefit from the cache. I think this is reasonable to keep the caching mechanism simple, and there is an easy developer workaround: pool an instance of the SymmetricAlgorithm.

This does mean though that there is a performance penalty, albeit a small one, for when the one-shot is used exactly once. The extra bookkeeping for the caching is wasted. However, it is far outweighed if you use it more than once.

Full data is below, but some examples:

Aes.EncryptCbc with a single call when from 579.0ns to 606.2ns.
Aes.EncryptCbc with three calls when from 1,395.5ns to 757.2ns.

TripleDES.EncryptCbc with a single call when from 11,311.6ns to 11,294.2ns (this "improvement" is within error)
TripleDES.EncryptCbc with three calls when from 33,826.3ns to 13,004.6ns.

The performance improvements can be substantial when it's used multiple times. For single uses, the performance loss is two Interlocked.Exchange calls.

This pull request addresses ECB and CBC only. Absent is CFB. CFB, because of feedback size, could be addressed differently, which is why I am leaving it out of this pull request. ECB and CBC are the far more common modes as well.

Benchmarks are in this Gist so they don't wrap funny in this comment. https://gist.github.com/vcsjones/b79889c4c6196bec50d783ff6e9aba0d

CBC benchmark source
ECB benchmark source

ghost · 2023-10-06T14:23:05Z

Tagging subscribers to this area: @dotnet/area-system-security, @bartonjs, @vcsjones
See info in area-owners.md if you want to be subscribed.

Issue Details

Today, when you use the symmetric one-shots to decrypt or decrypt, it creates a new lite cipher, uses it, then clean it up every time.

(Context: "lite cipher" is the internal facade over the platform's native encryption components)

Creating a lite cipher is expensive from various things like creating handles, validating keys and inputs, and simply the natural of the algorithm itself from activities like key expansion.

This change allows caching the lite cipher for the duration of the key, so long as the key doesn't change. If the key changes, then the lite cipher is invalidated.

The scenario this improves is this:

using Aes aes = aes.Create();
aes.Key = GetTheKey();
byte[] iv1 = RandomNumberGenerator.GetBytes(16);
byte[] iv2 = RandomNumberGenerator.GetBytes(16);
byte[] ciphertext1 = aes.EncryptCbc(plaintext1, iv1);
byte[] ciphertext2 = aes.EncryptCbc(plaintext2, iv1);
// N or more calls to EncryptCbc / DecryptCbc, etc.

Because the one-shots were previously more-or-less fully contained, they were thread safe - such that calling the one shots from a single instance of SymmetricAlgorithm would work and produce correct results. This is trait I would like to keep and I am not comfortable breaking. Because of this then, the caching uses interlocked exchanges of handles. Since we are only caching one instance at a given time, using an instance from multiple threads may not benefit from the cache. I think this is reasonable to keep the caching mechanism simple, and there is an easy developer workaround: pool an instance of the SymmetricAlgorithm.

This does mean though that there is a performance penalty, albeit a small one, for when the one-shot is used exactly once. The extra bookkeeping for the caching is wasted. However, it is far outweighed if you use it more than once.

Full data is below, but some examples:

Aes.EncryptCbc with a single call when from 579.0ns to 606.2ns.
Aes.EncryptCbc with three calls when from 1,395.5ns to 757.2ns.

TripleDES.EncryptCbc with a single call when from 11,311.6ns to 11,294.2ns (this "improvement" is within error)
TripleDES.EncryptCbc with three calls when from 33,826.3ns to 13,004.6ns.

The performance improvements can be substantial when it's used multiple times. For single uses, the performance loss is two Interlocked.Exchange calls.

This pull request addresses ECB and CBC only. Absent is CFB. CFB, because of feedback size, could be addressed differently, which is why I am leaving it out of this pull request. ECB and CBC are the far more common modes as well.

CBC Benchmarks

Method	Job	Toolchain	DataSize	NumberOfOperations	Mode	Algorithm	Mean	Error	StdDev	Ratio	Gen0	Allocated	Alloc Ratio
Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	Aes	606.2 ns	4.99 ns	4.67 ns	1.05	0.0658	416 B	1.08
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	1	PKCS7	Aes	579.0 ns	2.76 ns	2.58 ns	1.00	0.0610	384 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	Aes	605.5 ns	3.22 ns	3.01 ns	1.01	0.0658	416 B	1.08
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	1	PKCS7	Aes	596.8 ns	2.64 ns	2.47 ns	1.00	0.0610	384 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	DES	4,121.0 ns	6.31 ns	5.27 ns	1.01	0.0763	496 B	1.07
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	1	PKCS7	DES	4,086.9 ns	4.03 ns	3.57 ns	1.00	0.0687	464 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	DES	3,960.6 ns	18.83 ns	17.62 ns	0.97	0.0763	496 B	1.07
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	1	PKCS7	DES	4,094.3 ns	5.70 ns	5.33 ns	1.00	0.0687	464 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	RC2	1,992.9 ns	4.16 ns	3.69 ns	1.02	0.0725	456 B	1.08
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	1	PKCS7	RC2	1,956.9 ns	5.76 ns	5.39 ns	1.00	0.0648	424 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	RC2	1,827.2 ns	3.67 ns	3.25 ns	1.01	0.0725	456 B	1.08
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	1	PKCS7	RC2	1,813.8 ns	3.15 ns	2.80 ns	1.00	0.0668	424 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	TripleDES	11,294.2 ns	33.64 ns	31.46 ns	1.00	0.0763	496 B	1.07
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	1	PKCS7	TripleDES	11,311.6 ns	23.85 ns	21.15 ns	1.00	0.0610	464 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	TripleDES	11,463.8 ns	48.48 ns	45.35 ns	1.03	0.0763	496 B	1.07
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	1	PKCS7	TripleDES	11,145.7 ns	27.60 ns	25.82 ns	1.00	0.0610	464 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	Aes	678.0 ns	2.62 ns	2.45 ns	0.68	0.0658	416 B	0.81
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	2	PKCS7	Aes	999.1 ns	4.37 ns	4.09 ns	1.00	0.0801	512 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	Aes	683.3 ns	3.93 ns	3.67 ns	0.67	0.0658	416 B	0.81
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	2	PKCS7	Aes	1,026.2 ns	5.32 ns	4.97 ns	1.00	0.0801	512 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	DES	4,504.2 ns	5.09 ns	4.25 ns	0.57	0.0763	496 B	0.78
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	2	PKCS7	DES	7,907.5 ns	14.60 ns	12.19 ns	1.00	0.0916	632 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	DES	4,317.5 ns	11.72 ns	10.39 ns	0.54	0.0763	496 B	0.78
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	2	PKCS7	DES	8,013.6 ns	14.96 ns	12.49 ns	1.00	0.0916	632 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	RC2	2,484.2 ns	5.44 ns	5.09 ns	0.66	0.0839	528 B	0.87
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	2	PKCS7	RC2	3,753.9 ns	3.20 ns	2.84 ns	1.00	0.0954	608 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	RC2	2,173.9 ns	6.72 ns	6.28 ns	0.63	0.0839	528 B	0.87
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	2	PKCS7	RC2	3,461.0 ns	8.15 ns	7.22 ns	1.00	0.0954	608 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	TripleDES	12,109.4 ns	35.04 ns	32.78 ns	0.54	0.0763	496 B	0.78
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	2	PKCS7	TripleDES	22,458.3 ns	84.69 ns	79.22 ns	1.00	0.0916	632 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	TripleDES	11,931.6 ns	43.27 ns	38.36 ns	0.54	0.0763	496 B	0.78
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	2	PKCS7	TripleDES	22,154.8 ns	114.97 ns	101.92 ns	1.00	0.0916	632 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	Aes	757.2 ns	2.76 ns	2.58 ns	0.54	0.0658	416 B	0.65
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	3	PKCS7	Aes	1,395.5 ns	12.07 ns	11.29 ns	1.00	0.1011	640 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	Aes	752.4 ns	4.49 ns	3.98 ns	0.51	0.0658	416 B	0.65
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	3	PKCS7	Aes	1,471.2 ns	3.67 ns	3.43 ns	1.00	0.1011	640 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	DES	4,788.3 ns	13.78 ns	12.89 ns	0.40	0.0763	496 B	0.62
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	3	PKCS7	DES	11,966.2 ns	11.29 ns	9.43 ns	1.00	0.1221	800 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	DES	4,657.8 ns	6.53 ns	5.78 ns	0.40	0.0763	496 B	0.62
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	3	PKCS7	DES	11,743.8 ns	30.78 ns	27.28 ns	1.00	0.1221	800 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	RC2	2,953.6 ns	7.39 ns	6.91 ns	0.53	0.0954	600 B	0.76
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	3	PKCS7	RC2	5,536.0 ns	12.08 ns	11.30 ns	1.00	0.1221	792 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	RC2	2,496.5 ns	3.43 ns	3.04 ns	0.49	0.0954	600 B	0.76
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	3	PKCS7	RC2	5,096.1 ns	19.47 ns	18.21 ns	1.00	0.1221	792 B	1.00

Encrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	TripleDES	13,004.6 ns	50.16 ns	39.16 ns	0.38	0.0763	496 B	0.62
Encrypt_Cbc_ToSpan	Job-QKOWKL	main	16	3	PKCS7	TripleDES	33,826.3 ns	100.39 ns	88.99 ns	1.00	0.1221	800 B	1.00

Decrypt_Cbc_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	TripleDES	12,696.4 ns	61.61 ns	54.62 ns	0.38	0.0763	496 B	0.62
Decrypt_Cbc_ToSpan	Job-QKOWKL	main	16	3	PKCS7	TripleDES	33,138.3 ns	115.95 ns	108.46 ns	1.00	0.1221	800 B	1.00

ECB Benchmarks

Method	Job	Toolchain	DataSize	NumberOfOperations	Mode	Algorithm	Mean	Error	StdDev	Ratio	Gen0	Allocated	Alloc Ratio
Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	Aes	572.5 ns	2.21 ns	1.96 ns	1.02	0.0658	416 B	1.08
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	1	PKCS7	Aes	561.4 ns	2.82 ns	2.64 ns	1.00	0.0610	384 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	Aes	586.5 ns	4.27 ns	4.00 ns	1.02	0.0658	416 B	1.08
DecryptEcb_ToSpan	Job-QKOWKL	main	16	1	PKCS7	Aes	572.4 ns	3.30 ns	3.08 ns	1.00	0.0610	384 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	DES	3,987.4 ns	4.04 ns	3.78 ns	1.03	0.0763	496 B	1.07
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	1	PKCS7	DES	3,858.4 ns	5.13 ns	4.55 ns	1.00	0.0687	464 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	DES	3,936.6 ns	9.21 ns	8.17 ns	1.01	0.0763	496 B	1.07
DecryptEcb_ToSpan	Job-QKOWKL	main	16	1	PKCS7	DES	3,886.0 ns	9.01 ns	7.52 ns	1.00	0.0687	464 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	RC2	1,831.0 ns	3.45 ns	2.88 ns	1.01	0.0725	456 B	1.08
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	1	PKCS7	RC2	1,813.0 ns	3.39 ns	2.83 ns	1.00	0.0668	424 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	RC2	1,708.4 ns	3.00 ns	2.81 ns	1.01	0.0725	456 B	1.08
DecryptEcb_ToSpan	Job-QKOWKL	main	16	1	PKCS7	RC2	1,693.1 ns	5.26 ns	4.92 ns	1.00	0.0668	424 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	TripleDES	11,218.4 ns	34.75 ns	32.51 ns	1.02	0.0763	496 B	1.07
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	1	PKCS7	TripleDES	11,007.1 ns	44.17 ns	41.31 ns	1.00	0.0610	464 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	1	PKCS7	TripleDES	10,853.4 ns	50.29 ns	47.04 ns	0.96	0.0763	496 B	1.07
DecryptEcb_ToSpan	Job-QKOWKL	main	16	1	PKCS7	TripleDES	11,289.6 ns	23.50 ns	19.62 ns	1.00	0.0610	464 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	Aes	660.3 ns	5.69 ns	5.32 ns	0.69	0.0658	416 B	0.81
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	2	PKCS7	Aes	960.1 ns	6.02 ns	5.64 ns	1.00	0.0811	512 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	Aes	665.8 ns	3.13 ns	2.45 ns	0.67	0.0658	416 B	0.81
DecryptEcb_ToSpan	Job-QKOWKL	main	16	2	PKCS7	Aes	986.4 ns	4.80 ns	4.49 ns	1.00	0.0801	512 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	DES	4,334.5 ns	6.67 ns	5.91 ns	0.56	0.0763	496 B	0.78
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	2	PKCS7	DES	7,709.4 ns	5.02 ns	3.92 ns	1.00	0.0916	632 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	DES	4,196.8 ns	3.90 ns	3.26 ns	0.54	0.0763	496 B	0.78
DecryptEcb_ToSpan	Job-QKOWKL	main	16	2	PKCS7	DES	7,776.3 ns	13.75 ns	12.19 ns	1.00	0.0916	632 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	RC2	2,273.5 ns	3.72 ns	3.48 ns	0.66	0.0839	528 B	0.87
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	2	PKCS7	RC2	3,461.7 ns	7.17 ns	6.71 ns	1.00	0.0954	608 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	RC2	2,037.3 ns	3.80 ns	3.37 ns	0.63	0.0839	528 B	0.87
DecryptEcb_ToSpan	Job-QKOWKL	main	16	2	PKCS7	RC2	3,236.9 ns	5.67 ns	5.31 ns	1.00	0.0954	608 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	TripleDES	12,035.6 ns	56.44 ns	52.80 ns	0.54	0.0763	496 B	0.78
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	2	PKCS7	TripleDES	22,137.7 ns	72.31 ns	67.64 ns	1.00	0.0916	632 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	2	PKCS7	TripleDES	11,652.6 ns	29.61 ns	26.25 ns	0.53	0.0763	496 B	0.78
DecryptEcb_ToSpan	Job-QKOWKL	main	16	2	PKCS7	TripleDES	22,164.3 ns	57.66 ns	53.93 ns	1.00	0.0916	632 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	Aes	708.8 ns	2.30 ns	2.15 ns	0.53	0.0658	416 B	0.65
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	3	PKCS7	Aes	1,344.8 ns	16.80 ns	15.72 ns	1.00	0.1011	640 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	Aes	725.8 ns	4.55 ns	4.26 ns	0.52	0.0658	416 B	0.65
DecryptEcb_ToSpan	Job-QKOWKL	main	16	3	PKCS7	Aes	1,387.8 ns	13.76 ns	12.87 ns	1.00	0.1011	640 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	DES	4,443.4 ns	5.53 ns	4.31 ns	0.39	0.0763	496 B	0.62
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	3	PKCS7	DES	11,420.1 ns	16.38 ns	15.32 ns	1.00	0.1221	800 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	DES	4,508.5 ns	3.28 ns	2.91 ns	0.41	0.0763	496 B	0.62
DecryptEcb_ToSpan	Job-QKOWKL	main	16	3	PKCS7	DES	11,072.3 ns	16.52 ns	15.45 ns	1.00	0.1221	800 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	RC2	2,687.7 ns	6.07 ns	5.68 ns	0.53	0.0954	600 B	0.76
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	3	PKCS7	RC2	5,088.6 ns	10.13 ns	8.98 ns	1.00	0.1221	792 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	RC2	2,330.7 ns	4.45 ns	3.95 ns	0.49	0.0954	600 B	0.76
DecryptEcb_ToSpan	Job-QKOWKL	main	16	3	PKCS7	RC2	4,758.3 ns	12.44 ns	11.03 ns	1.00	0.1221	792 B	1.00

Encrypt_Ecb_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	TripleDES	12,663.7 ns	35.35 ns	33.06 ns	0.38	0.0763	496 B	0.62
Encrypt_Ecb_ToSpan	Job-QKOWKL	main	16	3	PKCS7	TripleDES	33,229.1 ns	215.92 ns	191.41 ns	1.00	0.1221	800 B	1.00

DecryptEcb_ToSpan	Job-VPVYBA	branch	16	3	PKCS7	TripleDES	12,796.3 ns	70.03 ns	65.50 ns	0.39	0.0763	496 B	0.62
DecryptEcb_ToSpan	Job-QKOWKL	main	16	3	PKCS7	TripleDES	32,748.0 ns	93.27 ns	87.24 ns	1.00	0.1221	800 B	1.00

CBC benchmark source
ECB benchmark source

Author:	vcsjones
Assignees:	-
Labels:	`area-System.Security`, `tenet-performance`
Milestone:	-

...libraries/System.Security.Cryptography/src/System/Security/Cryptography/AesImplementation.cs

bartonjs · 2023-10-06T20:51:30Z

...libraries/System.Security.Cryptography/src/System/Security/Cryptography/AesImplementation.cs

+                base.Key = value;
+                InvalidateOneShotAlgorithms();


This instance has been used, so there's a cached lite-cipher for the mode we care about.

Thread A calls into set_Key, hasn't yet made it to updating the field

Thread B starts encrypting (or decrypting, whatever), something long. Gigabyteish. Still has k0

Thread A makes it out to InvalidateOneShotAlgorithms. get_Key will now return k1

Thread B finishes. The field is null, so it puts back the instance it used at k0.

As long as parallel ops never happen again, the cache stays at k0 from one-borrow one-return.

Right?

Ah, right, yep.

vcsjones · 2023-10-06T21:01:08Z

Okay, going to close this for now, as the threading part needs a bit of work. Will re-open when that's in better shape, with new benchmarks.

vcsjones added 2 commits October 5, 2023 17:50

Cache CBC and ECB for AES

0db9f41

RC2, 3DES, DES

319135d

vcsjones added area-System.Security tenet-performance Performance related issue labels Oct 6, 2023

ghost assigned vcsjones Oct 6, 2023

Skip tests on Windows 7 because of CFB

53eaf7d

vcsjones marked this pull request as ready for review October 6, 2023 19:19

bartonjs reviewed Oct 6, 2023

View reviewed changes

...libraries/System.Security.Cryptography/src/System/Security/Cryptography/AesImplementation.cs Outdated Show resolved Hide resolved

Code review feedback

30cde47

bartonjs reviewed Oct 6, 2023

View reviewed changes

...libraries/System.Security.Cryptography/src/System/Security/Cryptography/AesImplementation.cs Show resolved Hide resolved

vcsjones marked this pull request as draft October 6, 2023 20:51

bartonjs reviewed Oct 6, 2023

View reviewed changes

vcsjones closed this Oct 6, 2023

This was referenced Oct 7, 2023

System.Net.Sockets.Tests failing in CI runs #90521

Closed

MSBuild crashing in the build #92290

Open

illink - Task host node exited prematurely #92951

Closed

ghost locked as resolved and limited conversation to collaborators Nov 6, 2023

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Improve performance of re-used symmetric one shots #93124

Improve performance of re-used symmetric one shots #93124

vcsjones commented Oct 6, 2023 •

edited

Loading

ghost commented Oct 6, 2023

bartonjs Oct 6, 2023

vcsjones Oct 6, 2023

vcsjones commented Oct 6, 2023

Improve performance of re-used symmetric one shots #93124

Improve performance of re-used symmetric one shots #93124

Conversation

vcsjones commented Oct 6, 2023 • edited Loading

ghost commented Oct 6, 2023

bartonjs Oct 6, 2023

Choose a reason for hiding this comment

vcsjones Oct 6, 2023

Choose a reason for hiding this comment

vcsjones commented Oct 6, 2023

vcsjones commented Oct 6, 2023 •

edited

Loading