===
dm-crypt optimization for larger block sizes
===
Currently, the iv generation algorithms are implemented in dm-crypt.c. The goal
is to move these algorithms from the dm layer to the kernel crypto layer by
implementing them as template ciphers so they can be used in relation with
algorithms like aes, and with multiple modes like cbc, ecb etc. As part of this
patchset, the iv-generation code is moved from the dm layer to the crypto layer
and adapt the dm-layer to send a whole 'bio' (as defined in the block layer)
at a time. Each bio contains the in memory representation of physically
contiguous disk blocks. Since the bio itself may not be contiguous in main
memory, the dm layer sets up a chained scatterlist of these blocks split into
physically contiguous segments in memory so that DMA can be performed.
One challenge in doing so is that the IVs are generated based on a 512-byte
sector number. This infact limits the block sizes to 512 bytes. But this should
not be a problem if a hardware with iv generation support is used. The geniv
itself splits the segments into sectors so it could choose the IV based on
sector number. But it could be modelled in hardware effectively by not
splitting up the segments in the bio.
Another challenge faced is that dm-crypt has an option to use multiple keys.
The key selection is done based on the sector number. If the whole bio is
encrypted / decrypted with the same key, the encrypted volumes will not be
compatible with the original dm-crypt [without the changes]. So, the key
selection code is moved to crypto layer so the neighboring sectors are
encrypted with a different key.
The dm layer allocates space for iv. The hardware drivers can choose to make
use of this space to generate their IVs sequentially or allocate it on their
own. This can be moved to crypto layer too. Postponing this decision until
the requirement to integrate milan's changes are clear.
Interface to the crypto layer - include/crypto/geniv.h
More information on test procedure can be found in v1.
---
Peformance comparison [Tests on 1 GiB Volume] on db410c
Test script:
https://github.com/binoyjayan/utilities/blob/master/utils/dmtest
dmtest -d -o out.log -s 1024 -r 384 -f 768
---
This includes tests done with dd, fio and bonnie++ with the original dm-crypt
and the proposed solution with algorithm 'essiv(cbc(aes-arm))' implemented
in software. The hardware is yet to be evaluated. These tests are to make sure
there is no drastic performance degradation on systems without hw crypto.
Tests with dd [direct i/o]
Sequential read-0.134 %
Sequential Write +0.091 %
Tests with fio [Aggregate bandwidth - aggrb]
Random Read+0.358 %
Random Write +0.010 %
Tests with bonnie++ [768 MiB File, 384 MiB Ram]
after mounting dm-crypt target as ext4
Sequential o/p [per-char] -2.876 %
Sequential o/p [per-blk] +0.992 %
Sequential o/p [re-write] +4.465 %
Sequential i/p [per-char] -0.453 %
Sequential i/p [per-blk] -0.740 %
Sequential create -0.255 %
Sequential delete +0.042 %
Random create -0.007 %
Random delete +0.454 %
NB: The '+' sign shows improvement and '-' shows degradation.
The tests were performed with minimal cpu load.
Tests with higher cpu load to be done
Revisions:
--
v1: https://patchwork.kernel.org/patch/9439175
v2: https://patchwork.kernel.org/patch/9471923
v3: https://lkml.org/lkml/2017/1/18/170
v4: https://patchwork.kernel.org/patch/9559665
v4 --> v5
--
1. Fix for the multiple instance issue in /proc/crypto
2. Few cosmetic changes including struct alignment
3. Simplified 'struct geniv_req_info'
v3 --> v4
--
Fix for the bug reported by Gilad Ben-Yossef.
The element '__ctx' in 'struct skcipher_request req' overflowed into the
element 'struct scatterlist src' which immediately follows 'req' in
'struct geniv_subreq' and corrupted src.
v2 --> v3
--
1. Moved iv algorithms in dm-crypt.c for control
2. Key management code moved from dm layer to cryto layer
so that cipher instance selection can be made depending on key_index
3. The revision v2 had scatterlist nodes created for every sector in the bio.
It is modified to create only once scatterlist node to reduce memory
foot print. Synchronous requests are processed sequentially. Asynchronous
requests are processed in parallel and is freed in the async callback.
4. Changed allocation for sub-requests using mempool
v1 --> v2
--
1. dm-crypt changes to process larger block sizes (one segment in a bio)
2. Incorporated changes w.r.t. comments from Herbert.
Binoy Jayan (1):
crypto: Add IV generation alg
