I.MX6 Cryptographic Acceleration

[johnlangworthy]

The I.MX6 is described as offering cryptographic acceleration in the CAAM. However, as far as I can see, there is no documentation avaialble to support this. Does anybody know what the capabilities are, and how we can access them?

Cheers

[kiyoshizen]

Thank you Yuri and everyone else for very useful thread,

I managed to use AF_ALG on my nitrogen6q to make OpenSSL work with CAAM drivers. but the result is not stable, when running the openssl test randomly cause "Internal error: Oops: 17"   anyone got any idea what might be the problem? please help

Example:

# openssl speed -evp md5

Doing md5 for 3s on 16 size blocks: 26876 md5's in 0.17s

Doing md5 for 3s on 64 size blocks: 15716 md5's in 0.22s

Doing md5 for 3s on 256 size blocks: 15474 md5's in 0.14s

Doing md5 for 3s on 1024 size blocks: 13861 md5's in 0.09s

Doing md5 for 3s on 8192 size blocks: 7063 md5's in 0.15s

OpenSSL 1.0.1i 6 Aug 2014

built on: Thu Sep 10 16:02:35 IDT 2015

options:bn(64,32) rc4(ptr,char) des(idx,cisc,16,long) aes(partial) blowfish(ptr)

compiler: /buildroot-2014.08/output/host/usr/bin/arm-buildroot-linux-gnueabi-gcc -fPIC -DOPENSSL_PIC -DZLIB_SHARED -DZLIB -DOPENSSL_THREADS -D_REENTRANT -DDSO_DLFCN -DHAVE_DLM

The 'numbers' are in 1000s of bytes per second processed.

type             16 bytes     64 bytes    256 bytes   1024 bytes   8192 bytes

md5               2529.51k     4571.93k    28295.31k   157707.38k   385733.97k

#

#

# openssl speed -evp md5

Doing md5 for 3s on 16 size block[ 1992.291028] Unable to handle kernel NULL pointer dereference at virtual address 00000e90

s: [ 1992.301706] pgd = dc76c000

[ 1992.304645] [00000e90] *pgd=2ca14831, *pte=00000000, *ppte=00000000

[ 1992.310985] Internal error: Oops: 17 [#17] SMP ARM

[ 1992.315779] Modules linked in: algif_hash algif_skcipher af_alg

[ 1992.321759] CPU: 3 PID: 3003 Comm: openssl Tainted: G      D      3.10.17+ #3

[ 1992.328898] task: dc7ae580 ti: dc5ee000 task.ti: dc5ee000

[ 1992.334310] PC is at memcpy+0x80/0x330

[ 1992.338071] LR is at ahash_update_first+0x7f8/0x8ec

[ 1992.342953] pc : [<c024a340>]    lr : [<c04bbc50>]    psr: 000f0013

[ 1992.342953] sp : dc5efd14  ip : 00000010  fp : dc5efd9c

[ 1992.354431] r10: 000000d0  r9 : dc526680  r8 : 00000000

[ 1992.359658] r7 : 00000000  r6 : dc526400  r5 : 00000010  r4 : dc526600

[ 1992.366186] r3 : c04b8a9c  r2 : fffffff0  r1 : 00000e90  r0 : dc526600

[ 1992.372717] Flags: nzcv  IRQs on  FIQs on  Mode SVC_32  ISA ARM  Segment user

[ 1992.379854] Control: 10c5387d  Table: 2c76c04a  DAC: 00000015

[ 1992.385601] Process openssl (pid: 3003, stack limit = 0xdc5ee238)

[ 1992.391697] Stack: (0xdc5efd14 to 0xdc5f0000)

[ 1992.396057] fd00:                                              00000010 dc526400 00000000

[ 1992.404240] fd20: 00000000 dc526600 dc526600 c04bbc50 c004b90c c0251e44 dc5efd84 dc5efd48

[ 1992.412421] fd40: c00a48c0 c004b908 00000000 00000000 dc526500 00000000 2c64c928 dc5efd88

[ 1992.420603] fd60: dc526600 00000000 dc64c840 c0259fec dc5efd88 00000010 01703ea0 00000010

[ 1992.428786] fd80: 00010000 dc526400 dc5ee000 00000010 dc5efdac dc5efda0 c04b5e94 c04bb464

[ 1992.436968] fda0: dc5efdf4 dc5efdb0 bf00b2b0 c04b5e8c dc5efecc 00000001 dc5efecc dc571000

[ 1992.445149] fdc0: dc526544 dc526580 dc64c800 00000001 00000010 dc5efecc dc7ae580 01703e90

[ 1992.453330] fde0: dc5ee000 00008000 dc5efeb4 dc5efdf8 c050800c bf00b1f8 00000001 00000000

[ 1992.461512] fe00: dc5efe1c 00000010 da169e00 00000000 00000000 dc5efecc dc5efe54 dc5efe28

[ 1992.469695] fe20: c00c8fd4 c05dfc60 000003d0 00000010 dc571000 dc5710a0 00000001 00000000

[ 1992.477875] fe40: 00000000 00000000 00000000 dc7ae580 00000000 00000000 00000000 00000000

[ 1992.486056] fe60: dc5efdf8 00000000 00000000 00000000 00000000 00000000 00000000 00000000

[ 1992.494237] fe80: 00000000 00000000 00000000 00000000 00000000 00000000 da169e00 00000010

[ 1992.502418] fea0: 00000000 00000000 dc5eff8c dc5efeb8 c0509e8c c0507f98 dc5efed4 fffffff7

[ 1992.510600] fec0: 00000000 01703e90 00000010 00000000 00000000 dc5efec4 00000001 00000000

[ 1992.518781] fee0: 00000000 00008000 0000011a c000f104 dc5effa4 dc5eff00 c05099b4 bf000350

[ 1992.526962] ff00: 00000000 00000000 61680026 00006873 00000000 00000000 00000000 00000000

[ 1992.535144] ff20: 0035646d 00000000 00000000 00000000 00000000 00000000 00000000 00000000

[ 1992.543324] ff40: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000

[ 1992.551506] ff60: c00e880c c00e878c 00000010 00000000 bec0b484 00000121 c000f104 00000000

[ 1992.559687] ff80: dc5effa4 dc5eff90 c0509ed0 c0509de0 00000000 00000000 00000000 dc5effa8

[ 1992.567869] ffa0: c000ef80 c0509ebc 00000010 00000000 00000004 01703e90 00000010 00008000

[ 1992.576050] ffc0: 00000010 00000000 bec0b484 00000121 01703e90 00000000 0006842c 0007773c

[ 1992.584231] ffe0: 00000000 bec0b284 b6cc6ac0 b6da1a7c 600f0010 00000004 ebff5a13 e51f15bc

[ 1992.592407] Backtrace:

[ 1992.594879] [<c04bb458>] (ahash_update_first+0x0/0x8ec) from [<c04b5e94>] (ahash_update+0x14/0x18)

[ 1992.603848] [<c04b5e80>] (ahash_update+0x0/0x18) from [<bf00b2b0>] (hash_sendmsg+0xc4/0x1d0 [algif_hash])

[ 1992.613430] [<bf00b1ec>] (hash_sendmsg+0x0/0x1d0 [algif_hash]) from [<c050800c>] (sock_sendmsg+0x80/0xa0)

[ 1992.623007] [<c0507f8c>] (sock_sendmsg+0x0/0xa0) from [<c0509e8c>] (SyS_sendto+0xb8/0xdc)

[ 1992.631184]  r7:00000000 r6:00000000 r5:00000010 r4:da169e00

[ 1992.636906] [<c0509dd4>] (SyS_sendto+0x0/0xdc) from [<c0509ed0>] (SyS_send+0x20/0x28)

[ 1992.644749] [<c0509eb0>] (SyS_send+0x0/0x28) from [<c000ef80>] (ret_fast_syscall+0x0/0x30)

[ 1992.653018] Code: e320f000 e4913004 e4914004 e4915004 (e4916004)

[ 1992.659153] ---[ end trace dd17fad707916035 ]---

Segmentation fault

# [ 1992.989147] rtc-isl12022 0-006f: voltage dropped below 85%, date and time is not reliable.

[Yuri]

Hello,

You may try the recent FSL BSPs, where the Cryptodev interface is supported

via CAAM.

Regards,

Yuri.

[_benjamindubb]

I tried using CAAM with ipsec and it seems like the kernel caam driver crashes. I'm on LTIB kernel 3.0.35-2310 version:

<8004a688>] (dma_sync_sg_for_device+0x18/0x4c) from [<803b290c>] (dma_map_sg_chained+0x48/0xa4)

[<803b290c>] (dma_map_sg_chained+0x48/0xa4) from [<803b4bfc>] (ablkcipher_edesc_alloc.constprop.27+0x3ac/0x3c8)

[<803b4bfc>] (ablkcipher_edesc_alloc.constprop.27+0x3ac/0x3c8) from [<803b4c34>] (ablkcipher_encrypt+0x1c/0x94)

[<803b4c34>] (ablkcipher_encrypt+0x1c/0x94) from [<8020b2bc>] (eseqiv_givencrypt+0x2c8/0x3d0)

[<8020b2bc>] (eseqiv_givencrypt+0x2c8/0x3d0) from [<8021ccc4>] (crypto_authenc_givencrypt+0x64/0x84)

[<8021ccc4>] (crypto_authenc_givencrypt+0x64/0x84) from [<8046daf8>] (esp_output+0x324/0x4bc)

[<8046daf8>] (esp_output+0x324/0x4bc) from [<8047a10c>] (xfrm_output_resume+0x188/0x278)

[<8047a10c>] (xfrm_output_resume+0x188/0x278) from [<80472034>] (xfrm4_output+0x20/0x24)

[<80472034>] (xfrm4_output+0x20/0x24) from [<80434f94>] (ip_local_out+0x28/0x2c)

[<80434f94>] (ip_local_out+0x28/0x2c) from [<8043610c>] (ip_send_skb+0xc/0xd8)

[<8043610c>] (ip_send_skb+0xc/0xd8) from [<80454560>] (udp_send_skb+0xf0/0x3c8)

[<80454560>] (udp_send_skb+0xf0/0x3c8) from [<80456008>] (udp_sendmsg+0x254/0x7f8)

[<80456008>] (udp_sendmsg+0x254/0x7f8) from [<8045d8e8>] (inet_sendmsg+0x94/0xb8)

[<8045d8e8>] (inet_sendmsg+0x94/0xb8) from [<803feb84>] (sock_aio_write+0x114/0x13c)

[<803feb84>] (sock_aio_write+0x114/0x13c) from [<800eff30>] (do_sync_write+0xa4/0xe4)

[<800eff30>] (do_sync_write+0xa4/0xe4) from [<800f06bc>] (vfs_write+0x130/0x138)

[<800f06bc>] (vfs_write+0x130/0x138) from [<800f08a4>] (sys_write+0x40/0x6c)

[<800f08a4>] (sys_write+0x40/0x6c) from [<80040f80>] (ret_fast_syscall+0x0/0x30)

Code: e1a07003 d8bd80f8 e1a04001 e3a05000 (e5940000)

---[ end trace 9693d9778bad1379 ]---

I'm just trying to see what type of improvement to expect from the CAAM driver. I expect the CPU should not be so busy with HW CAAM engine. Any idea how much faster the CAAM engine is compared to plain software implementation?

Thanks,

-Ben

[Yuri]

The patches are available on Freescales public GIT server at the following link:

http://git.freescale.com/git/cgit.cgi/imx/linux-2.6-imx.git/log/?h=imx_3.0.35_4.0.0

Direct link to the patches:

http://git.freescale.com/git/cgit.cgi/imx/linux-2.6-imx.git/commit/?h=imx_3.0.35_4.0.0&id=6068d7a77b...

http://git.freescale.com/git/cgit.cgi/imx/linux-2.6-imx.git/commit/?h=imx_3.0.35_4.0.0&id=b30237c790...

[_benjamindubb]

Freescale has shared with me some patches that address the CAAM driver crash. I can now pump traffic through ipsec tunnel setup with CAAM. These patches will hopefully be included in the next BSP update.

[Yuri]

Linux CAAM performance estimations may be found under the following discussion. 

< https://community.freescale.com/message/313424#313424 >

[_benjamindubb]

I get "Unauthorized" when I click on the link you provided. Do I need to ask my FAE for special access to this "secret" data?

[Yuri]

> Do I need to ask my FAE for special access to this "secret" data?

Yes, one way - to apply to FFAE.
Another way - just create SR (say, for me ) - I think I can provide the performance data under SR. 

[davehaynie]

I've been asking around for similar information. I have yet to meet anyone here from Freescale with crypto knowledge, until now. Original question is SR# 1-1062109201. Also a forum posting here.  Ideally after a benchmark of AES-GCM running on the i.MX6 (which, of course, can't be fully supported by the CAAM on this chip), but any performance metrics would be very useful in our selection process.

[AlbertT]

Hello Yuri,

So we can take advantage of the CAAM using OpenSSL ? Or it is restricted to some functions ?

Jocelyn

[Yuri]

> So we can take advantage of the CAAM using OpenSSL ?

Yes, if we find Open SSL implementation via scatterlist crypto API.

[AlbertT]

I don't fully understand, right now, if I use an OpenSSL function for AES encryption for instance, will it take advantage of CAAM or not ? Or should I code everything with Job Descriptor and so on ?

[Yuri]

If the OpenSSL uses /dev/crypto - we will use software implementation, no CAAM.

[AlbertT]

Oh ok, and is it possible to force OpenSSL to not use /dev/crypto or we have to wait for a new release of OSSL ?

[Yuri]

There is Security Reference Manual for i.MX6 - please apply to local Freescale FAE to get it.
(NDA is required.)