For more than 16 S/G entries, driver currently corrupts memory
on ARMv8, see below KASAN log.
Note: this does not reproduce on PowerPC due to different (smaller)
cache line size - 64 bytes on PPC vs. 128 bytes on ARMv8.
One such use case is one of the cbc(aes) test vectors - with 8 S/G
entries and src != dst. Driver needs 1 (IV) + 2 x 8 = 17 entries,
which goes over the 16 S/G entries limit:
(CAAM_QI_MEMCACHE_SIZE - offsetof(struct ablkcipher_edesc, sgt)) /
sizeof(struct qm_sg_entry) = 256 / 16 = 16 S/Gs
Fix this by:
-increasing object size in caamqicache pool from 512 to 768; this means
the maximum number of S/G entries grows from (at least) 16 to 32
(again, for ARMv8 case of 128-byte cache line)
-add checks in the driver to fail gracefully (ENOMEM) in case the 32 S/G
entries limit is exceeded
==================================================================
BUG: KASAN: slab-out-of-bounds in ablkcipher_edesc_alloc+0x4ec/0xf60
Write of size 1 at addr ffff800021cb6003 by task cryptomgr_test/1394
CPU: 3 PID: 1394 Comm: cryptomgr_test Not tainted 4.12.0-rc7-next-20170703-00023-g72badbcc1ea7-dirty #26
Hardware name: LS1046A RDB Board (DT)
Call trace:
[<ffff20000808ac6c>] dump_backtrace+0x0/0x290
[<ffff20000808b014>] show_stack+0x14/0x1c
[<ffff200008d62c00>] dump_stack+0xa4/0xc8
[<ffff200008264e40>] print_address_description+0x110/0x26c
[<ffff200008265224>] kasan_report+0x1d0/0x2fc
[<ffff2000082637b8>] __asan_store1+0x4c/0x54
[<ffff200008b4884c>] ablkcipher_edesc_alloc+0x4ec/0xf60
[<ffff200008b49304>] ablkcipher_encrypt+0x44/0xcc
[<ffff20000848a61c>] skcipher_encrypt_ablkcipher+0x120/0x138
[<ffff200008495014>] __test_skcipher+0xaec/0xe30
[<ffff200008497088>] test_skcipher+0x6c/0xd8
[<ffff200008497154>] alg_test_skcipher+0x60/0xe4
[<ffff2000084974c4>] alg_test.part.13+0x130/0x304
[<ffff2000084976d4>] alg_test+0x3c/0x68
[<ffff2000084938ac>] cryptomgr_test+0x54/0x5c
[<ffff20000810276c>] kthread+0x188/0x1c8
[<ffff2000080836c0>] ret_from_fork+0x10/0x50
Allocated by task 1394:
save_stack_trace_tsk+0x0/0x1ac
save_stack_trace+0x18/0x20
kasan_kmalloc.part.5+0x48/0x110
kasan_kmalloc+0x84/0xa0
kasan_slab_alloc+0x14/0x1c
kmem_cache_alloc+0x124/0x1e8
qi_cache_alloc+0x28/0x58
ablkcipher_edesc_alloc+0x244/0xf60
ablkcipher_encrypt+0x44/0xcc
skcipher_encrypt_ablkcipher+0x120/0x138
__test_skcipher+0xaec/0xe30
test_skcipher+0x6c/0xd8
alg_test_skcipher+0x60/0xe4
alg_test.part.13+0x130/0x304
alg_test+0x3c/0x68
cryptomgr_test+0x54/0x5c
kthread+0x188/0x1c8
ret_from_fork+0x10/0x50
Freed by task 0:
(stack is not available)
The buggy address belongs to the object at ffff800021cb5e00
which belongs to the cache caamqicache of size 512
The buggy address is located 3 bytes to the right of
512-byte region [ffff800021cb5e00, ffff800021cb6000)
The buggy address belongs to the page:
page:ffff7e0000872d00 count:1 mapcount:0 mapping: (null)
index:0x0 compound_mapcount: 0
flags: 0xfffc00000008100(slab|head)
raw: 0fffc00000008100 0000000000000000 0000000000000000 0000000180190019
raw: dead000000000100 dead000000000200 ffff800931268200 0000000000000000
page dumped because: kasan: bad access detected
Memory state around the buggy address:
ffff800021cb5f00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
ffff800021cb5f80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
>ffff800021cb6000: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
^
ffff800021cb6080: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
ffff800021cb6100: fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc fc
==================================================================
Fixes: b189817cf7 ("crypto: caam/qi - add ablkcipher and authenc algorithms")
Signed-off-by: Horia Geantă <horia.geanta@nxp.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
CAAM engine supports two interfaces for crypto job submission:
-job ring interface - already existing caam/jr driver
-Queue Interface (QI) - caam/qi driver added in current patch
QI is present in CAAM engines found on DPAA platforms.
QI gets its I/O (frame descriptors) from QMan (Queue Manager) queues.
This patch adds a platform device for accessing CAAM's queue interface.
The requests are submitted to CAAM using one frame queue per
cryptographic context. Each crypto context has one shared descriptor.
This shared descriptor is attached to frame queue associated with
corresponding driver context using context_a.
The driver hides the mechanics of FQ creation, initialisation from its
applications. Each cryptographic context needs to be associated with
driver context which houses the FQ to be used to transport the job to
CAAM. The driver provides API for:
(a) Context creation
(b) Job submission
(c) Context deletion
(d) Congestion indication - whether path to/from CAAM is congested
The driver supports affining its context to a particular CPU.
This means that any responses from CAAM for the context in question
would arrive at the given CPU. This helps in implementing one CPU
per packet round trip in IPsec application.
The driver processes CAAM responses under NAPI contexts.
NAPI contexts are instantiated only on cores with affined portals since
only cores having their own portal can receive responses from DQRR.
The responses from CAAM for all cryptographic contexts ride on a fixed
set of FQs. We use one response FQ per portal owning core. The response
FQ is configured in each core's and thus portal's dedicated channel.
This gives the flexibility to direct CAAM's responses for a crypto
context on a given core.
Signed-off-by: Vakul Garg <vakul.garg@nxp.com>
Signed-off-by: Alex Porosanu <alexandru.porosanu@nxp.com>
Signed-off-by: Horia Geantă <horia.geanta@nxp.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Horia Geantă