crypto: caam - fix DMA mapping of stack memory
authorHoria Geantă <horia.geanta@nxp.com>
Sat, 26 Jan 2019 18:02:15 +0000 (20:02 +0200)
committerHerbert Xu <herbert@gondor.apana.org.au>
Fri, 1 Feb 2019 06:44:38 +0000 (14:44 +0800)
commitc19650d6ea99bcd903d3e55dd61860026c701339
tree45170f328c5daac8a4bf21ee9cd63b75206af99f
parent5a22b198cd5274470d637d45f0e3bcdc0175a349
crypto: caam - fix DMA mapping of stack memory

Roland reports the following issue and provides a root cause analysis:

"On a v4.19 i.MX6 system with IMA and CONFIG_DMA_API_DEBUG enabled, a
warning is generated when accessing files on a filesystem for which IMA
measurement is enabled:

    ------------[ cut here ]------------
    WARNING: CPU: 0 PID: 1 at kernel/dma/debug.c:1181 check_for_stack.part.9+0xd0/0x120
    caam_jr 2101000.jr0: DMA-API: device driver maps memory from stack [addr=b668049e]
    Modules linked in:
    CPU: 0 PID: 1 Comm: switch_root Not tainted 4.19.0-20181214-1 #2
    Hardware name: Freescale i.MX6 Quad/DualLite (Device Tree)
    Backtrace:
    [<c010efb8>] (dump_backtrace) from [<c010f2d0>] (show_stack+0x20/0x24)
    [<c010f2b0>] (show_stack) from [<c08b04f4>] (dump_stack+0xa0/0xcc)
    [<c08b0454>] (dump_stack) from [<c012b610>] (__warn+0xf0/0x108)
    [<c012b520>] (__warn) from [<c012b680>] (warn_slowpath_fmt+0x58/0x74)
    [<c012b62c>] (warn_slowpath_fmt) from [<c0199acc>] (check_for_stack.part.9+0xd0/0x120)
    [<c01999fc>] (check_for_stack.part.9) from [<c019a040>] (debug_dma_map_page+0x144/0x174)
    [<c0199efc>] (debug_dma_map_page) from [<c065f7f4>] (ahash_final_ctx+0x5b4/0xcf0)
    [<c065f240>] (ahash_final_ctx) from [<c065b3c4>] (ahash_final+0x1c/0x20)
    [<c065b3a8>] (ahash_final) from [<c03fe278>] (crypto_ahash_op+0x38/0x80)
    [<c03fe240>] (crypto_ahash_op) from [<c03fe2e0>] (crypto_ahash_final+0x20/0x24)
    [<c03fe2c0>] (crypto_ahash_final) from [<c03f19a8>] (ima_calc_file_hash+0x29c/0xa40)
    [<c03f170c>] (ima_calc_file_hash) from [<c03f2b24>] (ima_collect_measurement+0x1dc/0x240)
    [<c03f2948>] (ima_collect_measurement) from [<c03f0a60>] (process_measurement+0x4c4/0x6b8)
    [<c03f059c>] (process_measurement) from [<c03f0cdc>] (ima_file_check+0x88/0xa4)
    [<c03f0c54>] (ima_file_check) from [<c02d8adc>] (path_openat+0x5d8/0x1364)
    [<c02d8504>] (path_openat) from [<c02dad24>] (do_filp_open+0x84/0xf0)
    [<c02daca0>] (do_filp_open) from [<c02cf50c>] (do_open_execat+0x84/0x1b0)
    [<c02cf488>] (do_open_execat) from [<c02d1058>] (__do_execve_file+0x43c/0x890)
    [<c02d0c1c>] (__do_execve_file) from [<c02d1770>] (sys_execve+0x44/0x4c)
    [<c02d172c>] (sys_execve) from [<c0101000>] (ret_fast_syscall+0x0/0x28)
    ---[ end trace 3455789a10e3aefd ]---

The cause is that the struct ahash_request *req is created as a
stack-local variable up in the stack (presumably somewhere in the IMA
implementation), then passed down into the CAAM driver, which tries to
dma_single_map the req->result (indirectly via map_seq_out_ptr_result)
in order to make that buffer available for the CAAM to store the result
of the following hash operation.

The calling code doesn't know how req will be used by the CAAM driver,
and there could be other such occurrences where stack memory is passed
down to the CAAM driver. Therefore we should rather fix this issue in
the CAAM driver where the requirements are known."

Fix this problem by:
-instructing the crypto engine to write the final hash in state->caam_ctx
-subsequently memcpy-ing the final hash into req->result

Cc: <stable@vger.kernel.org> # v4.19+
Reported-by: Roland Hieber <rhi@pengutronix.de>
Signed-off-by: Horia Geantă <horia.geanta@nxp.com>
Tested-by: Roland Hieber <rhi@pengutronix.de>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
drivers/crypto/caam/caamhash.c