[ Upstream commit d776b25495 ]
The callback function for RSA frees the memory allocated for the source
and destination buffers before unmapping them.
This sequence is wrong.
Change the cleanup sequence to unmap the buffers before freeing them.
Fixes: 3dfaf0071e ("crypto: qat - remove dma_free_coherent() for RSA")
Signed-off-by: Hareshx Sankar Raj <hareshx.sankar.raj@intel.com>
Co-developed-by: Bolemx Sivanagaleela <bolemx.sivanagaleela@intel.com>
Signed-off-by: Bolemx Sivanagaleela <bolemx.sivanagaleela@intel.com>
Reviewed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit eb7713f5ca ]
The callback function for DH frees the memory allocated for the
destination buffer before unmapping it.
This sequence is wrong.
Change the cleanup sequence to unmap the buffer before freeing it.
Fixes: 029aa4624a ("crypto: qat - remove dma_free_coherent() for DH")
Signed-off-by: Hareshx Sankar Raj <hareshx.sankar.raj@intel.com>
Co-developed-by: Bolemx Sivanagaleela <bolemx.sivanagaleela@intel.com>
Signed-off-by: Bolemx Sivanagaleela <bolemx.sivanagaleela@intel.com>
Reviewed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 80e62ad58d ]
The value of reqsize must only be changed through the helper.
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Reviewed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Stable-dep-of: eb7713f5ca ("crypto: qat - unmap buffer before free for DH")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 072d36eefd ]
When mapping the input and output parameters, the implementations of RSA
and DH pass to the function dma_map_single() a pointer to the first
member of the structure they want to map instead of a pointer to the
actual structure.
This results in set of warnings reported by the static analyser Smatch:
drivers/crypto/qat/qat_common/qat_asym_algs.c:335 qat_dh_compute_value() error: dma_map_single_attrs() '&qat_req->in.dh.in.b' too small (8 vs 64)
drivers/crypto/qat/qat_common/qat_asym_algs.c:341 qat_dh_compute_value() error: dma_map_single_attrs() '&qat_req->out.dh.r' too small (8 vs 64)
drivers/crypto/qat/qat_common/qat_asym_algs.c:732 qat_rsa_enc() error: dma_map_single_attrs() '&qat_req->in.rsa.enc.m' too small (8 vs 64)
drivers/crypto/qat/qat_common/qat_asym_algs.c:738 qat_rsa_enc() error: dma_map_single_attrs() '&qat_req->out.rsa.enc.c' too small (8 vs 64)
drivers/crypto/qat/qat_common/qat_asym_algs.c:878 qat_rsa_dec() error: dma_map_single_attrs() '&qat_req->in.rsa.dec.c' too small (8 vs 64)
drivers/crypto/qat/qat_common/qat_asym_algs.c:884 qat_rsa_dec() error: dma_map_single_attrs() '&qat_req->out.rsa.dec.m' too small (8 vs 64)
Where the address of the first element of a structure is used as an
input for the function dma_map_single(), replace it with the address of
the structure. This fix does not introduce any functional change as the
addresses are the same.
Signed-off-by: Damian Muszynski <damian.muszynski@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Adam Guerin <adam.guerin@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Stable-dep-of: eb7713f5ca ("crypto: qat - unmap buffer before free for DH")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit c2a1b91e47 ]
Currently the QAT driver code uses a self-defined wrapper function
called get_current_node() when it wants to learn the current NUMA node.
This implementation references the topology_physical_package_id[] array,
which more or less coincidentally contains the NUMA node id, at least
on x86.
Because this is not universal, and Linux offers a direct function to
learn the NUMA node ID, replace that function with a call to
numa_node_id(), which would work everywhere.
This fixes the QAT driver operation on arm64 machines.
Reported-by: Yoan Picchi <Yoan.Picchi@arm.com>
Signed-off-by: Andre Przywara <andre.przywara@arm.com>
Signed-off-by: Yoan Picchi <yoan.picchi@arm.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Stable-dep-of: eb7713f5ca ("crypto: qat - unmap buffer before free for DH")
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 8fb203c65a ]
If a request has the flag CRYPTO_TFM_REQ_MAY_SLEEP set, allocate memory
using the flag GFP_KERNEL otherwise use GFP_ATOMIC.
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Adam Guerin <adam.guerin@intel.com>
Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Stable-dep-of: eb7713f5ca ("crypto: qat - unmap buffer before free for DH")
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit f6044cc303 upstream.
When preparing an AER-CTR request, the driver copies the key provided by
the user into a data structure that is accessible by the firmware.
If the target device is QAT GEN4, the key size is rounded up by 16 since
a rounded up size is expected by the device.
If the key size is rounded up before the copy, the size used for copying
the key might be bigger than the size of the region containing the key,
causing an out-of-bounds read.
Fix by doing the copy first and then update the keylen.
This is to fix the following warning reported by KASAN:
[ 138.150574] BUG: KASAN: global-out-of-bounds in qat_alg_skcipher_init_com.isra.0+0x197/0x250 [intel_qat]
[ 138.150641] Read of size 32 at addr ffffffff88c402c0 by task cryptomgr_test/2340
[ 138.150651] CPU: 15 PID: 2340 Comm: cryptomgr_test Not tainted 6.2.0-rc1+ #45
[ 138.150659] Hardware name: Intel Corporation ArcherCity/ArcherCity, BIOS EGSDCRB1.86B.0087.D13.2208261706 08/26/2022
[ 138.150663] Call Trace:
[ 138.150668] <TASK>
[ 138.150922] kasan_check_range+0x13a/0x1c0
[ 138.150931] memcpy+0x1f/0x60
[ 138.150940] qat_alg_skcipher_init_com.isra.0+0x197/0x250 [intel_qat]
[ 138.151006] qat_alg_skcipher_init_sessions+0xc1/0x240 [intel_qat]
[ 138.151073] crypto_skcipher_setkey+0x82/0x160
[ 138.151085] ? prepare_keybuf+0xa2/0xd0
[ 138.151095] test_skcipher_vec_cfg+0x2b8/0x800
Fixes: 67916c9516 ("crypto: qat - add AES-CTR support for QAT GEN4 devices")
Cc: <stable@vger.kernel.org>
Reported-by: Vladis Dronov <vdronov@redhat.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Reviewed-by: Vladis Dronov <vdronov@redhat.com>
Tested-by: Vladis Dronov <vdronov@redhat.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit cf5bb835b7 ]
When CONFIG_DMA_API_DEBUG is selected, while running the crypto self
test on the QAT crypto algorithms, the function add_dma_entry() reports
a warning similar to the one below, saying that overlapping mappings
are not supported. This occurs in tests where the input and the output
scatter list point to the same buffers (i.e. two different scatter lists
which point to the same chunks of memory).
The logic that implements the mapping uses the flag DMA_BIDIRECTIONAL
for both the input and the output scatter lists which leads to
overlapped write mappings. These are not supported by the DMA layer.
Fix by specifying the correct DMA transfer directions when mapping
buffers. For in-place operations where the input scatter list
matches the output scatter list, buffers are mapped once with
DMA_BIDIRECTIONAL, otherwise input buffers are mapped using the flag
DMA_TO_DEVICE and output buffers are mapped with DMA_FROM_DEVICE.
Overlapping a read mapping with a write mapping is a valid case in
dma-coherent devices like QAT.
The function that frees and unmaps the buffers, qat_alg_free_bufl()
has been changed accordingly to the changes to the mapping function.
DMA-API: 4xxx 0000:06:00.0: cacheline tracking EEXIST, overlapping mappings aren't supported
WARNING: CPU: 53 PID: 4362 at kernel/dma/debug.c:570 add_dma_entry+0x1e9/0x270
...
Call Trace:
dma_map_page_attrs+0x82/0x2d0
? preempt_count_add+0x6a/0xa0
qat_alg_sgl_to_bufl+0x45b/0x990 [intel_qat]
qat_alg_aead_dec+0x71/0x250 [intel_qat]
crypto_aead_decrypt+0x3d/0x70
test_aead_vec_cfg+0x649/0x810
? number+0x310/0x3a0
? vsnprintf+0x2a3/0x550
? scnprintf+0x42/0x70
? valid_sg_divisions.constprop.0+0x86/0xa0
? test_aead_vec+0xdf/0x120
test_aead_vec+0xdf/0x120
alg_test_aead+0x185/0x400
alg_test+0x3d8/0x500
? crypto_acomp_scomp_free_ctx+0x30/0x30
? __schedule+0x32a/0x12a0
? ttwu_queue_wakelist+0xbf/0x110
? _raw_spin_unlock_irqrestore+0x23/0x40
? try_to_wake_up+0x83/0x570
? _raw_spin_unlock_irqrestore+0x23/0x40
? __set_cpus_allowed_ptr_locked+0xea/0x1b0
? crypto_acomp_scomp_free_ctx+0x30/0x30
cryptomgr_test+0x27/0x50
kthread+0xe6/0x110
? kthread_complete_and_exit+0x20/0x20
ret_from_fork+0x1f/0x30
Fixes: d370cec ("crypto: qat - Intel(R) QAT crypto interface")
Link: https://lore.kernel.org/linux-crypto/20220223080400.139367-1-gilad@benyossef.com/
Signed-off-by: Damian Muszynski <damian.muszynski@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit cc40b04c08 ]
The QAT HW supports an hardware mechanism to detect an accelerator hang.
The reporting of a hang occurs after a watchdog timer (WDT) expires.
The value of the WDT set previously was too small and was causing false
positives.
Change the default value of the WDT to 0x7000000ULL to avoid this.
Fixes: 1c4d9d5bbb ("crypto: qat - enable detection of accelerators hang")
Reviewed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Lucas Segarra Fernandez <lucas.segarra.fernandez@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 9c5f21b198 upstream.
This reverts commit e48767c177.
In an attempt to resolve a set of warnings reported by the static
analyzer Smatch, the reverted commit improperly reduced the sizes of the
DMA mappings used for the input and output parameters for both RSA and
DH creating a mismatch (map size=8 bytes, unmap size=64 bytes).
This issue is reported when CONFIG_DMA_API_DEBUG is selected, when the
crypto self test is run. The function dma_unmap_single() reports a
warning similar to the one below, saying that the `device driver frees
DMA memory with different size`.
DMA-API: 4xxx 0000:06:00.0: device driver frees DMA memory with different size [device address=0x0000000123206c80] [map size=8 bytes] [unmap size=64 bytes]
WARNING: CPU: 0 PID: 0 at kernel/dma/debug.c:973 check_unmap+0x3d0/0x8c0\
...
Call Trace:
<IRQ>
debug_dma_unmap_page+0x5c/0x60
qat_dh_cb+0xd7/0x110 [intel_qat]
qat_alg_asym_callback+0x1a/0x30 [intel_qat]
adf_response_handler+0xbd/0x1a0 [intel_qat]
tasklet_action_common.constprop.0+0xcd/0xe0
__do_softirq+0xf8/0x30c
__irq_exit_rcu+0xbf/0x140
common_interrupt+0xb9/0xd0
</IRQ>
<TASK>
The original commit was correct.
Cc: <stable@vger.kernel.org>
Reported-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit d091447459 ]
Re-enable the registration of algorithms after fixes to (1) use
pre-allocated buffers in the datapath and (2) support the
CRYPTO_TFM_REQ_MAY_BACKLOG flag.
This reverts commit 8893d27ffc.
Cc: stable@vger.kernel.org
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Adam Guerin <adam.guerin@intel.com>
Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 2acbb8771f ]
Reject requests with a source buffer that is bigger than the size of the
key. This is to prevent a possible integer underflow that might happen
when copying the source scatterlist into a linear buffer.
Cc: stable@vger.kernel.org
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Adam Guerin <adam.guerin@intel.com>
Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 9714061423 ]
Reject requests with a source buffer that is bigger than the size of the
key. This is to prevent a possible integer underflow that might happen
when copying the source scatterlist into a linear buffer.
Cc: stable@vger.kernel.org
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Adam Guerin <adam.guerin@intel.com>
Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 029aa4624a ]
The functions qat_dh_compute_value() allocates memory with
dma_alloc_coherent() if the source or the destination buffers are made
of multiple flat buffers or of a size that is not compatible with the
hardware.
This memory is then freed with dma_free_coherent() in the context of a
tasklet invoked to handle the response for the corresponding request.
According to Documentation/core-api/dma-api-howto.rst, the function
dma_free_coherent() cannot be called in an interrupt context.
Replace allocations with dma_alloc_coherent() in the function
qat_dh_compute_value() with kmalloc() + dma_map_single().
Cc: stable@vger.kernel.org
Fixes: c9839143eb ("crypto: qat - Add DH support")
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Adam Guerin <adam.guerin@intel.com>
Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 3dfaf0071e ]
After commit f5ff79fddf ("dma-mapping: remove CONFIG_DMA_REMAP"), if
the algorithms are enabled, the driver crashes with a BUG_ON while
executing vunmap() in the context of a tasklet. This is due to the fact
that the function dma_free_coherent() cannot be called in an interrupt
context (see Documentation/core-api/dma-api-howto.rst).
The functions qat_rsa_enc() and qat_rsa_dec() allocate memory with
dma_alloc_coherent() if the source or the destination buffers are made
of multiple flat buffers or of a size that is not compatible with the
hardware.
This memory is then freed with dma_free_coherent() in the context of a
tasklet invoked to handle the response for the corresponding request.
Replace allocations with dma_alloc_coherent() in the functions
qat_rsa_enc() and qat_rsa_dec() with kmalloc() + dma_map_single().
Cc: stable@vger.kernel.org
Fixes: a990532023 ("crypto: qat - Add support for RSA algorithm")
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Adam Guerin <adam.guerin@intel.com>
Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 80a52e1ee7 ]
When an RSA key represented in form 2 (as defined in PKCS #1 V2.1) is
used, some components of the private key persist even after the TFM is
released.
Replace the explicit calls to free the buffers in qat_rsa_exit_tfm()
with a call to qat_rsa_clear_ctx() which frees all buffers referenced in
the TFM context.
Cc: stable@vger.kernel.org
Fixes: 879f77e907 ("crypto: qat - Add RSA CRT mode")
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Adam Guerin <adam.guerin@intel.com>
Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 3868238397 ]
The implementations of the crypto algorithms (aead, skcipher, etc) in
the QAT driver do not properly support requests with the
CRYPTO_TFM_REQ_MAY_BACKLOG flag set. If the HW queue is full, the driver
returns -EBUSY but does not enqueue the request. This can result in
applications like dm-crypt waiting indefinitely for the completion of a
request that was never submitted to the hardware.
Fix this by adding a software backlog queue: if the ring buffer is more
than eighty percent full, then the request is enqueued to a backlog
list and the error code -EBUSY is returned back to the caller.
Requests in the backlog queue are resubmitted at a later time, in the
context of the callback of a previously submitted request.
The request for which -EBUSY is returned is then marked as -EINPROGRESS
once submitted to the HW queues.
The submission loop inside the function qat_alg_send_message() has been
modified to decide which submission policy to use based on the request
flags. If the request does not have the CRYPTO_TFM_REQ_MAY_BACKLOG set,
the previous behaviour has been preserved.
Based on a patch by
Vishnu Das Ramachandran <vishnu.dasx.ramachandran@intel.com>
Cc: stable@vger.kernel.org
Fixes: d370cec321 ("crypto: qat - Intel(R) QAT crypto interface")
Reported-by: Mikulas Patocka <mpatocka@redhat.com>
Reported-by: Kyle Sanderson <kyle.leet@gmail.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit af88d3c109 ]
All the algorithms in qat_algs.c and qat_asym_algs.c use the same
pattern to submit messages to the HW queues. Move the submission loop
to a new function, qat_alg_send_message(), and share it between the
symmetric and the asymmetric algorithms.
As part of this rework, since the number of retries before returning an
error is inconsistent between the symmetric and asymmetric
implementations, set it to a value that works for both (i.e. 20, was 10
in qat_algs.c and 100 in qat_asym_algs.c)
In addition fix the return code reported when the HW queues are full.
In that case return -ENOSPC instead of -EBUSY.
Including stable in CC since (1) the error code returned if the HW queues
are full is incorrect and (2) to facilitate the backport of the next fix
"crypto: qat - add backlog mechanism".
Cc: stable@vger.kernel.org
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit e0831e7af4 ]
In order to do DMAs, the QAT device requires that the scatterlist
structures are mapped and translated into a format that the firmware can
understand. This is defined as the composition of a scatter gather list
(SGL) descriptor header, the struct qat_alg_buf_list, plus a variable
number of flat buffer descriptors, the struct qat_alg_buf.
The allocation and mapping of these data structures is done each time a
request is received from the skcipher and aead APIs.
In an OOM situation, this behaviour might lead to a dead-lock if an
allocation fails.
Based on the conversation in [1], increase the size of the aead and
skcipher request contexts to include an SGL descriptor that can handle
a maximum of 4 flat buffers.
If requests exceed 4 entries buffers, memory is allocated dynamically.
[1] https://lore.kernel.org/linux-crypto/20200722072932.GA27544@gondor.apana.org.au/
Cc: stable@vger.kernel.org
Fixes: d370cec321 ("crypto: qat - Intel(R) QAT crypto interface")
Reported-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 1731160ff7 ]
Set to zero the context buffers containing the DH key before they are
freed.
This is a defense in depth measure that avoids keys to be recovered from
memory in case the system is compromised between the free of the buffer
and when that area of memory (containing keys) gets overwritten.
Cc: stable@vger.kernel.org
Fixes: c9839143eb ("crypto: qat - Add DH support")
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Adam Guerin <adam.guerin@intel.com>
Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 0eaa515432 ]
The capability detection logic clears bits for the features that are
disabled in a certain SKU. For example, if the bit associate to
compression is not present in the LEGFUSE register, the correspondent
bit is cleared in the capability mask.
This change adds the compression capability to the mask as this was
missing in the commit that enhanced the capability detection logic.
Fixes: cfe4894ecc ("crypto: qat - set COMPRESSION capability for QAT GEN2")
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 6a23804cb8 ]
Set the CIPHER capability for QAT DH895XCC devices if the hardware supports
it. This is done if both the CIPHER and the AUTHENTICATION engines are
available on the device.
Fixes: ad1332aa67 ("crypto: qat - add support for capability detection")
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit cfe4894ecc ]
Enhance the device capability detection for QAT GEN2 devices to detect if
a device supports the compression service.
This is done by checking both the fuse and the strap registers for c62x
and c3xxx and only the fuse register for dh895xcc.
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 547bde7bd4 ]
Set the CIPHER capability for QAT GEN2 devices if the hardware supports
it. This is done if both the CIPHER and the AUTHENTICATION engines are
available on the device.
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
commit 8893d27ffc upstream.
The implementations of aead and skcipher in the QAT driver do not
support properly requests with the CRYPTO_TFM_REQ_MAY_BACKLOG flag set.
If the HW queue is full, the driver returns -EBUSY but does not enqueue
the request.
This can result in applications like dm-crypt waiting indefinitely for a
completion of a request that was never submitted to the hardware.
To avoid this problem, disable the registration of all crypto algorithms
in the QAT driver by setting the number of crypto instances to 0 at
configuration time.
Cc: stable@vger.kernel.org
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 5002200b4f ]
If the remote function did not ACK the reception of a message, the
function __adf_iov_putmsg() could detect it as a collision.
This was due to the fact that the collision and the timeout checks after
the ACK loop were in the wrong order. The timeout must be checked at the
end of the loop, so fix by swapping the order of the two checks.
Fixes: 9b768e8a39 ("crypto: qat - detect PFVF collision after ACK")
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Co-developed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 6e680f94bc ]
The functions adf_iov_putmsg() and __adf_iov_putmsg() are shared by both
PF and VF. Any logging or documentation should not refer to any specific
direction.
Make comments and log messages direction agnostic by replacing PF2VF
with PFVF. Also fix the wording for some related comments.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit e17f49bb24 ]
The initial version of the PFVF protocol included an initial "carrier
sensing" to get ownership of the channel.
Collisions can happen anyway, the extra wait and test does not prevent
collisions, it instead slows the communication down, so remove it.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 18fcba469b ]
Upon receiving a PFVF message, check if the interrupt bit is set in the
message. If it is not, that means that the interrupt was probably
triggered by a collision. In this case, disregard the message and
re-enable the interrupts.
Fixes: ed8ccaef52 ("crypto: qat - Add support for SRIOV")
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit 9b768e8a39 ]
Detect a PFVF collision between the local and the remote function by
checking if the message on the PFVF CSR has been overwritten.
This is done after the remote function confirms that the message has
been received, by clearing the interrupt bit, or the maximum number of
attempts (ADF_IOV_MSG_ACK_MAX_RETRY) to check the CSR has been exceeded.
Fixes: ed8ccaef52 ("crypto: qat - Add support for SRIOV")
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Co-developed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
[ Upstream commit ca605f97da ]
After reset or boot, QAT 4xxx devices are inactive and require to be
explicitly activated.
This is done by writing the DRV_ACTIVE bit in the PM_INTERRUPT register
and polling the PM_INIT_STATE to make sure that the transaction has
completed properly.
If this is not done, the driver will fail the initialization sequence
reporting the following message:
[ 22.081193] 4xxx 0000:f7:00.0: enabling device (0140 -> 0142)
[ 22.720285] QAT: AE0 is inactive!!
[ 22.720287] QAT: failed to get device out of reset
[ 22.720288] 4xxx 0000:f7:00.0: qat_hal_clr_reset error
[ 22.720290] 4xxx 0000:f7:00.0: Failed to init the AEs
[ 22.720290] 4xxx 0000:f7:00.0: Failed to initialise Acceleration Engine
[ 22.720789] 4xxx 0000:f7:00.0: Resetting device qat_dev0
[ 22.825099] 4xxx: probe of 0000:f7:00.0 failed with error -14
The patch also temporarily disables the power management source of
interrupt, to avoid possible spurious interrupts as the power management
feature is not fully supported.
The device init function has been added to adf_dev_init(), and not in the
probe of 4xxx to make sure that the device is re-enabled in case of
reset.
Note that the error code reported by hw_data->init_device() in
adf_dev_init() has been shadowed for consistency with the other calls
in the same function.
Fixes: 8c8268166e ("crypto: qat - add qat_4xxx driver")
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Sasha Levin <sashal@kernel.org>
If the VF is newer than the PF, it decides whether it is compatible or
not. In case it is compatible, store that information in the
vf.compatible flag in the accel_dev structure.
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Suggested-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The function adf_iov_putmsg() is only used inside the intel_qat module
therefore should not be exported.
Remove EXPORT_SYMBOL for the function adf_iov_putmsg().
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
There is a race condition during shutdown in adf_disable_sriov() where
both the PF and the VF drivers are loaded on the host system.
The PF notifies a VF with a "RESTARTING" message due to which the VF
starts an asynchronous worker to stop and shutdown itself.
At the same time the PF calls pci_disable_sriov() which invokes the
remove() routine on the VF device driver triggering the shutdown flow
again.
This change fixes the problem by ensuring that the VF flushes the worker
that performs stop()/shutdown() before these two functions are called in
the remove(). To make sure that no additional PV/VF messages are
processed by the VF, interrupts are disabled before flushing the
workqueue.
Signed-off-by: Ahsan Atta <ahsan.atta@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
All QAT GEN2 devices share the same register offset for masking interrupts,
so they don't need any complex device specific infrastructure.
Remove this function in favor of a constant in order to simplify the code.
Also, future generations may require a more complex device specific
handling, making the current approach obsolete anyway.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Currently all the functions related to the activation of the PFVF
protocol, both on PF and VF, include the direction specific "vf2pf"
name.
Replace the existing naming schema with:
- a direction agnostic naming, that applies to both PF and VF, for the
function pointer ("pfvf")
- a direction specific naming schema for the implementations ("pf2vf" or
"vf2pf")
In particular this patch renames:
- adf_pf_enable_vf2pf_comms() in adf_enable_pf2vf_comms()
- enable_vf2pf_comms() in enable_pfvf_comms()
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Make sure all the steps in the initialization sequence are complete
before any completion event notification.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Move IOV functions at the end of hw_data so that PFVF functions related
functions are group together.
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
At start and shutdown, VFs notify the PF about their state. These
notifications are carried out through a message exchange using the PFVF
protocol.
Function names lead to believe they do perform init or shutdown logic.
This is to fix the naming to better reflect their purpose.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
In the PF interrupt handler, the interrupt is disabled for a set of VFs
by writing to the interrupt source mask register, ERRMSK.
The interrupt is re-enabled in the bottom half handler by writing to the
same CSR. This is done through the functions enable_vf2pf_interrupts()
and disable_vf2pf_interrupts() which perform a read-modify-write
operation on the ERRMSK registers to mask and unmask the source of
interrupt.
There can be a race condition where the top half handler for one VF
interrupt runs just as the bottom half for another VF is about to
re-enable the interrupt. Depending on whether the top or bottom half
updates the CSR first, this would result either in a spurious interrupt
or in the interrupt not being re-enabled.
This patch protects the access of ERRMSK with a spinlock.
The functions adf_enable_vf2pf_interrupts() and
adf_disable_vf2pf_interrupts() have been changed to acquire a spin lock
before accessing and modifying the ERRMSK registers. These functions use
spin_lock_irqsave() to disable IRQs and avoid potential deadlocks.
In addition, the function adf_disable_vf2pf_interrupts_irq() has been
added. This uses spin_lock() and it is meant to be used in the top half
only.
Signed-off-by: Kanchana Velusamy <kanchanax.velusamy@intel.com>
Co-developed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Interrupt code to enable interrupts from PF does not belong to the
protocol code, so move it to the interrupt handling specific file for
better code organization.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Use reinit_completion() to set to a clean state a completion variable,
used to coordinate the VF to PF request-response flow, before every
new VF request.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The PF driver uses the tasklet vf2pf_bh_tasklet to schedule a workqueue
to handle the vf2vf protocol (pf2vf_resp_wq).
Since the tasklet is only used to schedule the workqueue, this patch
removes it and schedules the pf2vf_resp_wq workqueue directly for the
top half.
Signed-off-by: Svyatoslav Pankratov <svyatoslav.pankratov@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Rename ADF_PFVF_COMPATIBILITY_VERSION in ADF_PFVF_COMPAT_THIS_VERSION
since it is used to indicate the current version of the PFVF protocol.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
There is a chance that the PFVF handler, adf_vf2pf_req_hndl(), runs
twice for the same request when multiple interrupts come simultaneously
from different VFs.
Since the source VF is identified by a positional bit set in the ERRSOU
registers and that is not cleared until the bottom half completes, new
top halves from other VFs may reschedule a second bottom half for
previous interrupts.
This patch solves the problem in the ISR handler by not considering
sources with already disabled interrupts (and processing pending), as
set in the ERRMSK registers.
Also, move some definitions where actually needed.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
QAT GEN2 devices suffer from a defect where the MSI interrupt can be
sent multiple times.
If the second (spurious) interrupt is handled before the bottom half
handler runs, then the extra interrupt is effectively ignored because
the bottom half is only scheduled once.
However, if the top half runs again after the bottom half runs, this
will appear as a spurious PF to VF interrupt.
This can be avoided by checking the interrupt mask register in addition
to the interrupt source register in the interrupt handler.
This patch is based on earlier work done by Conor McLoughlin.
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Co-developed-by: Marco Chiappero <marco.chiappero@intel.com>
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The top half of the VF drivers handled only a source at the time.
If an interrupt for PF2VF and bundle occurred at the same time, the ISR
scheduled only the bottom half for PF2VF.
This patch fixes the VF top half so that if both sources of interrupt
trigger at the same time, both bottom halves are scheduled.
This patch is based on earlier work done by Conor McLoughlin.
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
The function adf_dev_init() ignores the error code reported by
enable_vf2pf_comms(). If the latter fails, e.g. the VF is not compatible
with the pf, then the load of the VF driver progresses.
This patch changes adf_dev_init() so that the error code from
enable_vf2pf_comms() is returned to the caller.
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Marco Chiappero <marco.chiappero@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Enable device interrupts after the setup of the interrupt handlers.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Remove the empty implementation of sriov_configure() and set the
sriov_configure member of the pci_driver structure to NULL.
This way, if a user tries to enable VFs on a device, when kernel and
driver are built with CONFIG_PCI_IOV=n, the kernel reports an error
message saying that the driver does not support SRIOV configuration via
sysfs.
Signed-off-by: Marco Chiappero <marco.chiappero@intel.com>
Co-developed-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
Reviewed-by: Fiona Trahe <fiona.trahe@intel.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Hareshx Sankar Raj
Herbert Xu
Damian Muszynski
Andre Przywara
Giovanni Cabiddu
Lucas Segarra Fernandez
Marco Chiappero
Ahsan Atta
Kanchana Velusamy
Svyatoslav Pankratov