linux_dsm_epyc7002/drivers/crypto/caam/caamrng.c

262 lines
5.6 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0+
/*
* caam - Freescale FSL CAAM support for hw_random
*
* Copyright 2011 Freescale Semiconductor, Inc.
* Copyright 2018-2019 NXP
*
* Based on caamalg.c crypto API driver.
*
*/
#include <linux/hw_random.h>
#include <linux/completion.h>
#include <linux/atomic.h>
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
#include <linux/kfifo.h>
#include "compat.h"
#include "regs.h"
#include "intern.h"
#include "desc_constr.h"
#include "jr.h"
#include "error.h"
#define CAAM_RNG_MAX_FIFO_STORE_SIZE 16
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
/*
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
* Length of used descriptors, see caam_init_desc()
*/
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
#define CAAM_RNG_DESC_LEN (CAAM_CMD_SZ + \
CAAM_CMD_SZ + \
CAAM_CMD_SZ + CAAM_PTR_SZ_MAX)
/* rng per-device context */
struct caam_rng_ctx {
struct hwrng rng;
struct device *jrdev;
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
struct device *ctrldev;
void *desc_async;
void *desc_sync;
struct work_struct worker;
struct kfifo fifo;
};
struct caam_rng_job_ctx {
struct completion *done;
int *err;
};
static struct caam_rng_ctx *to_caam_rng_ctx(struct hwrng *r)
{
return (struct caam_rng_ctx *)r->priv;
}
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
static void caam_rng_done(struct device *jrdev, u32 *desc, u32 err,
void *context)
{
struct caam_rng_job_ctx *jctx = context;
if (err)
*jctx->err = caam_jr_strstatus(jrdev, err);
complete(jctx->done);
}
static u32 *caam_init_desc(u32 *desc, dma_addr_t dst_dma)
{
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
init_job_desc(desc, 0); /* + 1 cmd_sz */
/* Generate random bytes: + 1 cmd_sz */
append_operation(desc, OP_ALG_ALGSEL_RNG | OP_TYPE_CLASS1_ALG |
OP_ALG_PR_ON);
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
/* Store bytes: + 1 cmd_sz + caam_ptr_sz */
append_fifo_store(desc, dst_dma,
CAAM_RNG_MAX_FIFO_STORE_SIZE, FIFOST_TYPE_RNGSTORE);
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
print_hex_dump_debug("rng job desc@: ", DUMP_PREFIX_ADDRESS,
16, 4, desc, desc_bytes(desc), 1);
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
return desc;
}
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
static int caam_rng_read_one(struct device *jrdev,
void *dst, int len,
void *desc,
struct completion *done)
{
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
dma_addr_t dst_dma;
int err, ret = 0;
struct caam_rng_job_ctx jctx = {
.done = done,
.err = &ret,
};
len = CAAM_RNG_MAX_FIFO_STORE_SIZE;
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
dst_dma = dma_map_single(jrdev, dst, len, DMA_FROM_DEVICE);
if (dma_mapping_error(jrdev, dst_dma)) {
dev_err(jrdev, "unable to map destination memory\n");
return -ENOMEM;
}
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
init_completion(done);
err = caam_jr_enqueue(jrdev,
caam_init_desc(desc, dst_dma),
caam_rng_done, &jctx);
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
if (err == -EINPROGRESS) {
wait_for_completion(done);
err = 0;
}
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
dma_unmap_single(jrdev, dst_dma, len, DMA_FROM_DEVICE);
return err ?: (ret ?: len);
}
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
static void caam_rng_fill_async(struct caam_rng_ctx *ctx)
{
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
struct scatterlist sg[1];
struct completion done;
int len, nents;
sg_init_table(sg, ARRAY_SIZE(sg));
nents = kfifo_dma_in_prepare(&ctx->fifo, sg, ARRAY_SIZE(sg),
CAAM_RNG_MAX_FIFO_STORE_SIZE);
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
if (!nents)
return;
len = caam_rng_read_one(ctx->jrdev, sg_virt(&sg[0]),
sg[0].length,
ctx->desc_async,
&done);
if (len < 0)
return;
kfifo_dma_in_finish(&ctx->fifo, len);
}
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
static void caam_rng_worker(struct work_struct *work)
{
struct caam_rng_ctx *ctx = container_of(work, struct caam_rng_ctx,
worker);
caam_rng_fill_async(ctx);
}
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
static int caam_read(struct hwrng *rng, void *dst, size_t max, bool wait)
{
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
struct caam_rng_ctx *ctx = to_caam_rng_ctx(rng);
int out;
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
if (wait) {
struct completion done;
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
return caam_rng_read_one(ctx->jrdev, dst, max,
ctx->desc_sync, &done);
}
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
out = kfifo_out(&ctx->fifo, dst, max);
if (kfifo_is_empty(&ctx->fifo))
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
schedule_work(&ctx->worker);
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
return out;
}
static void caam_cleanup(struct hwrng *rng)
{
struct caam_rng_ctx *ctx = to_caam_rng_ctx(rng);
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
flush_work(&ctx->worker);
caam_jr_free(ctx->jrdev);
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
kfifo_free(&ctx->fifo);
}
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
static int caam_init(struct hwrng *rng)
{
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
struct caam_rng_ctx *ctx = to_caam_rng_ctx(rng);
int err;
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
ctx->desc_sync = devm_kzalloc(ctx->ctrldev, CAAM_RNG_DESC_LEN,
GFP_DMA | GFP_KERNEL);
if (!ctx->desc_sync)
return -ENOMEM;
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
ctx->desc_async = devm_kzalloc(ctx->ctrldev, CAAM_RNG_DESC_LEN,
GFP_DMA | GFP_KERNEL);
if (!ctx->desc_async)
return -ENOMEM;
if (kfifo_alloc(&ctx->fifo, CAAM_RNG_MAX_FIFO_STORE_SIZE,
GFP_DMA | GFP_KERNEL))
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
return -ENOMEM;
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
INIT_WORK(&ctx->worker, caam_rng_worker);
ctx->jrdev = caam_jr_alloc();
err = PTR_ERR_OR_ZERO(ctx->jrdev);
if (err) {
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
kfifo_free(&ctx->fifo);
pr_err("Job Ring Device allocation for transform failed\n");
return err;
}
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
/*
* Fill async buffer to have early randomness data for
* hw_random
*/
caam_rng_fill_async(ctx);
return 0;
}
int caam_rng_init(struct device *ctrldev);
void caam_rng_exit(struct device *ctrldev)
{
devres_release_group(ctrldev, caam_rng_init);
}
int caam_rng_init(struct device *ctrldev)
{
struct caam_rng_ctx *ctx;
u32 rng_inst;
struct caam_drv_private *priv = dev_get_drvdata(ctrldev);
int ret;
/* Check for an instantiated RNG before registration */
if (priv->era < 10)
rng_inst = (rd_reg32(&priv->ctrl->perfmon.cha_num_ls) &
CHA_ID_LS_RNG_MASK) >> CHA_ID_LS_RNG_SHIFT;
else
rng_inst = rd_reg32(&priv->ctrl->vreg.rng) & CHA_VER_NUM_MASK;
if (!rng_inst)
return 0;
if (!devres_open_group(ctrldev, caam_rng_init, GFP_KERNEL))
return -ENOMEM;
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
ctx = devm_kzalloc(ctrldev, sizeof(*ctx), GFP_KERNEL);
if (!ctx)
return -ENOMEM;
crypto: caam - simplify RNG implementation Rework CAAM RNG implementation as follows: - Make use of the fact that HWRNG supports partial reads and will handle such cases gracefully by removing recursion in caam_read() - Convert blocking caam_read() codepath to do a single blocking job read directly into requested buffer, bypassing any intermediary buffers - Convert async caam_read() codepath into a simple single reader/single writer FIFO use-case, thus simplifying concurrency handling and delegating buffer read/write position management to KFIFO subsystem. - Leverage the same low level RNG data extraction code for both async and blocking caam_read() scenarios, get rid of the shared job descriptor and make non-shared one as a simple as possible (just HEADER + ALGORITHM OPERATION + FIFO STORE) - Split private context from DMA related memory, so that the former could be allocated without GFP_DMA. NOTE: On its face value this commit decreased throughput numbers reported by dd if=/dev/hwrng of=/dev/null bs=1 count=100K [iflag=nonblock] by about 15%, however commits that enable prediction resistance and limit JR total size impact the performance so much and move the bottleneck such as to make this regression irrelevant. NOTE: On the bright side, this commit reduces RNG in kernel DMA buffer memory usage from 2 x RN_BUF_SIZE (~256K) to 32K. Signed-off-by: Andrey Smirnov <andrew.smirnov@gmail.com> Reviewed-by: Horia Geantă <horia.geanta@nxp.com> Cc: Chris Healy <cphealy@gmail.com> Cc: Lucas Stach <l.stach@pengutronix.de> Cc: Horia Geantă <horia.geanta@nxp.com> Cc: Herbert Xu <herbert@gondor.apana.org.au> Cc: Iuliana Prodan <iuliana.prodan@nxp.com> Cc: linux-crypto@vger.kernel.org Cc: linux-kernel@vger.kernel.org Cc: linux-imx@nxp.com Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
2020-03-19 23:12:28 +07:00
ctx->ctrldev = ctrldev;
ctx->rng.name = "rng-caam";
ctx->rng.init = caam_init;
ctx->rng.cleanup = caam_cleanup;
ctx->rng.read = caam_read;
ctx->rng.priv = (unsigned long)ctx;
ctx->rng.quality = 1024;
dev_info(ctrldev, "registering rng-caam\n");
ret = devm_hwrng_register(ctrldev, &ctx->rng);
if (ret) {
caam_rng_exit(ctrldev);
return ret;
}
devres_close_group(ctrldev, caam_rng_init);
return 0;
}