Skip to contentSkip to conversation

DM: crypto/fsl: fix unaligned exponent handling in fsl_rsa

PATCH In review Oleksandr Suvorov 2 participants 2 messages Apr 5 – Apr 9 Crypto API

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 DM: crypto/fsl: fix unaligned exponent handling in fsl_rsa

The RSA verification path loads the public exponent from the FIT key node with fdt_getprop(). That property buffer is not guaranteed to be naturally aligned. On ARM64 this can lead to a synchronous abort when fsl_rsa uses the exponent value directly during verified boot.

Fix the driver by copying the exponent into an aligned local buffer before building the CAAM RSA job descriptor. Read the value with the pointer-safe fdt64_to_cpu() helper and pass the aligned copy to the engine instead of the original FDT property buffer.

Also flush the cache for the aligned exponent buffer actually consumed by CAAM so the cache maintenance matches the data path used by the hardware.

This keeps RSA verification functional while avoiding crashes caused by unaligned FDT property access.

[1]
---

Trying to boot from MMC2

    sha256,rsa2048:dev"Synchronous Abort" handler, esr 0x96000021
    elr: 00000000007e737c lr : 00000000007e6e78
    x 0: 00000000007fd1e4 x 1: 0000000000000100
    x 2: 0000000000000800 x 3: 000000000091d420
    x 4: 000000000091d420 x 5: 000000000091d588
    x 6: 00000000401ff9b8 x 7: 0000000000000003
    x 8: 000000000091d49c x 9: 0000000000000002
...
    Code: f94008a0 2a0103f4 aa0303f5 b40005c0 (f9400003)
    Resetting CPU ...

---

Fixes 34276478f7b7 ("DM: crypto/fsl - Add Freescale rsa DM driver")
---
 drivers/crypto/fsl/fsl_rsa.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
drivers/crypto/fsl/fsl_rsa.c +3 -3
index 125a72ae6d3..053e1862bc4 100644
@@ -23,6 +23,7 @@ int fsl_mod_exp(struct udevice *dev, const uint8_t *sig, uint32_t sig_len,
 	struct pk_in_params pkin;
 	uint32_t desc[MAX_CAAM_DESCSIZE];
 	int ret;
+	fdt64_t exp = fdt64_to_cpu(prop->public_exponent);
 
 	/* Length in bytes */
 	keylen = prop->num_bits / 8;
@@ -31,15 +32,14 @@ int fsl_mod_exp(struct udevice *dev, const uint8_t *sig, uint32_t sig_len,
 	pkin.a_siz = sig_len;
 	pkin.n = prop->modulus;
 	pkin.n_siz = keylen;
-	pkin.e = prop->public_exponent;
+	pkin.e = (void *)&exp;
 	pkin.e_siz = prop->exp_len;
 
 	inline_cnstr_jobdesc_pkha_rsaexp(desc, &pkin, out, sig_len);
 
 	flush_dcache_range((ulong)sig, (ulong)sig + sig_len);
 	flush_dcache_range((ulong)prop->modulus, (ulong)(prop->modulus) + keylen);
-	flush_dcache_range((ulong)prop->public_exponent,
-			   (ulong)(prop->public_exponent) + prop->exp_len);
+	flush_dcache_range((ulong)&exp, (ulong)&exp + prop->exp_len);
 	flush_dcache_range((ulong)desc, (ulong)desc + (sizeof(uint32_t) * MAX_CAAM_DESCSIZE));
 	flush_dcache_range((ulong)out, (ulong)out + sig_len);
 
Signature · 4 lines
2.43.0

base-commit: 214aababe07de0598bd26ecb0c8fb2e2843ff7d5
branch: master-20260404-rsa-verify
#1 lore raw
Quoted 44 lines · Oleksandr Suvorov

The RSA verification path loads the public exponent from the FIT key node with fdt_getprop(). That property buffer is not guaranteed to be naturally aligned. On ARM64 this can lead to a synchronous abort when fsl_rsa uses the exponent value directly during verified boot.

Fix the driver by copying the exponent into an aligned local buffer before building the CAAM RSA job descriptor. Read the value with the pointer-safe fdt64_to_cpu() helper and pass the aligned copy to the engine instead of the original FDT property buffer.

Also flush the cache for the aligned exponent buffer actually consumed by CAAM so the cache maintenance matches the data path used by the hardware.

This keeps RSA verification functional while avoiding crashes caused by unaligned FDT property access.

[1]
---

Trying to boot from MMC2
sha256,rsa2048:dev"Synchronous Abort" handler, esr 0x96000021
elr: 00000000007e737c lr : 00000000007e6e78
x 0: 00000000007fd1e4 x 1: 0000000000000100
x 2: 0000000000000800 x 3: 000000000091d420
x 4: 000000000091d420 x 5: 000000000091d588
x 6: 00000000401ff9b8 x 7: 0000000000000003
x 8: 000000000091d49c x 9: 0000000000000002
...

   Code: f94008a0 2a0103f4 aa0303f5 b40005c0 (f9400003)
   Resetting CPU ...

---

Fixes 34276478f7b7 ("DM: crypto/fsl - Add Freescale rsa DM driver")

---
drivers/crypto/fsl/fsl_rsa.c | 6 +++---

1 file changed, 3 insertions(+), 3 deletions(-)
drivers/crypto/fsl/fsl_rsa.c +0 -0
index 125a72ae6d3..053e1862bc4 100644

drivers/crypto/fsl/fsl_rsa.c int fsl_mod_exp(struct udevice *dev, const uint8_t *sig, uint32_t sig_len, · line 26

Quoted 4 lines
drivers/crypto/fsl/fsl_rsa.c +1 -0
@@ -23,6 +23,7 @@ int fsl_mod_exp(struct udevice *dev, const uint8_t *sig, uint32_t sig_len,
 	struct pk_in_params pkin;
 	uint32_t desc[MAX_CAAM_DESCSIZE];
 	int ret;
+	fdt64_t exp = fdt64_to_cpu(prop->public_exponent);

Build warning->error.

+In file included from include/linux/byteorder/little_endian.h:14,

+                 from arch/arm/include/asm/byteorder.h:28,
+                 from include/linux/libfdt_env.h:15,
+                 from include/linux/libfdt.h:6,
+                 from include/fdtdec.h:17,
+                 from include/dm/ofnode.h:11,
+                 from include/dm/device.h:13,
+                 from include/dm.h:13,
+                 from drivers/crypto/fsl/fsl_rsa.c:9:

+drivers/crypto/fsl/fsl_rsa.c: In function 'fsl_mod_exp':
+include/linux/byteorder/little_endian.h:39:50: error: cast from pointer to integer of different size [-Werror=pointer-to-int-cast]
+ 39 | #define __be64_to_cpu(x) __swab64((__force __u64)(__be64)(x))
+ | ^
+include/linux/byteorder/swab.h:92:31: note: in definition of macro '__swab64'
+ 92 | (__builtin_constant_p((__u64)(x)) ? \
+ | ^
+include/linux/byteorder/generic.h:93:21: note: in expansion of macro '__be64_to_cpu'
+ 93 | #define be64_to_cpu __be64_to_cpu
+ | ^~~~~~~~~~~~~
+include/linux/libfdt_env.h:24:25: note: in expansion of macro 'be64_to_cpu'
+ 24 | #define fdt64_to_cpu(x) be64_to_cpu(x)
+ | ^~~~~~~~~~~
+drivers/crypto/fsl/fsl_rsa.c:26:23: note: in expansion of macro 'fdt64_to_cpu'
+ 26 | fdt64_t exp = fdt64_to_cpu(prop->public_exponent);
+ | ^~~~~~~~~~~~
+include/linux/byteorder/swab.h:33:34: note: in definition of macro '___swab64'
+ 33 | (__u64)(((__u64)(x) & (__u64)0x00000000000000ffULL) << 56) | \
+ | ^
+include/linux/byteorder/little_endian.h:39:26: note: in expansion of macro '__swab64'
+ | ^~~~~~~~
+include/linux/byteorder/swab.h:34:34: note: in definition of macro '___swab64'
+ 34 | (__u64)(((__u64)(x) & (__u64)0x000000000000ff00ULL) << 40) | \
+include/linux/byteorder/swab.h:35:34: note: in definition of macro '___swab64'
+ 35 | (__u64)(((__u64)(x) & (__u64)0x0000000000ff0000ULL) << 24) | \
+include/linux/byteorder/swab.h:36:34: note: in definition of macro '___swab64'
+ 36 | (__u64)(((__u64)(x) & (__u64)0x00000000ff000000ULL) << 8) | \
+include/linux/byteorder/swab.h:37:34: note: in definition of macro '___swab64'
+ 37 | (__u64)(((__u64)(x) & (__u64)0x000000ff00000000ULL) >> 8) | \
+include/linux/byteorder/swab.h:38:34: note: in definition of macro '___swab64'
+ 38 | (__u64)(((__u64)(x) & (__u64)0x0000ff0000000000ULL) >> 24) | \
+include/linux/byteorder/swab.h:93:2: note: in expansion of macro '___swab64'

+   93 |  ___swab64((x)) : \
+      |  ^~~~~~~~~

+include/linux/byteorder/swab.h:40:34: note: in definition of macro '___swab64'
+ 40 | (__u64)(((__u64)(x) & (__u64)0xff00000000000000ULL) >> 56) ))
+include/linux/byteorder/swab.h:94:13: note: in definition of macro '__swab64'
+ 94 | __fswab64((x)))
+ | ^
+cc1: all warnings being treated as errors
+make[4]: *** [scripts/Makefile.build:271: drivers/crypto/fsl/fsl_rsa.o] Error

Regards
Peng