@@ -25,6 +25,8 @@
#define GHASH_BLOCK_SIZE 16
#define GHASH_DIGEST_SIZE 16
+#define FPU_BYTES 4096U /* avoid kernel_fpu_begin/end scheduler/rcu stalls */
+
void clmul_ghash_mul(char *dst, const u128 *shash);
void clmul_ghash_update(char *dst, const char *src, unsigned int srclen,
@@ -81,10 +83,11 @@ static int ghash_update(struct shash_desc *desc,
struct ghash_desc_ctx *dctx = shash_desc_ctx(desc);
struct ghash_ctx *ctx = crypto_shash_ctx(desc->tfm);
u8 *dst = dctx->buffer;
+ unsigned int fpulen;
if (dctx->bytes) {
int n = min(srclen, dctx->bytes);
- u8 *pos = dst + (GHASH_BLOCK_SIZE - dctx->bytes);
+ u8 *pos = dst + GHASH_BLOCK_SIZE - dctx->bytes;
dctx->bytes -= n;
srclen -= n;
@@ -99,13 +102,23 @@ static int ghash_update(struct shash_desc *desc,
}
}
- kernel_fpu_begin();
- clmul_ghash_update(dst, src, srclen, &ctx->shash);
- kernel_fpu_end();
+ while (srclen >= GHASH_BLOCK_SIZE) {
+ fpulen = min(srclen, FPU_BYTES);
+
+ kernel_fpu_begin();
+ while (fpulen >= GHASH_BLOCK_SIZE) {
+ int n = min_t(unsigned int, fpulen, GHASH_BLOCK_SIZE);
+
+ clmul_ghash_update(dst, src, n, &ctx->shash);
+
+ srclen -= n;
+ fpulen -= n;
+ src += n;
+ }
+ kernel_fpu_end();
+ }
- if (srclen & 0xf) {
- src += srclen - (srclen & 0xf);
- srclen &= 0xf;
+ if (srclen) {
dctx->bytes = GHASH_BLOCK_SIZE - srclen;
while (srclen--)
*dst++ ^= *src++;
As done by the ECB and CBC helpers in arch/x86/crypt/ecb_cbc_helpers.h, limit the number of bytes processed between kernel_fpu_begin() and kernel_fpu_end() calls. Those functions call preempt_disable() and preempt_enable(), so the CPU core is unavailable for scheduling while running, leading to: rcu: INFO: rcu_preempt detected expedited stalls on CPUs/tasks: {12-... } 22 jiffies s: 277 root: 0x1/. Fixes: 0e1227d356e9 ("crypto: ghash - Add PCLMULQDQ accelerated implementation") Suggested-by: Herbert Xu <herbert@gondor.apana.org.au> Signed-off-by: Robert Elliott <elliott@hpe.com> --- arch/x86/crypto/ghash-clmulni-intel_glue.c | 27 ++++++++++++++++------ 1 file changed, 20 insertions(+), 7 deletions(-)