From: Chuck Ebbert Subject: [patch 1/3] crypto: padlock-aes: work around Nano CPU errata in ECB mode Date: Tue, 9 Jun 2009 10:36:50 -0400 Message-ID: <20090609103650.0e4b0405@dhcp-100-2-144.bos.redhat.com> References: <20090609103533.0c334594@dhcp-100-2-144.bos.redhat.com> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Cc: Herbert Xu , linux-kernel@vger.kernel.org, linux-crypto@vger.kernel.org, Sebastian Andrzej Siewior To: Harald Welte Return-path: Received: from mx2.redhat.com ([66.187.237.31]:48568 "EHLO mx2.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753213AbZFIOhn (ORCPT ); Tue, 9 Jun 2009 10:37:43 -0400 In-Reply-To: <20090609103533.0c334594@dhcp-100-2-144.bos.redhat.com> Sender: linux-crypto-owner@vger.kernel.org List-ID: From: Chuck Ebbert crypto: padlock-aes: work around Nano CPU errata in ECB mode The VIA Nano processor has a bug that makes it prefetch extra data during encryption operations, causing spurious page faults. Extend existing workarounds for ECB mode to copy the data to an temporary buffer to avoid the problem. Signed-off-by: Chuck Ebbert --- work-2.6.29.4.orig/drivers/crypto/padlock-aes.c +++ work-2.6.29.4/drivers/crypto/padlock-aes.c @@ -18,9 +18,17 @@ #include #include #include +#include #include #include "padlock.h" +/* number of data blocks actually fetched for each xcrypt insn */ +static unsigned int ecb_fetch_blocks = 2; +static unsigned int cbc_fetch_blocks = 1; + +#define ecb_fetch_bytes (ecb_fetch_blocks * AES_BLOCK_SIZE) +#define cbc_fetch_bytes (cbc_fetch_blocks * AES_BLOCK_SIZE) + /* Control word. */ struct cword { unsigned int __attribute__ ((__packed__)) @@ -169,54 +177,59 @@ static inline void padlock_store_cword(s */ static inline void padlock_xcrypt(const u8 *input, u8 *output, void *key, - struct cword *control_word) + struct cword *control_word, int count) { asm volatile (".byte 0xf3,0x0f,0xa7,0xc8" /* rep xcryptecb */ : "+S"(input), "+D"(output) - : "d"(control_word), "b"(key), "c"(1)); + : "d"(control_word), "b"(key), "c"(count)); } -static void aes_crypt_copy(const u8 *in, u8 *out, u32 *key, struct cword *cword) +static void aes_crypt_copy(const u8 *in, u8 *out, u32 *key, + struct cword *cword, int count) { - u8 buf[AES_BLOCK_SIZE * 2 + PADLOCK_ALIGNMENT - 1]; + /* + * Padlock prefetches extra data so we must provide mapped input buffers. + * Assume there are at least 16 bytes of stack already in use. + */ + u8 buf[AES_BLOCK_SIZE * 7 + PADLOCK_ALIGNMENT - 1]; u8 *tmp = PTR_ALIGN(&buf[0], PADLOCK_ALIGNMENT); - memcpy(tmp, in, AES_BLOCK_SIZE); - padlock_xcrypt(tmp, out, key, cword); + memcpy(tmp, in, count * AES_BLOCK_SIZE); + padlock_xcrypt(tmp, out, key, cword, count); } static inline void aes_crypt(const u8 *in, u8 *out, u32 *key, - struct cword *cword) + struct cword *cword, int count) { - /* padlock_xcrypt requires at least two blocks of data. */ - if (unlikely(!(((unsigned long)in ^ (PAGE_SIZE - AES_BLOCK_SIZE)) & - (PAGE_SIZE - 1)))) { - aes_crypt_copy(in, out, key, cword); + /* Padlock in ECB mode fetches at least ecb_fetch_bytes of data. + * We could avoid some copying here but it's probably not worth it. + */ + if (unlikely(((unsigned long)in & PAGE_SIZE) + ecb_fetch_bytes > PAGE_SIZE)) { + aes_crypt_copy(in, out, key, cword, count); return; } - padlock_xcrypt(in, out, key, cword); + padlock_xcrypt(in, out, key, cword, count); } static inline void padlock_xcrypt_ecb(const u8 *input, u8 *output, void *key, void *control_word, u32 count) { - if (count == 1) { - aes_crypt(input, output, key, control_word); + u32 initial = count & (ecb_fetch_blocks - 1); + + if (count < ecb_fetch_blocks) { + aes_crypt(input, output, key, control_word, count); return; } - asm volatile ("test $1, %%cl;" - "je 1f;" - "lea -1(%%ecx), %%eax;" - "mov $1, %%ecx;" - ".byte 0xf3,0x0f,0xa7,0xc8;" /* rep xcryptecb */ - "mov %%eax, %%ecx;" - "1:" - ".byte 0xf3,0x0f,0xa7,0xc8" /* rep xcryptecb */ + if (initial) + asm volatile (".byte 0xf3,0x0f,0xa7,0xc8" /* rep xcryptecb */ + : "+S"(input), "+D"(output) + : "d"(control_word), "b"(key), "c"(initial)); + + asm volatile (".byte 0xf3,0x0f,0xa7,0xc8" /* rep xcryptecb */ : "+S"(input), "+D"(output) - : "d"(control_word), "b"(key), "c"(count) - : "ax"); + : "d"(control_word), "b"(key), "c"(count - initial)); } static inline u8 *padlock_xcrypt_cbc(const u8 *input, u8 *output, void *key, @@ -236,7 +249,7 @@ static void aes_encrypt(struct crypto_tf padlock_reset_key(&ctx->cword.encrypt); ts_state = irq_ts_save(); - aes_crypt(in, out, ctx->E, &ctx->cword.encrypt); + aes_crypt(in, out, ctx->E, &ctx->cword.encrypt, 1); irq_ts_restore(ts_state); padlock_store_cword(&ctx->cword.encrypt); } @@ -248,7 +261,7 @@ static void aes_decrypt(struct crypto_tf padlock_reset_key(&ctx->cword.encrypt); ts_state = irq_ts_save(); - aes_crypt(in, out, ctx->D, &ctx->cword.decrypt); + aes_crypt(in, out, ctx->D, &ctx->cword.decrypt, 1); irq_ts_restore(ts_state); padlock_store_cword(&ctx->cword.encrypt); } @@ -441,6 +454,7 @@ static struct crypto_alg cbc_aes_alg = { static int __init padlock_init(void) { int ret; + struct cpuinfo_x86 *c = &cpu_data(0); if (!cpu_has_xcrypt) { printk(KERN_NOTICE PFX "VIA PadLock not detected.\n"); @@ -463,6 +477,12 @@ static int __init padlock_init(void) printk(KERN_NOTICE PFX "Using VIA PadLock ACE for AES algorithm.\n"); + if (c->x86 == 6 && c->x86_model == 15 && c->x86_mask == 2) { + ecb_fetch_blocks = 8; + cbc_fetch_blocks = 4; /* NOTE: notused */ + printk(KERN_NOTICE PFX "VIA Nano stepping 2 detected: enabling workaround.\n"); + } + out: return ret;