linuxkm/linuxkm_wc_port.h: add default setup for LINUXKM_LKCAPI_REGISTER_HASH_DRBG_DEFAULT, to make visibility in random.c;

linuxkm/lkcapi_sha_glue.c: revert f7c7ac2 (get_drbg() DISABLE_VECTOR_REGISTERS() for crypto_default_rng) -- compiler/inlining bug makes it break on at least one target, so caller needs to retain responsibility;

linuxkm/x86_vector_register_glue.c: in wc_save_vector_registers_x86(), always return WC_ACCEL_INHIBIT_E if already fpu_state & WC_FPU_INHIBITED_FLAG, for safe+correct dynamics on recursive calls.

Author: douzzer

linuxkm/linuxkm_wc_port.h

@@ -422,6 +422,17 @@
             #define WC_AES_XTS_SUPPORT_SIMULTANEOUS_ENC_AND_DEC_KEYS
         #endif
 
+        /* setup for LINUXKM_LKCAPI_REGISTER_HASH_DRBG_DEFAULT needs to be here
+         * to assure that calls to get_random_bytes() in random.c are gated out
+         * (they would recurse, potentially infinitely).
+         */
+        #if (defined(LINUXKM_LKCAPI_REGISTER_ALL) && \
+             !defined(LINUXKM_LKCAPI_DONT_REGISTER_HASH_DRBG) && \
+             !defined(LINUXKM_LKCAPI_DONT_REGISTER_HASH_DRBG_DEFAULT)) && \
+            !defined(LINUXKM_LKCAPI_REGISTER_HASH_DRBG_DEFAULT)
+            #define LINUXKM_LKCAPI_REGISTER_HASH_DRBG_DEFAULT
+        #endif
+
         #ifndef __PIE__
             #include <linux/crypto.h>
             #include <linux/scatterlist.h>

linuxkm/lkcapi_sha_glue.c

@@ -374,10 +374,7 @@
         !defined(LINUXKM_LKCAPI_REGISTER_HASH_DRBG)
         #define LINUXKM_LKCAPI_REGISTER_HASH_DRBG
     #endif
-    #if (defined(LINUXKM_LKCAPI_REGISTER_ALL) && !defined(LINUXKM_LKCAPI_DONT_REGISTER_HASH_DRBG_DEFAULT)) && \
-        !defined(LINUXKM_LKCAPI_REGISTER_HASH_DRBG_DEFAULT)
-        #define LINUXKM_LKCAPI_REGISTER_HASH_DRBG_DEFAULT
-    #endif
+    /* setup for LINUXKM_LKCAPI_REGISTER_HASH_DRBG_DEFAULT is in linuxkm_wc_port.h */
 #else
     #undef LINUXKM_LKCAPI_REGISTER_HASH_DRBG
 #endif
@@ -968,7 +965,6 @@ struct wc_linuxkm_drbg_ctx {
     struct wc_rng_inst {
         wolfSSL_Atomic_Int lock;
         WC_RNG rng;
-        int disabled_vec_ops;
     } *rngs; /* one per CPU ID */
 };
 
@@ -1090,14 +1086,8 @@ static inline struct wc_rng_inst *get_drbg(struct crypto_rng *tfm) {
 
     for (;;) {
         int expected = 0;
-        if (likely(__atomic_compare_exchange_n(&ctx->rngs[n].lock, &expected, new_lock_value, 0, __ATOMIC_SEQ_CST, __ATOMIC_ACQUIRE))) {
-            struct wc_rng_inst *drbg = &ctx->rngs[n];
-            if (tfm == crypto_default_rng)
-                drbg->disabled_vec_ops = (DISABLE_VECTOR_REGISTERS() == 0);
-            else
-                drbg->disabled_vec_ops = 0;
-            return drbg;
-        }
+        if (likely(__atomic_compare_exchange_n(&ctx->rngs[n].lock, &expected, new_lock_value, 0, __ATOMIC_SEQ_CST, __ATOMIC_ACQUIRE)))
+            return &ctx->rngs[n];
         ++n;
         if (n >= (int)ctx->n_rngs)
             n = 0;
@@ -1115,11 +1105,8 @@ static inline struct wc_rng_inst *get_drbg_n(struct wc_linuxkm_drbg_ctx *ctx, in
 
     for (;;) {
         int expected = 0;
-        if (likely(__atomic_compare_exchange_n(&ctx->rngs[n].lock, &expected, 1, 0, __ATOMIC_SEQ_CST, __ATOMIC_ACQUIRE))) {
-            struct wc_rng_inst *drbg = &ctx->rngs[n];
-            drbg->disabled_vec_ops = 0;
-            return drbg;
-        }
+        if (likely(__atomic_compare_exchange_n(&ctx->rngs[n].lock, &expected, 1, 0, __ATOMIC_SEQ_CST, __ATOMIC_ACQUIRE)))
+            return &ctx->rngs[n];
         if (can_sleep) {
             if (signal_pending(current))
                 return NULL;
@@ -1137,10 +1124,6 @@ static inline void put_drbg(struct wc_rng_inst *drbg) {
         (LINUX_VERSION_CODE >= KERNEL_VERSION(5, 7, 0))
     int migration_disabled = (drbg->lock == 2);
     #endif
-    if (drbg->disabled_vec_ops) {
-        REENABLE_VECTOR_REGISTERS();
-        drbg->disabled_vec_ops = 0;
-    }
     __atomic_store_n(&(drbg->lock),0,__ATOMIC_RELEASE);
     #if defined(CONFIG_SMP) && !defined(CONFIG_PREEMPT_COUNT) && \
         (LINUX_VERSION_CODE >= KERNEL_VERSION(5, 7, 0))
@@ -1154,13 +1137,19 @@ static int wc_linuxkm_drbg_generate(struct crypto_rng *tfm,
                         u8 *dst, unsigned int dlen)
 {
     int ret, retried = 0;
+    int need_fpu_restore;
     struct wc_rng_inst *drbg = get_drbg(tfm);
 
     if (! drbg) {
         pr_err_once("BUG: get_drbg() failed.");
         return -EFAULT;
     }
 
+    /* for the default RNG, make sure we don't cache an underlying SHA256
+     * method that uses vector insns (forbidden from irq handlers).
+     */
+    need_fpu_restore = (tfm == crypto_default_rng) ? (DISABLE_VECTOR_REGISTERS() == 0) : 0;
+
 retry:
 
     if (slen > 0) {
@@ -1194,6 +1183,8 @@ static int wc_linuxkm_drbg_generate(struct crypto_rng *tfm,
 
 out:
 
+    if (need_fpu_restore)
+        REENABLE_VECTOR_REGISTERS();
     put_drbg(drbg);
 
     return ret;

linuxkm/x86_vector_register_glue.c

@@ -346,24 +346,20 @@ WARN_UNUSED_RESULT int wc_save_vector_registers_x86(enum wc_svr_flags flags)
 
     /* allow for nested calls */
     if (pstate && (pstate->fpu_state != 0U)) {
+        if (pstate->fpu_state & WC_FPU_INHIBITED_FLAG) {
+            /* don't allow recursive inhibit calls when already inhibited --
+             * it would add no functionality and require keeping a separate
+             * count of inhibit recursions.
+             */
+            return WC_ACCEL_INHIBIT_E;
+        }
         if (unlikely((pstate->fpu_state & WC_FPU_COUNT_MASK)
                      == WC_FPU_COUNT_MASK))
         {
             pr_err("ERROR: wc_save_vector_registers_x86 recursion register overflow for "
                    "pid %d on CPU %d.\n", pstate->pid, raw_smp_processor_id());
             return BAD_STATE_E;
         }
-        if (pstate->fpu_state & WC_FPU_INHIBITED_FLAG) {
-            if (flags & WC_SVR_FLAG_INHIBIT) {
-                /* allow recursive inhibit calls as long as the whole stack of
-                 * them is inhibiting.
-                 */
-                ++pstate->fpu_state;
-                return 0;
-            }
-            else
-                return WC_ACCEL_INHIBIT_E;
-        }
         if (flags & WC_SVR_FLAG_INHIBIT) {
             ++pstate->fpu_state;
             pstate->fpu_state |= WC_FPU_INHIBITED_FLAG;