XTS Streaming .S and _c.c implementations, streaming callers & SVR brackets for aarch64

.wolfssl_known_macro_extras

@@ -143,6 +143,7 @@ CONFIG_IDF_TARGET_ESP32S3
 CONFIG_IDF_TARGET_ESP8266
 CONFIG_IDF_TARGET_ESP8684
 CONFIG_KASAN
+CONFIG_KERNEL_MODE_NEON
 CONFIG_KMSAN
 CONFIG_KPROBES
 CONFIG_MAIN_TASK_STACK_SIZE

configure.ac

@@ -12213,6 +12213,39 @@ then
 fi
 AC_SUBST([ENABLED_LINUXKM_LKCAPI_REGISTER])
 
+# The aarch64 ChaCha20 and Poly1305 assembly takes no vector register claim,
+# so a kernel module must not carry it.  Outside FIPS these algorithms are on
+# by default, so back them off with a warning when the user did not ask for
+# them, and only fail the build when they did.
+if test "$ENABLED_LINUXKM" = "yes" && test "$ENABLED_ARMASM" = "yes"
+then
+    case $host_cpu in
+    *aarch64*|*arm64*)
+        if test "$ENABLED_CHACHA" != "no" && test "$ENABLED_CHACHA" != "noasm" \
+           && test "$ENABLED_ASM" != "no"
+        then
+            if test "x$enable_chacha" = "x"
+            then
+                AC_MSG_WARN([ChaCha20 assembly holds no vector register claim on aarch64; building ChaCha20 without assembly for the kernel module.])
+                ENABLED_CHACHA=noasm
+            else
+                AC_MSG_ERROR([ChaCha20 assembly is unusable in a kernel module on aarch64: it holds no vector register claim.  Use --enable-chacha=noasm or --disable-chacha.])
+            fi
+        fi
+        if test "$ENABLED_POLY1305" != "no" && test "$ENABLED_ASM" != "no"
+        then
+            if test "x$enable_poly1305" = "x"
+            then
+                AC_MSG_WARN([Poly1305 assembly holds no vector register claim on aarch64 and has no assembly-free build; disabling Poly1305 for the kernel module.])
+                ENABLED_POLY1305=no
+            else
+                AC_MSG_ERROR([Poly1305 assembly is unusable in a kernel module on aarch64: it holds no vector register claim.  Use --disable-poly1305.])
+            fi
+        fi
+        ;;
+    esac
+fi
+
 # Library Suffix
 LIBSUFFIX=""
 AC_ARG_WITH([libsuffix],

linuxkm/Kbuild

@@ -264,6 +264,17 @@ $(obj)/wolfcrypt/src/wc_mldsa_asm.o: OBJECT_FILES_NON_STANDARD := y
 $(obj)/wolfcrypt/src/wc_falcon_fpr_x86_64_asm.o: asflags-y := $(WOLFSSL_ASFLAGS) $(ASFLAGS_FPU_DISABLE_SIMD_ENABLE)
 $(obj)/wolfcrypt/src/wc_falcon_fpr_x86_64_asm.o: OBJECT_FILES_NON_STANDARD := y
 
+# The arm64 kernel builds with a baseline -march that rejects the crypto and
+# NEON instructions in port/arm/*.S, and with -mgeneral-regs-only, which clang
+# also applies to inline asm.  Widen both for these objects.  An -march only
+# widens what the assembler accepts; it does not change what is emitted.
+ifeq ($(CONFIG_ARM64),y)
+$(obj)/wolfcrypt/src/port/arm/%.o: asflags-y := $(WOLFSSL_ASFLAGS) -march=armv8.2-a+crypto+sha3
+$(obj)/wolfcrypt/src/port/arm/%-asm_c.o: ccflags-y += -march=armv8.2-a+crypto+sha3
+$(obj)/wolfcrypt/src/port/arm/%-asm_c.o: ccflags-remove-y += -mgeneral-regs-only
+$(obj)/wolfcrypt/src/port/arm/%.o: OBJECT_FILES_NON_STANDARD := y
+endif
+
 ifndef READELF
     READELF := readelf
 endif

linuxkm/Makefile

@@ -62,7 +62,10 @@ endif
 
 WOLFSSL_ASFLAGS=-DHAVE_CONFIG_H -I$(SRC_TOP) -DBUILDING_WOLFSSL $(AM_CCASFLAGS) $(CCASFLAGS)
 
-WOLFSSL_OBJ_FILES=$(patsubst %.lo, %.o, $(patsubst src/src_libwolfssl_la-%, src/%, $(patsubst src/libwolfssl_la-%, src/%, $(patsubst wolfcrypt/src/src_libwolfssl_la-%, wolfcrypt/src/%, $(src_libwolfssl_la_OBJECTS)))))
+# Strip libtool's per-target object prefix so Kbuild sees the real object
+# names.  The innermost patsubst covers the wolfcrypt/src/port/arm/ assembly,
+# which the others do not reach.
+WOLFSSL_OBJ_FILES=$(patsubst %.lo, %.o, $(patsubst src/src_libwolfssl_la-%, src/%, $(patsubst src/libwolfssl_la-%, src/%, $(patsubst wolfcrypt/src/src_libwolfssl_la-%, wolfcrypt/src/%, $(patsubst wolfcrypt/src/port/arm/src_libwolfssl_la-%, wolfcrypt/src/port/arm/%, $(src_libwolfssl_la_OBJECTS))))))
 
 ifeq "$(ENABLED_CRYPT_TESTS)" "yes"
     WOLFSSL_OBJ_FILES+=wolfcrypt/test/test.o

linuxkm/arm64_vector_register_glue.c

@@ -0,0 +1,244 @@
+/* arm64_vector_register_glue.c: glue logic to claim and release the FPSIMD
+ * and NEON registers on arm64
+ *
+ * Copyright (C) 2006-2026 wolfSSL Inc.
+ *
+ * This file is part of wolfSSL.
+ *
+ * wolfSSL is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License as published by
+ * the Free Software Foundation; either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * wolfSSL is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1335, USA
+ */
+
+/* included by linuxkm/module_hooks.c */
+#ifndef WC_SKIP_INCLUDED_C_FILES
+
+#if !defined(WOLFSSL_USE_SAVE_VECTOR_REGISTERS) || !defined(CONFIG_ARM64)
+    #error arm64 vector register glue included in non-vectorized or non-arm64 project.
+#endif
+
+#ifndef CONFIG_KERNEL_MODE_NEON
+    /* Without this option the kernel exports no kernel_neon_begin() and
+     * may_use_simd() is always false (linux-6.6.99 fpsimd.c:1904, simd.h:46). */
+    #error wolfSSL linuxkm on arm64 requires CONFIG_KERNEL_MODE_NEON.
+#endif
+
+#if LINUX_VERSION_CODE >= KERNEL_VERSION(6, 19, 0)
+    /* Written against the void kernel_neon_begin() of linux-6.6.99.  6.19
+     * changes that signature and has not been read here. */
+    #error arm64 vector register glue does not yet support kernel_neon_begin() with a state buffer (6.19+).
+#endif
+
+#ifdef DEBUG_VECTOR_REGISTER_ACCESS_FUZZING
+    #error DEBUG_VECTOR_REGISTER_ACCESS_FUZZING is not implemented by the arm64 vector register glue.
+#endif
+
+/* kernel_neon_begin() BUGs unless may_use_simd(), then takes this CPU's FPSIMD
+ * context with bottom halves off (linux-6.6.99 fpsimd.c:1904 and :239,
+ * simd.h:26).  One record per CPU per context counts claim and inhibit depth. */
+
+struct wc_svr_arm64_ctx_state {
+    unsigned int depth;      /* open claims in this context on this CPU */
+    unsigned int inhibit_at; /* depth of the open inhibit claim, 0 when none */
+    unsigned int neon_held;  /* 1 between kernel_neon_begin() and _end() */
+    unsigned int bh_held;    /* 1 while a pin-only section holds local_bh_disable() */
+};
+
+struct wc_svr_arm64_cpu_state {
+    struct wc_svr_arm64_ctx_state ctx[2]; /* [0] task, [1] softirq */
+};
+
+static DEFINE_PER_CPU(struct wc_svr_arm64_cpu_state, wc_svr_arm64_state);
+static atomic64_t wc_svr_disallowed_count = ATOMIC64_INIT(0);
+
+#if LINUX_VERSION_CODE < KERNEL_VERSION(5, 11, 0)
+    #define wc_svr_arm64_in_hardirq() in_irq()
+#else
+    #define wc_svr_arm64_in_hardirq() in_hardirq()
+#endif
+
+/* Keep softirqs off a section that holds no registers, as the x86 glue does.
+ * With interrupts already off, nothing can arrive anyway. */
+static inline void wc_svr_arm64_pin_bh(struct wc_svr_arm64_ctx_state *st)
+{
+    if (! irqs_disabled()) {
+        local_bh_disable();
+        st->bh_held = 1;
+    }
+}
+
+/* The record for this context on this CPU.  Every caller has the CPU held,
+ * by preempt_disable() here or by the open claim's bottom-half disable. */
+static inline struct wc_svr_arm64_ctx_state *wc_svr_arm64_here(void)
+{
+    return &this_cpu_ptr(&wc_svr_arm64_state)->ctx[in_serving_softirq() ? 1 : 0];
+}
+
+/* The records are static.  These entry points exist because wc_port.c and
+ * the redirect table expect them. */
+__must_check int wc_linuxkm_allocate_svr_states(void)
+{
+    return 0;
+}
+
+void wc_linuxkm_free_svr_states(void)
+{
+}
+
+void wc_svr_disallowed_count_reset(void)
+{
+    atomic64_set(&wc_svr_disallowed_count, 0);
+}
+
+__must_check unsigned long long int wc_svr_disallowed_count_current(void)
+{
+    return (unsigned long long int)atomic64_read(&wc_svr_disallowed_count);
+}
+
+/* Nonzero when a claim made now would succeed. */
+__must_check int wc_can_save_vector_registers_x86(void)
+{
+    struct wc_svr_arm64_ctx_state *st;
+    int ret;
+
+    if (in_nmi() || wc_svr_arm64_in_hardirq())
+        return 0;
+
+    preempt_disable();
+    st = wc_svr_arm64_here();
+    if (st->depth > 0)
+        ret = (st->inhibit_at == 0);
+    else
+        ret = may_use_simd() ? 1 : 0;
+    preempt_enable();
+
+    return ret;
+}
+
+__must_check int wc_save_vector_registers_x86(enum wc_svr_flags flags)
+{
+    struct wc_svr_arm64_ctx_state *st;
+
+    if (in_nmi() || wc_svr_arm64_in_hardirq()) {
+        /* may_use_simd() is false here, and any open record belongs to the
+         * context this interrupt landed on. */
+        atomic64_inc(&wc_svr_disallowed_count);
+        return WC_ACCEL_INHIBIT_E;
+    }
+
+    preempt_disable();
+    st = wc_svr_arm64_here();
+
+    if (st->depth > 0) {
+        /* Nested in this context's own section, which already holds the
+         * CPU, so the preempt_disable() above is balanced, not carried. */
+        if (flags & WC_SVR_FLAG_MAYBE_INHIBIT) {
+            /* Pin-only claims are outermost only, as in the x86 glue. */
+            preempt_enable();
+            atomic64_inc(&wc_svr_disallowed_count);
+            return BAD_STATE_E;
+        }
+        if (st->inhibit_at != 0) {
+            preempt_enable();
+            atomic64_inc(&wc_svr_disallowed_count);
+            return WC_ACCEL_INHIBIT_E;
+        }
+        ++st->depth;
+        if (flags & WC_SVR_FLAG_INHIBIT)
+            st->inhibit_at = st->depth;
+        preempt_enable();
+        return 0;
+    }
+
+    /* Outermost claim.  The bottom-half disable below is what holds the CPU,
+     * taken by kernel_neon_begin() or wc_svr_arm64_pin_bh(). */
+    if (flags & WC_SVR_FLAG_INHIBIT) {
+        wc_svr_arm64_pin_bh(st);
+        st->depth = 1;
+        st->inhibit_at = 1;
+        /* Counted as the x86 glue counts it: a span in which claims are
+         * refused, opened on request. */
+        atomic64_inc(&wc_svr_disallowed_count);
+        return 0;
+    }
+
+    if (! may_use_simd()) {
+        if (flags & WC_SVR_FLAG_MAYBE_INHIBIT) {
+            /* Held without the registers: nested claims are refused. */
+            wc_svr_arm64_pin_bh(st);
+            st->depth = 1;
+            st->inhibit_at = 1;
+            atomic64_inc(&wc_svr_disallowed_count);
+            return 0;
+        }
+        preempt_enable();
+        atomic64_inc(&wc_svr_disallowed_count);
+        return WC_ACCEL_INHIBIT_E;
+    }
+
+    kernel_neon_begin();
+    st->depth = 1;
+    st->neon_held = 1;
+    return 0;
+}
+
+void wc_restore_vector_registers_x86(enum wc_svr_flags flags)
+{
+    struct wc_svr_arm64_ctx_state *st;
+
+    if (in_nmi() || wc_svr_arm64_in_hardirq()) {
+        wc_linuxkm_pr_err_ratelimited("BUG: wc_restore_vector_registers_x86() "
+            "called from %s on CPU %d, where no claim can be open.\n",
+            in_nmi() ? "NMI" : "hardirq", raw_smp_processor_id());
+        return;
+    }
+
+    preempt_disable();
+    st = wc_svr_arm64_here();
+    if (st->depth == 0) {
+        preempt_enable();
+        wc_linuxkm_pr_err_ratelimited("BUG: wc_restore_vector_registers_x86() "
+            "with no open claim on CPU %d.\n", raw_smp_processor_id());
+        return;
+    }
+    /* The open claim already holds the CPU; balance the disable above. */
+    preempt_enable();
+
+    if (st->inhibit_at == st->depth) {
+        if (! (flags & (WC_SVR_FLAG_INHIBIT | WC_SVR_FLAG_MAYBE_INHIBIT))) {
+            wc_linuxkm_pr_err_ratelimited("BUG: wc_restore_vector_registers_x86() "
+                "closing an inhibit claim without an inhibit flag on CPU %d.\n",
+                raw_smp_processor_id());
+        }
+        st->inhibit_at = 0;
+    }
+    else if (flags & WC_SVR_FLAG_INHIBIT) {
+        wc_linuxkm_pr_err_ratelimited("BUG: wc_restore_vector_registers_x86() "
+            "with the inhibit flag but no matching inhibit claim on CPU %d.\n",
+            raw_smp_processor_id());
+    }
+
+    if (--st->depth == 0) {
+        if (st->neon_held) {
+            st->neon_held = 0;
+            kernel_neon_end();
+        }
+        if (st->bh_held) {
+            st->bh_held = 0;
+            local_bh_enable();
+        }
+        preempt_enable();
+    }
+}
+
+#endif /* !WC_SKIP_INCLUDED_C_FILES */

linuxkm/include.am

@@ -16,6 +16,7 @@ EXTRA_DIST += m4/ax_linuxkm.m4 \
 	      linuxkm/linuxkm_memory.h \
 	      linuxkm/linuxkm_wc_port.h \
 	      linuxkm/x86_vector_register_glue.c \
+	      linuxkm/arm64_vector_register_glue.c \
 	      linuxkm/lkcapi_glue.c \
 	      linuxkm/lkcapi_aes_glue.c \
 	      linuxkm/lkcapi_sha_glue.c \