Linux 4.14, 4.19, 5.0+ compat: SIMD save/restore

Contrary to initial testing we cannot rely on these kernels to
invalidate the per-cpu FPU state and restore the FPU registers.
Nor can we guarantee that the kernel won't modify the FPU state
which we saved in the task struck.

Therefore, the kfpu_begin() and kfpu_end() functions have been
updated to save and restore the FPU state using our own dedicated
per-cpu FPU state variables.

This has the additional advantage of allowing us to use the FPU
again in user threads.  So we remove the code which was added to
use task queues to ensure some functions ran in kernel threads.

Reviewed-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #9346
Closes #9403

config/kernel-fpu.m4

@@ -2,15 +2,9 @@ dnl #
 dnl # Handle differences in kernel FPU code.
 dnl #
 dnl # Kernel
-dnl # 5.2:	The fpu->initialized flag was replaced by TIF_NEED_FPU_LOAD.
+dnl # 5.0:	Wrappers have been introduced to save/restore the FPU state.
-dnl #		HAVE_KERNEL_TIF_NEED_FPU_LOAD
-dnl #
-dnl # 5.0:	As an optimization SIMD operations performed by kernel
-dnl #		threads can skip saving and restoring their FPU context.
-dnl #		Wrappers have been introduced to determine the running
-dnl #		context and use either the SIMD or generic implementation.
 dnl #		This change was made to the 4.19.38 and 4.14.120 LTS kernels.
-dnl #		HAVE_KERNEL_FPU_INITIALIZED
+dnl #		HAVE_KERNEL_FPU_INTERNAL
 dnl #
 dnl # 4.2:	Use __kernel_fpu_{begin,end}()
 dnl #		HAVE_UNDERSCORE_KERNEL_FPU & KERNEL_EXPORTS_X86_FPU
@@ -38,6 +32,7 @@ AC_DEFUN([ZFS_AC_KERNEL_FPU_HEADER], [
 
 AC_DEFUN([ZFS_AC_KERNEL_SRC_FPU], [
 	ZFS_LINUX_TEST_SRC([kernel_fpu], [
+		#include <linux/types.h>
 		#ifdef HAVE_KERNEL_FPU_API_HEADER
 		#include <asm/fpu/api.h>
 		#else
@@ -50,6 +45,7 @@ AC_DEFUN([ZFS_AC_KERNEL_SRC_FPU], [
 	], [], [$ZFS_META_LICENSE])
 
 	ZFS_LINUX_TEST_SRC([__kernel_fpu], [
+		#include <linux/types.h>
 		#ifdef HAVE_KERNEL_FPU_API_HEADER
 		#include <asm/fpu/api.h>
 		#else
@@ -61,22 +57,41 @@ AC_DEFUN([ZFS_AC_KERNEL_SRC_FPU], [
 		__kernel_fpu_end();
 	], [], [$ZFS_META_LICENSE])
 
-	ZFS_LINUX_TEST_SRC([fpu_initialized], [
+	ZFS_LINUX_TEST_SRC([fpu_internal], [
-		#include <linux/module.h>
+		#if defined(__x86_64) || defined(__x86_64__) || \
-		#include <linux/sched.h>
+		    defined(__i386) || defined(__i386__)
+		#if !defined(__x86)
+		#define __x86
+		#endif
+		#endif
+
+		#if !defined(__x86)
+		#error Unsupported architecture
+		#endif
+
+		#include <linux/types.h>
+		#ifdef HAVE_KERNEL_FPU_API_HEADER
+		#include <asm/fpu/api.h>
+		#include <asm/fpu/internal.h>
+		#else
+		#include <asm/i387.h>
+		#include <asm/xcr.h>
+		#endif
+
+		#if !defined(XSTATE_XSAVE)
+		#error XSTATE_XSAVE not defined
+		#endif
+
+		#if !defined(XSTATE_XRESTORE)
+		#error XSTATE_XRESTORE not defined
+		#endif
 	],[
 		struct fpu *fpu = &current->thread.fpu;
-		if (fpu->initialized) { return (0); };
+		union fpregs_state *st = &fpu->state;
+		struct fregs_state *fr __attribute__ ((unused)) = &st->fsave;
+		struct fxregs_state *fxr __attribute__ ((unused)) = &st->fxsave;
+		struct xregs_state *xr __attribute__ ((unused)) = &st->xsave;
 	])
-
-	ZFS_LINUX_TEST_SRC([tif_need_fpu_load], [
-		#include <linux/module.h>
-		#include <asm/thread_info.h>
-
-		#if !defined(TIF_NEED_FPU_LOAD)
-		#error "TIF_NEED_FPU_LOAD undefined"
-		#endif
-	],[])
 ])
 
 AC_DEFUN([ZFS_AC_KERNEL_FPU], [
@@ -104,25 +119,12 @@ AC_DEFUN([ZFS_AC_KERNEL_FPU], [
 			AC_DEFINE(KERNEL_EXPORTS_X86_FPU, 1,
 			    [kernel exports FPU functions])
 		],[
-			dnl #
+			ZFS_LINUX_TEST_RESULT([fpu_internal], [
-			dnl # Linux 5.0 kernel
+				AC_MSG_RESULT(internal)
-			dnl #
+				AC_DEFINE(HAVE_KERNEL_FPU_INTERNAL, 1,
-			ZFS_LINUX_TEST_RESULT([fpu_initialized], [
+				    [kernel fpu internal])
-				AC_MSG_RESULT(fpu.initialized)
-				AC_DEFINE(HAVE_KERNEL_FPU_INITIALIZED, 1,
-				    [kernel fpu.initialized exists])
 			],[
-				dnl #
+				AC_MSG_RESULT(unavailable)
-				dnl # Linux 5.2 kernel
-				dnl #
-				ZFS_LINUX_TEST_RESULT([tif_need_fpu_load], [
-					AC_MSG_RESULT(TIF_NEED_FPU_LOAD)
-					AC_DEFINE(
-					    HAVE_KERNEL_TIF_NEED_FPU_LOAD, 1,
-					    [kernel TIF_NEED_FPU_LOAD exists])
-				],[
-					AC_MSG_RESULT(unavailable)
-				])
 			])
 		])
 	])

include/os/linux/kernel/linux/simd.h

@@ -33,9 +33,10 @@
 #else
 
 #define	kfpu_allowed()		0
-#define	kfpu_initialize(tsk)	do {} while (0)
 #define	kfpu_begin()		do {} while (0)
 #define	kfpu_end()		do {} while (0)
+#define	kfpu_init()		0
+#define	kfpu_fini()		((void) 0)
 
 #endif
 #endif /* _LINUX_SIMD_H */

include/os/linux/kernel/linux/simd_aarch64.h

@@ -27,9 +27,10 @@
  *
  * Kernel fpu methods:
  *	kfpu_allowed()
- *	kfpu_initialize()
  *	kfpu_begin()
  *	kfpu_end()
+ *	kfpu_init()
+ *	kfpu_fini()
  */
 
 #ifndef _LINUX_SIMD_AARCH64_H
@@ -43,9 +44,10 @@
 #include <asm/neon.h>
 
 #define	kfpu_allowed()		1
-#define	kfpu_initialize(tsk)	do {} while (0)
 #define	kfpu_begin()		kernel_neon_begin()
 #define	kfpu_end()		kernel_neon_end()
+#define	kfpu_init()		0
+#define	kfpu_fini()		((void) 0)
 
 #endif /* __aarch64__ */
 

include/os/linux/kernel/linux/simd_x86.h

@@ -27,9 +27,10 @@
  *
  * Kernel fpu methods:
  *	kfpu_allowed()
- *	kfpu_initialize()
  *	kfpu_begin()
  *	kfpu_end()
+ *	kfpu_init()
+ *	kfpu_fini()
  *
  * SIMD support:
  *
@@ -99,7 +100,8 @@
 #if defined(KERNEL_EXPORTS_X86_FPU)
 
 #define	kfpu_allowed()		1
-#define	kfpu_initialize(tsk)	do {} while (0)
+#define	kfpu_init()		0
+#define	kfpu_fini()		((void) 0)
 
 #if defined(HAVE_UNDERSCORE_KERNEL_FPU)
 #define	kfpu_begin()		\
@@ -126,45 +128,100 @@
 #endif
 
 #else /* defined(KERNEL_EXPORTS_X86_FPU) */
+
 /*
  * When the kernel_fpu_* symbols are unavailable then provide our own
- * versions which allow the FPU to be safely used in kernel threads.
+ * versions which allow the FPU to be safely used.
- * In practice, this is not a significant restriction for ZFS since the
- * vast majority of SIMD operations are performed by the IO pipeline.
  */
+#if defined(HAVE_KERNEL_FPU_INTERNAL)
+
+extern union fpregs_state **zfs_kfpu_fpregs;
 
 /*
- * Returns non-zero if FPU operations are allowed in the current context.
+ * Initialize per-cpu variables to store FPU state.
  */
-#if defined(HAVE_KERNEL_TIF_NEED_FPU_LOAD)
+static inline void
-#define	kfpu_allowed()		((current->flags & PF_KTHREAD) && \
+kfpu_fini(void)
-				test_thread_flag(TIF_NEED_FPU_LOAD))
+{
-#elif defined(HAVE_KERNEL_FPU_INITIALIZED)
+	int cpu;
-#define	kfpu_allowed()		((current->flags & PF_KTHREAD) && \
+
-				current->thread.fpu.initialized)
+	for_each_possible_cpu(cpu) {
-#else
+		if (zfs_kfpu_fpregs[cpu] != NULL) {
-#define	kfpu_allowed()		0
+			kfree(zfs_kfpu_fpregs[cpu]);
-#endif
+		}
+	}
+
+	kfree(zfs_kfpu_fpregs);
+}
+
+static inline int
+kfpu_init(void)
+{
+	int cpu;
+
+	zfs_kfpu_fpregs = kzalloc(num_possible_cpus() *
+	    sizeof (union fpregs_state *), GFP_KERNEL);
+	if (zfs_kfpu_fpregs == NULL)
+		return (-ENOMEM);
+
+	for_each_possible_cpu(cpu) {
+		zfs_kfpu_fpregs[cpu] = kmalloc_node(sizeof (union fpregs_state),
+		    GFP_KERNEL | __GFP_ZERO, cpu_to_node(cpu));
+		if (zfs_kfpu_fpregs[cpu] == NULL) {
+			kfpu_fini();
+			return (-ENOMEM);
+		}
+	}
+
+	return (0);
+}
+
+#define	kfpu_allowed()		1
+#define	ex_handler_fprestore	ex_handler_default
+
+/*
+ * FPU save and restore instructions.
+ */
+#define	__asm			__asm__ __volatile__
+#define	kfpu_fxsave(addr)	__asm("fxsave %0" : "=m" (*(addr)))
+#define	kfpu_fxsaveq(addr)	__asm("fxsaveq %0" : "=m" (*(addr)))
+#define	kfpu_fnsave(addr)	__asm("fnsave %0; fwait" : "=m" (*(addr)))
+#define	kfpu_fxrstor(addr)	__asm("fxrstor %0" : : "m" (*(addr)))
+#define	kfpu_fxrstorq(addr)	__asm("fxrstorq %0" : : "m" (*(addr)))
+#define	kfpu_frstor(addr)	__asm("frstor %0" : : "m" (*(addr)))
+#define	kfpu_fxsr_clean(rval)	__asm("fnclex; emms; fildl %P[addr]" \
+				    : : [addr] "m" (rval));
 
 static inline void
-kfpu_initialize(void)
+kfpu_save_xsave(struct xregs_state *addr, uint64_t mask)
 {
-	WARN_ON_ONCE(!(current->flags & PF_KTHREAD));
+	uint32_t low, hi;
+	int err;
 
-#if defined(HAVE_KERNEL_TIF_NEED_FPU_LOAD)
+	low = mask;
-	__fpu_invalidate_fpregs_state(&current->thread.fpu);
+	hi = mask >> 32;
-	set_thread_flag(TIF_NEED_FPU_LOAD);
+	XSTATE_XSAVE(addr, low, hi, err);
-#elif defined(HAVE_KERNEL_FPU_INITIALIZED)
+	WARN_ON_ONCE(err);
-	__fpu_invalidate_fpregs_state(&current->thread.fpu);
+}
-	current->thread.fpu.initialized = 1;
+
-#endif
+static inline void
+kfpu_save_fxsr(struct fxregs_state *addr)
+{
+	if (IS_ENABLED(CONFIG_X86_32))
+		kfpu_fxsave(addr);
+	else
+		kfpu_fxsaveq(addr);
+}
+
+static inline void
+kfpu_save_fsave(struct fregs_state *addr)
+{
+	kfpu_fnsave(addr);
 }
 
 static inline void
 kfpu_begin(void)
 {
-	WARN_ON_ONCE(!kfpu_allowed());
-
 	/*
 	 * Preemption and interrupts must be disabled for the critical
 	 * region where the FPU state is being modified.
@@ -172,50 +229,92 @@ kfpu_begin(void)
 	preempt_disable();
 	local_irq_disable();
 
-#if defined(HAVE_KERNEL_TIF_NEED_FPU_LOAD)
 	/*
 	 * The current FPU registers need to be preserved by kfpu_begin()
-	 * and restored by kfpu_end().  This is required because we can
+	 * and restored by kfpu_end().  They are stored in a dedicated
-	 * not call __cpu_invalidate_fpregs_state() to invalidate the
+	 * per-cpu variable, not in the task struct, this allows any user
-	 * per-cpu FPU state and force them to be restored during a
+	 * FPU state to be correctly preserved and restored.
-	 * context switch.
 	 */
-	copy_fpregs_to_fpstate(&current->thread.fpu);
+	union fpregs_state *state = zfs_kfpu_fpregs[smp_processor_id()];
-#elif defined(HAVE_KERNEL_FPU_INITIALIZED)
+
+	if (static_cpu_has(X86_FEATURE_XSAVE)) {
+		kfpu_save_xsave(&state->xsave, ~0);
+	} else if (static_cpu_has(X86_FEATURE_FXSR)) {
+		kfpu_save_fxsr(&state->fxsave);
+	} else {
+		kfpu_save_fsave(&state->fsave);
+	}
+}
+
+static inline void
+kfpu_restore_xsave(struct xregs_state *addr, uint64_t mask)
+{
+	uint32_t low, hi;
+
+	low = mask;
+	hi = mask >> 32;
+	XSTATE_XRESTORE(addr, low, hi);
+}
+
+static inline void
+kfpu_restore_fxsr(struct fxregs_state *addr)
+{
 	/*
-	 * There is no need to preserve and restore the FPU registers.
+	 * On AuthenticAMD K7 and K8 processors the fxrstor instruction only
-	 * They will always be restored from the task's stored FPU state
+	 * restores the _x87 FOP, FIP, and FDP registers when an exception
-	 * when switching contexts.
+	 * is pending.  Clean the _x87 state to force the restore.
 	 */
-	WARN_ON_ONCE(current->thread.fpu.initialized == 0);
+	if (unlikely(static_cpu_has_bug(X86_BUG_FXSAVE_LEAK)))
-#endif
+		kfpu_fxsr_clean(addr);
+
+	if (IS_ENABLED(CONFIG_X86_32)) {
+		kfpu_fxrstor(addr);
+	} else {
+		kfpu_fxrstorq(addr);
+	}
+}
+
+static inline void
+kfpu_restore_fsave(struct fregs_state *addr)
+{
+	kfpu_frstor(addr);
 }
 
 static inline void
 kfpu_end(void)
 {
-#if defined(HAVE_KERNEL_TIF_NEED_FPU_LOAD)
+	union fpregs_state *state = zfs_kfpu_fpregs[smp_processor_id()];
-	union fpregs_state *state = &current->thread.fpu.state;
-	int error;
 
-	if (use_xsave()) {
+	if (static_cpu_has(X86_FEATURE_XSAVE)) {
-		error = copy_kernel_to_xregs_err(&state->xsave, -1);
+		kfpu_restore_xsave(&state->xsave, ~0);
-	} else if (use_fxsr()) {
+	} else if (static_cpu_has(X86_FEATURE_FXSR)) {
-		error = copy_kernel_to_fxregs_err(&state->fxsave);
+		kfpu_restore_fxsr(&state->fxsave);
 	} else {
-		error = copy_kernel_to_fregs_err(&state->fsave);
+		kfpu_restore_fsave(&state->fsave);
 	}
-	WARN_ON_ONCE(error);
-#endif
 
 	local_irq_enable();
 	preempt_enable();
 }
-#endif /* defined(HAVE_KERNEL_FPU) */
+
+#else
+
+/*
+ * FPU support is unavailable.
+ */
+#define	kfpu_allowed()		0
+#define	kfpu_begin()		do {} while (0)
+#define	kfpu_end()		do {} while (0)
+#define	kfpu_init()		0
+#define	kfpu_fini()		((void) 0)
+
+#endif /* defined(HAVE_KERNEL_FPU_INTERNAL) */
+#endif /* defined(KERNEL_EXPORTS_X86_FPU) */
 
 /*
  * Linux kernel provides an interface for CPU feature testing.
  */
+
 /*
  * Detect register set support
  */

include/sys/zio_crypt.h

@@ -107,11 +107,11 @@ void zio_crypt_key_destroy(zio_crypt_key_t *key);
 int zio_crypt_key_init(uint64_t crypt, zio_crypt_key_t *key);
 int zio_crypt_key_get_salt(zio_crypt_key_t *key, uint8_t *salt_out);
 
-int zio_crypt_key_wrap(spa_t *spa, crypto_key_t *cwkey, zio_crypt_key_t *key,
+int zio_crypt_key_wrap(crypto_key_t *cwkey, zio_crypt_key_t *key, uint8_t *iv,
-    uint8_t *iv, uint8_t *mac, uint8_t *keydata_out, uint8_t *hmac_keydata_out);
+    uint8_t *mac, uint8_t *keydata_out, uint8_t *hmac_keydata_out);
-int zio_crypt_key_unwrap(spa_t *spa, crypto_key_t *cwkey, uint64_t crypt,
+int zio_crypt_key_unwrap(crypto_key_t *cwkey, uint64_t crypt, uint64_t version,
-    uint64_t version, uint64_t guid, uint8_t *keydata, uint8_t *hmac_keydata,
+    uint64_t guid, uint8_t *keydata, uint8_t *hmac_keydata, uint8_t *iv,
-    uint8_t *iv, uint8_t *mac, zio_crypt_key_t *key);
+    uint8_t *mac, zio_crypt_key_t *key);
 int zio_crypt_generate_iv(uint8_t *ivbuf);
 int zio_crypt_generate_iv_salt_dedup(zio_crypt_key_t *key, uint8_t *data,
     uint_t datalen, uint8_t *ivbuf, uint8_t *salt);
@@ -132,11 +132,11 @@ int zio_crypt_do_hmac(zio_crypt_key_t *key, uint8_t *data, uint_t datalen,
     uint8_t *digestbuf, uint_t digestlen);
 int zio_crypt_do_objset_hmacs(zio_crypt_key_t *key, void *data, uint_t datalen,
     boolean_t byteswap, uint8_t *portable_mac, uint8_t *local_mac);
-int zio_do_crypt_data(spa_t *spa, boolean_t encrypt, zio_crypt_key_t *key,
+int zio_do_crypt_data(boolean_t encrypt, zio_crypt_key_t *key,
     dmu_object_type_t ot, boolean_t byteswap, uint8_t *salt, uint8_t *iv,
     uint8_t *mac, uint_t datalen, uint8_t *plainbuf, uint8_t *cipherbuf,
     boolean_t *no_crypt);
-int zio_do_crypt_abd(spa_t *spa, boolean_t encrypt, zio_crypt_key_t *key,
+int zio_do_crypt_abd(boolean_t encrypt, zio_crypt_key_t *key,
     dmu_object_type_t ot, boolean_t byteswap, uint8_t *salt, uint8_t *iv,
     uint8_t *mac, uint_t datalen, abd_t *pabd, abd_t *cabd,
     boolean_t *no_crypt);

lib/libspl/include/sys/simd.h

@@ -34,9 +34,10 @@
 #include <cpuid.h>
 
 #define	kfpu_allowed()		1
-#define	kfpu_initialize(tsk)	do {} while (0)
 #define	kfpu_begin()		do {} while (0)
 #define	kfpu_end()		do {} while (0)
+#define	kfpu_init()		0
+#define	kfpu_fini()		((void) 0)
 
 /*
  * CPUID feature tests for user-space.

module/icp/algs/aes/aes_impl.c

@@ -295,9 +295,8 @@ aes_impl_get_ops(void)
 /*
  * Initialize all supported implementations.
  */
-/* ARGSUSED */
 void
-aes_impl_init(void *arg)
+aes_impl_init(void)
 {
 	aes_impl_ops_t *curr_impl;
 	int i, c;

module/icp/algs/modes/gcm.c

@@ -703,9 +703,8 @@ gcm_impl_get_ops()
 /*
  * Initialize all supported implementations.
  */
-/* ARGSUSED */
 void
-gcm_impl_init(void *arg)
+gcm_impl_init(void)
 {
 	gcm_impl_ops_t *curr_impl;
 	int i, c;

module/icp/include/aes/aes_impl.h

@@ -198,7 +198,7 @@ extern const aes_impl_ops_t aes_aesni_impl;
 /*
  * Initializes fastest implementation
  */
-void aes_impl_init(void *arg);
+void aes_impl_init(void);
 
 /*
  * Returns optimal allowed AES implementation

module/icp/include/modes/gcm_impl.h

@@ -61,7 +61,7 @@ extern const gcm_impl_ops_t gcm_pclmulqdq_impl;
 /*
  * Initializes fastest implementation
  */
-void gcm_impl_init(void *arg);
+void gcm_impl_init(void);
 
 /*
  * Returns optimal allowed GCM implementation

module/icp/io/aes.c

@@ -206,35 +206,9 @@ aes_mod_init(void)
 {
 	int ret;
 
-#if defined(_KERNEL)
+	/* Determine the fastest available implementation. */
-	/*
+	aes_impl_init();
-	 * Determine the fastest available implementation.  The benchmarks
+	gcm_impl_init();
-	 * are run in dedicated kernel threads to allow Linux 5.0+ kernels
-	 * to use SIMD operations.  If for some reason this isn't possible,
-	 * fallback to the generic implementations.  See the comment in
-	 * linux/simd_x86.h for additional details.  Additionally, this has
-	 * the benefit of allowing them to be run in parallel.
-	 */
-	taskqid_t aes_id = taskq_dispatch(system_taskq, aes_impl_init,
-	    NULL, TQ_SLEEP);
-	taskqid_t gcm_id = taskq_dispatch(system_taskq, gcm_impl_init,
-	    NULL, TQ_SLEEP);
-
-	if (aes_id != TASKQID_INVALID) {
-		taskq_wait_id(system_taskq, aes_id);
-	} else {
-		aes_impl_init(NULL);
-	}
-
-	if (gcm_id != TASKQID_INVALID) {
-		taskq_wait_id(system_taskq, gcm_id);
-	} else {
-		gcm_impl_init(NULL);
-	}
-#else
-	aes_impl_init(NULL);
-	gcm_impl_init(NULL);
-#endif
 
 	if ((ret = mod_install(&modlinkage)) != 0)
 		return (ret);

module/os/linux/spl/spl-taskq.c

@@ -28,7 +28,6 @@
 #include <sys/taskq.h>
 #include <sys/kmem.h>
 #include <sys/tsd.h>
-#include <sys/simd.h>
 
 int spl_taskq_thread_bind = 0;
 module_param(spl_taskq_thread_bind, int, 0644);
@@ -854,7 +853,6 @@ taskq_thread(void *args)
 	sigfillset(&blocked);
 	sigprocmask(SIG_BLOCK, &blocked, NULL);
 	flush_signals(current);
-	kfpu_initialize();
 
 	tsd_set(taskq_tsd, tq);
 	spin_lock_irqsave_nested(&tq->tq_lock, flags, tq->tq_lock_class);

module/os/linux/spl/spl-thread.c

@@ -27,7 +27,6 @@
 #include <sys/thread.h>
 #include <sys/kmem.h>
 #include <sys/tsd.h>
-#include <sys/simd.h>
 
 /*
  * Thread interfaces
@@ -55,7 +54,6 @@ thread_generic_wrapper(void *arg)
 	args = tp->tp_args;
 	set_current_state(tp->tp_state);
 	set_user_nice((kthread_t *)current, PRIO_TO_NICE(tp->tp_pri));
-	kfpu_initialize();
 	kmem_free(tp->tp_name, tp->tp_name_size);
 	kmem_free(tp, sizeof (thread_priv_t));
 

module/os/linux/zfs/zio_crypt.c

@@ -25,8 +25,6 @@
 #include <sys/zio.h>
 #include <sys/zil.h>
 #include <sys/sha2.h>
-#include <sys/simd.h>
-#include <sys/spa_impl.h>
 #include <sys/hkdf.h>
 #include <sys/qat.h>
 
@@ -376,7 +374,7 @@ error:
  * plaintext / ciphertext alone.
  */
 static int
-zio_do_crypt_uio_impl(boolean_t encrypt, uint64_t crypt, crypto_key_t *key,
+zio_do_crypt_uio(boolean_t encrypt, uint64_t crypt, crypto_key_t *key,
     crypto_ctx_template_t tmpl, uint8_t *ivbuf, uint_t datalen,
     uio_t *puio, uio_t *cuio, uint8_t *authbuf, uint_t auth_len)
 {
@@ -476,75 +474,9 @@ error:
 	return (ret);
 }
 
-typedef struct crypt_uio_arg {
-	boolean_t cu_encrypt;
-	uint64_t cu_crypt;
-	crypto_key_t *cu_key;
-	crypto_ctx_template_t cu_tmpl;
-	uint8_t *cu_ivbuf;
-	uint_t cu_datalen;
-	uio_t *cu_puio;
-	uio_t *cu_cuio;
-	uint8_t *cu_authbuf;
-	uint_t cu_auth_len;
-	int cu_error;
-} crypt_uio_arg_t;
-
-static void
-zio_do_crypt_uio_func(void *arg)
-{
-	crypt_uio_arg_t *cu = (crypt_uio_arg_t *)arg;
-
-	cu->cu_error = zio_do_crypt_uio_impl(cu->cu_encrypt, cu->cu_crypt,
-	    cu->cu_key, cu->cu_tmpl, cu->cu_ivbuf, cu->cu_datalen,
-	    cu->cu_puio, cu->cu_cuio, cu->cu_authbuf, cu->cu_auth_len);
-}
-
-static int
-zio_do_crypt_uio(spa_t *spa, boolean_t encrypt, uint64_t crypt,
-    crypto_key_t *key, crypto_ctx_template_t tmpl, uint8_t *ivbuf,
-    uint_t datalen, uio_t *puio, uio_t *cuio, uint8_t *authbuf,
-    uint_t auth_len)
-{
-	int error;
-
-	/*
-	 * Dispatch to the I/O pipeline as required by the context in order
-	 * to take advantage of the SIMD optimization when available.
-	 */
-	if (kfpu_allowed()) {
-		error = zio_do_crypt_uio_impl(encrypt, crypt, key, tmpl,
-		    ivbuf, datalen, puio, cuio, authbuf, auth_len);
-	} else {
-		crypt_uio_arg_t *cu;
-
-		cu = kmem_alloc(sizeof (*cu), KM_SLEEP);
-		cu->cu_encrypt = encrypt;
-		cu->cu_crypt = crypt;
-		cu->cu_key = key;
-		cu->cu_tmpl = tmpl;
-		cu->cu_ivbuf = ivbuf;
-		cu->cu_datalen = datalen;
-		cu->cu_puio = puio;
-		cu->cu_cuio = cuio;
-		cu->cu_authbuf = authbuf;
-		cu->cu_auth_len = auth_len;
-		cu->cu_error = 0;
-
-		spa_taskq_dispatch_sync(spa,
-		    encrypt ? ZIO_TYPE_WRITE : ZIO_TYPE_READ,
-		    ZIO_TASKQ_ISSUE, zio_do_crypt_uio_func, cu, TQ_SLEEP);
-
-		error = cu->cu_error;
-		kmem_free(cu, sizeof (*cu));
-	}
-
-	return (error);
-}
-
 int
-zio_crypt_key_wrap(spa_t *spa, crypto_key_t *cwkey, zio_crypt_key_t *key,
+zio_crypt_key_wrap(crypto_key_t *cwkey, zio_crypt_key_t *key, uint8_t *iv,
-    uint8_t *iv, uint8_t *mac, uint8_t *keydata_out, uint8_t *hmac_keydata_out)
+    uint8_t *mac, uint8_t *keydata_out, uint8_t *hmac_keydata_out)
 {
 	int ret;
 	uio_t puio, cuio;
@@ -601,7 +533,7 @@ zio_crypt_key_wrap(spa_t *spa, crypto_key_t *cwkey, zio_crypt_key_t *key,
 	cuio.uio_segflg = UIO_SYSSPACE;
 
 	/* encrypt the keys and store the resulting ciphertext and mac */
-	ret = zio_do_crypt_uio(spa, B_TRUE, crypt, cwkey, NULL, iv, enc_len,
+	ret = zio_do_crypt_uio(B_TRUE, crypt, cwkey, NULL, iv, enc_len,
 	    &puio, &cuio, (uint8_t *)aad, aad_len);
 	if (ret != 0)
 		goto error;
@@ -612,33 +544,12 @@ error:
 	return (ret);
 }
 
-static void
-zio_crypt_create_ctx_templates(void *arg)
-{
-	zio_crypt_key_t *key = (zio_crypt_key_t *)arg;
-	crypto_mechanism_t mech;
-	int ret;
-
-	mech.cm_type = crypto_mech2id(
-	    zio_crypt_table[key->zk_crypt].ci_mechname);
-
-	ret = crypto_create_ctx_template(&mech, &key->zk_current_key,
-	    &key->zk_current_tmpl, KM_SLEEP);
-	if (ret != CRYPTO_SUCCESS)
-		key->zk_current_tmpl = NULL;
-
-	mech.cm_type = crypto_mech2id(SUN_CKM_SHA512_HMAC);
-	ret = crypto_create_ctx_template(&mech, &key->zk_hmac_key,
-	    &key->zk_hmac_tmpl, KM_SLEEP);
-	if (ret != CRYPTO_SUCCESS)
-		key->zk_hmac_tmpl = NULL;
-}
-
 int
-zio_crypt_key_unwrap(spa_t *spa, crypto_key_t *cwkey, uint64_t crypt,
+zio_crypt_key_unwrap(crypto_key_t *cwkey, uint64_t crypt, uint64_t version,
-    uint64_t version, uint64_t guid, uint8_t *keydata, uint8_t *hmac_keydata,
+    uint64_t guid, uint8_t *keydata, uint8_t *hmac_keydata, uint8_t *iv,
-    uint8_t *iv, uint8_t *mac, zio_crypt_key_t *key)
+    uint8_t *mac, zio_crypt_key_t *key)
 {
+	crypto_mechanism_t mech;
 	uio_t puio, cuio;
 	uint64_t aad[3];
 	iovec_t plain_iovecs[2], cipher_iovecs[3];
@@ -685,7 +596,7 @@ zio_crypt_key_unwrap(spa_t *spa, crypto_key_t *cwkey, uint64_t crypt,
 	cuio.uio_segflg = UIO_SYSSPACE;
 
 	/* decrypt the keys and store the result in the output buffers */
-	ret = zio_do_crypt_uio(spa, B_FALSE, crypt, cwkey, NULL, iv, enc_len,
+	ret = zio_do_crypt_uio(B_FALSE, crypt, cwkey, NULL, iv, enc_len,
 	    &puio, &cuio, (uint8_t *)aad, aad_len);
 	if (ret != 0)
 		goto error;
@@ -711,18 +622,27 @@ zio_crypt_key_unwrap(spa_t *spa, crypto_key_t *cwkey, uint64_t crypt,
 	key->zk_hmac_key.ck_data = key->zk_hmac_keydata;
 	key->zk_hmac_key.ck_length = CRYPTO_BYTES2BITS(SHA512_HMAC_KEYLEN);
 
+	/*
+	 * Initialize the crypto templates. It's ok if this fails because
+	 * this is just an optimization.
+	 */
+	mech.cm_type = crypto_mech2id(zio_crypt_table[crypt].ci_mechname);
+	ret = crypto_create_ctx_template(&mech, &key->zk_current_key,
+	    &key->zk_current_tmpl, KM_SLEEP);
+	if (ret != CRYPTO_SUCCESS)
+		key->zk_current_tmpl = NULL;
+
+	mech.cm_type = crypto_mech2id(SUN_CKM_SHA512_HMAC);
+	ret = crypto_create_ctx_template(&mech, &key->zk_hmac_key,
+	    &key->zk_hmac_tmpl, KM_SLEEP);
+	if (ret != CRYPTO_SUCCESS)
+		key->zk_hmac_tmpl = NULL;
+
 	key->zk_crypt = crypt;
 	key->zk_version = version;
 	key->zk_guid = guid;
 	key->zk_salt_count = 0;
 
-	/*
-	 * Initialize the crypto templates in the context they will be
-	 * primarily used. It's ok if this fails, it's just an optimization.
-	 */
-	spa_taskq_dispatch_sync(spa, ZIO_TYPE_READ, ZIO_TASKQ_ISSUE,
-	    zio_crypt_create_ctx_templates, key, TQ_SLEEP);
-
 	return (0);
 
 error:
@@ -1941,7 +1861,7 @@ error:
  * Primary encryption / decryption entrypoint for zio data.
  */
 int
-zio_do_crypt_data(spa_t *spa, boolean_t encrypt, zio_crypt_key_t *key,
+zio_do_crypt_data(boolean_t encrypt, zio_crypt_key_t *key,
     dmu_object_type_t ot, boolean_t byteswap, uint8_t *salt, uint8_t *iv,
     uint8_t *mac, uint_t datalen, uint8_t *plainbuf, uint8_t *cipherbuf,
     boolean_t *no_crypt)
@@ -2028,8 +1948,8 @@ zio_do_crypt_data(spa_t *spa, boolean_t encrypt, zio_crypt_key_t *key,
 		goto error;
 
 	/* perform the encryption / decryption in software */
-	ret = zio_do_crypt_uio(spa, encrypt, key->zk_crypt, ckey, tmpl, iv,
+	ret = zio_do_crypt_uio(encrypt, key->zk_crypt, ckey, tmpl, iv, enc_len,
-	    enc_len, &puio, &cuio, authbuf, auth_len);
+	    &puio, &cuio, authbuf, auth_len);
 	if (ret != 0)
 		goto error;
 
@@ -2065,10 +1985,9 @@ error:
  * linear buffers.
  */
 int
-zio_do_crypt_abd(spa_t *spa, boolean_t encrypt, zio_crypt_key_t *key,
+zio_do_crypt_abd(boolean_t encrypt, zio_crypt_key_t *key, dmu_object_type_t ot,
-    dmu_object_type_t ot, boolean_t byteswap, uint8_t *salt, uint8_t *iv,
+    boolean_t byteswap, uint8_t *salt, uint8_t *iv, uint8_t *mac,
-    uint8_t *mac, uint_t datalen, abd_t *pabd, abd_t *cabd,
+    uint_t datalen, abd_t *pabd, abd_t *cabd, boolean_t *no_crypt)
-    boolean_t *no_crypt)
 {
 	int ret;
 	void *ptmp, *ctmp;
@@ -2081,7 +2000,7 @@ zio_do_crypt_abd(spa_t *spa, boolean_t encrypt, zio_crypt_key_t *key,
 		ctmp = abd_borrow_buf_copy(cabd, datalen);
 	}
 
-	ret = zio_do_crypt_data(spa, encrypt, key, ot, byteswap, salt, iv, mac,
+	ret = zio_do_crypt_data(encrypt, key, ot, byteswap, salt, iv, mac,
 	    datalen, ptmp, ctmp, no_crypt);
 	if (ret != 0)
 		goto error;

module/zcommon/zfs_fletcher.c

@@ -726,7 +726,7 @@ fletcher_4_benchmark_impl(boolean_t native, char *data, uint64_t data_size)
  * Initialize and benchmark all supported implementations.
  */
 static void
-fletcher_4_benchmark(void *arg)
+fletcher_4_benchmark(void)
 {
 	fletcher_4_ops_t *curr_impl;
 	int i, c;
@@ -769,20 +769,10 @@ fletcher_4_benchmark(void *arg)
 void
 fletcher_4_init(void)
 {
-#if defined(_KERNEL)
+	/* Determine the fastest available implementation. */
-	/*
+	fletcher_4_benchmark();
-	 * For 5.0 and latter Linux kernels the fletcher 4 benchmarks are
-	 * run in a kernel threads.  This is needed to take advantage of the
-	 * SIMD functionality, see linux/simd_x86.h for details.
-	 */
-	taskqid_t id = taskq_dispatch(system_taskq, fletcher_4_benchmark,
-	    NULL, TQ_SLEEP);
-	if (id != TASKQID_INVALID) {
-		taskq_wait_id(system_taskq, id);
-	} else {
-		fletcher_4_benchmark(NULL);
-	}
 
+#if defined(_KERNEL)
 	/* Install kstats for all implementations */
 	fletcher_4_kstat = kstat_create("zfs", 0, "fletcher_4_bench", "misc",
 	    KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VIRTUAL);
@@ -795,8 +785,6 @@ fletcher_4_init(void)
 		    fletcher_4_kstat_addr);
 		kstat_install(fletcher_4_kstat);
 	}
-#else
-	fletcher_4_benchmark(NULL);
 #endif
 
 	/* Finish initialization */

module/zcommon/zfs_prop.c

@@ -865,10 +865,23 @@ zfs_prop_align_right(zfs_prop_t prop)
 #endif
 
 #if defined(_KERNEL)
+
+#include <sys/simd.h>
+
+#if defined(HAVE_KERNEL_FPU_INTERNAL)
+union fpregs_state **zfs_kfpu_fpregs;
+EXPORT_SYMBOL(zfs_kfpu_fpregs);
+#endif /* HAVE_KERNEL_FPU_INTERNAL */
+
 static int __init
 zcommon_init(void)
 {
+	int error = kfpu_init();
+	if (error)
+		return (error);
+
 	fletcher_4_init();
+
 	return (0);
 }
 
@@ -876,6 +889,7 @@ static void __exit
 zcommon_fini(void)
 {
 	fletcher_4_fini();
+	kfpu_fini();
 }
 
 module_init(zcommon_init);

module/zfs/arc.c

@@ -8136,7 +8136,7 @@ l2arc_apply_transforms(spa_t *spa, arc_buf_hdr_t *hdr, uint64_t asize,
 		if (ret != 0)
 			goto error;
 
-		ret = zio_do_crypt_abd(spa, B_TRUE, &dck->dck_key,
+		ret = zio_do_crypt_abd(B_TRUE, &dck->dck_key,
 		    hdr->b_crypt_hdr.b_ot, bswap, hdr->b_crypt_hdr.b_salt,
 		    hdr->b_crypt_hdr.b_iv, mac, psize, to_write, eabd,
 		    &no_crypt);

module/zfs/dsl_crypt.c

@@ -601,8 +601,8 @@ dsl_crypto_key_open(objset_t *mos, dsl_wrapping_key_t *wkey,
 	 * Unwrap the keys. If there is an error return EACCES to indicate
 	 * an authentication failure.
 	 */
-	ret = zio_crypt_key_unwrap(mos->os_spa, &wkey->wk_key, crypt, version,
+	ret = zio_crypt_key_unwrap(&wkey->wk_key, crypt, version, guid,
-	    guid, raw_keydata, raw_hmac_keydata, iv, mac, &dck->dck_key);
+	    raw_keydata, raw_hmac_keydata, iv, mac, &dck->dck_key);
 	if (ret != 0) {
 		ret = SET_ERROR(EACCES);
 		goto error;
@@ -1221,7 +1221,6 @@ dsl_crypto_key_sync(dsl_crypto_key_t *dck, dmu_tx_t *tx)
 {
 	zio_crypt_key_t *key = &dck->dck_key;
 	dsl_wrapping_key_t *wkey = dck->dck_wkey;
-	objset_t *mos = tx->tx_pool->dp_meta_objset;
 	uint8_t keydata[MASTER_KEY_MAX_LEN];
 	uint8_t hmac_keydata[SHA512_HMAC_KEYLEN];
 	uint8_t iv[WRAPPING_IV_LEN];
@@ -1231,13 +1230,14 @@ dsl_crypto_key_sync(dsl_crypto_key_t *dck, dmu_tx_t *tx)
 	ASSERT3U(key->zk_crypt, <, ZIO_CRYPT_FUNCTIONS);
 
 	/* encrypt and store the keys along with the IV and MAC */
-	VERIFY0(zio_crypt_key_wrap(mos->os_spa, &dck->dck_wkey->wk_key, key,
+	VERIFY0(zio_crypt_key_wrap(&dck->dck_wkey->wk_key, key, iv, mac,
-	    iv, mac, keydata, hmac_keydata));
+	    keydata, hmac_keydata));
 
 	/* update the ZAP with the obtained values */
-	dsl_crypto_key_sync_impl(mos, dck->dck_obj, key->zk_crypt,
+	dsl_crypto_key_sync_impl(tx->tx_pool->dp_meta_objset, dck->dck_obj,
-	    wkey->wk_ddobj, key->zk_guid, iv, mac, keydata, hmac_keydata,
+	    key->zk_crypt, wkey->wk_ddobj, key->zk_guid, iv, mac, keydata,
-	    wkey->wk_keyformat, wkey->wk_salt, wkey->wk_iters, tx);
+	    hmac_keydata, wkey->wk_keyformat, wkey->wk_salt, wkey->wk_iters,
+	    tx);
 }
 
 typedef struct spa_keystore_change_key_args {
@@ -2846,8 +2846,8 @@ spa_do_crypt_abd(boolean_t encrypt, spa_t *spa, const zbookmark_phys_t *zb,
 	}
 
 	/* call lower level function to perform encryption / decryption */
-	ret = zio_do_crypt_data(spa, encrypt, &dck->dck_key, ot, bswap, salt,
+	ret = zio_do_crypt_data(encrypt, &dck->dck_key, ot, bswap, salt, iv,
-	    iv, mac, datalen, plainbuf, cipherbuf, no_crypt);
+	    mac, datalen, plainbuf, cipherbuf, no_crypt);
 
 	/*
 	 * Handle injected decryption faults. Unfortunately, we cannot inject

module/zfs/vdev_raidz_math.c

@@ -445,7 +445,7 @@ benchmark_raidz_impl(raidz_map_t *bench_rm, const int fn, benchmark_fn bench_fn)
  * Initialize and benchmark all supported implementations.
  */
 static void
-benchmark_raidz(void *arg)
+benchmark_raidz(void)
 {
 	raidz_impl_ops_t *curr_impl;
 	int i, c;
@@ -515,20 +515,10 @@ benchmark_raidz(void *arg)
 void
 vdev_raidz_math_init(void)
 {
-#if defined(_KERNEL)
+	/* Determine the fastest available implementation. */
-	/*
+	benchmark_raidz();
-	 * For 5.0 and latter Linux kernels the fletcher 4 benchmarks are
-	 * run in a kernel threads.  This is needed to take advantage of the
-	 * SIMD functionality, see include/linux/simd_x86.h for details.
-	 */
-	taskqid_t id = taskq_dispatch(system_taskq, benchmark_raidz,
-	    NULL, TQ_SLEEP);
-	if (id != TASKQID_INVALID) {
-		taskq_wait_id(system_taskq, id);
-	} else {
-		benchmark_raidz(NULL);
-	}
 
+#if defined(_KERNEL)
 	/* Install kstats for all implementations */
 	raidz_math_kstat = kstat_create("zfs", 0, "vdev_raidz_bench", "misc",
 	    KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VIRTUAL);
@@ -541,8 +531,6 @@ vdev_raidz_math_init(void)
 		    raidz_math_kstat_addr);
 		kstat_install(raidz_math_kstat);
 	}
-#else
-	benchmark_raidz(NULL);
 #endif
 
 	/* Finish initialization */