Linux 5.0 compat: SIMD compatibility

Restore the SIMD optimization for 4.19.38 LTS, 4.14.120 LTS,
and 5.0 and newer kernels.  This is accomplished by leveraging
the fact that by definition dedicated kernel threads never need
to concern themselves with saving and restoring the user FPU state.
Therefore, they may use the FPU as long as we can guarantee user
tasks always restore their FPU state before context switching back
to user space.

For the 5.0 and 5.1 kernels disabling preemption and local
interrupts is sufficient to allow the FPU to be used.  All non-kernel
threads will restore the preserved user FPU state.

For 5.2 and latter kernels the user FPU state restoration will be
skipped if the kernel determines the registers have not changed.
Therefore, for these kernels we need to perform the additional
step of saving and restoring the FPU registers.  Invalidating the
per-cpu global tracking the FPU state would force a restore but
that functionality is private to the core x86 FPU implementation
and unavailable.

In practice, restricting SIMD to kernel threads is not a major
restriction for ZFS.  The vast majority of SIMD operations are
already performed by the IO pipeline.  The remaining cases are
relatively infrequent and can be handled by the generic code
without significant impact.  The two most noteworthy cases are:

  1) Decrypting the wrapping key for an encrypted dataset,
     i.e. `zfs load-key`.  All other encryption and decryption
     operations will use the SIMD optimized implementations.

  2) Generating the payload checksums for a `zfs send` stream.

In order to avoid making any changes to the higher layers of ZFS
all of the `*_get_ops()` functions were updated to take in to
consideration the calling context.  This allows for the fastest
implementation to be used as appropriate (see kfpu_allowed()).

The only other notable instance of SIMD operations being used
outside a kernel thread was at module load time.  This code
was moved in to a taskq in order to accommodate the new kernel
thread restriction.

Finally, a few other modifications were made in order to further
harden this code and facilitate testing.  They include updating
each implementations operations structure to be declared as a
constant.  And allowing "cycle" to be set when selecting the
preferred ops in the kernel as well as user space.

Reviewed-by: Tony Hutter <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Closes #8754 
Closes #8793 
Closes #8965 

1 files changed, 58 insertions, 30 deletions

diff --git a/module/zcommon/zfs_fletcher.c b/module/zcommon/zfs_fletcher.c
index 5a991ba60..b75d8ab00 100644
--- a/module/zcommon/zfs_fletcher.c
+++ b/module/zcommon/zfs_fletcher.c
@@ -140,6 +140,7 @@
 #include <sys/zio_checksum.h>
 #include <sys/zfs_context.h>
 #include <zfs_fletcher.h>
+#include <linux/simd.h>
 
 #define	FLETCHER_MIN_SIMD_SIZE	64
 
@@ -205,21 +206,19 @@ static struct fletcher_4_impl_selector {
 	const char	*fis_name;
 	uint32_t	fis_sel;
 } fletcher_4_impl_selectors[] = {
-#if !defined(_KERNEL)
 	{ "cycle",	IMPL_CYCLE },
-#endif
 	{ "fastest",	IMPL_FASTEST },
 	{ "scalar",	IMPL_SCALAR }
 };
 
 #if defined(_KERNEL)
 static kstat_t *fletcher_4_kstat;
-#endif
 
 static struct fletcher_4_kstat {
 	uint64_t native;
 	uint64_t byteswap;
 } fletcher_4_stat_data[ARRAY_SIZE(fletcher_4_impls) + 1];
+#endif
 
 /* Indicate that benchmark has been completed */
 static boolean_t fletcher_4_initialized = B_FALSE;
@@ -408,32 +407,36 @@ fletcher_4_impl_set(const char *val)
 	return (err);
 }
 
+/*
+ * Returns the Fletcher 4 operations for checksums.   When a SIMD
+ * implementation is not allowed in the current context, then fallback
+ * to the fastest generic implementation.
+ */
 static inline const fletcher_4_ops_t *
 fletcher_4_impl_get(void)
 {
-	fletcher_4_ops_t *ops = NULL;
-	const uint32_t impl = IMPL_READ(fletcher_4_impl_chosen);
+	if (!kfpu_allowed())
+		return (&fletcher_4_superscalar4_ops);
+
+	const fletcher_4_ops_t *ops = NULL;
+	uint32_t impl = IMPL_READ(fletcher_4_impl_chosen);
 
 	switch (impl) {
 	case IMPL_FASTEST:
 		ASSERT(fletcher_4_initialized);
 		ops = &fletcher_4_fastest_impl;
 		break;
-#if !defined(_KERNEL)
-	case IMPL_CYCLE: {
+	case IMPL_CYCLE:
+		/* Cycle through supported implementations */
 		ASSERT(fletcher_4_initialized);
 		ASSERT3U(fletcher_4_supp_impls_cnt, >, 0);
-
 		static uint32_t cycle_count = 0;
 		uint32_t idx = (++cycle_count) % fletcher_4_supp_impls_cnt;
 		ops = fletcher_4_supp_impls[idx];
-	}
-	break;
-#endif
+		break;
 	default:
 		ASSERT3U(fletcher_4_supp_impls_cnt, >, 0);
 		ASSERT3U(impl, <, fletcher_4_supp_impls_cnt);
-
 		ops = fletcher_4_supp_impls[impl];
 		break;
 	}
@@ -658,6 +661,7 @@ fletcher_4_kstat_addr(kstat_t *ksp, loff_t n)
 typedef void fletcher_checksum_func_t(const void *, uint64_t, const void *,
 					zio_cksum_t *);
 
+#if defined(_KERNEL)
 static void
 fletcher_4_benchmark_impl(boolean_t native, char *data, uint64_t data_size)
 {
@@ -716,16 +720,18 @@ fletcher_4_benchmark_impl(boolean_t native, char *data, uint64_t data_size)
 	/* restore original selection */
 	atomic_swap_32(&fletcher_4_impl_chosen, sel_save);
 }
+#endif /* _KERNEL */
 
-void
-fletcher_4_init(void)
+/*
+ * Initialize and benchmark all supported implementations.
+ */
+static void
+fletcher_4_benchmark(void *arg)
 {
-	static const size_t data_size = 1 << SPA_OLD_MAXBLOCKSHIFT; /* 128kiB */
 	fletcher_4_ops_t *curr_impl;
-	char *databuf;
 	int i, c;
 
-	/* move supported impl into fletcher_4_supp_impls */
+	/* Move supported implementations into fletcher_4_supp_impls */
 	for (i = 0, c = 0; i < ARRAY_SIZE(fletcher_4_impls); i++) {
 		curr_impl = (fletcher_4_ops_t *)fletcher_4_impls[i];
 
@@ -735,19 +741,10 @@ fletcher_4_init(void)
 	membar_producer();	/* complete fletcher_4_supp_impls[] init */
 	fletcher_4_supp_impls_cnt = c;	/* number of supported impl */
 
-#if !defined(_KERNEL)
-	/* Skip benchmarking and use last implementation as fastest */
-	memcpy(&fletcher_4_fastest_impl,
-	    fletcher_4_supp_impls[fletcher_4_supp_impls_cnt-1],
-	    sizeof (fletcher_4_fastest_impl));
-	fletcher_4_fastest_impl.name = "fastest";
-	membar_producer();
+#if defined(_KERNEL)
+	static const size_t data_size = 1 << SPA_OLD_MAXBLOCKSHIFT; /* 128kiB */
+	char *databuf = vmem_alloc(data_size, KM_SLEEP);
 
-	fletcher_4_initialized = B_TRUE;
-	return;
-#endif
-	/* Benchmark all supported implementations */
-	databuf = vmem_alloc(data_size, KM_SLEEP);
 	for (i = 0; i < data_size / sizeof (uint64_t); i++)
 		((uint64_t *)databuf)[i] = (uintptr_t)(databuf+i); /* warm-up */
 
@@ -755,9 +752,38 @@ fletcher_4_init(void)
 	fletcher_4_benchmark_impl(B_TRUE, databuf, data_size);
 
 	vmem_free(databuf, data_size);
+#else
+	/*
+	 * Skip the benchmark in user space to avoid impacting libzpool
+	 * consumers (zdb, zhack, zinject, ztest).  The last implementation
+	 * is assumed to be the fastest and used by default.
+	 */
+	memcpy(&fletcher_4_fastest_impl,
+	    fletcher_4_supp_impls[fletcher_4_supp_impls_cnt - 1],
+	    sizeof (fletcher_4_fastest_impl));
+	fletcher_4_fastest_impl.name = "fastest";
+	membar_producer();
+#endif /* _KERNEL */
+}
 
+void
+fletcher_4_init(void)
+{
 #if defined(_KERNEL)
-	/* install kstats for all implementations */
+	/*
+	 * 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, fletcher_4_benchmark,
+	    NULL, TQ_SLEEP);
+	if (id != TASKQID_INVALID) {
+		taskq_wait_id(system_taskq, id);
+	} else {
+		fletcher_4_benchmark(NULL);
+	}
+
+	/* Install kstats for all implementations */
 	fletcher_4_kstat = kstat_create("zfs", 0, "fletcher_4_bench", "misc",
 	    KSTAT_TYPE_RAW, 0, KSTAT_FLAG_VIRTUAL);
 	if (fletcher_4_kstat != NULL) {
@@ -769,6 +795,8 @@ fletcher_4_init(void)
 		    fletcher_4_kstat_addr);
 		kstat_install(fletcher_4_kstat);
 	}
+#else
+	fletcher_4_benchmark(NULL);
 #endif
 
 	/* Finish initialization */