MFC r281026, r281108, r281109:

Make ZFS ARC track both KVA usage and fragmentation.

Even on Illumos, with its much larger KVA, ZFS ARC steps back if KVA usage
reaches certain threshold (3/4 on i386 or 16/17 otherwise).  FreeBSD has
even less KVA, but had no such limit on archs with direct map as amd64.
As result, on machines with a lot of RAM, during load with very small user-
space memory pressure, such as `zfs send`, it was possible to reach state,
when there is enough both physical RAM and KVA (I've seen up to 25-30%),
but no continuous KVA range to allocate even single 128KB I/O request.

Address this situation from two sides:
 - restore KVA usage limitations in a way the most close to Illumos;
 - introduce new requirement for KVA fragmentation, specifying that we
should have at least one sequential KVA range of zfs_max_recordsize bytes.

Experiments show that first limitation done alone is not sufficient.  On
machine with 64GB of RAM it is sometimes needed to drop up to half of ARC
size to get at leats one 1MB KVA chunk.  Statically limiting ARC to half
of KVA/RAM is too strict, so second limitation makes it to work in cycles:
accumulate trash up to certain critical mass, do massive spring-cleaning,
and then start littering again.

3 files changed, 25 insertions, 2 deletions

diff --git a/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c b/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c
index 255b5e0..a8f13d2 100644
--- a/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c
+++ b/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c
@@ -2606,8 +2606,11 @@ arc_reclaim_needed(void)
 		    (vmem_size(heap_arena, VMEM_FREE | VMEM_ALLOC)) >> 2);
 		return (1);
 	}
+#define	zio_arena	NULL
+#else
+#define	zio_arena	heap_arena
 #endif
-#ifdef sun
+
 	/*
 	 * If zio data pages are being allocated out of a separate heap segment,
 	 * then enforce that the size of available vmem for this arena remains
@@ -2621,7 +2624,14 @@ arc_reclaim_needed(void)
 	    vmem_size(zio_arena, VMEM_FREE) <
 	    (vmem_size(zio_arena, VMEM_ALLOC) >> 4))
 		return (1);
-#endif	/* sun */
+
+	/*
+	 * Above limits know nothing about real level of KVA fragmentation.
+	 * Start aggressive reclamation if too little sequential KVA left.
+	 */
+	if (vmem_size(heap_arena, VMEM_MAXFREE) < zfs_max_recordsize)
+		return (1);
+
 #else	/* _KERNEL */
 	if (spa_get_random(100) == 0)
 		return (1);
diff --git a/sys/kern/subr_vmem.c b/sys/kern/subr_vmem.c
index baaee84..8cc020a 100644
--- a/sys/kern/subr_vmem.c
+++ b/sys/kern/subr_vmem.c
@@ -1319,6 +1319,7 @@ vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags)
 vmem_size_t
 vmem_size(vmem_t *vm, int typemask)
 {
+	int i;
 
 	switch (typemask) {
 	case VMEM_ALLOC:
@@ -1327,6 +1328,17 @@ vmem_size(vmem_t *vm, int typemask)
 		return vm->vm_size - vm->vm_inuse;
 	case VMEM_FREE|VMEM_ALLOC:
 		return vm->vm_size;
+	case VMEM_MAXFREE:
+		VMEM_LOCK(vm);
+		for (i = VMEM_MAXORDER - 1; i >= 0; i--) {
+			if (LIST_EMPTY(&vm->vm_freelist[i]))
+				continue;
+			VMEM_UNLOCK(vm);
+			return ((vmem_size_t)ORDER2SIZE(i) <<
+			    vm->vm_quantum_shift);
+		}
+		VMEM_UNLOCK(vm);
+		return (0);
 	default:
 		panic("vmem_size");
 	}
diff --git a/sys/sys/vmem.h b/sys/sys/vmem.h
index 47f55fb..f6b960a 100644
--- a/sys/sys/vmem.h
+++ b/sys/sys/vmem.h
@@ -129,6 +129,7 @@ void vmem_startup(void);
 /* vmem_size typemask */
 #define VMEM_ALLOC	0x01
 #define VMEM_FREE	0x02
+#define VMEM_MAXFREE	0x10
 
 #endif /* _KERNEL */