OS-8005 bhyve memory pressure needs to target ARC better (#354)

Reviewed by: Dan McDonald <danmcd@joyent.com>
Reviewed by: Mike Zeller <mike.zeller@joyent.com>
Approved by: Brian Bennett <brian.bennett@joyent.com>

10 files changed, 116 insertions, 9 deletions

diff --git a/usr/src/cmd/bhyve/bhyverun.c b/usr/src/cmd/bhyve/bhyverun.c
index 0cbb2c78eb..53e10bbb55 100644
--- a/usr/src/cmd/bhyve/bhyverun.c
+++ b/usr/src/cmd/bhyve/bhyverun.c
@@ -1446,6 +1446,13 @@ main(int argc, char *argv[])
 #ifdef	__FreeBSD__
 	err = vm_setup_memory(ctx, memsize, VM_MMAP_ALL);
 #else
+	err = vm_arc_resv(ctx, memsize);
+	if (err != 0) {
+		(void) fprintf(stderr, "Could not shrink ARC: %s\n",
+		    strerror(err));
+		exit(4);
+	}
+
 	do {
 		errno = 0;
 		err = vm_setup_memory(ctx, memsize, VM_MMAP_ALL);
diff --git a/usr/src/lib/libvmmapi/common/mapfile-vers b/usr/src/lib/libvmmapi/common/mapfile-vers
index be0a055490..e6468c36fb 100644
--- a/usr/src/lib/libvmmapi/common/mapfile-vers
+++ b/usr/src/lib/libvmmapi/common/mapfile-vers
@@ -11,7 +11,7 @@
 
 #
 # Copyright 2013 Pluribus Networks Inc.
-# Copyright 2019 Joyent, Inc.
+# Copyright 2020 Joyent, Inc.
 # Copyright 2020 Oxide Computer Company
 #
 
@@ -37,6 +37,7 @@ SYMBOL_VERSION ILLUMOSprivate {
 		vm_activate_cpu;
 		vm_active_cpus;
 		vm_apicid2vcpu;
+		vm_arc_resv;
 		vm_assign_pptdev;
 		vm_capability_name2type;
 		vm_capability_type2name;
diff --git a/usr/src/lib/libvmmapi/common/vmmapi.c b/usr/src/lib/libvmmapi/common/vmmapi.c
index fcb098a74f..0d084b9cd5 100644
--- a/usr/src/lib/libvmmapi/common/vmmapi.c
+++ b/usr/src/lib/libvmmapi/common/vmmapi.c
@@ -1909,6 +1909,14 @@ vm_set_run_state(struct vmctx *ctx, int vcpu, enum vcpu_run_state state,
 	return (0);
 }
 
+int
+vm_arc_resv(struct vmctx *ctx, size_t len)
+{
+	if (ioctl(ctx->fd, VM_ARC_RESV, (uint64_t)len) != 0) {
+		return (errno);
+	}
+	return (0);
+}
 #endif /* __FreeBSD__ */
 
 #ifdef __FreeBSD__
diff --git a/usr/src/lib/libvmmapi/common/vmmapi.h b/usr/src/lib/libvmmapi/common/vmmapi.h
index 72e43a4e3d..cad555e498 100644
--- a/usr/src/lib/libvmmapi/common/vmmapi.h
+++ b/usr/src/lib/libvmmapi/common/vmmapi.h
@@ -38,7 +38,7 @@
  * http://www.illumos.org/license/CDDL.
  *
  * Copyright 2015 Pluribus Networks Inc.
- * Copyright 2019 Joyent, Inc.
+ * Copyright 2020 Joyent, Inc.
  * Copyright 2020 Oxide Computer Company
  */
 
@@ -310,6 +310,7 @@ int vm_get_run_state(struct vmctx *ctx, int vcpu, enum vcpu_run_state *state,
     uint8_t *sipi_vector);
 int vm_set_run_state(struct vmctx *ctx, int vcpu, enum vcpu_run_state state,
     uint8_t sipi_vector);
+int	vm_arc_resv(struct vmctx *ctx, size_t);
 #endif	/* __FreeBSD__ */
 
 #ifdef	__FreeBSD__
diff --git a/usr/src/uts/common/fs/zfs/arc.c b/usr/src/uts/common/fs/zfs/arc.c
index e88213a491..cc5143e4df 100644
--- a/usr/src/uts/common/fs/zfs/arc.c
+++ b/usr/src/uts/common/fs/zfs/arc.c
@@ -20,7 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2019, Joyent, Inc.
+ * Copyright 2020 Joyent, Inc.
  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
  * Copyright (c) 2014 by Saso Kiselkov. All rights reserved.
  * Copyright 2017 Nexenta Systems, Inc.  All rights reserved.
@@ -4399,6 +4399,7 @@ typedef enum free_memory_reason_t {
 	FMR_PAGES_PP_MAXIMUM,
 	FMR_HEAP_ARENA,
 	FMR_ZIO_ARENA,
+	FMR_VIRT_MACHINE,	/* 'VM' seems ambiguous in this context */
 } free_memory_reason_t;
 
 int64_t last_free_memory;
@@ -4414,6 +4415,48 @@ int64_t arc_pages_pp_reserve = 64;
  */
 int64_t arc_swapfs_reserve = 64;
 
+static volatile uint64_t arc_virt_machine_reserved;
+
+/*
+ * XXX: A possible concern is that we allow arc_virt_machine_reserved to
+ * get so large that we cause the arc to perform a lot of additional
+ * work to keep the arc extremely small. We may want to set limits to
+ * the size of arc_virt_machine_reserved and disallow reservations
+ * beyond that limit.
+ */
+int
+arc_virt_machine_reserve(size_t pages)
+{
+	uint64_t newv;
+
+	newv = atomic_add_64_nv(&arc_virt_machine_reserved, pages);
+
+	/*
+	 * Since arc_virt_machine_reserved effectively lowers arc_c_max
+	 * as needed for vmm memory, if this request would put the arc
+	 * under arc_c_min, we reject it.  arc_c_min should be a value that
+	 * ensures reasonable performance for non-VMM stuff, as well as keep
+	 * us from dipping below lotsfree, which could trigger the pager
+	 * (and send the system toa grinding halt while it pages).
+	 *
+	 * XXX: This is a bit hacky and might be better done w/ a mutex
+	 * instead of atomic ops.
+	 */
+	if (newv + arc_c_min > arc_c_max) {
+		atomic_add_64(&arc_virt_machine_reserved, -(int64_t)pages);
+		return (ENOMEM);
+	}
+
+	zthr_wakeup(arc_reap_zthr);
+	return (0);
+}
+
+void
+arc_virt_machine_release(size_t pages)
+{
+	atomic_add_64(&arc_virt_machine_reserved, -(int64_t)pages);
+}
+
 /*
  * Return the amount of memory that can be consumed before reclaim will be
  * needed.  Positive if there is sufficient free memory, negative indicates
@@ -4477,6 +4520,17 @@ arc_available_memory(void)
 		r = FMR_PAGES_PP_MAXIMUM;
 	}
 
+	/*
+	 * Check that we have enough memory for any virtual machines that
+	 * are running or starting. We add desfree to keep us out of
+	 * particularly dire circumstances.
+	 */
+	n = PAGESIZE * (availrmem - arc_virt_machine_reserved - desfree);
+	if (n < lowest) {
+		lowest = n;
+		r = FMR_VIRT_MACHINE;
+	}
+
 #if defined(__i386)
 	/*
 	 * If we're on an i386 platform, it's possible that we'll exhaust the
diff --git a/usr/src/uts/i86pc/io/vmm/sys/vmm_kernel.h b/usr/src/uts/i86pc/io/vmm/sys/vmm_kernel.h
index 09e9afd8a8..ec1707ea79 100644
--- a/usr/src/uts/i86pc/io/vmm/sys/vmm_kernel.h
+++ b/usr/src/uts/i86pc/io/vmm/sys/vmm_kernel.h
@@ -203,6 +203,10 @@ int vm_service_mmio_read(struct vm *vm, int cpuid, uint64_t gpa, uint64_t *rval,
 int vm_service_mmio_write(struct vm *vm, int cpuid, uint64_t gpa, uint64_t wval,
     int wsize);
 
+#ifndef __FreeBSD__
+int vm_arc_resv(struct vm *vm, size_t);
+#endif
+
 #ifdef _SYS__CPUSET_H_
 cpuset_t vm_active_cpus(struct vm *vm);
 cpuset_t vm_debug_cpus(struct vm *vm);
diff --git a/usr/src/uts/i86pc/io/vmm/vmm.c b/usr/src/uts/i86pc/io/vmm/vmm.c
index 1cd0b23a1c..ad3a9f548f 100644
--- a/usr/src/uts/i86pc/io/vmm/vmm.c
+++ b/usr/src/uts/i86pc/io/vmm/vmm.c
@@ -38,7 +38,7 @@
  * http://www.illumos.org/license/CDDL.
  *
  * Copyright 2015 Pluribus Networks Inc.
- * Copyright 2018 Joyent, Inc.
+ * Copyright 2021 Joyent, Inc.
  * Copyright 2020 Oxide Computer Company
  */
 
@@ -196,6 +196,7 @@ struct vm {
 	uint16_t	threads;		/* (o) num of threads/core */
 	uint16_t	maxcpus;		/* (o) max pluggable cpus */
 	uint64_t	boot_tsc_offset;	/* (i) TSC offset at VM boot */
+	size_t		arc_resv;		/* # of pages take from ARC */
 
 	struct ioport_config ioports;		/* (o) ioport handling */
 };
@@ -287,6 +288,9 @@ static int vcpu_vector_sipi(struct vm *vm, int vcpuid, uint8_t vector);
 #ifndef __FreeBSD__
 static void vm_clear_memseg(struct vm *, int);
 
+extern int arc_virt_machine_reserve(size_t);
+extern void arc_virt_machine_release(size_t);
+
 /* Flags for vtc_status */
 #define	VTCS_FPU_RESTORED	1 /* guest FPU restored, host FPU saved */
 #define	VTCS_FPU_CTX_CRITICAL	2 /* in ctx where FPU restore cannot be lazy */
@@ -296,6 +300,7 @@ typedef struct vm_thread_ctx {
 	int		vtc_vcpuid;
 	uint_t		vtc_status;
 } vm_thread_ctx_t;
+
 #endif /* __FreeBSD__ */
 
 #ifdef KTR
@@ -645,6 +650,12 @@ vm_cleanup(struct vm *vm, bool destroy)
 
 		VMSPACE_FREE(vm->vmspace);
 		vm->vmspace = NULL;
+
+#ifndef __FreeBSD__
+		arc_virt_machine_release(vm->arc_resv);
+		vm->arc_resv = 0;
+#endif
+
 	}
 #ifndef __FreeBSD__
 	else {
@@ -3775,3 +3786,20 @@ vm_ioport_unhook(struct vm *vm, void **cookie)
 
 	*cookie = NULL;
 }
+
+#ifndef __FreeBSD__
+int
+vm_arc_resv(struct vm *vm, uint64_t len)
+{
+	/* Since we already have the compat macros included, we use those */
+	size_t pages = (size_t)roundup2(len, PAGE_SIZE) >> PAGE_SHIFT;
+	int err = 0;
+
+	err = arc_virt_machine_reserve(pages);
+	if (err != 0)
+		return (err);
+
+	vm->arc_resv += pages;
+	return (0);
+}
+#endif /* __FreeBSD__ */
diff --git a/usr/src/uts/i86pc/io/vmm/vmm_sol_dev.c b/usr/src/uts/i86pc/io/vmm/vmm_sol_dev.c
index 7b04aed61a..9267befd19 100644
--- a/usr/src/uts/i86pc/io/vmm/vmm_sol_dev.c
+++ b/usr/src/uts/i86pc/io/vmm/vmm_sol_dev.c
@@ -12,7 +12,7 @@
 
 /*
  * Copyright 2015 Pluribus Networks Inc.
- * Copyright 2019 Joyent, Inc.
+ * Copyright 2020 Joyent, Inc.
  * Copyright 2020 OmniOS Community Edition (OmniOSce) Association.
  * Copyright 2020 Oxide Computer Company
  */
@@ -471,6 +471,7 @@ vmmdev_do_ioctl(vmm_softc_t *sc, int cmd, intptr_t arg, int md,
 	case VM_MMAP_MEMSEG:
 	case VM_WRLOCK_CYCLE:
 	case VM_PMTMR_LOCATE:
+	case VM_ARC_RESV:
 		vmm_write_lock(sc);
 		lock_type = LOCK_WRITE_HOLD;
 		break;
@@ -1396,6 +1397,9 @@ vmmdev_do_ioctl(vmm_softc_t *sc, int cmd, intptr_t arg, int md,
 		 */
 		break;
 	}
+	case VM_ARC_RESV:
+		error = vm_arc_resv(sc->vmm_vm, (uint64_t)arg);
+		break;
 #endif
 	default:
 		error = ENOTTY;
diff --git a/usr/src/uts/i86pc/sys/vmm_dev.h b/usr/src/uts/i86pc/sys/vmm_dev.h
index f4a68636b3..3f20f1d5ee 100644
--- a/usr/src/uts/i86pc/sys/vmm_dev.h
+++ b/usr/src/uts/i86pc/sys/vmm_dev.h
@@ -38,7 +38,7 @@
  * http://www.illumos.org/license/CDDL.
  *
  * Copyright 2015 Pluribus Networks Inc.
- * Copyright 2019 Joyent, Inc.
+ * Copyright 2020 Joyent, Inc.
  */
 
 #ifndef	_VMM_DEV_H_
@@ -369,6 +369,7 @@ struct vm_run_state {
 #define	VM_RESUME_CPU			(VMM_IOC_BASE | 0x1e)
 
 #define	VM_PPTDEV_DISABLE_MSIX		(VMM_IOC_BASE | 0x1f)
+#define	VM_ARC_RESV			(VMM_IOC_BASE | 0xfe)
 
 #define	VM_DEVMEM_GETOFFSET		(VMM_IOC_BASE | 0xff)
 
diff --git a/usr/src/uts/i86pc/vmm/Makefile b/usr/src/uts/i86pc/vmm/Makefile
index 0106dd0a0f..2c0843f165 100644
--- a/usr/src/uts/i86pc/vmm/Makefile
+++ b/usr/src/uts/i86pc/vmm/Makefile
@@ -11,7 +11,7 @@
 
 #
 # Copyright 2013 Pluribus Networks Inc.
-# Copyright 2019 Joyent, Inc.
+# Copyright 2020 Joyent, Inc.
 #
 
 #
@@ -43,7 +43,6 @@ INSTALL_TARGET	= $(BINARY) $(ROOTMODULE) $(ROOT_CONFFILE)
 #	Overrides and additions
 #
 
-
 # 3rd party code
 SMOFF += all_func_returns
 
@@ -60,7 +59,7 @@ AS_INC_PATH	+= -I$(UTSBASE)/i86pc/io/vmm -I$(OBJS_DIR)
 # enable collection of VMM statistics
 CFLAGS		+= -DVMM_KEEP_STATS
 
-LDFLAGS         += -N misc/acpica -N misc/pcie -N fs/dev
+LDFLAGS         += -N misc/acpica -N misc/pcie -N fs/dev -N fs/zfs
 LDFLAGS         += -z type=kmod -M $(MAPFILE)
 
 OFFSETS_VMX	= $(CONF_SRCDIR)/intel/offsets.in