The branch main has been updated by bnovkov:
URL:
https://cgit.FreeBSD.org/src/commit/?id=08c7dd2fbe4fb7ae5cd6943afef04bd4cb350c1f
commit 08c7dd2fbe4fb7ae5cd6943afef04bd4cb350c1f
Author: Bojan Novković <bnov...@freebsd.org>
AuthorDate: 2024-09-08 15:57:55 +0000
Commit: Bojan Novković <bnov...@freebsd.org>
CommitDate: 2025-07-27 16:31:48 +0000
libvmmapi: Add support for setting up and configuring guest NUMA domains
This patch reworks libvmmapi to provide support for emulating NUMA
domains in guests.
More specifically, it reworks 'vm_setup_memory' to setup system memory
segments for each guest NUMA domain.
An emulated NUMA domain is described by a 'struct vmdom' in vmmapi.h.
Aside from its size in bytes, each domain can be configured to use a
specific domainset(9) policy and domain mask.
'vm_setup_memory' now takes two additional arguments - an array of
struct vmdoms and the array's size. It then proceeds to set up a memory
segment for each specified domain using the existing memory mapping
scheme. If no domain info is passed, the memory setup falls back to
the original, non-NUMA behaviour.
Differential Revision: https://reviews.freebsd.org/D44566
Reviewed by: markj
---
lib/libvmmapi/internal.h | 11 ++-
lib/libvmmapi/vmmapi.c | 181 ++++++++++++++++++++++++++++++++++-------------
lib/libvmmapi/vmmapi.h | 12 +++-
3 files changed, 145 insertions(+), 59 deletions(-)
diff --git a/lib/libvmmapi/internal.h b/lib/libvmmapi/internal.h
index aa7b1d8e6a93..4afe1cab3460 100644
--- a/lib/libvmmapi/internal.h
+++ b/lib/libvmmapi/internal.h
@@ -8,12 +8,7 @@
#define __VMMAPI_INTERNAL_H__
#include <sys/types.h>
-
-enum {
- VM_MEMSEG_LOW,
- VM_MEMSEG_HIGH,
- VM_MEMSEG_COUNT,
-};
+#include <dev/vmm/vmm_mem.h>
struct vmctx {
int fd; /* device file descriptor */
@@ -21,7 +16,9 @@ struct vmctx {
struct {
vm_paddr_t base;
vm_size_t size;
- } memsegs[VM_MEMSEG_COUNT];
+ } memsegs[VM_MAX_MEMSEGS];
+ size_t lowmem_size;
+ size_t highmem_size;
int memflags;
char *baseaddr;
char *name;
diff --git a/lib/libvmmapi/vmmapi.c b/lib/libvmmapi/vmmapi.c
index a1a5d56ff8a2..77f0f8f5c581 100644
--- a/lib/libvmmapi/vmmapi.c
+++ b/lib/libvmmapi/vmmapi.c
@@ -28,13 +28,14 @@
#include <sys/param.h>
#include <sys/capsicum.h>
+#include <sys/cpuset.h>
+#include <sys/domainset.h>
#include <sys/sysctl.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/linker.h>
#include <sys/module.h>
#include <sys/_iovec.h>
-#include <sys/cpuset.h>
#include <capsicum_helpers.h>
#include <err.h>
@@ -322,8 +323,8 @@ vm_get_guestmem_from_ctx(struct vmctx *ctx, char
**guest_baseaddr,
{
*guest_baseaddr = ctx->baseaddr;
- *lowmem_size = ctx->memsegs[VM_MEMSEG_LOW].size;
- *highmem_size = ctx->memsegs[VM_MEMSEG_HIGH].size;
+ *lowmem_size = ctx->lowmem_size;
+ *highmem_size = ctx->highmem_size;
return (0);
}
@@ -379,7 +380,8 @@ cmpseg(size_t len, const char *str, size_t len2, const char
*str2)
}
static int
-vm_alloc_memseg(struct vmctx *ctx, int segid, size_t len, const char *name)
+vm_alloc_memseg(struct vmctx *ctx, int segid, size_t len, const char *name,
+ int ds_policy, domainset_t *ds_mask, size_t ds_size)
{
struct vm_memseg memseg;
size_t n;
@@ -407,6 +409,13 @@ vm_alloc_memseg(struct vmctx *ctx, int segid, size_t len,
const char *name)
bzero(&memseg, sizeof(struct vm_memseg));
memseg.segid = segid;
memseg.len = len;
+ if (ds_mask == NULL) {
+ memseg.ds_policy = DOMAINSET_POLICY_INVALID;
+ } else {
+ memseg.ds_policy = ds_policy;
+ memseg.ds_mask = ds_mask;
+ memseg.ds_mask_size = ds_size;
+ }
if (name != NULL) {
n = strlcpy(memseg.name, name, sizeof(memseg.name));
if (n >= sizeof(memseg.name)) {
@@ -442,13 +451,14 @@ vm_get_memseg(struct vmctx *ctx, int segid, size_t *lenp,
char *namebuf,
}
static int
-setup_memory_segment(struct vmctx *ctx, vm_paddr_t gpa, size_t len, char *base)
+map_memory_segment(struct vmctx *ctx, int segid, vm_paddr_t gpa, size_t len,
+ size_t segoff, char *base)
{
char *ptr;
int error, flags;
/* Map 'len' bytes starting at 'gpa' in the guest address space */
- error = vm_mmap_memseg(ctx, gpa, VM_SYSMEM, gpa, len, PROT_ALL);
+ error = vm_mmap_memseg(ctx, gpa, segid, segoff, len, PROT_ALL);
if (error)
return (error);
@@ -464,65 +474,136 @@ setup_memory_segment(struct vmctx *ctx, vm_paddr_t gpa,
size_t len, char *base)
return (0);
}
+/*
+ * Allocates and maps virtual machine memory segments according
+ * to the NUMA topology specified by the 'doms' array.
+ *
+ * The domains are laid out sequentially in the guest's physical address space.
+ * The [VM_LOWMEM_LIMIT, VM_HIGHMEM_BASE) address range is skipped and
+ * left unmapped.
+ */
int
-vm_setup_memory(struct vmctx *ctx, size_t memsize, enum vm_mmap_style vms)
+vm_setup_memory_domains(struct vmctx *ctx, enum vm_mmap_style vms,
+ struct vm_mem_domain *doms, int ndoms)
{
- size_t objsize, len;
- vm_paddr_t gpa;
+ size_t low_len, len, totalsize;
+ struct vm_mem_domain *dom;
+ struct vm_memseg memseg;
char *baseaddr, *ptr;
- int error;
+ int error, i, segid;
+ vm_paddr_t gpa;
+ /* Sanity checks. */
assert(vms == VM_MMAP_ALL);
-
- /*
- * If 'memsize' cannot fit entirely in the 'lowmem' segment then create
- * another 'highmem' segment above VM_HIGHMEM_BASE for the remainder.
- */
- if (memsize > VM_LOWMEM_LIMIT) {
- ctx->memsegs[VM_MEMSEG_LOW].size = VM_LOWMEM_LIMIT;
- ctx->memsegs[VM_MEMSEG_HIGH].size = memsize - VM_LOWMEM_LIMIT;
- objsize = VM_HIGHMEM_BASE + ctx->memsegs[VM_MEMSEG_HIGH].size;
- } else {
- ctx->memsegs[VM_MEMSEG_LOW].size = memsize;
- ctx->memsegs[VM_MEMSEG_HIGH].size = 0;
- objsize = memsize;
+ if (doms == NULL || ndoms <= 0 || ndoms > VM_MAXMEMDOM) {
+ errno = EINVAL;
+ return (-1);
}
- error = vm_alloc_memseg(ctx, VM_SYSMEM, objsize, NULL);
- if (error)
- return (error);
+ /* Calculate total memory size. */
+ totalsize = 0;
+ for (i = 0; i < ndoms; i++)
+ totalsize += doms[i].size;
+
+ if (totalsize > VM_LOWMEM_LIMIT)
+ totalsize = VM_HIGHMEM_BASE + (totalsize - VM_LOWMEM_LIMIT);
/*
* Stake out a contiguous region covering the guest physical memory
* and the adjoining guard regions.
*/
- len = VM_MMAP_GUARD_SIZE + objsize + VM_MMAP_GUARD_SIZE;
+ len = VM_MMAP_GUARD_SIZE + totalsize + VM_MMAP_GUARD_SIZE;
ptr = mmap(NULL, len, PROT_NONE, MAP_GUARD | MAP_ALIGNED_SUPER, -1, 0);
if (ptr == MAP_FAILED)
return (-1);
-
baseaddr = ptr + VM_MMAP_GUARD_SIZE;
- if (ctx->memsegs[VM_MEMSEG_HIGH].size > 0) {
- gpa = VM_HIGHMEM_BASE;
- len = ctx->memsegs[VM_MEMSEG_HIGH].size;
- error = setup_memory_segment(ctx, gpa, len, baseaddr);
- if (error)
- return (error);
- }
- if (ctx->memsegs[VM_MEMSEG_LOW].size > 0) {
- gpa = 0;
- len = ctx->memsegs[VM_MEMSEG_LOW].size;
- error = setup_memory_segment(ctx, gpa, len, baseaddr);
- if (error)
- return (error);
- }
+ /*
+ * Allocate and map memory segments for the virtual machine.
+ */
+ gpa = VM_LOWMEM_LIMIT > 0 ? 0 : VM_HIGHMEM_BASE;
+ ctx->lowmem_size = 0;
+ ctx->highmem_size = 0;
+ for (i = 0; i < ndoms; i++) {
+ segid = VM_SYSMEM + i;
+ dom = &doms[i];
+
+ /*
+ * Check if the memory segment already exists.
+ * If 'ndoms' is greater than one, refuse to proceed if the
+ * memseg already exists. If only one domain was requested, use
+ * the existing segment to preserve the behaviour of the
previous
+ * implementation.
+ *
+ * Splitting existing memory segments is tedious and
+ * error-prone, which is why we don't support NUMA
+ * domains for bhyveload(8)-loaded VMs.
+ */
+ error = vm_get_memseg(ctx, segid, &len, memseg.name,
+ sizeof(memseg.name));
+ if (error == 0 && len != 0) {
+ if (ndoms != 1) {
+ errno = EEXIST;
+ return (-1);
+ } else
+ doms[0].size = len;
+ } else {
+ error = vm_alloc_memseg(ctx, segid, dom->size, NULL,
+ dom->ds_policy, dom->ds_mask, dom->ds_size);
+ if (error)
+ return (error);
+ }
+ /*
+ * If a domain is split by VM_LOWMEM_LIMIT then break
+ * its segment mapping into two parts, one below VM_LOWMEM_LIMIT
+ * and one above VM_HIGHMEM_BASE.
+ */
+ if (gpa <= VM_LOWMEM_LIMIT &&
+ gpa + dom->size > VM_LOWMEM_LIMIT) {
+ low_len = VM_LOWMEM_LIMIT - gpa;
+ error = map_memory_segment(ctx, segid, gpa, low_len, 0,
+ baseaddr);
+ if (error)
+ return (error);
+ ctx->lowmem_size = VM_LOWMEM_LIMIT;
+ /* Map the remainder. */
+ gpa = VM_HIGHMEM_BASE;
+ len = dom->size - low_len;
+ error = map_memory_segment(ctx, segid, gpa, len,
+ low_len, baseaddr);
+ if (error)
+ return (error);
+ } else {
+ len = dom->size;
+ error = map_memory_segment(ctx, segid, gpa, len, 0,
+ baseaddr);
+ if (error)
+ return (error);
+ }
+ if (gpa <= VM_LOWMEM_LIMIT)
+ ctx->lowmem_size += len;
+ else
+ ctx->highmem_size += len;
+ gpa += len;
+ }
ctx->baseaddr = baseaddr;
return (0);
}
+int
+vm_setup_memory(struct vmctx *ctx, size_t memsize, enum vm_mmap_style vms)
+{
+ struct vm_mem_domain dom0;
+
+ memset(&dom0, 0, sizeof(dom0));
+ dom0.ds_policy = DOMAINSET_POLICY_INVALID;
+ dom0.size = memsize;
+
+ return (vm_setup_memory_domains(ctx, vms, &dom0, 1));
+}
+
/*
* Returns a non-NULL pointer if [gaddr, gaddr+len) is entirely contained in
* the lowmem or highmem regions.
@@ -535,13 +616,13 @@ vm_map_gpa(struct vmctx *ctx, vm_paddr_t gaddr, size_t
len)
{
vm_size_t lowsize, highsize;
- lowsize = ctx->memsegs[VM_MEMSEG_LOW].size;
+ lowsize = ctx->lowmem_size;
if (lowsize > 0) {
if (gaddr < lowsize && len <= lowsize && gaddr + len <= lowsize)
return (ctx->baseaddr + gaddr);
}
- highsize = ctx->memsegs[VM_MEMSEG_HIGH].size;
+ highsize = ctx->highmem_size;
if (highsize > 0 && gaddr >= VM_HIGHMEM_BASE) {
if (gaddr < VM_HIGHMEM_BASE + highsize && len <= highsize &&
gaddr + len <= VM_HIGHMEM_BASE + highsize)
@@ -559,12 +640,12 @@ vm_rev_map_gpa(struct vmctx *ctx, void *addr)
offaddr = (char *)addr - ctx->baseaddr;
- lowsize = ctx->memsegs[VM_MEMSEG_LOW].size;
+ lowsize = ctx->lowmem_size;
if (lowsize > 0)
if (offaddr <= lowsize)
return (offaddr);
- highsize = ctx->memsegs[VM_MEMSEG_HIGH].size;
+ highsize = ctx->highmem_size;
if (highsize > 0)
if (offaddr >= VM_HIGHMEM_BASE &&
offaddr < VM_HIGHMEM_BASE + highsize)
@@ -583,8 +664,7 @@ vm_get_name(struct vmctx *ctx)
size_t
vm_get_lowmem_size(struct vmctx *ctx)
{
-
- return (ctx->memsegs[VM_MEMSEG_LOW].size);
+ return (ctx->lowmem_size);
}
vm_paddr_t
@@ -597,8 +677,7 @@ vm_get_highmem_base(struct vmctx *ctx __unused)
size_t
vm_get_highmem_size(struct vmctx *ctx)
{
-
- return (ctx->memsegs[VM_MEMSEG_HIGH].size);
+ return (ctx->highmem_size);
}
void *
@@ -616,7 +695,7 @@ vm_create_devmem(struct vmctx *ctx, int segid, const char
*name, size_t len)
goto done;
}
- error = vm_alloc_memseg(ctx, segid, len, name);
+ error = vm_alloc_memseg(ctx, segid, len, name, 0, NULL, 0);
if (error)
goto done;
diff --git a/lib/libvmmapi/vmmapi.h b/lib/libvmmapi/vmmapi.h
index 2072c0105e37..b637c45d1eff 100644
--- a/lib/libvmmapi/vmmapi.h
+++ b/lib/libvmmapi/vmmapi.h
@@ -64,6 +64,14 @@ enum vm_mmap_style {
#define VM_MEM_F_INCORE 0x01 /* include guest memory in core file */
#define VM_MEM_F_WIRED 0x02 /* guest memory is wired */
+/* Memory size and allocation policy for a single NUMA domain. */
+struct vm_mem_domain {
+ size_t size;
+ int ds_policy;
+ domainset_t *ds_mask;
+ size_t ds_size;
+};
+
__BEGIN_DECLS
/*
* Get the length and name of the memory segment identified by 'segid'.
@@ -115,7 +123,9 @@ struct vcpu *vm_vcpu_open(struct vmctx *ctx, int vcpuid);
void vm_vcpu_close(struct vcpu *vcpu);
int vcpu_id(struct vcpu *vcpu);
int vm_parse_memsize(const char *optarg, size_t *memsize);
-int vm_setup_memory(struct vmctx *ctx, size_t len, enum vm_mmap_style s);
+int vm_setup_memory(struct vmctx *ctx, size_t len, enum vm_mmap_style s);
+int vm_setup_memory_domains(struct vmctx *ctx, enum vm_mmap_style s,
+ struct vm_mem_domain *doms, int ndoms);
void *vm_map_gpa(struct vmctx *ctx, vm_paddr_t gaddr, size_t len);
/* inverse operation to vm_map_gpa - extract guest address from host pointer */
vm_paddr_t vm_rev_map_gpa(struct vmctx *ctx, void *addr);
