[PATCH kvmtool v2 0/4] Add guest_memfd support for arm64

[PATCH kvmtool v2 0/4] Add guest_memfd support for arm64

[Fuad Tabba]

Hi folks,

Changes since v1 [1]:
- Use kvm_userspace_memory_region2 only for guest_memfd slots (Suzuki)
- Check KVM_CAP_GUEST_MEMFD before KVM_CAP_GUEST_MEMFD_FLAGS (Alexandru)
- Close guest_memfd fd after unmapping at teardown (Alexandru)
- Carried Reviewed-bys

Due to popular demand, i.e., Alexandru asking me about it, I repolished my
guest_memfd support series. The one I sent years ago was stale and
over-engineered, so this is a fresh start instead.

This series adds support for backing guest RAM with guest_memfd on
arm64, for non-protected VMs under both regular KVM and pKVM.

The first two patches are standalone fixes in the code this series
builds on: an uninitialized return value in kvm__for_each_mem_bank(),
and doubled newlines in die() messages. Patch 3 adds guest_memfd
plumbing to kvm__register_mem(), using kvm_userspace_memory_region2
only for guest_memfd-backed slots.

A new --guest-memfd option creates guest RAM from a guest_memfd with
mmap support (GUEST_MEMFD_FLAG_MMAP and GUEST_MEMFD_FLAG_INIT_SHARED,
Linux 6.18+), maps it to provide the userspace mapping, and registers
it with KVM_SET_USER_MEMORY_REGION2.

Protected VMs need in-place shared/private conversion of guest_memfd,
which is still in progress in the kernel [2], so --protected rejects
the option.

Based on kvmtool master (ca0ddafa941d).

Cheers,
/fuad

[1] https://lore.kernel.org/all/20260706083555.302972-1-fuad.tabba@linux.dev/
[2] https://lore.kernel.org/all/20260618-gmem-inplace-conversion-v8-0-9d2959357853@google.com/

Fuad Tabba (4):
  Initialize the return value in kvm__for_each_mem_bank()
  Remove newline from end of die() aborts
  Add support for registering guest_memfd-backed memory regions
  arm64: Add --guest-memfd option to back guest RAM with guest_memfd

 arm64/include/kvm/kvm-config-arch.h |  5 ++-
 arm64/kvm.c                         | 43 +++++++++++++++++++++-----
 hw/cfi_flash.c                      |  2 +-
 include/kvm/kvm.h                   | 16 ++++++++--
 include/kvm/util.h                  |  1 +
 kvm.c                               | 42 ++++++++++++++++++--------
 util/util.c                         | 47 ++++++++++++++++++++++++++---
 7 files changed, 127 insertions(+), 29 deletions(-)

-- 
2.39.5

[PATCH kvmtool v2 1/4] Initialize the return value in kvm__for_each_mem_bank()

[Fuad Tabba]

If none of the bank types match, the function would return an
uninitialized value.

Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
---
 kvm.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/kvm.c b/kvm.c
index 7a52c6a..95ccfcb 100644
--- a/kvm.c
+++ b/kvm.c
@@ -383,14 +383,15 @@ u64 host_to_guest_flat(struct kvm *kvm, void *ptr)
  * their type.
  *
  * If one call to @fun returns a non-zero value, stop iterating and return the
- * value. Otherwise, return zero.
+ * value. If none of the bank types match, return -ENODEV. Otherwise, return
+ * zero.
  */
 int kvm__for_each_mem_bank(struct kvm *kvm, enum kvm_mem_type type,
 			   int (*fun)(struct kvm *kvm, struct kvm_mem_bank *bank, void *data),
 			   void *data)
 {
-	int ret;
 	struct kvm_mem_bank *bank;
+	int ret = -ENODEV;
 
 	list_for_each_entry(bank, &kvm->mem_banks, list) {
 		if (type != KVM_MEM_TYPE_ALL && !(bank->type & type))
-- 
2.39.5

[PATCH kvmtool v2 2/4] Remove newline from end of die() aborts

[Fuad Tabba]

die() already adds the newline at the end of the string.

Reviewed-by: Suzuki K Poulose <suzuki.poulose@arm.com>
Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
---
 util/util.c | 10 +++++-----
 1 file changed, 5 insertions(+), 5 deletions(-)

diff --git a/util/util.c b/util/util.c
index 6f41e2c..05d49c4 100644
--- a/util/util.c
+++ b/util/util.c
@@ -128,14 +128,14 @@ void *mmap_hugetlbfs(struct kvm *kvm, const char *htlbfs_path, u64 size)
 	unsigned long blk_size;
 
 	if (statfs(htlbfs_path, &sfs) < 0)
-		die("Can't stat %s\n", htlbfs_path);
+		die("Can't stat %s", htlbfs_path);
 
 	if ((unsigned int)sfs.f_type != HUGETLBFS_MAGIC)
-		die("%s is not hugetlbfs!\n", htlbfs_path);
+		die("%s is not hugetlbfs!", htlbfs_path);
 
 	blk_size = (unsigned long)sfs.f_bsize;
 	if (sfs.f_bsize == 0 || blk_size > size) {
-		die("Can't use hugetlbfs pagesize %ld for mem size %lld\n",
+		die("Can't use hugetlbfs pagesize %ld for mem size %lld",
 			blk_size, (unsigned long long)size);
 	}
 
@@ -144,10 +144,10 @@ void *mmap_hugetlbfs(struct kvm *kvm, const char *htlbfs_path, u64 size)
 	snprintf(mpath, PATH_MAX, "%s/kvmtoolXXXXXX", htlbfs_path);
 	fd = mkstemp(mpath);
 	if (fd < 0)
-		die("Can't open %s for hugetlbfs map\n", mpath);
+		die("Can't open %s for hugetlbfs map", mpath);
 	unlink(mpath);
 	if (ftruncate(fd, size) < 0)
-		die("Can't ftruncate for mem mapping size %lld\n",
+		die("Can't ftruncate for mem mapping size %lld",
 			(unsigned long long)size);
 	addr = mmap(NULL, size, PROT_RW, MAP_PRIVATE, fd, 0);
 	close(fd);
-- 
2.39.5

[PATCH kvmtool v2 3/4] Add support for registering guest_memfd-backed memory regions

[Fuad Tabba]

Extend kvm__register_mem() with a guest_memfd file descriptor and
offset. A bank with a guest_memfd is registered with
KVM_SET_USER_MEMORY_REGION2 and KVM_MEM_GUEST_MEMFD, binding the
memslot to the file. This prepares for backing guest RAM with
guest_memfd on arm64.

Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
---
 hw/cfi_flash.c    |  2 +-
 include/kvm/kvm.h | 15 +++++++++++++--
 kvm.c             | 36 ++++++++++++++++++++++++++----------
 3 files changed, 40 insertions(+), 13 deletions(-)

diff --git a/hw/cfi_flash.c b/hw/cfi_flash.c
index b68d9e0..4413955 100644
--- a/hw/cfi_flash.c
+++ b/hw/cfi_flash.c
@@ -451,7 +451,7 @@ static int map_flash_memory(struct kvm *kvm, struct cfi_flash_device *sfdev)
 	int ret;
 
 	ret = kvm__register_mem(kvm, sfdev->base_addr, sfdev->size,
-				sfdev->flash_memory,
+				sfdev->flash_memory, -1, 0,
 				KVM_MEM_TYPE_RAM | KVM_MEM_TYPE_READONLY);
 	if (!ret)
 		sfdev->is_mapped = true;
diff --git a/include/kvm/kvm.h b/include/kvm/kvm.h
index eb23e2f..5e7b5eb 100644
--- a/include/kvm/kvm.h
+++ b/include/kvm/kvm.h
@@ -128,11 +128,22 @@ bool kvm__emulate_io(struct kvm_cpu *vcpu, u16 port, void *data, int direction,
 bool kvm__emulate_mmio(struct kvm_cpu *vcpu, u64 phys_addr, u8 *data, u32 len, u8 is_write);
 int kvm__destroy_mem(struct kvm *kvm, u64 guest_phys, u64 size, void *userspace_addr);
 int kvm__register_mem(struct kvm *kvm, u64 guest_phys, u64 size, void *userspace_addr,
+		      int guest_memfd, u64 guest_memfd_offset,
 		      enum kvm_mem_type type);
 static inline int kvm__register_ram(struct kvm *kvm, u64 guest_phys, u64 size,
 				    void *userspace_addr)
 {
 	return kvm__register_mem(kvm, guest_phys, size, userspace_addr,
+				 -1, 0, KVM_MEM_TYPE_RAM);
+}
+
+static inline int kvm__register_ram_guest_memfd(struct kvm *kvm, u64 guest_phys,
+						u64 size, void *userspace_addr,
+						int guest_memfd,
+						u64 guest_memfd_offset)
+{
+	return kvm__register_mem(kvm, guest_phys, size, userspace_addr,
+				 guest_memfd, guest_memfd_offset,
 				 KVM_MEM_TYPE_RAM);
 }
 
@@ -140,13 +151,13 @@ static inline int kvm__register_dev_mem(struct kvm *kvm, u64 guest_phys,
 					u64 size, void *userspace_addr)
 {
 	return kvm__register_mem(kvm, guest_phys, size, userspace_addr,
-				 KVM_MEM_TYPE_DEVICE);
+				 -1, 0, KVM_MEM_TYPE_DEVICE);
 }
 
 static inline int kvm__reserve_mem(struct kvm *kvm, u64 guest_phys, u64 size)
 {
 	return kvm__register_mem(kvm, guest_phys, size, NULL,
-				 KVM_MEM_TYPE_RESERVED);
+				 -1, 0, KVM_MEM_TYPE_RESERVED);
 }
 
 int __must_check kvm__register_iotrap(struct kvm *kvm, u64 phys_addr, u64 len,
diff --git a/kvm.c b/kvm.c
index 95ccfcb..947721b 100644
--- a/kvm.c
+++ b/kvm.c
@@ -237,9 +237,9 @@ out:
 }
 
 int kvm__register_mem(struct kvm *kvm, u64 guest_phys, u64 size,
-		      void *userspace_addr, enum kvm_mem_type type)
+		      void *userspace_addr, int guest_memfd,
+		      u64 guest_memfd_offset, enum kvm_mem_type type)
 {
-	struct kvm_userspace_memory_region mem;
 	struct kvm_mem_bank *merged = NULL;
 	struct kvm_mem_bank *bank;
 	struct list_head *prev_entry;
@@ -320,15 +320,31 @@ int kvm__register_mem(struct kvm *kvm, u64 guest_phys, u64 size,
 		flags |= KVM_MEM_READONLY;
 
 	if (type != KVM_MEM_TYPE_RESERVED) {
-		mem = (struct kvm_userspace_memory_region) {
-			.slot			= slot,
-			.flags			= flags,
-			.guest_phys_addr	= guest_phys,
-			.memory_size		= size,
-			.userspace_addr		= (unsigned long)userspace_addr,
-		};
+		if (guest_memfd >= 0) {
+			struct kvm_userspace_memory_region2 mem2 = {
+				.slot			= slot,
+				.flags			= flags | KVM_MEM_GUEST_MEMFD,
+				.guest_phys_addr	= guest_phys,
+				.memory_size		= size,
+				.userspace_addr		= (unsigned long)userspace_addr,
+				.guest_memfd		= guest_memfd,
+				.guest_memfd_offset	= guest_memfd_offset,
+			};
 
-		ret = ioctl(kvm->vm_fd, KVM_SET_USER_MEMORY_REGION, &mem);
+			ret = ioctl(kvm->vm_fd, KVM_SET_USER_MEMORY_REGION2,
+				    &mem2);
+		} else {
+			struct kvm_userspace_memory_region mem = {
+				.slot			= slot,
+				.flags			= flags,
+				.guest_phys_addr	= guest_phys,
+				.memory_size		= size,
+				.userspace_addr		= (unsigned long)userspace_addr,
+			};
+
+			ret = ioctl(kvm->vm_fd, KVM_SET_USER_MEMORY_REGION,
+				    &mem);
+		}
 		if (ret < 0) {
 			ret = -errno;
 			goto out;
-- 
2.39.5

[PATCH kvmtool v2 4/4] arm64: Add --guest-memfd option to back guest RAM with guest_memfd

[Fuad Tabba]

Back guest RAM with a guest_memfd created with GUEST_MEMFD_FLAG_MMAP
and GUEST_MEMFD_FLAG_INIT_SHARED, mmapped MAP_SHARED to provide the
userspace mapping. The guest_memfd file descriptor is closed at
teardown, after the mapping is removed.

Protected VMs need in-place shared/private conversion of guest_memfd,
which the kernel does not support yet, so reject the combination of
--protected and --guest-memfd.

Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
---
 arm64/include/kvm/kvm-config-arch.h |  5 +++-
 arm64/kvm.c                         | 43 +++++++++++++++++++++++------
 include/kvm/kvm.h                   |  1 +
 include/kvm/util.h                  |  1 +
 kvm.c                               |  1 +
 util/util.c                         | 37 +++++++++++++++++++++++++
 6 files changed, 79 insertions(+), 9 deletions(-)

diff --git a/arm64/include/kvm/kvm-config-arch.h b/arm64/include/kvm/kvm-config-arch.h
index c2702d5..d8a8ef7 100644
--- a/arm64/include/kvm/kvm-config-arch.h
+++ b/arm64/include/kvm/kvm-config-arch.h
@@ -20,6 +20,7 @@ struct kvm_config_arch {
 	bool		no_pvtime;
 	bool		psci;
 	bool		protected;
+	bool		guest_memfd;
 };
 
 int irqchip_parser(const struct option *opt, const char *arg, int unset);
@@ -73,6 +74,8 @@ int sve_vl_parser(const struct option *opt, const char *arg, int unset);
 	OPT_BOOLEAN('\0', "e2h0", &(cfg)->e2h0,					\
 		    "Create guest without VHE support"),			\
 	OPT_BOOLEAN('\0', "protected", &(cfg)->protected,			\
-			"Create a protected VM when pKVM is enabled"),
+			"Create a protected VM when pKVM is enabled"),		\
+	OPT_BOOLEAN('\0', "guest-memfd", &(cfg)->guest_memfd,			\
+			"Use guest_memfd to back guest RAM"),
 
 #endif /* ARM_COMMON__KVM_CONFIG_ARCH_H */
diff --git a/arm64/kvm.c b/arm64/kvm.c
index fb0b98d..bf4e34e 100644
--- a/arm64/kvm.c
+++ b/arm64/kvm.c
@@ -49,9 +49,14 @@ void kvm__init_ram(struct kvm *kvm)
 	kvm->arch.ram_alloc_size = kvm->ram_size;
 	if (!kvm->cfg.hugetlbfs_path)
 		kvm->arch.ram_alloc_size += SZ_2M;
-	kvm->arch.ram_alloc_start = mmap_anon_or_hugetlbfs(kvm,
-						kvm->cfg.hugetlbfs_path,
-						kvm->arch.ram_alloc_size);
+
+	if (kvm->cfg.arch.guest_memfd)
+		kvm->arch.ram_alloc_start =
+			mmap_guest_memfd(kvm, kvm->arch.ram_alloc_size);
+	else
+		kvm->arch.ram_alloc_start =
+			mmap_anon_or_hugetlbfs(kvm, kvm->cfg.hugetlbfs_path,
+					       kvm->arch.ram_alloc_size);
 
 	if (kvm->arch.ram_alloc_start == MAP_FAILED)
 		die("Failed to map %lld bytes for guest memory (%d)",
@@ -60,17 +65,25 @@ void kvm__init_ram(struct kvm *kvm)
 	kvm->ram_start = (void *)ALIGN((unsigned long)kvm->arch.ram_alloc_start,
 					SZ_2M);
 
-	madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
-		MADV_MERGEABLE);
+	if (!kvm->cfg.arch.guest_memfd) {
+		madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
+			MADV_MERGEABLE);
 
-	madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
-		MADV_HUGEPAGE);
+		madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
+			MADV_HUGEPAGE);
+	}
 
 	phys_start	= kvm->cfg.ram_addr;
 	phys_size	= kvm->ram_size;
 	host_mem	= kvm->ram_start;
 
-	err = kvm__register_ram(kvm, phys_start, phys_size, host_mem);
+	if (kvm->cfg.arch.guest_memfd)
+		err = kvm__register_ram_guest_memfd(kvm, phys_start,
+						    phys_size, host_mem,
+						    kvm->ram_guest_memfd,
+						    host_mem - kvm->arch.ram_alloc_start);
+	else
+		err = kvm__register_ram(kvm, phys_start, phys_size, host_mem);
 	if (err)
 		die("Failed to register %lld bytes of memory at physical "
 		    "address 0x%llx [err %d]", phys_size, phys_start, err);
@@ -84,6 +97,11 @@ void kvm__init_ram(struct kvm *kvm)
 void kvm__arch_delete_ram(struct kvm *kvm)
 {
 	munmap(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size);
+
+	if (kvm->ram_guest_memfd >= 0) {
+		close(kvm->ram_guest_memfd);
+		kvm->ram_guest_memfd = -1;
+	}
 }
 
 void kvm__arch_read_term(struct kvm *kvm)
@@ -468,6 +486,15 @@ void kvm__arch_validate_cfg(struct kvm *kvm)
 	if (kvm->cfg.arch.e2h0 && !kvm->cfg.arch.nested_virt)
 		pr_warning("--e2h0 requires --nested, ignoring");
 
+	if (kvm->cfg.arch.guest_memfd) {
+		/* Requires in-place shared/private conversion of guest_memfd. */
+		if (kvm->cfg.arch.protected)
+			die("Protected VMs do not yet support guest_memfd");
+
+		if (kvm->cfg.hugetlbfs_path)
+			die("guest_memfd cannot be backed by hugetlbfs");
+	}
+
 	if (kvm->cfg.arch.protected) {
 		if (kvm->cfg.ram_size &&
 		    kvm->cfg.ram_size < DMA_MEM_REGION_SIZE) {
diff --git a/include/kvm/kvm.h b/include/kvm/kvm.h
index 5e7b5eb..95dc52f 100644
--- a/include/kvm/kvm.h
+++ b/include/kvm/kvm.h
@@ -90,6 +90,7 @@ struct kvm {
 	u64			ram_size;	/* Guest memory size, in bytes */
 	void			*ram_start;
 	u64			ram_pagesize;
+	int			ram_guest_memfd; /* guest_memfd backing RAM, -1 if none */
 	struct mutex		mem_banks_lock;
 	struct list_head	mem_banks;
 
diff --git a/include/kvm/util.h b/include/kvm/util.h
index a6dba11..9e23431 100644
--- a/include/kvm/util.h
+++ b/include/kvm/util.h
@@ -148,5 +148,6 @@ static inline int pow2_size(unsigned long x)
 struct kvm;
 void *mmap_hugetlbfs(struct kvm *kvm, const char *htlbfs_path, u64 size);
 void *mmap_anon_or_hugetlbfs(struct kvm *kvm, const char *hugetlbfs_path, u64 size);
+void *mmap_guest_memfd(struct kvm *kvm, u64 size);
 
 #endif /* KVM__UTIL_H */
diff --git a/kvm.c b/kvm.c
index 947721b..96583f9 100644
--- a/kvm.c
+++ b/kvm.c
@@ -162,6 +162,7 @@ struct kvm *kvm__new(void)
 	mutex_init(&kvm->mem_banks_lock);
 	kvm->sys_fd = -1;
 	kvm->vm_fd = -1;
+	kvm->ram_guest_memfd = -1;
 
 #ifdef KVM_BRLOCK_DEBUG
 	kvm->brlock_sem = (pthread_rwlock_t) PTHREAD_RWLOCK_INITIALIZER;
diff --git a/util/util.c b/util/util.c
index 05d49c4..51a4591 100644
--- a/util/util.c
+++ b/util/util.c
@@ -6,7 +6,9 @@
 #include "kvm/util.h"
 
 #include <kvm/kvm.h>
+#include <linux/kvm.h>
 #include <linux/magic.h>	/* For HUGETLBFS_MAGIC */
+#include <sys/ioctl.h>
 #include <sys/mman.h>
 #include <sys/stat.h>
 #include <sys/statfs.h>
@@ -169,3 +171,38 @@ void *mmap_anon_or_hugetlbfs(struct kvm *kvm, const char *hugetlbfs_path, u64 si
 		return mmap(NULL, size, PROT_RW, MAP_ANON_NORESERVE, -1, 0);
 	}
 }
+
+/*
+ * Map guest RAM backed by a guest_memfd. The file descriptor is kept open
+ * until teardown (kvm__arch_delete_ram), after the mapping is removed.
+ */
+void *mmap_guest_memfd(struct kvm *kvm, u64 size)
+{
+	u64 flags = GUEST_MEMFD_FLAG_MMAP | GUEST_MEMFD_FLAG_INIT_SHARED;
+	struct kvm_create_guest_memfd gmem = {
+		.size	= size,
+		.flags	= flags,
+	};
+	void *addr;
+	int ret;
+
+	if (!kvm__supports_extension(kvm, KVM_CAP_GUEST_MEMFD))
+		die("kernel does not support guest_memfd");
+
+	ret = ioctl(kvm->vm_fd, KVM_CHECK_EXTENSION, KVM_CAP_GUEST_MEMFD_FLAGS);
+	if (ret < 0 || ((u64)ret & flags) != flags)
+		die("VM type does not support guest_memfd with mmap");
+
+	ret = ioctl(kvm->vm_fd, KVM_CREATE_GUEST_MEMFD, &gmem);
+	if (ret < 0)
+		die_perror("KVM_CREATE_GUEST_MEMFD");
+
+	addr = mmap(NULL, size, PROT_RW, MAP_SHARED, ret, 0);
+	if (addr == MAP_FAILED)
+		die_perror("guest_memfd mmap");
+
+	kvm->ram_guest_memfd = ret;
+	kvm->ram_pagesize = getpagesize();
+
+	return addr;
+}
-- 
2.39.5

Re: [PATCH kvmtool v2 3/4] Add support for registering guest_memfd-backed memory regions

[Alexandru Elisei]

Hi Fuad,

On Wed, Jul 08, 2026 at 07:25:09PM +0100, Fuad Tabba wrote:
> Extend kvm__register_mem() with a guest_memfd file descriptor and
> offset. A bank with a guest_memfd is registered with
> KVM_SET_USER_MEMORY_REGION2 and KVM_MEM_GUEST_MEMFD, binding the
> memslot to the file. This prepares for backing guest RAM with
> guest_memfd on arm64.
> 
> Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
> ---
>  hw/cfi_flash.c    |  2 +-
>  include/kvm/kvm.h | 15 +++++++++++++--
>  kvm.c             | 36 ++++++++++++++++++++++++++----------
>  3 files changed, 40 insertions(+), 13 deletions(-)
> 
> diff --git a/hw/cfi_flash.c b/hw/cfi_flash.c
> index b68d9e0..4413955 100644
> --- a/hw/cfi_flash.c
> +++ b/hw/cfi_flash.c
> @@ -451,7 +451,7 @@ static int map_flash_memory(struct kvm *kvm, struct cfi_flash_device *sfdev)
>  	int ret;
>  
>  	ret = kvm__register_mem(kvm, sfdev->base_addr, sfdev->size,
> -				sfdev->flash_memory,
> +				sfdev->flash_memory, -1, 0,
>  				KVM_MEM_TYPE_RAM | KVM_MEM_TYPE_READONLY);
>  	if (!ret)
>  		sfdev->is_mapped = true;
> diff --git a/include/kvm/kvm.h b/include/kvm/kvm.h
> index eb23e2f..5e7b5eb 100644
> --- a/include/kvm/kvm.h
> +++ b/include/kvm/kvm.h
> @@ -128,11 +128,22 @@ bool kvm__emulate_io(struct kvm_cpu *vcpu, u16 port, void *data, int direction,
>  bool kvm__emulate_mmio(struct kvm_cpu *vcpu, u64 phys_addr, u8 *data, u32 len, u8 is_write);
>  int kvm__destroy_mem(struct kvm *kvm, u64 guest_phys, u64 size, void *userspace_addr);
>  int kvm__register_mem(struct kvm *kvm, u64 guest_phys, u64 size, void *userspace_addr,
> +		      int guest_memfd, u64 guest_memfd_offset,
>  		      enum kvm_mem_type type);
>  static inline int kvm__register_ram(struct kvm *kvm, u64 guest_phys, u64 size,
>  				    void *userspace_addr)
>  {
>  	return kvm__register_mem(kvm, guest_phys, size, userspace_addr,
> +				 -1, 0, KVM_MEM_TYPE_RAM);
> +}
> +
> +static inline int kvm__register_ram_guest_memfd(struct kvm *kvm, u64 guest_phys,
> +						u64 size, void *userspace_addr,
> +						int guest_memfd,
> +						u64 guest_memfd_offset)
> +{
> +	return kvm__register_mem(kvm, guest_phys, size, userspace_addr,
> +				 guest_memfd, guest_memfd_offset,
>  				 KVM_MEM_TYPE_RAM);

One small nitpick here, KVM_CREATE_GUEST_MEMFD returns -1 on error. If
callers forget to check the result of KVM_CREATE_GUEST_MEMFD and pass the
result to kvm__register_ram_guest_memfd(), the error won't be detected and
kvmtool will end up creating a non-guest_memfd backed memslot. Definitely
unexpected. It might be useful to someone, someday, to add a check here for
guest_memfd >= 0, but it's not terribly important.

Regardless:

Reviewed-by: Alexandru Elisei <alexandru.elisei@arm.com>

Thanks,
Alex

>  }
>  
> @@ -140,13 +151,13 @@ static inline int kvm__register_dev_mem(struct kvm *kvm, u64 guest_phys,
>  					u64 size, void *userspace_addr)
>  {
>  	return kvm__register_mem(kvm, guest_phys, size, userspace_addr,
> -				 KVM_MEM_TYPE_DEVICE);
> +				 -1, 0, KVM_MEM_TYPE_DEVICE);
>  }
>  
>  static inline int kvm__reserve_mem(struct kvm *kvm, u64 guest_phys, u64 size)
>  {
>  	return kvm__register_mem(kvm, guest_phys, size, NULL,
> -				 KVM_MEM_TYPE_RESERVED);
> +				 -1, 0, KVM_MEM_TYPE_RESERVED);
>  }
>  
>  int __must_check kvm__register_iotrap(struct kvm *kvm, u64 phys_addr, u64 len,
> diff --git a/kvm.c b/kvm.c
> index 95ccfcb..947721b 100644
> --- a/kvm.c
> +++ b/kvm.c
> @@ -237,9 +237,9 @@ out:
>  }
>  
>  int kvm__register_mem(struct kvm *kvm, u64 guest_phys, u64 size,
> -		      void *userspace_addr, enum kvm_mem_type type)
> +		      void *userspace_addr, int guest_memfd,
> +		      u64 guest_memfd_offset, enum kvm_mem_type type)
>  {
> -	struct kvm_userspace_memory_region mem;
>  	struct kvm_mem_bank *merged = NULL;
>  	struct kvm_mem_bank *bank;
>  	struct list_head *prev_entry;
> @@ -320,15 +320,31 @@ int kvm__register_mem(struct kvm *kvm, u64 guest_phys, u64 size,
>  		flags |= KVM_MEM_READONLY;
>  
>  	if (type != KVM_MEM_TYPE_RESERVED) {
> -		mem = (struct kvm_userspace_memory_region) {
> -			.slot			= slot,
> -			.flags			= flags,
> -			.guest_phys_addr	= guest_phys,
> -			.memory_size		= size,
> -			.userspace_addr		= (unsigned long)userspace_addr,
> -		};
> +		if (guest_memfd >= 0) {
> +			struct kvm_userspace_memory_region2 mem2 = {
> +				.slot			= slot,
> +				.flags			= flags | KVM_MEM_GUEST_MEMFD,
> +				.guest_phys_addr	= guest_phys,
> +				.memory_size		= size,
> +				.userspace_addr		= (unsigned long)userspace_addr,
> +				.guest_memfd		= guest_memfd,
> +				.guest_memfd_offset	= guest_memfd_offset,
> +			};
>  
> -		ret = ioctl(kvm->vm_fd, KVM_SET_USER_MEMORY_REGION, &mem);
> +			ret = ioctl(kvm->vm_fd, KVM_SET_USER_MEMORY_REGION2,
> +				    &mem2);
> +		} else {
> +			struct kvm_userspace_memory_region mem = {
> +				.slot			= slot,
> +				.flags			= flags,
> +				.guest_phys_addr	= guest_phys,
> +				.memory_size		= size,
> +				.userspace_addr		= (unsigned long)userspace_addr,
> +			};
> +
> +			ret = ioctl(kvm->vm_fd, KVM_SET_USER_MEMORY_REGION,
> +				    &mem);
> +		}
>  		if (ret < 0) {
>  			ret = -errno;
>  			goto out;
> -- 
> 2.39.5
> 

Re: [PATCH kvmtool v2 4/4] arm64: Add --guest-memfd option to back guest RAM with guest_memfd

[Alexandru Elisei]

Hi Fuad,

On Wed, Jul 08, 2026 at 07:25:10PM +0100, Fuad Tabba wrote:
> Back guest RAM with a guest_memfd created with GUEST_MEMFD_FLAG_MMAP
> and GUEST_MEMFD_FLAG_INIT_SHARED, mmapped MAP_SHARED to provide the
> userspace mapping. The guest_memfd file descriptor is closed at
> teardown, after the mapping is removed.
> 
> Protected VMs need in-place shared/private conversion of guest_memfd,
> which the kernel does not support yet, so reject the combination of
> --protected and --guest-memfd.
> 
> Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
> ---
>  arm64/include/kvm/kvm-config-arch.h |  5 +++-
>  arm64/kvm.c                         | 43 +++++++++++++++++++++++------
>  include/kvm/kvm.h                   |  1 +
>  include/kvm/util.h                  |  1 +
>  kvm.c                               |  1 +
>  util/util.c                         | 37 +++++++++++++++++++++++++
>  6 files changed, 79 insertions(+), 9 deletions(-)
> 
> diff --git a/arm64/include/kvm/kvm-config-arch.h b/arm64/include/kvm/kvm-config-arch.h
> index c2702d5..d8a8ef7 100644
> --- a/arm64/include/kvm/kvm-config-arch.h
> +++ b/arm64/include/kvm/kvm-config-arch.h
> @@ -20,6 +20,7 @@ struct kvm_config_arch {
>  	bool		no_pvtime;
>  	bool		psci;
>  	bool		protected;
> +	bool		guest_memfd;
>  };
>  
>  int irqchip_parser(const struct option *opt, const char *arg, int unset);
> @@ -73,6 +74,8 @@ int sve_vl_parser(const struct option *opt, const char *arg, int unset);
>  	OPT_BOOLEAN('\0', "e2h0", &(cfg)->e2h0,					\
>  		    "Create guest without VHE support"),			\
>  	OPT_BOOLEAN('\0', "protected", &(cfg)->protected,			\
> -			"Create a protected VM when pKVM is enabled"),
> +			"Create a protected VM when pKVM is enabled"),		\
> +	OPT_BOOLEAN('\0', "guest-memfd", &(cfg)->guest_memfd,			\
> +			"Use guest_memfd to back guest RAM"),
>  
>  #endif /* ARM_COMMON__KVM_CONFIG_ARCH_H */
> diff --git a/arm64/kvm.c b/arm64/kvm.c
> index fb0b98d..bf4e34e 100644
> --- a/arm64/kvm.c
> +++ b/arm64/kvm.c
> @@ -49,9 +49,14 @@ void kvm__init_ram(struct kvm *kvm)
>  	kvm->arch.ram_alloc_size = kvm->ram_size;
>  	if (!kvm->cfg.hugetlbfs_path)
>  		kvm->arch.ram_alloc_size += SZ_2M;
> -	kvm->arch.ram_alloc_start = mmap_anon_or_hugetlbfs(kvm,
> -						kvm->cfg.hugetlbfs_path,
> -						kvm->arch.ram_alloc_size);
> +
> +	if (kvm->cfg.arch.guest_memfd)
> +		kvm->arch.ram_alloc_start =
> +			mmap_guest_memfd(kvm, kvm->arch.ram_alloc_size);
> +	else
> +		kvm->arch.ram_alloc_start =
> +			mmap_anon_or_hugetlbfs(kvm, kvm->cfg.hugetlbfs_path,
> +					       kvm->arch.ram_alloc_size);

Would you mind not splitting the line after the equal sign? It makes
parsing the code really hard (at least for me).

>  
>  	if (kvm->arch.ram_alloc_start == MAP_FAILED)
>  		die("Failed to map %lld bytes for guest memory (%d)",
> @@ -60,17 +65,25 @@ void kvm__init_ram(struct kvm *kvm)
>  	kvm->ram_start = (void *)ALIGN((unsigned long)kvm->arch.ram_alloc_start,
>  					SZ_2M);
>  
> -	madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
> -		MADV_MERGEABLE);
> +	if (!kvm->cfg.arch.guest_memfd) {
> +		madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
> +			MADV_MERGEABLE);
>  
> -	madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
> -		MADV_HUGEPAGE);
> +		madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
> +			MADV_HUGEPAGE);
> +	}
>  
>  	phys_start	= kvm->cfg.ram_addr;
>  	phys_size	= kvm->ram_size;
>  	host_mem	= kvm->ram_start;
>  
> -	err = kvm__register_ram(kvm, phys_start, phys_size, host_mem);
> +	if (kvm->cfg.arch.guest_memfd)
> +		err = kvm__register_ram_guest_memfd(kvm, phys_start,
> +						    phys_size, host_mem,
> +						    kvm->ram_guest_memfd,
> +						    host_mem - kvm->arch.ram_alloc_start);
> +	else
> +		err = kvm__register_ram(kvm, phys_start, phys_size, host_mem);
>  	if (err)
>  		die("Failed to register %lld bytes of memory at physical "
>  		    "address 0x%llx [err %d]", phys_size, phys_start, err);
> @@ -84,6 +97,11 @@ void kvm__init_ram(struct kvm *kvm)
>  void kvm__arch_delete_ram(struct kvm *kvm)
>  {
>  	munmap(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size);
> +
> +	if (kvm->ram_guest_memfd >= 0) {
> +		close(kvm->ram_guest_memfd);
> +		kvm->ram_guest_memfd = -1;
> +	}
>  }
>  
>  void kvm__arch_read_term(struct kvm *kvm)
> @@ -468,6 +486,15 @@ void kvm__arch_validate_cfg(struct kvm *kvm)
>  	if (kvm->cfg.arch.e2h0 && !kvm->cfg.arch.nested_virt)
>  		pr_warning("--e2h0 requires --nested, ignoring");
>  
> +	if (kvm->cfg.arch.guest_memfd) {
> +		/* Requires in-place shared/private conversion of guest_memfd. */
> +		if (kvm->cfg.arch.protected)
> +			die("Protected VMs do not yet support guest_memfd");
> +
> +		if (kvm->cfg.hugetlbfs_path)
> +			die("guest_memfd cannot be backed by hugetlbfs");
> +	}

I think MTE is incompatible with guest_memfd, KVM certainly returns an
error from kvm_arch_prepare_memory_region() when both are set. Since MTE is
enabled by default, it might be useful to let the user know that MTE has
been disabled because of guest_memfd.

Otherwise, everything looks good.

Thanks,
Alex

> +
>  	if (kvm->cfg.arch.protected) {
>  		if (kvm->cfg.ram_size &&
>  		    kvm->cfg.ram_size < DMA_MEM_REGION_SIZE) {
> diff --git a/include/kvm/kvm.h b/include/kvm/kvm.h
> index 5e7b5eb..95dc52f 100644
> --- a/include/kvm/kvm.h
> +++ b/include/kvm/kvm.h
> @@ -90,6 +90,7 @@ struct kvm {
>  	u64			ram_size;	/* Guest memory size, in bytes */
>  	void			*ram_start;
>  	u64			ram_pagesize;
> +	int			ram_guest_memfd; /* guest_memfd backing RAM, -1 if none */
>  	struct mutex		mem_banks_lock;
>  	struct list_head	mem_banks;
>  
> diff --git a/include/kvm/util.h b/include/kvm/util.h
> index a6dba11..9e23431 100644
> --- a/include/kvm/util.h
> +++ b/include/kvm/util.h
> @@ -148,5 +148,6 @@ static inline int pow2_size(unsigned long x)
>  struct kvm;
>  void *mmap_hugetlbfs(struct kvm *kvm, const char *htlbfs_path, u64 size);
>  void *mmap_anon_or_hugetlbfs(struct kvm *kvm, const char *hugetlbfs_path, u64 size);
> +void *mmap_guest_memfd(struct kvm *kvm, u64 size);
>  
>  #endif /* KVM__UTIL_H */
> diff --git a/kvm.c b/kvm.c
> index 947721b..96583f9 100644
> --- a/kvm.c
> +++ b/kvm.c
> @@ -162,6 +162,7 @@ struct kvm *kvm__new(void)
>  	mutex_init(&kvm->mem_banks_lock);
>  	kvm->sys_fd = -1;
>  	kvm->vm_fd = -1;
> +	kvm->ram_guest_memfd = -1;
>  
>  #ifdef KVM_BRLOCK_DEBUG
>  	kvm->brlock_sem = (pthread_rwlock_t) PTHREAD_RWLOCK_INITIALIZER;
> diff --git a/util/util.c b/util/util.c
> index 05d49c4..51a4591 100644
> --- a/util/util.c
> +++ b/util/util.c
> @@ -6,7 +6,9 @@
>  #include "kvm/util.h"
>  
>  #include <kvm/kvm.h>
> +#include <linux/kvm.h>
>  #include <linux/magic.h>	/* For HUGETLBFS_MAGIC */
> +#include <sys/ioctl.h>
>  #include <sys/mman.h>
>  #include <sys/stat.h>
>  #include <sys/statfs.h>
> @@ -169,3 +171,38 @@ void *mmap_anon_or_hugetlbfs(struct kvm *kvm, const char *hugetlbfs_path, u64 si
>  		return mmap(NULL, size, PROT_RW, MAP_ANON_NORESERVE, -1, 0);
>  	}
>  }
> +
> +/*
> + * Map guest RAM backed by a guest_memfd. The file descriptor is kept open
> + * until teardown (kvm__arch_delete_ram), after the mapping is removed.
> + */
> +void *mmap_guest_memfd(struct kvm *kvm, u64 size)
> +{
> +	u64 flags = GUEST_MEMFD_FLAG_MMAP | GUEST_MEMFD_FLAG_INIT_SHARED;
> +	struct kvm_create_guest_memfd gmem = {
> +		.size	= size,
> +		.flags	= flags,
> +	};
> +	void *addr;
> +	int ret;
> +
> +	if (!kvm__supports_extension(kvm, KVM_CAP_GUEST_MEMFD))
> +		die("kernel does not support guest_memfd");
> +
> +	ret = ioctl(kvm->vm_fd, KVM_CHECK_EXTENSION, KVM_CAP_GUEST_MEMFD_FLAGS);
> +	if (ret < 0 || ((u64)ret & flags) != flags)
> +		die("VM type does not support guest_memfd with mmap");
> +
> +	ret = ioctl(kvm->vm_fd, KVM_CREATE_GUEST_MEMFD, &gmem);
> +	if (ret < 0)
> +		die_perror("KVM_CREATE_GUEST_MEMFD");
> +
> +	addr = mmap(NULL, size, PROT_RW, MAP_SHARED, ret, 0);
> +	if (addr == MAP_FAILED)
> +		die_perror("guest_memfd mmap");
> +
> +	kvm->ram_guest_memfd = ret;
> +	kvm->ram_pagesize = getpagesize();
> +
> +	return addr;
> +}
> -- 
> 2.39.5
> 

Re: [PATCH kvmtool v2 3/4] Add support for registering guest_memfd-backed memory regions

[Fuad Tabba]

On Thu, 9 Jul 2026 at 16:07, Alexandru Elisei <alexandru.elisei@arm.com> wrote:
>
> Hi Fuad,
>
> On Wed, Jul 08, 2026 at 07:25:09PM +0100, Fuad Tabba wrote:
> > Extend kvm__register_mem() with a guest_memfd file descriptor and
> > offset. A bank with a guest_memfd is registered with
> > KVM_SET_USER_MEMORY_REGION2 and KVM_MEM_GUEST_MEMFD, binding the
> > memslot to the file. This prepares for backing guest RAM with
> > guest_memfd on arm64.
> >
> > Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
> > ---
> >  hw/cfi_flash.c    |  2 +-
> >  include/kvm/kvm.h | 15 +++++++++++++--
> >  kvm.c             | 36 ++++++++++++++++++++++++++----------
> >  3 files changed, 40 insertions(+), 13 deletions(-)
> >
> > diff --git a/hw/cfi_flash.c b/hw/cfi_flash.c
> > index b68d9e0..4413955 100644
> > --- a/hw/cfi_flash.c
> > +++ b/hw/cfi_flash.c
> > @@ -451,7 +451,7 @@ static int map_flash_memory(struct kvm *kvm, struct cfi_flash_device *sfdev)
> >       int ret;
> >
> >       ret = kvm__register_mem(kvm, sfdev->base_addr, sfdev->size,
> > -                             sfdev->flash_memory,
> > +                             sfdev->flash_memory, -1, 0,
> >                               KVM_MEM_TYPE_RAM | KVM_MEM_TYPE_READONLY);
> >       if (!ret)
> >               sfdev->is_mapped = true;
> > diff --git a/include/kvm/kvm.h b/include/kvm/kvm.h
> > index eb23e2f..5e7b5eb 100644
> > --- a/include/kvm/kvm.h
> > +++ b/include/kvm/kvm.h
> > @@ -128,11 +128,22 @@ bool kvm__emulate_io(struct kvm_cpu *vcpu, u16 port, void *data, int direction,
> >  bool kvm__emulate_mmio(struct kvm_cpu *vcpu, u64 phys_addr, u8 *data, u32 len, u8 is_write);
> >  int kvm__destroy_mem(struct kvm *kvm, u64 guest_phys, u64 size, void *userspace_addr);
> >  int kvm__register_mem(struct kvm *kvm, u64 guest_phys, u64 size, void *userspace_addr,
> > +                   int guest_memfd, u64 guest_memfd_offset,
> >                     enum kvm_mem_type type);
> >  static inline int kvm__register_ram(struct kvm *kvm, u64 guest_phys, u64 size,
> >                                   void *userspace_addr)
> >  {
> >       return kvm__register_mem(kvm, guest_phys, size, userspace_addr,
> > +                              -1, 0, KVM_MEM_TYPE_RAM);
> > +}
> > +
> > +static inline int kvm__register_ram_guest_memfd(struct kvm *kvm, u64 guest_phys,
> > +                                             u64 size, void *userspace_addr,
> > +                                             int guest_memfd,
> > +                                             u64 guest_memfd_offset)
> > +{
> > +     return kvm__register_mem(kvm, guest_phys, size, userspace_addr,
> > +                              guest_memfd, guest_memfd_offset,
> >                                KVM_MEM_TYPE_RAM);
>
> One small nitpick here, KVM_CREATE_GUEST_MEMFD returns -1 on error. If
> callers forget to check the result of KVM_CREATE_GUEST_MEMFD and pass the
> result to kvm__register_ram_guest_memfd(), the error won't be detected and
> kvmtool will end up creating a non-guest_memfd backed memslot. Definitely
> unexpected. It might be useful to someone, someday, to add a check here for
> guest_memfd >= 0, but it's not terribly important.

Good point. The only caller today (mmap_guest_memfd) already dies on
error, but a defensive check in kvm__register_mem makes the contract
explicit. I'll add one in v3.

> Regardless:
>
> Reviewed-by: Alexandru Elisei <alexandru.elisei@arm.com>

Thanks!
/fuad

>
> Thanks,
> Alex
>
> >  }
> >
> > @@ -140,13 +151,13 @@ static inline int kvm__register_dev_mem(struct kvm *kvm, u64 guest_phys,
> >                                       u64 size, void *userspace_addr)
> >  {
> >       return kvm__register_mem(kvm, guest_phys, size, userspace_addr,
> > -                              KVM_MEM_TYPE_DEVICE);
> > +                              -1, 0, KVM_MEM_TYPE_DEVICE);
> >  }
> >
> >  static inline int kvm__reserve_mem(struct kvm *kvm, u64 guest_phys, u64 size)
> >  {
> >       return kvm__register_mem(kvm, guest_phys, size, NULL,
> > -                              KVM_MEM_TYPE_RESERVED);
> > +                              -1, 0, KVM_MEM_TYPE_RESERVED);
> >  }
> >
> >  int __must_check kvm__register_iotrap(struct kvm *kvm, u64 phys_addr, u64 len,
> > diff --git a/kvm.c b/kvm.c
> > index 95ccfcb..947721b 100644
> > --- a/kvm.c
> > +++ b/kvm.c
> > @@ -237,9 +237,9 @@ out:
> >  }
> >
> >  int kvm__register_mem(struct kvm *kvm, u64 guest_phys, u64 size,
> > -                   void *userspace_addr, enum kvm_mem_type type)
> > +                   void *userspace_addr, int guest_memfd,
> > +                   u64 guest_memfd_offset, enum kvm_mem_type type)
> >  {
> > -     struct kvm_userspace_memory_region mem;
> >       struct kvm_mem_bank *merged = NULL;
> >       struct kvm_mem_bank *bank;
> >       struct list_head *prev_entry;
> > @@ -320,15 +320,31 @@ int kvm__register_mem(struct kvm *kvm, u64 guest_phys, u64 size,
> >               flags |= KVM_MEM_READONLY;
> >
> >       if (type != KVM_MEM_TYPE_RESERVED) {
> > -             mem = (struct kvm_userspace_memory_region) {
> > -                     .slot                   = slot,
> > -                     .flags                  = flags,
> > -                     .guest_phys_addr        = guest_phys,
> > -                     .memory_size            = size,
> > -                     .userspace_addr         = (unsigned long)userspace_addr,
> > -             };
> > +             if (guest_memfd >= 0) {
> > +                     struct kvm_userspace_memory_region2 mem2 = {
> > +                             .slot                   = slot,
> > +                             .flags                  = flags | KVM_MEM_GUEST_MEMFD,
> > +                             .guest_phys_addr        = guest_phys,
> > +                             .memory_size            = size,
> > +                             .userspace_addr         = (unsigned long)userspace_addr,
> > +                             .guest_memfd            = guest_memfd,
> > +                             .guest_memfd_offset     = guest_memfd_offset,
> > +                     };
> >
> > -             ret = ioctl(kvm->vm_fd, KVM_SET_USER_MEMORY_REGION, &mem);
> > +                     ret = ioctl(kvm->vm_fd, KVM_SET_USER_MEMORY_REGION2,
> > +                                 &mem2);
> > +             } else {
> > +                     struct kvm_userspace_memory_region mem = {
> > +                             .slot                   = slot,
> > +                             .flags                  = flags,
> > +                             .guest_phys_addr        = guest_phys,
> > +                             .memory_size            = size,
> > +                             .userspace_addr         = (unsigned long)userspace_addr,
> > +                     };
> > +
> > +                     ret = ioctl(kvm->vm_fd, KVM_SET_USER_MEMORY_REGION,
> > +                                 &mem);
> > +             }
> >               if (ret < 0) {
> >                       ret = -errno;
> >                       goto out;
> > --
> > 2.39.5
> >

Re: [PATCH kvmtool v2 4/4] arm64: Add --guest-memfd option to back guest RAM with guest_memfd

[Fuad Tabba]

On Thu, 9 Jul 2026 at 16:08, Alexandru Elisei <alexandru.elisei@arm.com> wrote:
>
> Hi Fuad,
>
> On Wed, Jul 08, 2026 at 07:25:10PM +0100, Fuad Tabba wrote:
> > Back guest RAM with a guest_memfd created with GUEST_MEMFD_FLAG_MMAP
> > and GUEST_MEMFD_FLAG_INIT_SHARED, mmapped MAP_SHARED to provide the
> > userspace mapping. The guest_memfd file descriptor is closed at
> > teardown, after the mapping is removed.
> >
> > Protected VMs need in-place shared/private conversion of guest_memfd,
> > which the kernel does not support yet, so reject the combination of
> > --protected and --guest-memfd.
> >
> > Signed-off-by: Fuad Tabba <fuad.tabba@linux.dev>
> > ---
> >  arm64/include/kvm/kvm-config-arch.h |  5 +++-
> >  arm64/kvm.c                         | 43 +++++++++++++++++++++++------
> >  include/kvm/kvm.h                   |  1 +
> >  include/kvm/util.h                  |  1 +
> >  kvm.c                               |  1 +
> >  util/util.c                         | 37 +++++++++++++++++++++++++
> >  6 files changed, 79 insertions(+), 9 deletions(-)
> >
> > diff --git a/arm64/include/kvm/kvm-config-arch.h b/arm64/include/kvm/kvm-config-arch.h
> > index c2702d5..d8a8ef7 100644
> > --- a/arm64/include/kvm/kvm-config-arch.h
> > +++ b/arm64/include/kvm/kvm-config-arch.h
> > @@ -20,6 +20,7 @@ struct kvm_config_arch {
> >       bool            no_pvtime;
> >       bool            psci;
> >       bool            protected;
> > +     bool            guest_memfd;
> >  };
> >
> >  int irqchip_parser(const struct option *opt, const char *arg, int unset);
> > @@ -73,6 +74,8 @@ int sve_vl_parser(const struct option *opt, const char *arg, int unset);
> >       OPT_BOOLEAN('\0', "e2h0", &(cfg)->e2h0,                                 \
> >                   "Create guest without VHE support"),                        \
> >       OPT_BOOLEAN('\0', "protected", &(cfg)->protected,                       \
> > -                     "Create a protected VM when pKVM is enabled"),
> > +                     "Create a protected VM when pKVM is enabled"),          \
> > +     OPT_BOOLEAN('\0', "guest-memfd", &(cfg)->guest_memfd,                   \
> > +                     "Use guest_memfd to back guest RAM"),
> >
> >  #endif /* ARM_COMMON__KVM_CONFIG_ARCH_H */
> > diff --git a/arm64/kvm.c b/arm64/kvm.c
> > index fb0b98d..bf4e34e 100644
> > --- a/arm64/kvm.c
> > +++ b/arm64/kvm.c
> > @@ -49,9 +49,14 @@ void kvm__init_ram(struct kvm *kvm)
> >       kvm->arch.ram_alloc_size = kvm->ram_size;
> >       if (!kvm->cfg.hugetlbfs_path)
> >               kvm->arch.ram_alloc_size += SZ_2M;
> > -     kvm->arch.ram_alloc_start = mmap_anon_or_hugetlbfs(kvm,
> > -                                             kvm->cfg.hugetlbfs_path,
> > -                                             kvm->arch.ram_alloc_size);
> > +
> > +     if (kvm->cfg.arch.guest_memfd)
> > +             kvm->arch.ram_alloc_start =
> > +                     mmap_guest_memfd(kvm, kvm->arch.ram_alloc_size);
> > +     else
> > +             kvm->arch.ram_alloc_start =
> > +                     mmap_anon_or_hugetlbfs(kvm, kvm->cfg.hugetlbfs_path,
> > +                                            kvm->arch.ram_alloc_size);
>
> Would you mind not splitting the line after the equal sign? It makes
> parsing the code really hard (at least for me).

Done.


>
> >
> >       if (kvm->arch.ram_alloc_start == MAP_FAILED)
> >               die("Failed to map %lld bytes for guest memory (%d)",
> > @@ -60,17 +65,25 @@ void kvm__init_ram(struct kvm *kvm)
> >       kvm->ram_start = (void *)ALIGN((unsigned long)kvm->arch.ram_alloc_start,
> >                                       SZ_2M);
> >
> > -     madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
> > -             MADV_MERGEABLE);
> > +     if (!kvm->cfg.arch.guest_memfd) {
> > +             madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
> > +                     MADV_MERGEABLE);
> >
> > -     madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
> > -             MADV_HUGEPAGE);
> > +             madvise(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size,
> > +                     MADV_HUGEPAGE);
> > +     }
> >
> >       phys_start      = kvm->cfg.ram_addr;
> >       phys_size       = kvm->ram_size;
> >       host_mem        = kvm->ram_start;
> >
> > -     err = kvm__register_ram(kvm, phys_start, phys_size, host_mem);
> > +     if (kvm->cfg.arch.guest_memfd)
> > +             err = kvm__register_ram_guest_memfd(kvm, phys_start,
> > +                                                 phys_size, host_mem,
> > +                                                 kvm->ram_guest_memfd,
> > +                                                 host_mem - kvm->arch.ram_alloc_start);
> > +     else
> > +             err = kvm__register_ram(kvm, phys_start, phys_size, host_mem);
> >       if (err)
> >               die("Failed to register %lld bytes of memory at physical "
> >                   "address 0x%llx [err %d]", phys_size, phys_start, err);
> > @@ -84,6 +97,11 @@ void kvm__init_ram(struct kvm *kvm)
> >  void kvm__arch_delete_ram(struct kvm *kvm)
> >  {
> >       munmap(kvm->arch.ram_alloc_start, kvm->arch.ram_alloc_size);
> > +
> > +     if (kvm->ram_guest_memfd >= 0) {
> > +             close(kvm->ram_guest_memfd);
> > +             kvm->ram_guest_memfd = -1;
> > +     }
> >  }
> >
> >  void kvm__arch_read_term(struct kvm *kvm)
> > @@ -468,6 +486,15 @@ void kvm__arch_validate_cfg(struct kvm *kvm)
> >       if (kvm->cfg.arch.e2h0 && !kvm->cfg.arch.nested_virt)
> >               pr_warning("--e2h0 requires --nested, ignoring");
> >
> > +     if (kvm->cfg.arch.guest_memfd) {
> > +             /* Requires in-place shared/private conversion of guest_memfd. */
> > +             if (kvm->cfg.arch.protected)
> > +                     die("Protected VMs do not yet support guest_memfd");
> > +
> > +             if (kvm->cfg.hugetlbfs_path)
> > +                     die("guest_memfd cannot be backed by hugetlbfs");
> > +     }
>
> I think MTE is incompatible with guest_memfd, KVM certainly returns an
> error from kvm_arch_prepare_memory_region() when both are set. Since MTE is
> enabled by default, it might be useful to let the user know that MTE has
> been disabled because of guest_memfd.

You're right, VMAs don't get VM_MTE_ALLOWED, so
kvm_arch_prepare_memory_region rejects the memslot when MTE is
enabled.

I'll disable MTE when --guest-memfd is used, with a message to let the
user know.

Cheers,
/fuad

>
> Otherwise, everything looks good.
>
> Thanks,
> Alex
>
> > +
> >       if (kvm->cfg.arch.protected) {
> >               if (kvm->cfg.ram_size &&
> >                   kvm->cfg.ram_size < DMA_MEM_REGION_SIZE) {
> > diff --git a/include/kvm/kvm.h b/include/kvm/kvm.h
> > index 5e7b5eb..95dc52f 100644
> > --- a/include/kvm/kvm.h
> > +++ b/include/kvm/kvm.h
> > @@ -90,6 +90,7 @@ struct kvm {
> >       u64                     ram_size;       /* Guest memory size, in bytes */
> >       void                    *ram_start;
> >       u64                     ram_pagesize;
> > +     int                     ram_guest_memfd; /* guest_memfd backing RAM, -1 if none */
> >       struct mutex            mem_banks_lock;
> >       struct list_head        mem_banks;
> >
> > diff --git a/include/kvm/util.h b/include/kvm/util.h
> > index a6dba11..9e23431 100644
> > --- a/include/kvm/util.h
> > +++ b/include/kvm/util.h
> > @@ -148,5 +148,6 @@ static inline int pow2_size(unsigned long x)
> >  struct kvm;
> >  void *mmap_hugetlbfs(struct kvm *kvm, const char *htlbfs_path, u64 size);
> >  void *mmap_anon_or_hugetlbfs(struct kvm *kvm, const char *hugetlbfs_path, u64 size);
> > +void *mmap_guest_memfd(struct kvm *kvm, u64 size);
> >
> >  #endif /* KVM__UTIL_H */
> > diff --git a/kvm.c b/kvm.c
> > index 947721b..96583f9 100644
> > --- a/kvm.c
> > +++ b/kvm.c
> > @@ -162,6 +162,7 @@ struct kvm *kvm__new(void)
> >       mutex_init(&kvm->mem_banks_lock);
> >       kvm->sys_fd = -1;
> >       kvm->vm_fd = -1;
> > +     kvm->ram_guest_memfd = -1;
> >
> >  #ifdef KVM_BRLOCK_DEBUG
> >       kvm->brlock_sem = (pthread_rwlock_t) PTHREAD_RWLOCK_INITIALIZER;
> > diff --git a/util/util.c b/util/util.c
> > index 05d49c4..51a4591 100644
> > --- a/util/util.c
> > +++ b/util/util.c
> > @@ -6,7 +6,9 @@
> >  #include "kvm/util.h"
> >
> >  #include <kvm/kvm.h>
> > +#include <linux/kvm.h>
> >  #include <linux/magic.h>     /* For HUGETLBFS_MAGIC */
> > +#include <sys/ioctl.h>
> >  #include <sys/mman.h>
> >  #include <sys/stat.h>
> >  #include <sys/statfs.h>
> > @@ -169,3 +171,38 @@ void *mmap_anon_or_hugetlbfs(struct kvm *kvm, const char *hugetlbfs_path, u64 si
> >               return mmap(NULL, size, PROT_RW, MAP_ANON_NORESERVE, -1, 0);
> >       }
> >  }
> > +
> > +/*
> > + * Map guest RAM backed by a guest_memfd. The file descriptor is kept open
> > + * until teardown (kvm__arch_delete_ram), after the mapping is removed.
> > + */
> > +void *mmap_guest_memfd(struct kvm *kvm, u64 size)
> > +{
> > +     u64 flags = GUEST_MEMFD_FLAG_MMAP | GUEST_MEMFD_FLAG_INIT_SHARED;
> > +     struct kvm_create_guest_memfd gmem = {
> > +             .size   = size,
> > +             .flags  = flags,
> > +     };
> > +     void *addr;
> > +     int ret;
> > +
> > +     if (!kvm__supports_extension(kvm, KVM_CAP_GUEST_MEMFD))
> > +             die("kernel does not support guest_memfd");
> > +
> > +     ret = ioctl(kvm->vm_fd, KVM_CHECK_EXTENSION, KVM_CAP_GUEST_MEMFD_FLAGS);
> > +     if (ret < 0 || ((u64)ret & flags) != flags)
> > +             die("VM type does not support guest_memfd with mmap");
> > +
> > +     ret = ioctl(kvm->vm_fd, KVM_CREATE_GUEST_MEMFD, &gmem);
> > +     if (ret < 0)
> > +             die_perror("KVM_CREATE_GUEST_MEMFD");
> > +
> > +     addr = mmap(NULL, size, PROT_RW, MAP_SHARED, ret, 0);
> > +     if (addr == MAP_FAILED)
> > +             die_perror("guest_memfd mmap");
> > +
> > +     kvm->ram_guest_memfd = ret;
> > +     kvm->ram_pagesize = getpagesize();
> > +
> > +     return addr;
> > +}
> > --
> > 2.39.5
> >