Fix a variety of found-by-inspection bugs in KVMGT, and overhaul KVM's
page-track APIs to provide a leaner and cleaner interface. The motivation
for this series is to (significantly) reduce the number of KVM APIs that
KVMGT uses, with a long-term goal of making all kvm_host.h headers
KVM-internal. That said, I think the cleanup itself is worthwhile,
e.g. KVMGT really shouldn't be touching kvm->mmu_lock.
Note! The KVMGT changes are compile tested only as I don't have the
necessary hardware (AFAIK). Testing, and lots of it, on the KVMGT side
of things is needed and any help on that front would be much appreciated.
Sean Christopherson (24):
drm/i915/gvt: Verify pfn is "valid" before dereferencing "struct page"
KVM: x86/mmu: Factor out helper to get max mapping size of a memslot
drm/i915/gvt: Incorporate KVM memslot info into check for 2MiB GTT
entry
drm/i915/gvt: Verify VFIO-pinned page is THP when shadowing 2M gtt
entry
drm/i915/gvt: Put the page reference obtained by KVM's gfn_to_pfn()
drm/i915/gvt: Don't rely on KVM's gfn_to_pfn() to query possible 2M
GTT
drm/i915/gvt: Use an "unsigned long" to iterate over memslot gfns
drm/i915/gvt: Hoist acquisition of vgpu_lock out to
kvmgt_page_track_write()
drm/i915/gvt: Protect gfn hash table with dedicated mutex
KVM: x86/mmu: Don't rely on page-track mechanism to flush on memslot
change
KVM: x86/mmu: Don't bounce through page-track mechanism for guest PTEs
KVM: drm/i915/gvt: Drop @vcpu from KVM's ->track_write() hook
KVM: x86: Reject memslot MOVE operations if KVMGT is attached
drm/i915/gvt: Don't bother removing write-protection on to-be-deleted
slot
KVM: x86/mmu: Move KVM-only page-track declarations to internal header
KVM: x86/mmu: Use page-track notifiers iff there are external users
KVM: x86/mmu: Drop infrastructure for multiple page-track modes
KVM: x86/mmu: Rename page-track APIs to reflect the new reality
KVM: x86/mmu: Assert that correct locks are held for page
write-tracking
KVM: x86/mmu: Bug the VM if write-tracking is used but not enabled
KVM: x86/mmu: Drop @slot param from exported/external page-track APIs
KVM: x86/mmu: Handle KVM bookkeeping in page-track APIs, not callers
KVM: x86/mmu: Add page-track API to query if a gfn is valid
drm/i915/gvt: Drop final dependencies on KVM internal details
Yan Zhao (3):
KVM: x86: Add a new page-track hook to handle memslot deletion
drm/i915/gvt: switch from ->track_flush_slot() to
->track_remove_region()
KVM: x86: Remove the unused page-track hook track_flush_slot()
arch/x86/include/asm/kvm_host.h | 16 +-
arch/x86/include/asm/kvm_page_track.h | 67 +++---
arch/x86/kvm/mmu.h | 2 +
arch/x86/kvm/mmu/mmu.c | 61 +++---
arch/x86/kvm/mmu/mmu_internal.h | 2 +
arch/x86/kvm/mmu/page_track.c | 283 +++++++++++++++-----------
arch/x86/kvm/mmu/page_track.h | 59 ++++++
arch/x86/kvm/x86.c | 13 +-
drivers/gpu/drm/i915/gvt/gtt.c | 45 ++--
drivers/gpu/drm/i915/gvt/gvt.h | 4 +-
drivers/gpu/drm/i915/gvt/kvmgt.c | 138 ++++++-------
drivers/gpu/drm/i915/gvt/page_track.c | 10 +-
drivers/gpu/drm/i915/gvt/vgpu.c | 1 +
13 files changed, 386 insertions(+), 315 deletions(-)
create mode 100644 arch/x86/kvm/mmu/page_track.h
base-commit: 9d75a3251adfbcf444681474511b58042a364863
--
2.39.0.314.g84b9a713c41-goog
Bug the VM if something attempts to write-track a gfn, but write-tracking
isn't enabled. The VM is doomed (and KVM has an egregious bug) if KVM or
KVMGT wants to shadow guest page tables but can't because write-tracking
isn't enabled.
Signed-off-by: Sean Christopherson <[email protected]>
---
arch/x86/kvm/mmu/page_track.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
index 209f6beba5ac..d4c3bd6642b3 100644
--- a/arch/x86/kvm/mmu/page_track.c
+++ b/arch/x86/kvm/mmu/page_track.c
@@ -89,7 +89,7 @@ void kvm_write_track_add_gfn(struct kvm *kvm, struct kvm_memory_slot *slot,
lockdep_assert_once(lockdep_is_held(&kvm->slots_lock) ||
srcu_read_lock_held(&kvm->srcu));
- if (WARN_ON(!kvm_page_track_write_tracking_enabled(kvm)))
+ if (KVM_BUG_ON(!kvm_page_track_write_tracking_enabled(kvm), kvm))
return;
update_gfn_write_track(slot, gfn, 1);
@@ -121,7 +121,7 @@ void kvm_write_track_remove_gfn(struct kvm *kvm,
lockdep_assert_once(lockdep_is_held(&kvm->slots_lock) ||
srcu_read_lock_held(&kvm->srcu));
- if (WARN_ON(!kvm_page_track_write_tracking_enabled(kvm)))
+ if (KVM_BUG_ON(!kvm_page_track_write_tracking_enabled(kvm), kvm))
return;
update_gfn_write_track(slot, gfn, -1);
--
2.39.0.314.g84b9a713c41-goog
Disable the page-track notifier code at compile time if there are no
external users, i.e. if CONFIG_KVM_EXTERNAL_WRITE_TRACKING=n. KVM itself
now hooks emulated writes directly instead of relying on the page-track
mechanism.
Signed-off-by: Sean Christopherson <[email protected]>
---
arch/x86/include/asm/kvm_host.h | 2 ++
arch/x86/include/asm/kvm_page_track.h | 2 ++
arch/x86/kvm/mmu/page_track.c | 9 ++++----
arch/x86/kvm/mmu/page_track.h | 30 +++++++++++++++++++++++----
4 files changed, 35 insertions(+), 8 deletions(-)
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index eec424fac0ba..e8f8e1bd96c7 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -1223,7 +1223,9 @@ struct kvm_arch {
* create an NX huge page (without hanging the guest).
*/
struct list_head possible_nx_huge_pages;
+#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
struct kvm_page_track_notifier_head track_notifier_head;
+#endif
/*
* Protects marking pages unsync during page faults, as TDP MMU page
* faults only take mmu_lock for read. For simplicity, the unsync
diff --git a/arch/x86/include/asm/kvm_page_track.h b/arch/x86/include/asm/kvm_page_track.h
index deece45936a5..53c2adb25a07 100644
--- a/arch/x86/include/asm/kvm_page_track.h
+++ b/arch/x86/include/asm/kvm_page_track.h
@@ -55,6 +55,7 @@ void kvm_slot_page_track_remove_page(struct kvm *kvm,
struct kvm_memory_slot *slot, gfn_t gfn,
enum kvm_page_track_mode mode);
+#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
enum pg_level max_level);
@@ -64,5 +65,6 @@ kvm_page_track_register_notifier(struct kvm *kvm,
void
kvm_page_track_unregister_notifier(struct kvm *kvm,
struct kvm_page_track_notifier_node *n);
+#endif /* CONFIG_KVM_EXTERNAL_WRITE_TRACKING */
#endif
diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
index 2b302fd2c5dd..f932909aa9b5 100644
--- a/arch/x86/kvm/mmu/page_track.c
+++ b/arch/x86/kvm/mmu/page_track.c
@@ -193,6 +193,7 @@ bool kvm_slot_page_track_is_active(struct kvm *kvm,
return !!READ_ONCE(slot->arch.gfn_track[mode][index]);
}
+#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
void kvm_page_track_cleanup(struct kvm *kvm)
{
struct kvm_page_track_notifier_head *head;
@@ -208,6 +209,7 @@ int kvm_page_track_init(struct kvm *kvm)
head = &kvm->arch.track_notifier_head;
INIT_HLIST_HEAD(&head->track_notifier_list);
return init_srcu_struct(&head->track_srcu);
+ return 0;
}
/*
@@ -254,8 +256,8 @@ EXPORT_SYMBOL_GPL(kvm_page_track_unregister_notifier);
* The node should figure out if the written page is the one that node is
* interested in by itself.
*/
-void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
- int bytes)
+void __kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
+ int bytes)
{
struct kvm_page_track_notifier_head *head;
struct kvm_page_track_notifier_node *n;
@@ -272,8 +274,6 @@ void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
if (n->track_write)
n->track_write(gpa, new, bytes, n);
srcu_read_unlock(&head->track_srcu, idx);
-
- kvm_mmu_track_write(vcpu, gpa, new, bytes);
}
/*
@@ -316,3 +316,4 @@ enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
return max_level;
}
EXPORT_SYMBOL_GPL(kvm_page_track_max_mapping_level);
+#endif
diff --git a/arch/x86/kvm/mmu/page_track.h b/arch/x86/kvm/mmu/page_track.h
index 89712f123ad3..1b363784aa4a 100644
--- a/arch/x86/kvm/mmu/page_track.h
+++ b/arch/x86/kvm/mmu/page_track.h
@@ -6,8 +6,6 @@
#include <asm/kvm_page_track.h>
-int kvm_page_track_init(struct kvm *kvm);
-void kvm_page_track_cleanup(struct kvm *kvm);
bool kvm_page_track_write_tracking_enabled(struct kvm *kvm);
int kvm_page_track_write_tracking_alloc(struct kvm_memory_slot *slot);
@@ -21,13 +19,37 @@ bool kvm_slot_page_track_is_active(struct kvm *kvm,
const struct kvm_memory_slot *slot,
gfn_t gfn, enum kvm_page_track_mode mode);
-void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
- int bytes);
+#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
+int kvm_page_track_init(struct kvm *kvm);
+void kvm_page_track_cleanup(struct kvm *kvm);
+
+void __kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
+ int bytes);
void kvm_page_track_delete_slot(struct kvm *kvm, struct kvm_memory_slot *slot);
static inline bool kvm_page_track_has_external_user(struct kvm *kvm)
{
return hlist_empty(&kvm->arch.track_notifier_head.track_notifier_list);
}
+#else
+static inline int kvm_page_track_init(struct kvm *kvm) { return 0; }
+static inline void kvm_page_track_cleanup(struct kvm *kvm) { }
+
+static inline void __kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
+ const u8 *new, int bytes) { }
+static inline void kvm_page_track_delete_slot(struct kvm *kvm,
+ struct kvm_memory_slot *slot) { }
+
+static inline bool kvm_page_track_has_external_user(struct kvm *kvm) { return false; }
+
+#endif /* CONFIG_KVM_EXTERNAL_WRITE_TRACKING */
+
+static inline void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
+ const u8 *new, int bytes)
+{
+ __kvm_page_track_write(vcpu, gpa, new, bytes);
+
+ kvm_mmu_track_write(vcpu, gpa, new, bytes);
+}
#endif /* __KVM_X86_PAGE_TRACK_H */
--
2.39.0.314.g84b9a713c41-goog
When shadowing a GTT entry with a 2M page, explicitly verify that the
first page pinned by VFIO is a transparent hugepage instead of assuming
that page observed by is_2MB_gtt_possible() is the same page pinned by
vfio_pin_pages(). E.g. if userspace is doing something funky with the
guest's memslots, or if the page is demoted between is_2MB_gtt_possible()
and vfio_pin_pages().
This is more of a performance optimization than a bug fix as the check
for contiguous struct pages should guard against incorrect mapping (even
though assuming struct pages are virtually contiguous is wrong).
The real motivation for explicitly checking for a transparent hugepage
after pinning is that it will reduce the risk of introducing a bug in a
future fix for a page refcount leak (KVMGT doesn't put the reference
acquired by gfn_to_pfn()), and eventually will allow KVMGT to stop using
KVM's gfn_to_pfn() altogether.
Signed-off-by: Sean Christopherson <[email protected]>
---
drivers/gpu/drm/i915/gvt/kvmgt.c | 18 ++++++++++++++++--
1 file changed, 16 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/i915/gvt/kvmgt.c b/drivers/gpu/drm/i915/gvt/kvmgt.c
index 714221f9a131..6f358b4fe406 100644
--- a/drivers/gpu/drm/i915/gvt/kvmgt.c
+++ b/drivers/gpu/drm/i915/gvt/kvmgt.c
@@ -159,11 +159,25 @@ static int gvt_pin_guest_page(struct intel_vgpu *vgpu, unsigned long gfn,
goto err;
}
- if (npage == 0)
- base_page = cur_page;
+ if (npage == 0) {
+ /*
+ * Bail immediately to avoid unnecessary pinning when
+ * trying to shadow a 2M page and the host page isn't
+ * a transparent hugepage.
+ *
+ * TODO: support other type hugepages, e.g. HugeTLB.
+ */
+ if (size == I915_GTT_PAGE_SIZE_2M &&
+ !PageTransHuge(cur_page))
+ ret = -EIO;
+ else
+ base_page = cur_page;
+ }
else if (base_page + npage != cur_page) {
gvt_vgpu_err("The pages are not continuous\n");
ret = -EINVAL;
+ }
+ if (ret < 0) {
npage++;
goto err;
}
--
2.39.0.314.g84b9a713c41-goog
Bury the declaration of the page-track helpers that are intended only for
internal KVM use in a "private" header. In addition to guarding against
unwanted usage of the internal-only helpers, dropping their definitions
avoids exposing other structures that should be KVM-internal, e.g. for
memslots. This is a baby step toward making kvm_host.h a KVM-internal
header in the very distant future.
Signed-off-by: Sean Christopherson <[email protected]>
---
arch/x86/include/asm/kvm_page_track.h | 26 ++++-----------------
arch/x86/kvm/mmu/mmu.c | 3 ++-
arch/x86/kvm/mmu/page_track.c | 8 +------
arch/x86/kvm/mmu/page_track.h | 33 +++++++++++++++++++++++++++
arch/x86/kvm/x86.c | 1 +
5 files changed, 42 insertions(+), 29 deletions(-)
create mode 100644 arch/x86/kvm/mmu/page_track.h
diff --git a/arch/x86/include/asm/kvm_page_track.h b/arch/x86/include/asm/kvm_page_track.h
index e5eb98ca4fce..deece45936a5 100644
--- a/arch/x86/include/asm/kvm_page_track.h
+++ b/arch/x86/include/asm/kvm_page_track.h
@@ -2,6 +2,8 @@
#ifndef _ASM_X86_KVM_PAGE_TRACK_H
#define _ASM_X86_KVM_PAGE_TRACK_H
+#include <linux/kvm_types.h>
+
enum kvm_page_track_mode {
KVM_PAGE_TRACK_WRITE,
KVM_PAGE_TRACK_MAX,
@@ -46,28 +48,15 @@ struct kvm_page_track_notifier_node {
struct kvm_page_track_notifier_node *node);
};
-int kvm_page_track_init(struct kvm *kvm);
-void kvm_page_track_cleanup(struct kvm *kvm);
-
-bool kvm_page_track_write_tracking_enabled(struct kvm *kvm);
-int kvm_page_track_write_tracking_alloc(struct kvm_memory_slot *slot);
-enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
- enum pg_level max_level);
-
-void kvm_page_track_free_memslot(struct kvm_memory_slot *slot);
-int kvm_page_track_create_memslot(struct kvm *kvm,
- struct kvm_memory_slot *slot,
- unsigned long npages);
-
void kvm_slot_page_track_add_page(struct kvm *kvm,
struct kvm_memory_slot *slot, gfn_t gfn,
enum kvm_page_track_mode mode);
void kvm_slot_page_track_remove_page(struct kvm *kvm,
struct kvm_memory_slot *slot, gfn_t gfn,
enum kvm_page_track_mode mode);
-bool kvm_slot_page_track_is_active(struct kvm *kvm,
- const struct kvm_memory_slot *slot,
- gfn_t gfn, enum kvm_page_track_mode mode);
+
+enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
+ enum pg_level max_level);
void
kvm_page_track_register_notifier(struct kvm *kvm,
@@ -75,10 +64,5 @@ kvm_page_track_register_notifier(struct kvm *kvm,
void
kvm_page_track_unregister_notifier(struct kvm *kvm,
struct kvm_page_track_notifier_node *n);
-void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
- int bytes);
-void kvm_page_track_delete_slot(struct kvm *kvm, struct kvm_memory_slot *slot);
-
-bool kvm_page_track_has_external_user(struct kvm *kvm);
#endif
diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index dfeddea8148a..6477ef435575 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -24,6 +24,7 @@
#include "kvm_cache_regs.h"
#include "smm.h"
#include "kvm_emulate.h"
+#include "page_track.h"
#include "cpuid.h"
#include "spte.h"
@@ -51,7 +52,7 @@
#include <asm/io.h>
#include <asm/set_memory.h>
#include <asm/vmx.h>
-#include <asm/kvm_page_track.h>
+
#include "trace.h"
extern bool itlb_multihit_kvm_mitigation;
diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
index d2b9f7f183cc..2b302fd2c5dd 100644
--- a/arch/x86/kvm/mmu/page_track.c
+++ b/arch/x86/kvm/mmu/page_track.c
@@ -14,10 +14,9 @@
#include <linux/kvm_host.h>
#include <linux/rculist.h>
-#include <asm/kvm_page_track.h>
-
#include "mmu.h"
#include "mmu_internal.h"
+#include "page_track.h"
bool kvm_page_track_write_tracking_enabled(struct kvm *kvm)
{
@@ -317,8 +316,3 @@ enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
return max_level;
}
EXPORT_SYMBOL_GPL(kvm_page_track_max_mapping_level);
-
-bool kvm_page_track_has_external_user(struct kvm *kvm)
-{
- return hlist_empty(&kvm->arch.track_notifier_head.track_notifier_list);
-}
diff --git a/arch/x86/kvm/mmu/page_track.h b/arch/x86/kvm/mmu/page_track.h
new file mode 100644
index 000000000000..89712f123ad3
--- /dev/null
+++ b/arch/x86/kvm/mmu/page_track.h
@@ -0,0 +1,33 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __KVM_X86_PAGE_TRACK_H
+#define __KVM_X86_PAGE_TRACK_H
+
+#include <linux/kvm_host.h>
+
+#include <asm/kvm_page_track.h>
+
+int kvm_page_track_init(struct kvm *kvm);
+void kvm_page_track_cleanup(struct kvm *kvm);
+
+bool kvm_page_track_write_tracking_enabled(struct kvm *kvm);
+int kvm_page_track_write_tracking_alloc(struct kvm_memory_slot *slot);
+
+void kvm_page_track_free_memslot(struct kvm_memory_slot *slot);
+int kvm_page_track_create_memslot(struct kvm *kvm,
+ struct kvm_memory_slot *slot,
+ unsigned long npages);
+
+bool kvm_slot_page_track_is_active(struct kvm *kvm,
+ const struct kvm_memory_slot *slot,
+ gfn_t gfn, enum kvm_page_track_mode mode);
+
+void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
+ int bytes);
+void kvm_page_track_delete_slot(struct kvm *kvm, struct kvm_memory_slot *slot);
+
+static inline bool kvm_page_track_has_external_user(struct kvm *kvm)
+{
+ return hlist_empty(&kvm->arch.track_notifier_head.track_notifier_list);
+}
+
+#endif /* __KVM_X86_PAGE_TRACK_H */
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index f372c41ee2c2..41d47a23396c 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -24,6 +24,7 @@
#include "tss.h"
#include "kvm_cache_regs.h"
#include "kvm_emulate.h"
+#include "mmu/page_track.h"
#include "x86.h"
#include "cpuid.h"
#include "pmu.h"
--
2.39.0.314.g84b9a713c41-goog
Check that the pfn found by gfn_to_pfn() is actually backed by "struct
page" memory prior to retrieving and dereferencing the page. KVM
supports backing guest memory with VM_PFNMAP, VM_IO, etc., and so
there is no guarantee the pfn returned by gfn_to_pfn() has an associated
"struct page".
Fixes: b901b252b6cf ("drm/i915/gvt: Add 2M huge gtt support")
Signed-off-by: Sean Christopherson <[email protected]>
---
drivers/gpu/drm/i915/gvt/gtt.c | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/drivers/gpu/drm/i915/gvt/gtt.c b/drivers/gpu/drm/i915/gvt/gtt.c
index ce0eb03709c3..d0fca53a3563 100644
--- a/drivers/gpu/drm/i915/gvt/gtt.c
+++ b/drivers/gpu/drm/i915/gvt/gtt.c
@@ -1188,6 +1188,10 @@ static int is_2MB_gtt_possible(struct intel_vgpu *vgpu,
pfn = gfn_to_pfn(vgpu->vfio_device.kvm, ops->get_pfn(entry));
if (is_error_noslot_pfn(pfn))
return -EINVAL;
+
+ if (!pfn_valid(pfn))
+ return -EINVAL;
+
return PageTransHuge(pfn_to_page(pfn));
}
--
2.39.0.314.g84b9a713c41-goog
Disallow moving memslots if the VM has external page-track users, i.e. if
KVMGT is being used to expose a virtual GPU to the guest, as KVM doesn't
correctly handle moving memory regions.
Note, this is potential ABI breakage! E.g. userspace could move regions
that aren't shadowed by KVMGT without harming the guest. However, the
only known user of KVMGT is QEMU, and QEMU doesn't move generic memory
regions. KVM's own support for moving memory regions was also broken for
multiple years (albeit for an edge case, but arguably moving RAM is
itself an edge case), e.g. see commit edd4fa37baa6 ("KVM: x86: Allocate
new rmap and large page tracking when moving memslot").
Signed-off-by: Sean Christopherson <[email protected]>
---
arch/x86/include/asm/kvm_page_track.h | 3 +++
arch/x86/kvm/mmu/page_track.c | 5 +++++
arch/x86/kvm/x86.c | 7 +++++++
3 files changed, 15 insertions(+)
diff --git a/arch/x86/include/asm/kvm_page_track.h b/arch/x86/include/asm/kvm_page_track.h
index 0d65ae203fd6..6a287bcbe8a9 100644
--- a/arch/x86/include/asm/kvm_page_track.h
+++ b/arch/x86/include/asm/kvm_page_track.h
@@ -77,4 +77,7 @@ kvm_page_track_unregister_notifier(struct kvm *kvm,
void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
int bytes);
void kvm_page_track_flush_slot(struct kvm *kvm, struct kvm_memory_slot *slot);
+
+bool kvm_page_track_has_external_user(struct kvm *kvm);
+
#endif
diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
index 32357599cb09..c474a0ff24ba 100644
--- a/arch/x86/kvm/mmu/page_track.c
+++ b/arch/x86/kvm/mmu/page_track.c
@@ -320,3 +320,8 @@ enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
return max_level;
}
EXPORT_SYMBOL_GPL(kvm_page_track_max_mapping_level);
+
+bool kvm_page_track_has_external_user(struct kvm *kvm)
+{
+ return hlist_empty(&kvm->arch.track_notifier_head.track_notifier_list);
+}
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index af0d83e33bc4..b587858e878e 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -12419,6 +12419,13 @@ int kvm_arch_prepare_memory_region(struct kvm *kvm,
struct kvm_memory_slot *new,
enum kvm_mr_change change)
{
+ /*
+ * KVM doesn't support moving memslots when there are external page
+ * trackers attached to the VM, i.e. if KVMGT is in use.
+ */
+ if (change == KVM_MR_MOVE && kvm_page_track_has_external_user(kvm))
+ return -EINVAL;
+
if (change == KVM_MR_CREATE || change == KVM_MR_MOVE) {
if ((new->base_gfn + new->npages - 1) > kvm_mmu_max_gfn())
return -EINVAL;
--
2.39.0.314.g84b9a713c41-goog
Honor KVM's max allowed page size when determining whether or not a 2MiB
GTT shadow page can be created for the guest. Querying KVM's max allowed
size is somewhat odd as there's no strict requirement that KVM's memslots
and VFIO's mappings are configured with the same gfn=>hva mapping, but
the check will be accurate if userspace wants to have a functional guest,
and at the very least checking KVM's memslots guarantees that the entire
2MiB range has been exposed to the guest.
Note, KVM may also restrict the mapping size for reasons that aren't
relevant to KVMGT, e.g. for KVM's iTLB multi-hit workaround or if the gfn
is write-tracked (KVM's write-tracking only handles writes from vCPUs).
However, such scenarios are unlikely to occur with a well-behaved guest,
and at worst will result in sub-optimal performance.
Fixes: b901b252b6cf ("drm/i915/gvt: Add 2M huge gtt support")
Signed-off-by: Sean Christopherson <[email protected]>
---
arch/x86/include/asm/kvm_page_track.h | 2 ++
arch/x86/kvm/mmu/page_track.c | 18 ++++++++++++++++++
drivers/gpu/drm/i915/gvt/gtt.c | 10 +++++++++-
3 files changed, 29 insertions(+), 1 deletion(-)
diff --git a/arch/x86/include/asm/kvm_page_track.h b/arch/x86/include/asm/kvm_page_track.h
index eb186bc57f6a..3f72c7a172fc 100644
--- a/arch/x86/include/asm/kvm_page_track.h
+++ b/arch/x86/include/asm/kvm_page_track.h
@@ -51,6 +51,8 @@ void kvm_page_track_cleanup(struct kvm *kvm);
bool kvm_page_track_write_tracking_enabled(struct kvm *kvm);
int kvm_page_track_write_tracking_alloc(struct kvm_memory_slot *slot);
+enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
+ enum pg_level max_level);
void kvm_page_track_free_memslot(struct kvm_memory_slot *slot);
int kvm_page_track_create_memslot(struct kvm *kvm,
diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
index 2e09d1b6249f..69ea16c31859 100644
--- a/arch/x86/kvm/mmu/page_track.c
+++ b/arch/x86/kvm/mmu/page_track.c
@@ -300,3 +300,21 @@ void kvm_page_track_flush_slot(struct kvm *kvm, struct kvm_memory_slot *slot)
n->track_flush_slot(kvm, slot, n);
srcu_read_unlock(&head->track_srcu, idx);
}
+
+enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
+ enum pg_level max_level)
+{
+ struct kvm_memory_slot *slot;
+ int idx;
+
+ idx = srcu_read_lock(&kvm->srcu);
+ slot = gfn_to_memslot(kvm, gfn);
+ if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
+ max_level = PG_LEVEL_4K;
+ else
+ max_level = kvm_mmu_max_slot_mapping_level(slot, gfn, max_level);
+ srcu_read_unlock(&kvm->srcu, idx);
+
+ return max_level;
+}
+EXPORT_SYMBOL_GPL(kvm_page_track_max_mapping_level);
diff --git a/drivers/gpu/drm/i915/gvt/gtt.c b/drivers/gpu/drm/i915/gvt/gtt.c
index d0fca53a3563..6736d7bd94ea 100644
--- a/drivers/gpu/drm/i915/gvt/gtt.c
+++ b/drivers/gpu/drm/i915/gvt/gtt.c
@@ -1178,14 +1178,22 @@ static int is_2MB_gtt_possible(struct intel_vgpu *vgpu,
struct intel_gvt_gtt_entry *entry)
{
const struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
+ unsigned long gfn = ops->get_pfn(entry);
kvm_pfn_t pfn;
+ int max_level;
if (!HAS_PAGE_SIZES(vgpu->gvt->gt->i915, I915_GTT_PAGE_SIZE_2M))
return 0;
if (!vgpu->attached)
return -EINVAL;
- pfn = gfn_to_pfn(vgpu->vfio_device.kvm, ops->get_pfn(entry));
+
+ max_level = kvm_page_track_max_mapping_level(vgpu->vfio_device.kvm,
+ gfn, PG_LEVEL_2M);
+ if (max_level < PG_LEVEL_2M)
+ return 0;
+
+ pfn = gfn_to_pfn(vgpu->vfio_device.kvm, gfn);
if (is_error_noslot_pfn(pfn))
return -EINVAL;
--
2.39.0.314.g84b9a713c41-goog
When handling a slot "flush", don't call back into KVM to drop write
protection for gfns in the slot. Now that KVM rejects attempts to move
memory slots while KVMGT is attached, the only time a slot is "flushed"
is when it's being removed, i.e. the memslot and all its write-tracking
metadata is about to be deleted.
Signed-off-by: Sean Christopherson <[email protected]>
---
drivers/gpu/drm/i915/gvt/kvmgt.c | 8 +-------
1 file changed, 1 insertion(+), 7 deletions(-)
diff --git a/drivers/gpu/drm/i915/gvt/kvmgt.c b/drivers/gpu/drm/i915/gvt/kvmgt.c
index 5ff17a212107..3c59e7cd75d9 100644
--- a/drivers/gpu/drm/i915/gvt/kvmgt.c
+++ b/drivers/gpu/drm/i915/gvt/kvmgt.c
@@ -1659,14 +1659,8 @@ static void kvmgt_page_track_flush_slot(struct kvm *kvm,
mutex_lock(&info->gfn_lock);
for (i = 0; i < slot->npages; i++) {
gfn = slot->base_gfn + i;
- if (kvmgt_gfn_is_write_protected(info, gfn)) {
- write_lock(&kvm->mmu_lock);
- kvm_slot_page_track_remove_page(kvm, slot, gfn,
- KVM_PAGE_TRACK_WRITE);
- write_unlock(&kvm->mmu_lock);
-
+ if (kvmgt_gfn_is_write_protected(info, gfn))
kvmgt_protect_table_del(info, gfn);
- }
}
mutex_unlock(&info->gfn_lock);
}
--
2.39.0.314.g84b9a713c41-goog
Call kvm_mmu_zap_all_fast() directly when flushing a memslot instead of
bounding through the page-track mechanism. KVM (unfortunately) needs to
zap and flush all page tables on memslot DELETE/MOVE irrespective of
whether KVM is shadowing guest page tables.
This will allow changing KVM to register a page-track notifier on the
first shadow root allocation, and will also allow deleting the misguided
kvm_page_track_flush_slot() hook itself once KVM-GT also moves to a
different method for reacting to memslot changes.
No functional change intended.
Cc: Yan Zhao <[email protected]>
Signed-off-by: Sean Christopherson <[email protected]>
Link: https://lore.kernel.org/r/[email protected]
---
arch/x86/include/asm/kvm_host.h | 1 +
arch/x86/kvm/mmu/mmu.c | 10 +---------
arch/x86/kvm/x86.c | 2 ++
3 files changed, 4 insertions(+), 9 deletions(-)
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index aa4eb8cfcd7e..fcb042f971ee 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -1798,6 +1798,7 @@ void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
const struct kvm_memory_slot *memslot);
void kvm_mmu_zap_all(struct kvm *kvm);
+void kvm_mmu_zap_all_fast(struct kvm *kvm);
void kvm_mmu_invalidate_mmio_sptes(struct kvm *kvm, u64 gen);
void kvm_mmu_change_mmu_pages(struct kvm *kvm, unsigned long kvm_nr_mmu_pages);
diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index ca7428b68eba..8c3a453554ed 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -6009,7 +6009,7 @@ static void kvm_zap_obsolete_pages(struct kvm *kvm)
* not use any resource of the being-deleted slot or all slots
* after calling the function.
*/
-static void kvm_mmu_zap_all_fast(struct kvm *kvm)
+void kvm_mmu_zap_all_fast(struct kvm *kvm)
{
lockdep_assert_held(&kvm->slots_lock);
@@ -6065,13 +6065,6 @@ static bool kvm_has_zapped_obsolete_pages(struct kvm *kvm)
return unlikely(!list_empty_careful(&kvm->arch.zapped_obsolete_pages));
}
-static void kvm_mmu_invalidate_zap_pages_in_memslot(struct kvm *kvm,
- struct kvm_memory_slot *slot,
- struct kvm_page_track_notifier_node *node)
-{
- kvm_mmu_zap_all_fast(kvm);
-}
-
int kvm_mmu_init_vm(struct kvm *kvm)
{
struct kvm_page_track_notifier_node *node = &kvm->arch.mmu_sp_tracker;
@@ -6089,7 +6082,6 @@ int kvm_mmu_init_vm(struct kvm *kvm)
}
node->track_write = kvm_mmu_pte_write;
- node->track_flush_slot = kvm_mmu_invalidate_zap_pages_in_memslot;
kvm_page_track_register_notifier(kvm, node);
kvm->arch.split_page_header_cache.kmem_cache = mmu_page_header_cache;
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 312aea1854ae..af0d83e33bc4 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -12599,6 +12599,8 @@ void kvm_arch_flush_shadow_all(struct kvm *kvm)
void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
struct kvm_memory_slot *slot)
{
+ kvm_mmu_zap_all_fast(kvm);
+
kvm_page_track_flush_slot(kvm, slot);
}
--
2.39.0.314.g84b9a713c41-goog
Rename the page-track APIs to capture that they're all about tracking
writes, now that the facade of supporting multiple modes is gone.
Opportunstically replace "slot" with "gfn" in anticipation of removing
the @slot param from the external APIs.
No functional change intended.
Signed-off-by: Sean Christopherson <[email protected]>
---
arch/x86/include/asm/kvm_page_track.h | 8 ++++----
arch/x86/kvm/mmu/mmu.c | 8 ++++----
arch/x86/kvm/mmu/page_track.c | 21 +++++++++------------
arch/x86/kvm/mmu/page_track.h | 4 ++--
drivers/gpu/drm/i915/gvt/kvmgt.c | 4 ++--
5 files changed, 21 insertions(+), 24 deletions(-)
diff --git a/arch/x86/include/asm/kvm_page_track.h b/arch/x86/include/asm/kvm_page_track.h
index 42a4ae451d36..20055064793a 100644
--- a/arch/x86/include/asm/kvm_page_track.h
+++ b/arch/x86/include/asm/kvm_page_track.h
@@ -43,10 +43,10 @@ struct kvm_page_track_notifier_node {
struct kvm_page_track_notifier_node *node);
};
-void kvm_slot_page_track_add_page(struct kvm *kvm,
- struct kvm_memory_slot *slot, gfn_t gfn);
-void kvm_slot_page_track_remove_page(struct kvm *kvm,
- struct kvm_memory_slot *slot, gfn_t gfn);
+void kvm_write_track_add_gfn(struct kvm *kvm,
+ struct kvm_memory_slot *slot, gfn_t gfn);
+void kvm_write_track_remove_gfn(struct kvm *kvm, struct kvm_memory_slot *slot,
+ gfn_t gfn);
#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index ffcfc75cd4c1..b4cc762cfe11 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -807,7 +807,7 @@ static void account_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
/* the non-leaf shadow pages are keeping readonly. */
if (sp->role.level > PG_LEVEL_4K)
- return kvm_slot_page_track_add_page(kvm, slot, gfn);
+ return kvm_write_track_add_gfn(kvm, slot, gfn);
kvm_mmu_gfn_disallow_lpage(slot, gfn);
@@ -853,7 +853,7 @@ static void unaccount_shadowed(struct kvm *kvm, struct kvm_mmu_page *sp)
slots = kvm_memslots_for_spte_role(kvm, sp->role);
slot = __gfn_to_memslot(slots, gfn);
if (sp->role.level > PG_LEVEL_4K)
- return kvm_slot_page_track_remove_page(kvm, slot, gfn);
+ return kvm_write_track_remove_gfn(kvm, slot, gfn);
kvm_mmu_gfn_allow_lpage(slot, gfn);
}
@@ -2725,7 +2725,7 @@ int mmu_try_to_unsync_pages(struct kvm *kvm, const struct kvm_memory_slot *slot,
* track machinery is used to write-protect upper-level shadow pages,
* i.e. this guards the role.level == 4K assertion below!
*/
- if (kvm_slot_page_track_is_active(kvm, slot, gfn))
+ if (kvm_gfn_is_write_tracked(kvm, slot, gfn))
return -EPERM;
/*
@@ -4135,7 +4135,7 @@ static bool page_fault_handle_page_track(struct kvm_vcpu *vcpu,
* guest is writing the page which is write tracked which can
* not be fixed by page fault handler.
*/
- if (kvm_slot_page_track_is_active(vcpu->kvm, fault->slot, fault->gfn))
+ if (kvm_gfn_is_write_tracked(vcpu->kvm, fault->slot, fault->gfn))
return true;
return false;
diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
index 4077aa6d6ff4..1eb516119fdb 100644
--- a/arch/x86/kvm/mmu/page_track.c
+++ b/arch/x86/kvm/mmu/page_track.c
@@ -83,10 +83,9 @@ static void update_gfn_write_track(struct kvm_memory_slot *slot, gfn_t gfn,
* @slot: the @gfn belongs to.
* @gfn: the guest page.
*/
-void kvm_slot_page_track_add_page(struct kvm *kvm,
- struct kvm_memory_slot *slot, gfn_t gfn)
+void kvm_write_track_add_gfn(struct kvm *kvm, struct kvm_memory_slot *slot,
+ gfn_t gfn)
{
-
if (WARN_ON(!kvm_page_track_write_tracking_enabled(kvm)))
return;
@@ -101,12 +100,11 @@ void kvm_slot_page_track_add_page(struct kvm *kvm,
if (kvm_mmu_slot_gfn_write_protect(kvm, slot, gfn, PG_LEVEL_4K))
kvm_flush_remote_tlbs(kvm);
}
-EXPORT_SYMBOL_GPL(kvm_slot_page_track_add_page);
+EXPORT_SYMBOL_GPL(kvm_write_track_add_gfn);
/*
* remove the guest page from the tracking pool which stops the interception
- * of corresponding access on that page. It is the opposed operation of
- * kvm_slot_page_track_add_page().
+ * of corresponding access on that page.
*
* It should be called under the protection both of mmu-lock and kvm->srcu
* or kvm->slots_lock.
@@ -115,8 +113,8 @@ EXPORT_SYMBOL_GPL(kvm_slot_page_track_add_page);
* @slot: the @gfn belongs to.
* @gfn: the guest page.
*/
-void kvm_slot_page_track_remove_page(struct kvm *kvm,
- struct kvm_memory_slot *slot, gfn_t gfn)
+void kvm_write_track_remove_gfn(struct kvm *kvm,
+ struct kvm_memory_slot *slot, gfn_t gfn)
{
if (WARN_ON(!kvm_page_track_write_tracking_enabled(kvm)))
return;
@@ -129,14 +127,13 @@ void kvm_slot_page_track_remove_page(struct kvm *kvm,
*/
kvm_mmu_gfn_allow_lpage(slot, gfn);
}
-EXPORT_SYMBOL_GPL(kvm_slot_page_track_remove_page);
+EXPORT_SYMBOL_GPL(kvm_write_track_remove_gfn);
/*
* check if the corresponding access on the specified guest page is tracked.
*/
-bool kvm_slot_page_track_is_active(struct kvm *kvm,
- const struct kvm_memory_slot *slot,
- gfn_t gfn)
+bool kvm_gfn_is_write_tracked(struct kvm *kvm,
+ const struct kvm_memory_slot *slot, gfn_t gfn)
{
int index;
diff --git a/arch/x86/kvm/mmu/page_track.h b/arch/x86/kvm/mmu/page_track.h
index ae2860bdf560..b27ccc588648 100644
--- a/arch/x86/kvm/mmu/page_track.h
+++ b/arch/x86/kvm/mmu/page_track.h
@@ -15,8 +15,8 @@ int kvm_page_track_create_memslot(struct kvm *kvm,
struct kvm_memory_slot *slot,
unsigned long npages);
-bool kvm_slot_page_track_is_active(struct kvm *kvm,
- const struct kvm_memory_slot *slot, gfn_t gfn);
+bool kvm_gfn_is_write_tracked(struct kvm *kvm,
+ const struct kvm_memory_slot *slot, gfn_t gfn);
#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
int kvm_page_track_init(struct kvm *kvm);
diff --git a/drivers/gpu/drm/i915/gvt/kvmgt.c b/drivers/gpu/drm/i915/gvt/kvmgt.c
index cabad0ff722c..325afeb1246c 100644
--- a/drivers/gpu/drm/i915/gvt/kvmgt.c
+++ b/drivers/gpu/drm/i915/gvt/kvmgt.c
@@ -1584,7 +1584,7 @@ int intel_gvt_page_track_add(struct intel_vgpu *info, u64 gfn)
}
write_lock(&kvm->mmu_lock);
- kvm_slot_page_track_add_page(kvm, slot, gfn);
+ kvm_write_track_add_gfn(kvm, slot, gfn);
write_unlock(&kvm->mmu_lock);
srcu_read_unlock(&kvm->srcu, idx);
@@ -1618,7 +1618,7 @@ int intel_gvt_page_track_remove(struct intel_vgpu *info, u64 gfn)
}
write_lock(&kvm->mmu_lock);
- kvm_slot_page_track_remove_page(kvm, slot, gfn);
+ kvm_write_track_remove_gfn(kvm, slot, gfn);
write_unlock(&kvm->mmu_lock);
srcu_read_unlock(&kvm->srcu, idx);
--
2.39.0.314.g84b9a713c41-goog
On Fri, Dec 23, 2022 at 12:57:12AM +0000, Sean Christopherson wrote:
> Fix a variety of found-by-inspection bugs in KVMGT, and overhaul KVM's
> page-track APIs to provide a leaner and cleaner interface. The motivation
> for this series is to (significantly) reduce the number of KVM APIs that
> KVMGT uses, with a long-term goal of making all kvm_host.h headers
> KVM-internal. That said, I think the cleanup itself is worthwhile,
> e.g. KVMGT really shouldn't be touching kvm->mmu_lock.
>
> Note! The KVMGT changes are compile tested only as I don't have the
> necessary hardware (AFAIK). Testing, and lots of it, on the KVMGT side
> of things is needed and any help on that front would be much appreciated.
hi Sean,
Thanks for the patch!
Could you also provide the commit id that this series is based on?
I applied them on top of latest master branch (6.1.0+,
8395ae05cb5a2e31d36106e8c85efa11cda849be) in repo
https://github.com/torvalds/linux.git, yet met some conflicts and I
fixed them manually. (patch 11 and patch 25).
A rough test shows that below mutex_init is missing.
But even with this fix, I still met guest hang during guest boots up.
Will look into it and have a detailed review next week.
diff --git a/drivers/gpu/drm/i915/gvt/vgpu.c b/drivers/gpu/drm/i915/gvt/vgpu.c
index a7ac2ec00196..c274b6a05555 100644
--- a/drivers/gpu/drm/i915/gvt/vgpu.c
+++ b/drivers/gpu/drm/i915/gvt/vgpu.c
@@ -331,6 +331,7 @@ int intel_gvt_create_vgpu(struct intel_vgpu *vgpu,
vgpu->id = ret;
vgpu->sched_ctl.weight = conf->weight;
mutex_init(&vgpu->vgpu_lock);
+ mutex_init(&vgpu->gfn_lock);
mutex_init(&vgpu->dmabuf_lock);
INIT_LIST_HEAD(&vgpu->dmabuf_obj_list_head);
INIT_RADIX_TREE(&vgpu->page_track_tree, GFP_KERNEL);
Thanks
Yan
On Fri, Dec 23, 2022 at 12:57:15AM +0000, Sean Christopherson wrote:
> Honor KVM's max allowed page size when determining whether or not a 2MiB
> GTT shadow page can be created for the guest. Querying KVM's max allowed
> size is somewhat odd as there's no strict requirement that KVM's memslots
> and VFIO's mappings are configured with the same gfn=>hva mapping, but
Without vIOMMU, VFIO's mapping is configured with the same as KVM's
memslots, i.e. with the same gfn==>HVA mapping
> the check will be accurate if userspace wants to have a functional guest,
> and at the very least checking KVM's memslots guarantees that the entire
> 2MiB range has been exposed to the guest.
I think just check the entrie 2MiB GFN range are all within KVM memslot is
enough.
If for some reason, KVM maps a 2MiB range in 4K sizes, KVMGT can still map
it in IOMMU size in 2MiB size as long as the PFNs are continous and the
whole range is all exposed to guest.
Actually normal device passthrough with VFIO-PCI also maps GFNs in a
similar way, i.e. maps a guest visible range in as large size as
possible as long as the PFN is continous.
>
> Note, KVM may also restrict the mapping size for reasons that aren't
> relevant to KVMGT, e.g. for KVM's iTLB multi-hit workaround or if the gfn
Will iTLB multi-hit affect DMA?
AFAIK, IOMMU mappings currently never sets exec bit (and I'm told this bit is
under discussion to be removed).
> is write-tracked (KVM's write-tracking only handles writes from vCPUs).
> However, such scenarios are unlikely to occur with a well-behaved guest,
> and at worst will result in sub-optimal performance.
> Fixes: b901b252b6cf ("drm/i915/gvt: Add 2M huge gtt support")
> Signed-off-by: Sean Christopherson <[email protected]>
> ---
> arch/x86/include/asm/kvm_page_track.h | 2 ++
> arch/x86/kvm/mmu/page_track.c | 18 ++++++++++++++++++
> drivers/gpu/drm/i915/gvt/gtt.c | 10 +++++++++-
> 3 files changed, 29 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/include/asm/kvm_page_track.h b/arch/x86/include/asm/kvm_page_track.h
> index eb186bc57f6a..3f72c7a172fc 100644
> --- a/arch/x86/include/asm/kvm_page_track.h
> +++ b/arch/x86/include/asm/kvm_page_track.h
> @@ -51,6 +51,8 @@ void kvm_page_track_cleanup(struct kvm *kvm);
>
> bool kvm_page_track_write_tracking_enabled(struct kvm *kvm);
> int kvm_page_track_write_tracking_alloc(struct kvm_memory_slot *slot);
> +enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
> + enum pg_level max_level);
>
> void kvm_page_track_free_memslot(struct kvm_memory_slot *slot);
> int kvm_page_track_create_memslot(struct kvm *kvm,
> diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
> index 2e09d1b6249f..69ea16c31859 100644
> --- a/arch/x86/kvm/mmu/page_track.c
> +++ b/arch/x86/kvm/mmu/page_track.c
> @@ -300,3 +300,21 @@ void kvm_page_track_flush_slot(struct kvm *kvm, struct kvm_memory_slot *slot)
> n->track_flush_slot(kvm, slot, n);
> srcu_read_unlock(&head->track_srcu, idx);
> }
> +
> +enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
> + enum pg_level max_level)
> +{
> + struct kvm_memory_slot *slot;
> + int idx;
> +
> + idx = srcu_read_lock(&kvm->srcu);
> + slot = gfn_to_memslot(kvm, gfn);
> + if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
> + max_level = PG_LEVEL_4K;
> + else
> + max_level = kvm_mmu_max_slot_mapping_level(slot, gfn, max_level);
> + srcu_read_unlock(&kvm->srcu, idx);
> +
> + return max_level;
> +}
> +EXPORT_SYMBOL_GPL(kvm_page_track_max_mapping_level);
> diff --git a/drivers/gpu/drm/i915/gvt/gtt.c b/drivers/gpu/drm/i915/gvt/gtt.c
> index d0fca53a3563..6736d7bd94ea 100644
> --- a/drivers/gpu/drm/i915/gvt/gtt.c
> +++ b/drivers/gpu/drm/i915/gvt/gtt.c
> @@ -1178,14 +1178,22 @@ static int is_2MB_gtt_possible(struct intel_vgpu *vgpu,
> struct intel_gvt_gtt_entry *entry)
> {
> const struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
> + unsigned long gfn = ops->get_pfn(entry);
> kvm_pfn_t pfn;
> + int max_level;
>
> if (!HAS_PAGE_SIZES(vgpu->gvt->gt->i915, I915_GTT_PAGE_SIZE_2M))
> return 0;
>
> if (!vgpu->attached)
> return -EINVAL;
> - pfn = gfn_to_pfn(vgpu->vfio_device.kvm, ops->get_pfn(entry));
> +
> + max_level = kvm_page_track_max_mapping_level(vgpu->vfio_device.kvm,
> + gfn, PG_LEVEL_2M);
> + if (max_level < PG_LEVEL_2M)
> + return 0;
> +
> + pfn = gfn_to_pfn(vgpu->vfio_device.kvm, gfn);
> if (is_error_noslot_pfn(pfn))
> return -EINVAL;
>
> --
> 2.39.0.314.g84b9a713c41-goog
>
On Fri, Dec 23, 2022 at 12:57:31AM +0000, Sean Christopherson wrote:
> Disable the page-track notifier code at compile time if there are no
> external users, i.e. if CONFIG_KVM_EXTERNAL_WRITE_TRACKING=n. KVM itself
> now hooks emulated writes directly instead of relying on the page-track
> mechanism.
>
> Signed-off-by: Sean Christopherson <[email protected]>
> ---
> arch/x86/include/asm/kvm_host.h | 2 ++
> arch/x86/include/asm/kvm_page_track.h | 2 ++
> arch/x86/kvm/mmu/page_track.c | 9 ++++----
> arch/x86/kvm/mmu/page_track.h | 30 +++++++++++++++++++++++----
> 4 files changed, 35 insertions(+), 8 deletions(-)
>
> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> index eec424fac0ba..e8f8e1bd96c7 100644
> --- a/arch/x86/include/asm/kvm_host.h
> +++ b/arch/x86/include/asm/kvm_host.h
> @@ -1223,7 +1223,9 @@ struct kvm_arch {
> * create an NX huge page (without hanging the guest).
> */
> struct list_head possible_nx_huge_pages;
> +#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
> struct kvm_page_track_notifier_head track_notifier_head;
> +#endif
> /*
> * Protects marking pages unsync during page faults, as TDP MMU page
> * faults only take mmu_lock for read. For simplicity, the unsync
> diff --git a/arch/x86/include/asm/kvm_page_track.h b/arch/x86/include/asm/kvm_page_track.h
> index deece45936a5..53c2adb25a07 100644
> --- a/arch/x86/include/asm/kvm_page_track.h
> +++ b/arch/x86/include/asm/kvm_page_track.h
> @@ -55,6 +55,7 @@ void kvm_slot_page_track_remove_page(struct kvm *kvm,
> struct kvm_memory_slot *slot, gfn_t gfn,
> enum kvm_page_track_mode mode);
>
> +#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
> enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
> enum pg_level max_level);
>
> @@ -64,5 +65,6 @@ kvm_page_track_register_notifier(struct kvm *kvm,
> void
> kvm_page_track_unregister_notifier(struct kvm *kvm,
> struct kvm_page_track_notifier_node *n);
> +#endif /* CONFIG_KVM_EXTERNAL_WRITE_TRACKING */
>
> #endif
> diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
> index 2b302fd2c5dd..f932909aa9b5 100644
> --- a/arch/x86/kvm/mmu/page_track.c
> +++ b/arch/x86/kvm/mmu/page_track.c
> @@ -193,6 +193,7 @@ bool kvm_slot_page_track_is_active(struct kvm *kvm,
> return !!READ_ONCE(slot->arch.gfn_track[mode][index]);
> }
>
> +#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
> void kvm_page_track_cleanup(struct kvm *kvm)
> {
> struct kvm_page_track_notifier_head *head;
> @@ -208,6 +209,7 @@ int kvm_page_track_init(struct kvm *kvm)
> head = &kvm->arch.track_notifier_head;
> INIT_HLIST_HEAD(&head->track_notifier_list);
> return init_srcu_struct(&head->track_srcu);
> + return 0;
Double "return"s.
> }
>
> /*
> @@ -254,8 +256,8 @@ EXPORT_SYMBOL_GPL(kvm_page_track_unregister_notifier);
> * The node should figure out if the written page is the one that node is
> * interested in by itself.
> */
> -void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
> - int bytes)
> +void __kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
> + int bytes)
> {
> struct kvm_page_track_notifier_head *head;
> struct kvm_page_track_notifier_node *n;
> @@ -272,8 +274,6 @@ void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
> if (n->track_write)
> n->track_write(gpa, new, bytes, n);
> srcu_read_unlock(&head->track_srcu, idx);
> -
> - kvm_mmu_track_write(vcpu, gpa, new, bytes);
> }
>
> /*
> @@ -316,3 +316,4 @@ enum pg_level kvm_page_track_max_mapping_level(struct kvm *kvm, gfn_t gfn,
> return max_level;
> }
> EXPORT_SYMBOL_GPL(kvm_page_track_max_mapping_level);
> +#endif
> diff --git a/arch/x86/kvm/mmu/page_track.h b/arch/x86/kvm/mmu/page_track.h
> index 89712f123ad3..1b363784aa4a 100644
> --- a/arch/x86/kvm/mmu/page_track.h
> +++ b/arch/x86/kvm/mmu/page_track.h
> @@ -6,8 +6,6 @@
>
> #include <asm/kvm_page_track.h>
>
> -int kvm_page_track_init(struct kvm *kvm);
> -void kvm_page_track_cleanup(struct kvm *kvm);
>
> bool kvm_page_track_write_tracking_enabled(struct kvm *kvm);
> int kvm_page_track_write_tracking_alloc(struct kvm_memory_slot *slot);
> @@ -21,13 +19,37 @@ bool kvm_slot_page_track_is_active(struct kvm *kvm,
> const struct kvm_memory_slot *slot,
> gfn_t gfn, enum kvm_page_track_mode mode);
>
> -void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
> - int bytes);
> +#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
> +int kvm_page_track_init(struct kvm *kvm);
> +void kvm_page_track_cleanup(struct kvm *kvm);
> +
> +void __kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa, const u8 *new,
> + int bytes);
> void kvm_page_track_delete_slot(struct kvm *kvm, struct kvm_memory_slot *slot);
>
> static inline bool kvm_page_track_has_external_user(struct kvm *kvm)
> {
> return hlist_empty(&kvm->arch.track_notifier_head.track_notifier_list);
> }
> +#else
> +static inline int kvm_page_track_init(struct kvm *kvm) { return 0; }
> +static inline void kvm_page_track_cleanup(struct kvm *kvm) { }
> +
> +static inline void __kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
> + const u8 *new, int bytes) { }
> +static inline void kvm_page_track_delete_slot(struct kvm *kvm,
> + struct kvm_memory_slot *slot) { }
> +
> +static inline bool kvm_page_track_has_external_user(struct kvm *kvm) { return false; }
> +
> +#endif /* CONFIG_KVM_EXTERNAL_WRITE_TRACKING */
> +
> +static inline void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
> + const u8 *new, int bytes)
> +{
> + __kvm_page_track_write(vcpu, gpa, new, bytes);
> +
Why not convert "vcpu" to "kvm" in __kvm_page_track_write() ?
i.e.
void __kvm_page_track_write(struct kvm *kvm, gpa_t gpa, const u8 *new, int bytes);
Thanks
Yan
> + kvm_mmu_track_write(vcpu, gpa, new, bytes);
> +}
>
> #endif /* __KVM_X86_PAGE_TRACK_H */
> --
> 2.39.0.314.g84b9a713c41-goog
>
On Wed, Dec 28, 2022, Yan Zhao wrote:
> On Fri, Dec 23, 2022 at 12:57:15AM +0000, Sean Christopherson wrote:
> > Honor KVM's max allowed page size when determining whether or not a 2MiB
> > GTT shadow page can be created for the guest. Querying KVM's max allowed
> > size is somewhat odd as there's no strict requirement that KVM's memslots
> > and VFIO's mappings are configured with the same gfn=>hva mapping, but
> Without vIOMMU, VFIO's mapping is configured with the same as KVM's
> memslots, i.e. with the same gfn==>HVA mapping
But that's controlled by userspace, correct?
> > the check will be accurate if userspace wants to have a functional guest,
> > and at the very least checking KVM's memslots guarantees that the entire
> > 2MiB range has been exposed to the guest.
>
> I think just check the entrie 2MiB GFN range are all within KVM memslot is
> enough.
Strictly speaking, no. E.g. if a 2MiB region is covered with multiple memslots
and the memslots have different properties.
> If for some reason, KVM maps a 2MiB range in 4K sizes, KVMGT can still map
> it in IOMMU size in 2MiB size as long as the PFNs are continous and the
> whole range is all exposed to guest.
I agree that practically speaking this will hold true, but if KVMGT wants to honor
KVM's memslots then checking that KVM allows a hugepage is correct. Hrm, but on
the flip side, KVMGT ignores read-only memslot flags, so KVMGT is already ignoring
pieces of KVM's memslots.
I have no objection to KVMGT defining its ABI such that KVMGT is allowed to create
2MiB so long as (a) the GFN is contiguous according to VFIO, and (b) that the entire
2MiB range is exposed to the guest.
That said, being fully permissive also seems wasteful, e.g. KVM would need to
explicitly support straddling multiple memslots.
As a middle ground, what about tweaking kvm_page_track_is_valid_gfn() to take a
range, and then checking that the range is contained in a single memslot?
E.g. something like:
bool kvm_page_track_is_contiguous_gfn_range(struct kvm *kvm, gfn_t gfn,
unsigned long nr_pages)
{
struct kvm_memory_slot *memslot;
bool ret;
int idx;
idx = srcu_read_lock(&kvm->srcu);
memslot = gfn_to_memslot(kvm, gfn);
ret = kvm_is_visible_memslot(memslot) &&
gfn + nr_pages <= memslot->base_gfn + memslot->npages;
srcu_read_unlock(&kvm->srcu, idx);
return ret;
}
> Actually normal device passthrough with VFIO-PCI also maps GFNs in a
> similar way, i.e. maps a guest visible range in as large size as
> possible as long as the PFN is continous.
> >
> > Note, KVM may also restrict the mapping size for reasons that aren't
> > relevant to KVMGT, e.g. for KVM's iTLB multi-hit workaround or if the gfn
> Will iTLB multi-hit affect DMA?
I highly doubt it, I can't imagine an IOMMU would have a dedicated instruction
TLB :-)
> AFAIK, IOMMU mappings currently never sets exec bit (and I'm told this bit is
> under discussion to be removed).
On Wed, Dec 28, 2022, Yan Zhao wrote:
> On Fri, Dec 23, 2022 at 12:57:31AM +0000, Sean Christopherson wrote:
> > diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
> > index 2b302fd2c5dd..f932909aa9b5 100644
> > --- a/arch/x86/kvm/mmu/page_track.c
> > +++ b/arch/x86/kvm/mmu/page_track.c
> > @@ -193,6 +193,7 @@ bool kvm_slot_page_track_is_active(struct kvm *kvm,
> > return !!READ_ONCE(slot->arch.gfn_track[mode][index]);
> > }
> >
> > +#ifdef CONFIG_KVM_EXTERNAL_WRITE_TRACKING
> > void kvm_page_track_cleanup(struct kvm *kvm)
> > {
> > struct kvm_page_track_notifier_head *head;
> > @@ -208,6 +209,7 @@ int kvm_page_track_init(struct kvm *kvm)
> > head = &kvm->arch.track_notifier_head;
> > INIT_HLIST_HEAD(&head->track_notifier_list);
> > return init_srcu_struct(&head->track_srcu);
> > + return 0;
> Double "return"s.
Huh, I'm surprised this didn't throw a warning. I'm pretty sure I screwed up a
refactoring, I originally had the "return 0" in an #else branch.
> > +#endif /* CONFIG_KVM_EXTERNAL_WRITE_TRACKING */
> > +
> > +static inline void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
> > + const u8 *new, int bytes)
> > +{
> > + __kvm_page_track_write(vcpu, gpa, new, bytes);
> > +
> Why not convert "vcpu" to "kvm" in __kvm_page_track_write() ?
No reason, I just overlooked the opportunistic cleanup. I'll do this in the next
version.
Thanks much for the reviews!
On Fri, Dec 23, 2022, Yan Zhao wrote:
> On Fri, Dec 23, 2022 at 12:57:12AM +0000, Sean Christopherson wrote:
> > Fix a variety of found-by-inspection bugs in KVMGT, and overhaul KVM's
> > page-track APIs to provide a leaner and cleaner interface. The motivation
> > for this series is to (significantly) reduce the number of KVM APIs that
> > KVMGT uses, with a long-term goal of making all kvm_host.h headers
> > KVM-internal. That said, I think the cleanup itself is worthwhile,
> > e.g. KVMGT really shouldn't be touching kvm->mmu_lock.
> >
> > Note! The KVMGT changes are compile tested only as I don't have the
> > necessary hardware (AFAIK). Testing, and lots of it, on the KVMGT side
> > of things is needed and any help on that front would be much appreciated.
> hi Sean,
> Thanks for the patch!
> Could you also provide the commit id that this series is based on?
The commit ID is provided in the cover letter:
base-commit: 9d75a3251adfbcf444681474511b58042a364863
Though you might have a hard time finding that commit as it's from an old
version of kvm/queue that's probably since been force pushed.
> I applied them on top of latest master branch (6.1.0+,
> 8395ae05cb5a2e31d36106e8c85efa11cda849be) in repo
> https://github.com/torvalds/linux.git, yet met some conflicts and I
> fixed them manually. (patch 11 and patch 25).
>
> A rough test shows that below mutex_init is missing.
> But even with this fix, I still met guest hang during guest boots up.
> Will look into it and have a detailed review next week.
Thanks again for the reviews and testing! I'll get a v2 out in the next week or
so (catching up from holidays) and will be more explicit in documenting the base
version.
On Tue, Jan 03, 2023 at 09:13:54PM +0000, Sean Christopherson wrote:
> On Wed, Dec 28, 2022, Yan Zhao wrote:
> > On Fri, Dec 23, 2022 at 12:57:15AM +0000, Sean Christopherson wrote:
> > > Honor KVM's max allowed page size when determining whether or not a 2MiB
> > > GTT shadow page can be created for the guest. Querying KVM's max allowed
> > > size is somewhat odd as there's no strict requirement that KVM's memslots
> > > and VFIO's mappings are configured with the same gfn=>hva mapping, but
> > Without vIOMMU, VFIO's mapping is configured with the same as KVM's
> > memslots, i.e. with the same gfn==>HVA mapping
>
> But that's controlled by userspace, correct?
Yes, controlled by QEMU.
VFIO in kernel has no idea of whether vIOMMU is enabled or not.
KVMGT currently is known not working with vIOMMU with shadow mode on
(in this mode, VFIO maps gIOVA ==> HVA ==> HPA) .
>
> > > the check will be accurate if userspace wants to have a functional guest,
> > > and at the very least checking KVM's memslots guarantees that the entire
> > > 2MiB range has been exposed to the guest.
> >
> > I think just check the entrie 2MiB GFN range are all within KVM memslot is
> > enough.
>
> Strictly speaking, no. E.g. if a 2MiB region is covered with multiple memslots
> and the memslots have different properties.
>
> > If for some reason, KVM maps a 2MiB range in 4K sizes, KVMGT can still map
> > it in IOMMU size in 2MiB size as long as the PFNs are continous and the
> > whole range is all exposed to guest.
>
> I agree that practically speaking this will hold true, but if KVMGT wants to honor
> KVM's memslots then checking that KVM allows a hugepage is correct. Hrm, but on
> the flip side, KVMGT ignores read-only memslot flags, so KVMGT is already ignoring
> pieces of KVM's memslots.
KVMGT calls dma_map_page() with DMA_BIDIRECTIONAL after checking gvt_pin_guest_page().
Though for a read-only memslot, DMA_TO_DEVICE should be used instead
(see dma_info_to_prot()),
as gvt_pin_guest_page() checks (IOMMU_READ | IOMMU_WRITE) permission for each page,
it actually ensures that the pinned GFN is not in a read-only memslot.
So, it should be fine.
>
> I have no objection to KVMGT defining its ABI such that KVMGT is allowed to create
> 2MiB so long as (a) the GFN is contiguous according to VFIO, and (b) that the entire
> 2MiB range is exposed to the guest.
>
sorry. I may not put it clearly enough.
for a normal device pass-through via VFIO-PCI, VFIO maps IOMMU mappings in this way:
(a) fault in PFNs in a GFN range within the same memslot (VFIO saves dma_list, which is
the same as memslot list when vIOMMU is not on or not in shadow mode).
(b) map continuous PFNs into iommu driver (honour ro attribute and can > 2MiB as long as
PFNs are continuous).
(c) IOMMU driver decides to map in 2MiB or in 4KiB according to its setting.
For KVMGT, gvt_dma_map_page() first calls gvt_pin_guest_page() which
(a) calls vfio_pin_pages() to check each GFN is within allowed dma_list with
(IOMMU_READ | IOMMU_WRITE) permission and fault-in page.
(b) checks PFNs are continuous in 2MiB,
Though checking kvm_page_track_max_mapping_level() is also fine, it makes DMA
mapping size unnecessarily smaller.
> That said, being fully permissive also seems wasteful, e.g. KVM would need to
> explicitly support straddling multiple memslots.
>
> As a middle ground, what about tweaking kvm_page_track_is_valid_gfn() to take a
> range, and then checking that the range is contained in a single memslot?
>
> E.g. something like:
>
> bool kvm_page_track_is_contiguous_gfn_range(struct kvm *kvm, gfn_t gfn,
> unsigned long nr_pages)
> {
> struct kvm_memory_slot *memslot;
> bool ret;
> int idx;
>
> idx = srcu_read_lock(&kvm->srcu);
> memslot = gfn_to_memslot(kvm, gfn);
> ret = kvm_is_visible_memslot(memslot) &&
> gfn + nr_pages <= memslot->base_gfn + memslot->npages;
> srcu_read_unlock(&kvm->srcu, idx);
>
> return ret;
> }
Yes, it's good.
But as explained above, gvt_dma_map_page() checks in an equivalent way.
Maybe checking kvm_page_track_is_contiguous_gfn_range() is also not
required?
>
> > Actually normal device passthrough with VFIO-PCI also maps GFNs in a
> > similar way, i.e. maps a guest visible range in as large size as
> > possible as long as the PFN is continous.
> > >
> > > Note, KVM may also restrict the mapping size for reasons that aren't
> > > relevant to KVMGT, e.g. for KVM's iTLB multi-hit workaround or if the gfn
> > Will iTLB multi-hit affect DMA?
>
> I highly doubt it, I can't imagine an IOMMU would have a dedicated instruction
> TLB :-)
I can double check it with IOMMU hardware experts.
But if DMA would tamper instruction TLB, it should have been reported
as an issue with normal VFIO pass-through?
> > AFAIK, IOMMU mappings currently never sets exec bit (and I'm told this bit is
> > under discussion to be removed).
On Wed, Jan 04, 2023 at 01:01:13AM +0000, Sean Christopherson wrote:
> On Fri, Dec 23, 2022, Yan Zhao wrote:
> > On Fri, Dec 23, 2022 at 12:57:12AM +0000, Sean Christopherson wrote:
> > > Fix a variety of found-by-inspection bugs in KVMGT, and overhaul KVM's
> > > page-track APIs to provide a leaner and cleaner interface. The motivation
> > > for this series is to (significantly) reduce the number of KVM APIs that
> > > KVMGT uses, with a long-term goal of making all kvm_host.h headers
> > > KVM-internal. That said, I think the cleanup itself is worthwhile,
> > > e.g. KVMGT really shouldn't be touching kvm->mmu_lock.
> > >
> > > Note! The KVMGT changes are compile tested only as I don't have the
> > > necessary hardware (AFAIK). Testing, and lots of it, on the KVMGT side
> > > of things is needed and any help on that front would be much appreciated.
> > hi Sean,
> > Thanks for the patch!
> > Could you also provide the commit id that this series is based on?
>
> The commit ID is provided in the cover letter:
>
> base-commit: 9d75a3251adfbcf444681474511b58042a364863
>
> Though you might have a hard time finding that commit as it's from an old
> version of kvm/queue that's probably since been force pushed.
>
> > I applied them on top of latest master branch (6.1.0+,
> > 8395ae05cb5a2e31d36106e8c85efa11cda849be) in repo
> > https://github.com/torvalds/linux.git, yet met some conflicts and I
> > fixed them manually. (patch 11 and patch 25).
> >
> > A rough test shows that below mutex_init is missing.
> > But even with this fix, I still met guest hang during guest boots up.
> > Will look into it and have a detailed review next week.
>
> Thanks again for the reviews and testing! I'll get a v2 out in the next week or
> so (catching up from holidays) and will be more explicit in documenting the base
> version.
That's fine and it's a pleasure to me :)
On Thu, Jan 05, 2023, Yan Zhao wrote:
> On Tue, Jan 03, 2023 at 09:13:54PM +0000, Sean Christopherson wrote:
> > On Wed, Dec 28, 2022, Yan Zhao wrote:
> > > On Fri, Dec 23, 2022 at 12:57:15AM +0000, Sean Christopherson wrote:
> > > > Honor KVM's max allowed page size when determining whether or not a 2MiB
> > > > GTT shadow page can be created for the guest. Querying KVM's max allowed
> > > > size is somewhat odd as there's no strict requirement that KVM's memslots
> > > > and VFIO's mappings are configured with the same gfn=>hva mapping, but
> > > Without vIOMMU, VFIO's mapping is configured with the same as KVM's
> > > memslots, i.e. with the same gfn==>HVA mapping
> >
> > But that's controlled by userspace, correct?
>
> Yes, controlled by QEMU.
...
> > Strictly speaking, no. E.g. if a 2MiB region is covered with multiple memslots
> > and the memslots have different properties.
> >
> > > If for some reason, KVM maps a 2MiB range in 4K sizes, KVMGT can still map
> > > it in IOMMU size in 2MiB size as long as the PFNs are continous and the
> > > whole range is all exposed to guest.
> >
> > I agree that practically speaking this will hold true, but if KVMGT wants to honor
> > KVM's memslots then checking that KVM allows a hugepage is correct. Hrm, but on
> > the flip side, KVMGT ignores read-only memslot flags, so KVMGT is already ignoring
> > pieces of KVM's memslots.
> KVMGT calls dma_map_page() with DMA_BIDIRECTIONAL after checking gvt_pin_guest_page().
> Though for a read-only memslot, DMA_TO_DEVICE should be used instead
> (see dma_info_to_prot()),
> as gvt_pin_guest_page() checks (IOMMU_READ | IOMMU_WRITE) permission for each page,
> it actually ensures that the pinned GFN is not in a read-only memslot.
> So, it should be fine.
>
> >
> > I have no objection to KVMGT defining its ABI such that KVMGT is allowed to create
> > 2MiB so long as (a) the GFN is contiguous according to VFIO, and (b) that the entire
> > 2MiB range is exposed to the guest.
> >
> sorry. I may not put it clearly enough.
> for a normal device pass-through via VFIO-PCI, VFIO maps IOMMU mappings in this way:
>
> (a) fault in PFNs in a GFN range within the same memslot (VFIO saves dma_list, which is
> the same as memslot list when vIOMMU is not on or not in shadow mode).
> (b) map continuous PFNs into iommu driver (honour ro attribute and can > 2MiB as long as
> PFNs are continuous).
> (c) IOMMU driver decides to map in 2MiB or in 4KiB according to its setting.
>
> For KVMGT, gvt_dma_map_page() first calls gvt_pin_guest_page() which
> (a) calls vfio_pin_pages() to check each GFN is within allowed dma_list with
> (IOMMU_READ | IOMMU_WRITE) permission and fault-in page.
> (b) checks PFNs are continuous in 2MiB,
>
> Though checking kvm_page_track_max_mapping_level() is also fine, it makes DMA
> mapping size unnecessarily smaller.
Yeah, I got all that. What I'm trying to say, and why I asked about whether or
not userspace controls the mappings, is that AFAIK there is nothing in the kernel
that coordinates mappings between VFIO and KVM. So, very technically, userspace
could map a 2MiB range contiguous in VFIO but not in KVM, or RW in VFIO but RO in KVM.
I can't imagine there's a real use case for doing so, and arguably there's no
requirement that KVMGT honor KVM's memslot. But because KVMGT taps into KVM's
page-tracking, KVMGT _does_ honor KVM's memslots to some extent because KVMGT
needs to know whether or not a given GFN can be write-protected.
I'm totally fine if KVMGT's ABI is that VFIO is the source of truth for mappings
and permissions, and that the only requirement for KVM memslots is that GTT page
tables need to be visible in KVM's memslots. But if that's the ABI, then
intel_gvt_is_valid_gfn() should be probing VFIO, not KVM (commit cc753fbe1ac4
("drm/i915/gvt: validate gfn before set shadow page entry").
In other words, pick either VFIO or KVM. Checking that X is valid according to
KVM and then mapping X through VFIO is confusing and makes assumptions about how
userspace configures KVM and VFIO. It works because QEMU always configures KVM
and VFIO as expected, but IMO it's unnecessarily fragile and again confusing for
unaware readers because the code is technically flawed.
On a related topic, ppgtt_populate_shadow_entry() should check the validity of the
gfn. If I'm reading the code correctly, checking only in ppgtt_populate_spt() fails
to handle the case where the guest creates a bogus mapping when writing an existing
GTT PT.
Combing all my trains of thought, what about this as an end state for this series?
(completely untested at this point). Get rid of the KVM mapping size checks,
verify the validity of the entire range being mapped, and add a FIXME to complain
about using KVM instead of VFIO to determine the validity of ranges.
static bool intel_gvt_is_valid_gfn(struct intel_vgpu *vgpu, unsigned long gfn,
enum intel_gvt_gtt_type type)
{
unsigned long nr_pages;
if (!vgpu->attached)
return false;
if (type == GTT_TYPE_PPGTT_PTE_64K_ENTRY)
nr_pages = I915_GTT_PAGE_SIZE_64K >> PAGE_SHIFT;
else if (type == GTT_TYPE_PPGTT_PTE_2M_ENTRY)
nr_pages = I915_GTT_PAGE_SIZE_2M >> PAGE_SHIFT;
else
nr_pages = 1;
/*
* FIXME: Probe VFIO, not KVM. VFIO is the source of truth for KVMGT
* mappings and permissions, KVM's involvement is purely to handle
* write-tracking of GTT page tables.
*/
return kvm_page_track_is_contiguous_gfn_range(vgpu->vfio_device.kvm,
gfn, nr_pages);
}
static int try_map_2MB_gtt_entry(struct intel_vgpu *vgpu, unsigned long gfn,
dma_addr_t *dma_addr)
{
if (!HAS_PAGE_SIZES(vgpu->gvt->gt->i915, I915_GTT_PAGE_SIZE_2M))
return 0;
return intel_gvt_dma_map_guest_page(vgpu, gfn,
I915_GTT_PAGE_SIZE_2M, dma_addr);
}
static int ppgtt_populate_shadow_entry(struct intel_vgpu *vgpu,
struct intel_vgpu_ppgtt_spt *spt, unsigned long index,
struct intel_gvt_gtt_entry *ge)
{
const struct intel_gvt_gtt_pte_ops *pte_ops = vgpu->gvt->gtt.pte_ops;
dma_addr_t dma_addr = vgpu->gvt->gtt.scratch_mfn << PAGE_SHIFT;
struct intel_gvt_gtt_entry se = *ge;
unsigned long gfn;
int ret;
if (!pte_ops->test_present(ge))
goto set_shadow_entry;
gfn = pte_ops->get_pfn(ge);
if (!intel_gvt_is_valid_gfn(vgpu, gfn, ge->type))
goto set_shadow_entry;
...
set_shadow_entry:
pte_ops->set_pfn(&se, dma_addr >> PAGE_SHIFT);
ppgtt_set_shadow_entry(spt, &se, index);
return 0;
}
On Thu, Jan 05, 2023 at 05:40:32PM +0000, Sean Christopherson wrote:
> On Thu, Jan 05, 2023, Yan Zhao wrote:
> > On Tue, Jan 03, 2023 at 09:13:54PM +0000, Sean Christopherson wrote:
> > > On Wed, Dec 28, 2022, Yan Zhao wrote:
> > > > On Fri, Dec 23, 2022 at 12:57:15AM +0000, Sean Christopherson wrote:
> > > > > Honor KVM's max allowed page size when determining whether or not a 2MiB
> > > > > GTT shadow page can be created for the guest. Querying KVM's max allowed
> > > > > size is somewhat odd as there's no strict requirement that KVM's memslots
> > > > > and VFIO's mappings are configured with the same gfn=>hva mapping, but
> > > > Without vIOMMU, VFIO's mapping is configured with the same as KVM's
> > > > memslots, i.e. with the same gfn==>HVA mapping
> > >
> > > But that's controlled by userspace, correct?
> >
> > Yes, controlled by QEMU.
>
> ...
>
> > > Strictly speaking, no. E.g. if a 2MiB region is covered with multiple memslots
> > > and the memslots have different properties.
> > >
> > > > If for some reason, KVM maps a 2MiB range in 4K sizes, KVMGT can still map
> > > > it in IOMMU size in 2MiB size as long as the PFNs are continous and the
> > > > whole range is all exposed to guest.
> > >
> > > I agree that practically speaking this will hold true, but if KVMGT wants to honor
> > > KVM's memslots then checking that KVM allows a hugepage is correct. Hrm, but on
> > > the flip side, KVMGT ignores read-only memslot flags, so KVMGT is already ignoring
> > > pieces of KVM's memslots.
> > KVMGT calls dma_map_page() with DMA_BIDIRECTIONAL after checking gvt_pin_guest_page().
> > Though for a read-only memslot, DMA_TO_DEVICE should be used instead
> > (see dma_info_to_prot()),
> > as gvt_pin_guest_page() checks (IOMMU_READ | IOMMU_WRITE) permission for each page,
> > it actually ensures that the pinned GFN is not in a read-only memslot.
> > So, it should be fine.
> >
> > >
> > > I have no objection to KVMGT defining its ABI such that KVMGT is allowed to create
> > > 2MiB so long as (a) the GFN is contiguous according to VFIO, and (b) that the entire
> > > 2MiB range is exposed to the guest.
> > >
> > sorry. I may not put it clearly enough.
> > for a normal device pass-through via VFIO-PCI, VFIO maps IOMMU mappings in this way:
> >
> > (a) fault in PFNs in a GFN range within the same memslot (VFIO saves dma_list, which is
> > the same as memslot list when vIOMMU is not on or not in shadow mode).
> > (b) map continuous PFNs into iommu driver (honour ro attribute and can > 2MiB as long as
> > PFNs are continuous).
> > (c) IOMMU driver decides to map in 2MiB or in 4KiB according to its setting.
> >
> > For KVMGT, gvt_dma_map_page() first calls gvt_pin_guest_page() which
> > (a) calls vfio_pin_pages() to check each GFN is within allowed dma_list with
> > (IOMMU_READ | IOMMU_WRITE) permission and fault-in page.
> > (b) checks PFNs are continuous in 2MiB,
> >
> > Though checking kvm_page_track_max_mapping_level() is also fine, it makes DMA
> > mapping size unnecessarily smaller.
>
> Yeah, I got all that. What I'm trying to say, and why I asked about whether or
> not userspace controls the mappings, is that AFAIK there is nothing in the kernel
> that coordinates mappings between VFIO and KVM. So, very technically, userspace
> could map a 2MiB range contiguous in VFIO but not in KVM, or RW in VFIO but RO in KVM.
>
> I can't imagine there's a real use case for doing so, and arguably there's no
> requirement that KVMGT honor KVM's memslot. But because KVMGT taps into KVM's
> page-tracking, KVMGT _does_ honor KVM's memslots to some extent because KVMGT
> needs to know whether or not a given GFN can be write-protected.
>
> I'm totally fine if KVMGT's ABI is that VFIO is the source of truth for mappings
> and permissions, and that the only requirement for KVM memslots is that GTT page
> tables need to be visible in KVM's memslots. But if that's the ABI, then
> intel_gvt_is_valid_gfn() should be probing VFIO, not KVM (commit cc753fbe1ac4
> ("drm/i915/gvt: validate gfn before set shadow page entry").
>
> In other words, pick either VFIO or KVM. Checking that X is valid according to
> KVM and then mapping X through VFIO is confusing and makes assumptions about how
> userspace configures KVM and VFIO. It works because QEMU always configures KVM
> and VFIO as expected, but IMO it's unnecessarily fragile and again confusing for
> unaware readers because the code is technically flawed.
>
Agreed.
Then after some further thought, I think maybe we can just remove
intel_gvt_is_valid_gfn() in KVMGT, because
(1) both intel_gvt_is_valid_gfn() in emulate_ggtt_mmio_write() and
ppgtt_populate_spt() are not for page track purpose, but to validate bogus
GFN.
(2) gvt_pin_guest_page() with gfn and size can do the validity checking,
which is called in intel_gvt_dma_map_guest_page(). So, we can move the
mapping of scratch page to the error path after intel_gvt_dma_map_guest_page().
As below,
diff --git a/drivers/gpu/drm/i915/gvt/gtt.c b/drivers/gpu/drm/i915/gvt/gtt.c
index 54b32ab843eb..5a85936df6d4 100644
--- a/drivers/gpu/drm/i915/gvt/gtt.c
+++ b/drivers/gpu/drm/i915/gvt/gtt.c
@@ -49,15 +49,6 @@
static bool enable_out_of_sync = false;
static int preallocated_oos_pages = 8192;
-static bool intel_gvt_is_valid_gfn(struct intel_vgpu *vgpu, unsigned long gfn)
-{
- if (!vgpu->attached)
- return false;
-
- /* FIXME: This doesn't properly handle guest entries larger than 4K. */
- return kvm_page_track_is_valid_gfn(vgpu->vfio_device.kvm, gfn);
-}
-
/*
* validate a gm address and related range size,
* translate it to host gm address
@@ -1340,16 +1331,12 @@ static int ppgtt_populate_spt(struct intel_vgpu_ppgtt_spt *spt)
ppgtt_generate_shadow_entry(&se, s, &ge);
ppgtt_set_shadow_entry(spt, &se, i);
} else {
- gfn = ops->get_pfn(&ge);
- if (!intel_gvt_is_valid_gfn(vgpu, gfn)) {
+ ret = ppgtt_populate_shadow_entry(vgpu, spt, i, &ge);
+ if (ret) {
ops->set_pfn(&se, gvt->gtt.scratch_mfn);
ppgtt_set_shadow_entry(spt, &se, i);
- continue;
- }
-
- ret = ppgtt_populate_shadow_entry(vgpu, spt, i, &ge);
- if (ret)
goto fail;
+ }
}
}
return 0;
@@ -2336,14 +2325,6 @@ static int emulate_ggtt_mmio_write(struct intel_vgpu *vgpu, unsigned int off,
m.val64 = e.val64;
m.type = e.type;
- /* one PTE update may be issued in multiple writes and the
- * first write may not construct a valid gfn
- */
- if (!intel_gvt_is_valid_gfn(vgpu, gfn)) {
- ops->set_pfn(&m, gvt->gtt.scratch_mfn);
- goto out;
- }
-
ret = intel_gvt_dma_map_guest_page(vgpu, gfn, PAGE_SIZE,
&dma_addr);
if (ret) {
> On a related topic, ppgtt_populate_shadow_entry() should check the validity of the
> gfn. If I'm reading the code correctly, checking only in ppgtt_populate_spt() fails
> to handle the case where the guest creates a bogus mapping when writing an existing
> GTT PT.
Don't get it here. Could you elaborate more?
>
> Combing all my trains of thought, what about this as an end state for this series?
> (completely untested at this point). Get rid of the KVM mapping size checks,
> verify the validity of the entire range being mapped, and add a FIXME to complain
> about using KVM instead of VFIO to determine the validity of ranges.
>
> static bool intel_gvt_is_valid_gfn(struct intel_vgpu *vgpu, unsigned long gfn,
> enum intel_gvt_gtt_type type)
> {
> unsigned long nr_pages;
>
> if (!vgpu->attached)
> return false;
>
> if (type == GTT_TYPE_PPGTT_PTE_64K_ENTRY)
> nr_pages = I915_GTT_PAGE_SIZE_64K >> PAGE_SHIFT;
> else if (type == GTT_TYPE_PPGTT_PTE_2M_ENTRY)
> nr_pages = I915_GTT_PAGE_SIZE_2M >> PAGE_SHIFT;
> else
> nr_pages = 1;
>
> /*
> * FIXME: Probe VFIO, not KVM. VFIO is the source of truth for KVMGT
> * mappings and permissions, KVM's involvement is purely to handle
> * write-tracking of GTT page tables.
> */
> return kvm_page_track_is_contiguous_gfn_range(vgpu->vfio_device.kvm,
> gfn, nr_pages);
> }
>
> static int try_map_2MB_gtt_entry(struct intel_vgpu *vgpu, unsigned long gfn,
> dma_addr_t *dma_addr)
> {
> if (!HAS_PAGE_SIZES(vgpu->gvt->gt->i915, I915_GTT_PAGE_SIZE_2M))
> return 0;
>
> return intel_gvt_dma_map_guest_page(vgpu, gfn,
> I915_GTT_PAGE_SIZE_2M, dma_addr);
> }
>
> static int ppgtt_populate_shadow_entry(struct intel_vgpu *vgpu,
> struct intel_vgpu_ppgtt_spt *spt, unsigned long index,
> struct intel_gvt_gtt_entry *ge)
> {
> const struct intel_gvt_gtt_pte_ops *pte_ops = vgpu->gvt->gtt.pte_ops;
> dma_addr_t dma_addr = vgpu->gvt->gtt.scratch_mfn << PAGE_SHIFT;
> struct intel_gvt_gtt_entry se = *ge;
> unsigned long gfn;
> int ret;
>
> if (!pte_ops->test_present(ge))
> goto set_shadow_entry;
>
> gfn = pte_ops->get_pfn(ge);
> if (!intel_gvt_is_valid_gfn(vgpu, gfn, ge->type))
> goto set_shadow_entry;
As KVMGT only tracks PPGTT page table pages, this check here is not for page
track purpose, but to check bogus GFN.
So, Just leave the bogus GFN check to intel_gvt_dma_map_guest_page() through
VFIO is all right.
On the other hand, for the GFN validity for page track purpose, we can
leave it to kvm_write_track_add_gfn().
Do you think it's ok?
> ...
>
>
> set_shadow_entry:
> pte_ops->set_pfn(&se, dma_addr >> PAGE_SHIFT);
> ppgtt_set_shadow_entry(spt, &se, index);
> return 0;
> }
On Fri, Jan 06, 2023, Yan Zhao wrote:
> On Thu, Jan 05, 2023 at 05:40:32PM +0000, Sean Christopherson wrote:
> > On Thu, Jan 05, 2023, Yan Zhao wrote:
> > I'm totally fine if KVMGT's ABI is that VFIO is the source of truth for mappings
> > and permissions, and that the only requirement for KVM memslots is that GTT page
> > tables need to be visible in KVM's memslots. But if that's the ABI, then
> > intel_gvt_is_valid_gfn() should be probing VFIO, not KVM (commit cc753fbe1ac4
> > ("drm/i915/gvt: validate gfn before set shadow page entry").
> >
> > In other words, pick either VFIO or KVM. Checking that X is valid according to
> > KVM and then mapping X through VFIO is confusing and makes assumptions about how
> > userspace configures KVM and VFIO. It works because QEMU always configures KVM
> > and VFIO as expected, but IMO it's unnecessarily fragile and again confusing for
> > unaware readers because the code is technically flawed.
> >
> Agreed.
> Then after some further thought, I think maybe we can just remove
> intel_gvt_is_valid_gfn() in KVMGT, because
>
> (1) both intel_gvt_is_valid_gfn() in emulate_ggtt_mmio_write() and
> ppgtt_populate_spt() are not for page track purpose, but to validate bogus
> GFN.
> (2) gvt_pin_guest_page() with gfn and size can do the validity checking,
> which is called in intel_gvt_dma_map_guest_page(). So, we can move the
> mapping of scratch page to the error path after intel_gvt_dma_map_guest_page().
IIUC, that will re-introduce the problem commit cc753fbe1ac4 ("drm/i915/gvt: validate
gfn before set shadow page entry") solved by poking into KVM. Lack of pre-validation
means that bogus GFNs will trigger error messages, e.g.
gvt_vgpu_err("vfio_pin_pages failed for iova %pad, ret %d\n",
&cur_iova, ret);
and
gvt_vgpu_err("fail to populate guest ggtt entry\n");
One thought would be to turn those printks into tracepoints to eliminate unwanted
noise, and to prevent the guest from spamming the host kernel log by programming
garbage into the GTT (gvt_vgpu_err() isn't ratelimited).
> > On a related topic, ppgtt_populate_shadow_entry() should check the validity of the
> > gfn. If I'm reading the code correctly, checking only in ppgtt_populate_spt() fails
> > to handle the case where the guest creates a bogus mapping when writing an existing
> > GTT PT.
> Don't get it here. Could you elaborate more?
AFAICT, KVMGT only pre-validates the GFN on the initial setup, not when the guest
modifies a write-tracked entry. I believe this is a moot point if the pre-validation
is removed entirely.
> > gfn = pte_ops->get_pfn(ge);
> > if (!intel_gvt_is_valid_gfn(vgpu, gfn, ge->type))
> > goto set_shadow_entry;
> As KVMGT only tracks PPGTT page table pages, this check here is not for page
> track purpose, but to check bogus GFN.
> So, Just leave the bogus GFN check to intel_gvt_dma_map_guest_page() through
> VFIO is all right.
>
> On the other hand, for the GFN validity for page track purpose, we can
> leave it to kvm_write_track_add_gfn().
>
> Do you think it's ok?
Yep, the only hiccup is the gvt_vgpu_err() calls that are guest-triggerable, and
converting those to a tracepoint seems like the right answer.
On Fri, Jan 06, 2023 at 11:01:53PM +0000, Sean Christopherson wrote:
> On Fri, Jan 06, 2023, Yan Zhao wrote:
> > On Thu, Jan 05, 2023 at 05:40:32PM +0000, Sean Christopherson wrote:
> > > On Thu, Jan 05, 2023, Yan Zhao wrote:
> > > I'm totally fine if KVMGT's ABI is that VFIO is the source of truth for mappings
> > > and permissions, and that the only requirement for KVM memslots is that GTT page
> > > tables need to be visible in KVM's memslots. But if that's the ABI, then
> > > intel_gvt_is_valid_gfn() should be probing VFIO, not KVM (commit cc753fbe1ac4
> > > ("drm/i915/gvt: validate gfn before set shadow page entry").
> > >
> > > In other words, pick either VFIO or KVM. Checking that X is valid according to
> > > KVM and then mapping X through VFIO is confusing and makes assumptions about how
> > > userspace configures KVM and VFIO. It works because QEMU always configures KVM
> > > and VFIO as expected, but IMO it's unnecessarily fragile and again confusing for
> > > unaware readers because the code is technically flawed.
> > >
> > Agreed.
> > Then after some further thought, I think maybe we can just remove
> > intel_gvt_is_valid_gfn() in KVMGT, because
> >
> > (1) both intel_gvt_is_valid_gfn() in emulate_ggtt_mmio_write() and
> > ppgtt_populate_spt() are not for page track purpose, but to validate bogus
> > GFN.
> > (2) gvt_pin_guest_page() with gfn and size can do the validity checking,
> > which is called in intel_gvt_dma_map_guest_page(). So, we can move the
> > mapping of scratch page to the error path after intel_gvt_dma_map_guest_page().
>
> IIUC, that will re-introduce the problem commit cc753fbe1ac4 ("drm/i915/gvt: validate
> gfn before set shadow page entry") solved by poking into KVM. Lack of pre-validation
> means that bogus GFNs will trigger error messages, e.g.
>
> gvt_vgpu_err("vfio_pin_pages failed for iova %pad, ret %d\n",
> &cur_iova, ret);
>
> and
>
> gvt_vgpu_err("fail to populate guest ggtt entry\n");
Thanks for pointing it out.
I checked this commit message and found below original intentions to introduce
pre-validation:
"GVT may receive partial write on one guest PTE update. Validate gfn
not to translate incomplete gfn. This avoids some unnecessary error
messages incurred by the incomplete gfn translating. Also fix the
bug that the whole PPGTT shadow page update is aborted on any invalid
gfn entry"
(1) first intention -- unnecessary error message came from GGTT partial write.
For guest GGTT writes, the guest calls writeq to an MMIO GPA, which is
8 bytes in length, while QEMU splits the MMIO write into 2 4-byte writes.
The splitted 2 writes can cause invalid GFN to be found.
But this partial write issue has been fixed by the two follow-up commits:
bc0686ff5fad drm/i915/gvt: support inconsecutive partial gtt entry write
510fe10b6180 drm/i915/gvt: fix a bug of partially write ggtt enties
so pre-validation to reduce noise is not necessary any more here.
(2) the second intention -- "the whole PPGTT shadow page update is aborted on any
invalid gfn entry"
As PPGTT resides in normal guest RAM and we only treat 8-byte writes
as valid page table writes, any invalid GPA found is regarded as
an error, either due to guest misbehavior/attack or bug in host
shadow code.
So, direct abort looks good too. Like below:
@@ -1340,13 +1338,6 @@ static int ppgtt_populate_spt(struct intel_vgpu_ppgtt_spt *spt)
ppgtt_generate_shadow_entry(&se, s, &ge);
ppgtt_set_shadow_entry(spt, &se, i);
} else {
- gfn = ops->get_pfn(&ge);
- if (!intel_gvt_is_valid_gfn(vgpu, gfn)) {
- ops->set_pfn(&se, gvt->gtt.scratch_mfn);
- ppgtt_set_shadow_entry(spt, &se, i);
- continue;
- }
-
ret = ppgtt_populate_shadow_entry(vgpu, spt, i, &ge);
if (ret)
goto fail;
(I actually found that the original code will print "invalid entry type"
warning which indicates it's broken for a while due to lack of test in
this invalid gfn path)
> One thought would be to turn those printks into tracepoints to eliminate unwanted
> noise, and to prevent the guest from spamming the host kernel log by programming
> garbage into the GTT (gvt_vgpu_err() isn't ratelimited).
As those printks would not happen in normal conditions and printks may have
some advantages to discover the attack or bug, could we just convert
gvt_vgpu_err() to be ratelimited ?
Thanks
Yan
>
> > > On a related topic, ppgtt_populate_shadow_entry() should check the validity of the
> > > gfn. If I'm reading the code correctly, checking only in ppgtt_populate_spt() fails
> > > to handle the case where the guest creates a bogus mapping when writing an existing
> > > GTT PT.
> > Don't get it here. Could you elaborate more?
>
> AFAICT, KVMGT only pre-validates the GFN on the initial setup, not when the guest
> modifies a write-tracked entry. I believe this is a moot point if the pre-validation
> is removed entirely.
>
> > > gfn = pte_ops->get_pfn(ge);
> > > if (!intel_gvt_is_valid_gfn(vgpu, gfn, ge->type))
> > > goto set_shadow_entry;
> > As KVMGT only tracks PPGTT page table pages, this check here is not for page
> > track purpose, but to check bogus GFN.
> > So, Just leave the bogus GFN check to intel_gvt_dma_map_guest_page() through
> > VFIO is all right.
> >
> > On the other hand, for the GFN validity for page track purpose, we can
> > leave it to kvm_write_track_add_gfn().
> >
> > Do you think it's ok?
>
> Yep, the only hiccup is the gvt_vgpu_err() calls that are guest-triggerable, and
> converting those to a tracepoint seems like the right answer.
On Mon, Jan 09, 2023, Yan Zhao wrote:
> On Fri, Jan 06, 2023 at 11:01:53PM +0000, Sean Christopherson wrote:
> > On Fri, Jan 06, 2023, Yan Zhao wrote:
> > > On Thu, Jan 05, 2023 at 05:40:32PM +0000, Sean Christopherson wrote:
> > > > On Thu, Jan 05, 2023, Yan Zhao wrote:
> > > > I'm totally fine if KVMGT's ABI is that VFIO is the source of truth for mappings
> > > > and permissions, and that the only requirement for KVM memslots is that GTT page
> > > > tables need to be visible in KVM's memslots. But if that's the ABI, then
> > > > intel_gvt_is_valid_gfn() should be probing VFIO, not KVM (commit cc753fbe1ac4
> > > > ("drm/i915/gvt: validate gfn before set shadow page entry").
> > > >
> > > > In other words, pick either VFIO or KVM. Checking that X is valid according to
> > > > KVM and then mapping X through VFIO is confusing and makes assumptions about how
> > > > userspace configures KVM and VFIO. It works because QEMU always configures KVM
> > > > and VFIO as expected, but IMO it's unnecessarily fragile and again confusing for
> > > > unaware readers because the code is technically flawed.
> > > >
> > > Agreed.
> > > Then after some further thought, I think maybe we can just remove
> > > intel_gvt_is_valid_gfn() in KVMGT, because
> > >
> > > (1) both intel_gvt_is_valid_gfn() in emulate_ggtt_mmio_write() and
> > > ppgtt_populate_spt() are not for page track purpose, but to validate bogus
> > > GFN.
> > > (2) gvt_pin_guest_page() with gfn and size can do the validity checking,
> > > which is called in intel_gvt_dma_map_guest_page(). So, we can move the
> > > mapping of scratch page to the error path after intel_gvt_dma_map_guest_page().
> >
> > IIUC, that will re-introduce the problem commit cc753fbe1ac4 ("drm/i915/gvt: validate
> > gfn before set shadow page entry") solved by poking into KVM. Lack of pre-validation
> > means that bogus GFNs will trigger error messages, e.g.
> >
> > gvt_vgpu_err("vfio_pin_pages failed for iova %pad, ret %d\n",
> > &cur_iova, ret);
> >
> > and
> >
> > gvt_vgpu_err("fail to populate guest ggtt entry\n");
>
> Thanks for pointing it out.
> I checked this commit message and found below original intentions to introduce
> pre-validation:
...
> (I actually found that the original code will print "invalid entry type"
> warning which indicates it's broken for a while due to lack of test in
> this invalid gfn path)
>
>
> > One thought would be to turn those printks into tracepoints to eliminate unwanted
> > noise, and to prevent the guest from spamming the host kernel log by programming
> > garbage into the GTT (gvt_vgpu_err() isn't ratelimited).
> As those printks would not happen in normal conditions and printks may have
> some advantages to discover the attack or bug, could we just convert
> gvt_vgpu_err() to be ratelimited ?
That's ultimately a decision that needs to be made by the GVT maintainers, as the
answer depends on the use case. E.g. if most users of KVMGT run a single VM and
the guest user is also the host admin, then pr_err_ratelimited() is likely an
acceptable/preferable choice as there's a decent chance a human will see the errors
in the host kernel logs and be able to take action.
But if there's unlikely to be a human monitoring the host logs, and/or the guest
user is unrelated to the host admin, then a ratelimited printk() is less useful.
E.g. if there's no one monitoring the logs, then losing messages due to
ratelimiting provides a worse debug experience overall than having to manually
enable tracepoints. And if there may be many tens of VMs (seems unlikely?), then
ratelimited printk() is even less useful because errors for a specific VM may be
lost, i.e. the printk() can't be relied upon in any way to detect issues.
FWIW, in KVM proper, use of printk() to capture guest "errors" is strongly discourage
for exactly these reasons.
> > > > Note, KVM may also restrict the mapping size for reasons that aren't
> > > > relevant to KVMGT, e.g. for KVM's iTLB multi-hit workaround or if the gfn
> > > Will iTLB multi-hit affect DMA?
> >
> > I highly doubt it, I can't imagine an IOMMU would have a dedicated instruction
> > TLB :-)
> I can double check it with IOMMU hardware experts.
> But if DMA would tamper instruction TLB, it should have been reported
> as an issue with normal VFIO pass-through?
hi Sean,
This is the feedback:
- CPU Instruction TLB is only filled when CPU fetches an instruction.
- IOMMU uses IOTLB to cache IOVA translation.
A remapping hardware may implement multiple IOTLBs, and some of these may
be for special purposes, e.g., only for instruction fetches.
There is no way for software to be aware that multiple
translations for smaller pages have been used for a large page. If software
modifies the paging structures so that the page size used for a 4-KByte range
of input-addresses changes, the IOTLBs may subsequently contain multiple
translations for the address range (one for each page size).
A reference to a input-address in the address range may use any of these
translations. Which translation is used may vary from one execution to
another, and the choice may be implementation-specific.
- Theres no similar bug related to DMA requests for instruction fetch hitting
multiple IOTLB entries reported in IOMMU side.
The X bit in IOMMU paging structure is to be removed in future and is
currently always unset.
Thanks
Yan
On 2023.01.11 17:55:04 +0000, Sean Christopherson wrote:
> On Mon, Jan 09, 2023, Yan Zhao wrote:
> > On Fri, Jan 06, 2023 at 11:01:53PM +0000, Sean Christopherson wrote:
> > > On Fri, Jan 06, 2023, Yan Zhao wrote:
> > > > On Thu, Jan 05, 2023 at 05:40:32PM +0000, Sean Christopherson wrote:
> > > > > On Thu, Jan 05, 2023, Yan Zhao wrote:
> > > > > I'm totally fine if KVMGT's ABI is that VFIO is the source of truth for mappings
> > > > > and permissions, and that the only requirement for KVM memslots is that GTT page
> > > > > tables need to be visible in KVM's memslots. But if that's the ABI, then
> > > > > intel_gvt_is_valid_gfn() should be probing VFIO, not KVM (commit cc753fbe1ac4
> > > > > ("drm/i915/gvt: validate gfn before set shadow page entry").
> > > > >
> > > > > In other words, pick either VFIO or KVM. Checking that X is valid according to
> > > > > KVM and then mapping X through VFIO is confusing and makes assumptions about how
> > > > > userspace configures KVM and VFIO. It works because QEMU always configures KVM
> > > > > and VFIO as expected, but IMO it's unnecessarily fragile and again confusing for
> > > > > unaware readers because the code is technically flawed.
> > > > >
> > > > Agreed.
> > > > Then after some further thought, I think maybe we can just remove
> > > > intel_gvt_is_valid_gfn() in KVMGT, because
> > > >
> > > > (1) both intel_gvt_is_valid_gfn() in emulate_ggtt_mmio_write() and
> > > > ppgtt_populate_spt() are not for page track purpose, but to validate bogus
> > > > GFN.
> > > > (2) gvt_pin_guest_page() with gfn and size can do the validity checking,
> > > > which is called in intel_gvt_dma_map_guest_page(). So, we can move the
> > > > mapping of scratch page to the error path after intel_gvt_dma_map_guest_page().
> > >
> > > IIUC, that will re-introduce the problem commit cc753fbe1ac4 ("drm/i915/gvt: validate
> > > gfn before set shadow page entry") solved by poking into KVM. Lack of pre-validation
> > > means that bogus GFNs will trigger error messages, e.g.
> > >
> > > gvt_vgpu_err("vfio_pin_pages failed for iova %pad, ret %d\n",
> > > &cur_iova, ret);
> > >
> > > and
> > >
> > > gvt_vgpu_err("fail to populate guest ggtt entry\n");
> >
> > Thanks for pointing it out.
> > I checked this commit message and found below original intentions to introduce
> > pre-validation:
>
> ...
>
> > (I actually found that the original code will print "invalid entry type"
> > warning which indicates it's broken for a while due to lack of test in
> > this invalid gfn path)
> >
> >
> > > One thought would be to turn those printks into tracepoints to eliminate unwanted
> > > noise, and to prevent the guest from spamming the host kernel log by programming
> > > garbage into the GTT (gvt_vgpu_err() isn't ratelimited).
> > As those printks would not happen in normal conditions and printks may have
> > some advantages to discover the attack or bug, could we just convert
> > gvt_vgpu_err() to be ratelimited ?
>
> That's ultimately a decision that needs to be made by the GVT maintainers, as the
> answer depends on the use case. E.g. if most users of KVMGT run a single VM and
> the guest user is also the host admin, then pr_err_ratelimited() is likely an
> acceptable/preferable choice as there's a decent chance a human will see the errors
> in the host kernel logs and be able to take action.
>
> But if there's unlikely to be a human monitoring the host logs, and/or the guest
> user is unrelated to the host admin, then a ratelimited printk() is less useful.
> E.g. if there's no one monitoring the logs, then losing messages due to
> ratelimiting provides a worse debug experience overall than having to manually
> enable tracepoints. And if there may be many tens of VMs (seems unlikely?), then
> ratelimited printk() is even less useful because errors for a specific VM may be
> lost, i.e. the printk() can't be relied upon in any way to detect issues.
>
> FWIW, in KVM proper, use of printk() to capture guest "errors" is strongly discourage
> for exactly these reasons.
Current KVMGT usage is mostly in controlled mode, either user is own host admin,
or host admin would pre-configure specific limited number of VMs for KVMGT use.
I think printk on error should be fine, we don't need rate limit, and adding
extra trace monitor for admin might not be necessary. So I'm towards to keep to
use current error message.
thanks
On Thu, Jan 19, 2023 at 10:58:42AM +0800, Zhenyu Wang wrote:
> On 2023.01.11 17:55:04 +0000, Sean Christopherson wrote:
> > On Mon, Jan 09, 2023, Yan Zhao wrote:
> > > On Fri, Jan 06, 2023 at 11:01:53PM +0000, Sean Christopherson wrote:
> > > > On Fri, Jan 06, 2023, Yan Zhao wrote:
> > > > > On Thu, Jan 05, 2023 at 05:40:32PM +0000, Sean Christopherson wrote:
> > > > > > On Thu, Jan 05, 2023, Yan Zhao wrote:
> > > > > > I'm totally fine if KVMGT's ABI is that VFIO is the source of truth for mappings
> > > > > > and permissions, and that the only requirement for KVM memslots is that GTT page
> > > > > > tables need to be visible in KVM's memslots. But if that's the ABI, then
> > > > > > intel_gvt_is_valid_gfn() should be probing VFIO, not KVM (commit cc753fbe1ac4
> > > > > > ("drm/i915/gvt: validate gfn before set shadow page entry").
> > > > > >
> > > > > > In other words, pick either VFIO or KVM. Checking that X is valid according to
> > > > > > KVM and then mapping X through VFIO is confusing and makes assumptions about how
> > > > > > userspace configures KVM and VFIO. It works because QEMU always configures KVM
> > > > > > and VFIO as expected, but IMO it's unnecessarily fragile and again confusing for
> > > > > > unaware readers because the code is technically flawed.
> > > > > >
> > > > > Agreed.
> > > > > Then after some further thought, I think maybe we can just remove
> > > > > intel_gvt_is_valid_gfn() in KVMGT, because
> > > > >
> > > > > (1) both intel_gvt_is_valid_gfn() in emulate_ggtt_mmio_write() and
> > > > > ppgtt_populate_spt() are not for page track purpose, but to validate bogus
> > > > > GFN.
> > > > > (2) gvt_pin_guest_page() with gfn and size can do the validity checking,
> > > > > which is called in intel_gvt_dma_map_guest_page(). So, we can move the
> > > > > mapping of scratch page to the error path after intel_gvt_dma_map_guest_page().
> > > >
> > > > IIUC, that will re-introduce the problem commit cc753fbe1ac4 ("drm/i915/gvt: validate
> > > > gfn before set shadow page entry") solved by poking into KVM. Lack of pre-validation
> > > > means that bogus GFNs will trigger error messages, e.g.
> > > >
> > > > gvt_vgpu_err("vfio_pin_pages failed for iova %pad, ret %d\n",
> > > > &cur_iova, ret);
> > > >
> > > > and
> > > >
> > > > gvt_vgpu_err("fail to populate guest ggtt entry\n");
> > >
> > > Thanks for pointing it out.
> > > I checked this commit message and found below original intentions to introduce
> > > pre-validation:
> >
> > ...
> >
> > > (I actually found that the original code will print "invalid entry type"
> > > warning which indicates it's broken for a while due to lack of test in
> > > this invalid gfn path)
> > >
> > >
> > > > One thought would be to turn those printks into tracepoints to eliminate unwanted
> > > > noise, and to prevent the guest from spamming the host kernel log by programming
> > > > garbage into the GTT (gvt_vgpu_err() isn't ratelimited).
> > > As those printks would not happen in normal conditions and printks may have
> > > some advantages to discover the attack or bug, could we just convert
> > > gvt_vgpu_err() to be ratelimited ?
> >
> > That's ultimately a decision that needs to be made by the GVT maintainers, as the
> > answer depends on the use case. E.g. if most users of KVMGT run a single VM and
> > the guest user is also the host admin, then pr_err_ratelimited() is likely an
> > acceptable/preferable choice as there's a decent chance a human will see the errors
> > in the host kernel logs and be able to take action.
> >
> > But if there's unlikely to be a human monitoring the host logs, and/or the guest
> > user is unrelated to the host admin, then a ratelimited printk() is less useful.
> > E.g. if there's no one monitoring the logs, then losing messages due to
> > ratelimiting provides a worse debug experience overall than having to manually
> > enable tracepoints. And if there may be many tens of VMs (seems unlikely?), then
> > ratelimited printk() is even less useful because errors for a specific VM may be
> > lost, i.e. the printk() can't be relied upon in any way to detect issues.
> >
> > FWIW, in KVM proper, use of printk() to capture guest "errors" is strongly discourage
> > for exactly these reasons.
>
> Current KVMGT usage is mostly in controlled mode, either user is own host admin,
> or host admin would pre-configure specific limited number of VMs for KVMGT use.
> I think printk on error should be fine, we don't need rate limit, and adding
> extra trace monitor for admin might not be necessary. So I'm towards to keep to
> use current error message.
>
Thanks, Sean and Zhenyu.
So, could I just post the final fix as below?
And, Sean, would you like to include it in this series or should I send it out
first?
From dcc931011da3712333f61684ebb20765dbf2fb46 Mon Sep 17 00:00:00 2001
From: Yan Zhao <[email protected]>
Date: Thu, 19 Jan 2023 11:15:54 +0800
Subject: [PATCH] drm/i915/gvt: remove interface intel_gvt_is_valid_gfn
Currently intel_gvt_is_valid_gfn() is called in two places:
(1) shadowing guest GGTT entry
(2) shadowing guest PPGTT leaf entry,
which was introduced in commit cc753fbe1ac4
("drm/i915/gvt: validate gfn before set shadow page entry").
However, now it's not necessary to call this interface any more, because
a. GGTT partial write issue has been fixed by
commit bc0686ff5fad
("drm/i915/gvt: support inconsecutive partial gtt entry write")
commit 510fe10b6180
("drm/i915/gvt: fix a bug of partially write ggtt enties")
b. PPGTT resides in normal guest RAM and we only treat 8-byte writes
as valid page table writes. Any invalid GPA found is regarded as
an error, either due to guest misbehavior/attack or bug in host
shadow code.
So,rather than do GFN pre-checking and replace invalid GFNs with
scratch GFN and continue silently, just remove the pre-checking and
abort PPGTT shadowing on error detected.
c. GFN validity check is still performed in
intel_gvt_dma_map_guest_page() --> gvt_pin_guest_page().
It's more desirable to call VFIO interface to do both validity check
and mapping.
Calling intel_gvt_is_valid_gfn() to do GFN validity check from KVM side
while later mapping the GFN through VFIO interface is unnecessarily
fragile and confusing for unaware readers.
Signed-off-by: Yan Zhao <[email protected]>
---
drivers/gpu/drm/i915/gvt/gtt.c | 31 -------------------------------
1 file changed, 31 deletions(-)
diff --git a/drivers/gpu/drm/i915/gvt/gtt.c b/drivers/gpu/drm/i915/gvt/gtt.c
index 445afecbe7ae..9b6c2ca1ee16 100644
--- a/drivers/gpu/drm/i915/gvt/gtt.c
+++ b/drivers/gpu/drm/i915/gvt/gtt.c
@@ -49,22 +49,6 @@
static bool enable_out_of_sync = false;
static int preallocated_oos_pages = 8192;
-static bool intel_gvt_is_valid_gfn(struct intel_vgpu *vgpu, unsigned long gfn)
-{
- struct kvm *kvm = vgpu->vfio_device.kvm;
- int idx;
- bool ret;
-
- if (!vgpu->attached)
- return false;
-
- idx = srcu_read_lock(&kvm->srcu);
- ret = kvm_is_visible_gfn(kvm, gfn);
- srcu_read_unlock(&kvm->srcu, idx);
-
- return ret;
-}
-
/*
* validate a gm address and related range size,
* translate it to host gm address
@@ -1345,13 +1329,6 @@ static int ppgtt_populate_spt(struct intel_vgpu_ppgtt_spt *spt)
ppgtt_generate_shadow_entry(&se, s, &ge);
ppgtt_set_shadow_entry(spt, &se, i);
} else {
- gfn = ops->get_pfn(&ge);
- if (!intel_gvt_is_valid_gfn(vgpu, gfn)) {
- ops->set_pfn(&se, gvt->gtt.scratch_mfn);
- ppgtt_set_shadow_entry(spt, &se, i);
- continue;
- }
-
ret = ppgtt_populate_shadow_entry(vgpu, spt, i, &ge);
if (ret)
goto fail;
@@ -2326,14 +2303,6 @@ static int emulate_ggtt_mmio_write(struct intel_vgpu *vgpu, unsigned int off,
m.val64 = e.val64;
m.type = e.type;
- /* one PTE update may be issued in multiple writes and the
- * first write may not construct a valid gfn
- */
- if (!intel_gvt_is_valid_gfn(vgpu, gfn)) {
- ops->set_pfn(&m, gvt->gtt.scratch_mfn);
- goto out;
- }
-
ret = intel_gvt_dma_map_guest_page(vgpu, gfn, PAGE_SIZE,
&dma_addr);
if (ret) {
--
2.17.1
Apologies for the super slow reply, I put this series on the backburner while I
caught up on other stuff and completely missed your questions.
On Thu, Jan 19, 2023, Yan Zhao wrote:
> On Thu, Jan 19, 2023 at 10:58:42AM +0800, Zhenyu Wang wrote:
> > Current KVMGT usage is mostly in controlled mode, either user is own host admin,
> > or host admin would pre-configure specific limited number of VMs for KVMGT use.
> > I think printk on error should be fine, we don't need rate limit, and adding
> > extra trace monitor for admin might not be necessary. So I'm towards to keep to
> > use current error message.
> >
>
> Thanks, Sean and Zhenyu.
> So, could I just post the final fix as below?
No objection here.
> And, Sean, would you like to include it in this series or should I send it out
> first?
I'd like to include it in this series as it's necessary (for some definitions of
necessary) to clean up KVM's APIs, and the main benefactor is KVM, i.e. getting
the patch merged sooner than later doesn't really benefit KVMGT itself.
Thanks much!
On Thu, Feb 23, 2023 at 12:41:28PM -0800, Sean Christopherson wrote:
> Apologies for the super slow reply, I put this series on the backburner while I
> caught up on other stuff and completely missed your questions.
>
Never mind :)
> On Thu, Jan 19, 2023, Yan Zhao wrote:
> > On Thu, Jan 19, 2023 at 10:58:42AM +0800, Zhenyu Wang wrote:
> > > Current KVMGT usage is mostly in controlled mode, either user is own host admin,
> > > or host admin would pre-configure specific limited number of VMs for KVMGT use.
> > > I think printk on error should be fine, we don't need rate limit, and adding
> > > extra trace monitor for admin might not be necessary. So I'm towards to keep to
> > > use current error message.
> > >
> >
> > Thanks, Sean and Zhenyu.
> > So, could I just post the final fix as below?
>
> No objection here.
>
> > And, Sean, would you like to include it in this series or should I send it out
> > first?
>
> I'd like to include it in this series as it's necessary (for some definitions of
> necessary) to clean up KVM's APIs, and the main benefactor is KVM, i.e. getting
> the patch merged sooner than later doesn't really benefit KVMGT itself.
>
> Thanks much!
Then please include it and I can help to test once you sending out next
version.
Thanks
Yan
On 23/12/2022 8:57 am, Sean Christopherson wrote:
> +static inline void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
> + const u8 *new, int bytes)
> +{
> + __kvm_page_track_write(vcpu, gpa, new, bytes);
> +
> + kvm_mmu_track_write(vcpu, gpa, new, bytes);
> +}
The kvm_mmu_track_write() is only used for x86, where the incoming parameter
"u8 *new" has not been required since 0e0fee5c539b ("kvm: mmu: Fix race in
emulated page table writes"), please help confirm if it's still needed ? Thanks.
A minor clean up is proposed.
On Mon, Aug 07, 2023, Like Xu wrote:
> On 23/12/2022 8:57 am, Sean Christopherson wrote:
> > +static inline void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
> > + const u8 *new, int bytes)
> > +{
> > + __kvm_page_track_write(vcpu, gpa, new, bytes);
> > +
> > + kvm_mmu_track_write(vcpu, gpa, new, bytes);
> > +}
>
> The kvm_mmu_track_write() is only used for x86, where the incoming parameter
> "u8 *new" has not been required since 0e0fee5c539b ("kvm: mmu: Fix race in
> emulated page table writes"), please help confirm if it's still needed ? Thanks.
> A minor clean up is proposed.
Hmm, unless I'm misreading things, KVMGT ultimately doesn't consume @new either.
So I think we can remove @new from kvm_page_track_write() entirely.
Feel free to send a patch, otherwise I'll get to it later this week.
On Mon, Aug 07, 2023 at 10:19:07AM -0700, Sean Christopherson wrote:
> On Mon, Aug 07, 2023, Like Xu wrote:
> > On 23/12/2022 8:57 am, Sean Christopherson wrote:
> > > +static inline void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
> > > + const u8 *new, int bytes)
> > > +{
> > > + __kvm_page_track_write(vcpu, gpa, new, bytes);
> > > +
> > > + kvm_mmu_track_write(vcpu, gpa, new, bytes);
> > > +}
> >
> > The kvm_mmu_track_write() is only used for x86, where the incoming parameter
> > "u8 *new" has not been required since 0e0fee5c539b ("kvm: mmu: Fix race in
> > emulated page table writes"), please help confirm if it's still needed ? Thanks.
> > A minor clean up is proposed.
>
> Hmm, unless I'm misreading things, KVMGT ultimately doesn't consume @new either.
> So I think we can remove @new from kvm_page_track_write() entirely.
Sorry for the late reply.
Yes, KVMGT does not consume @new and it reads the guest PTE again in the
page track write handler.
But I have a couple of questions related to the memtioned commit as
below:
(1) If "re-reading the current value of the guest PTE after the MMU lock has
been acquired", then should KVMGT also acquire the MMU lock too?
If so, could we move the MMU lock and unlock into kvm_page_track_write()
as it's common.
(2) Even if KVMGT consumes @new,
will kvm_page_track_write() be called for once or twice if there are two
concurent emulated write?
commit 0e0fee5c539b61fdd098332e0e2cc375d9073706
Author: Junaid Shahid <[email protected]>
Date: Wed Oct 31 14:53:57 2018 -0700
kvm: mmu: Fix race in emulated page table writes
When a guest page table is updated via an emulated write,
kvm_mmu_pte_write() is called to update the shadow PTE using the just
written guest PTE value. But if two emulated guest PTE writes happened
concurrently, it is possible that the guest PTE and the shadow PTE end
up being out of sync. Emulated writes do not mark the shadow page as
unsync-ed, so this inconsistency will not be resolved even by a guest TLB
flush (unless the page was marked as unsync-ed at some other point).
This is fixed by re-reading the current value of the guest PTE after the
MMU lock has been acquired instead of just using the value that was
written prior to calling kvm_mmu_pte_write().
On Wed, Aug 09, 2023, Yan Zhao wrote:
> On Mon, Aug 07, 2023 at 10:19:07AM -0700, Sean Christopherson wrote:
> > On Mon, Aug 07, 2023, Like Xu wrote:
> > > On 23/12/2022 8:57 am, Sean Christopherson wrote:
> > > > +static inline void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
> > > > + const u8 *new, int bytes)
> > > > +{
> > > > + __kvm_page_track_write(vcpu, gpa, new, bytes);
> > > > +
> > > > + kvm_mmu_track_write(vcpu, gpa, new, bytes);
> > > > +}
> > >
> > > The kvm_mmu_track_write() is only used for x86, where the incoming parameter
> > > "u8 *new" has not been required since 0e0fee5c539b ("kvm: mmu: Fix race in
> > > emulated page table writes"), please help confirm if it's still needed ? Thanks.
> > > A minor clean up is proposed.
> >
> > Hmm, unless I'm misreading things, KVMGT ultimately doesn't consume @new either.
> > So I think we can remove @new from kvm_page_track_write() entirely.
> Sorry for the late reply.
> Yes, KVMGT does not consume @new and it reads the guest PTE again in the
> page track write handler.
>
> But I have a couple of questions related to the memtioned commit as
> below:
>
> (1) If "re-reading the current value of the guest PTE after the MMU lock has
> been acquired", then should KVMGT also acquire the MMU lock too?
No. If applicable, KVMGT should read the new/current value after acquiring
whatever lock protects the generation (or update) of the shadow entries. I
suspect KVMGT already does this, but I don't have time to confirm that at this
exact memory.
The race that was fixed in KVM was:
vCPU0 vCPU1
write X
write Y
sync SPTE w/ Y
sync SPTE w/ X
Reading the value after acquiring mmu_lock ensures that both vCPUs will see whatever
value "loses" the race, i.e. whatever written value is processed second ('Y' in the
above sequence).
> If so, could we move the MMU lock and unlock into kvm_page_track_write()
> as it's common.
>
> (2) Even if KVMGT consumes @new,
> will kvm_page_track_write() be called for once or twice if there are two
> concurent emulated write?
Twice, kvm_page_track_write() is wired up directly to the emulation of the write,
i.e. there is no batching.
On Wed, Aug 09, 2023 at 07:33:45AM -0700, Sean Christopherson wrote:
> On Wed, Aug 09, 2023, Yan Zhao wrote:
> > On Mon, Aug 07, 2023 at 10:19:07AM -0700, Sean Christopherson wrote:
> > > On Mon, Aug 07, 2023, Like Xu wrote:
> > > > On 23/12/2022 8:57 am, Sean Christopherson wrote:
> > > > > +static inline void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
> > > > > + const u8 *new, int bytes)
> > > > > +{
> > > > > + __kvm_page_track_write(vcpu, gpa, new, bytes);
> > > > > +
> > > > > + kvm_mmu_track_write(vcpu, gpa, new, bytes);
> > > > > +}
> > > >
> > > > The kvm_mmu_track_write() is only used for x86, where the incoming parameter
> > > > "u8 *new" has not been required since 0e0fee5c539b ("kvm: mmu: Fix race in
> > > > emulated page table writes"), please help confirm if it's still needed ? Thanks.
> > > > A minor clean up is proposed.
> > >
> > > Hmm, unless I'm misreading things, KVMGT ultimately doesn't consume @new either.
> > > So I think we can remove @new from kvm_page_track_write() entirely.
> > Sorry for the late reply.
> > Yes, KVMGT does not consume @new and it reads the guest PTE again in the
> > page track write handler.
> >
> > But I have a couple of questions related to the memtioned commit as
> > below:
> >
> > (1) If "re-reading the current value of the guest PTE after the MMU lock has
> > been acquired", then should KVMGT also acquire the MMU lock too?
>
> No. If applicable, KVMGT should read the new/current value after acquiring
> whatever lock protects the generation (or update) of the shadow entries. I
> suspect KVMGT already does this, but I don't have time to confirm that at this
I think the mutex lock and unlock of info->vgpu_lock you added in
kvmgt_page_track_write() is the counterpart :)
> exact memory.
>
> The race that was fixed in KVM was:
>
> vCPU0 vCPU1
> write X
> write Y
> sync SPTE w/ Y
> sync SPTE w/ X
>
> Reading the value after acquiring mmu_lock ensures that both vCPUs will see whatever
> value "loses" the race, i.e. whatever written value is processed second ('Y' in the
> above sequence).
I suspect that vCPU0 may still generate a wrong SPTE if vCPU1 wrote 4
bytes while vCPU0 wrote 8 bytes, though the chances are very low.
>
> > If so, could we move the MMU lock and unlock into kvm_page_track_write()
> > as it's common.
> >
> > (2) Even if KVMGT consumes @new,
> > will kvm_page_track_write() be called for once or twice if there are two
> > concurent emulated write?
>
> Twice, kvm_page_track_write() is wired up directly to the emulation of the write,
> i.e. there is no batching.
On Thu, Aug 10, 2023 at 07:21:03AM +0800, Yan Zhao wrote:
> On Wed, Aug 09, 2023 at 07:33:45AM -0700, Sean Christopherson wrote:
> > On Wed, Aug 09, 2023, Yan Zhao wrote:
> > > On Mon, Aug 07, 2023 at 10:19:07AM -0700, Sean Christopherson wrote:
> > > > On Mon, Aug 07, 2023, Like Xu wrote:
> > > > > On 23/12/2022 8:57 am, Sean Christopherson wrote:
> > > > > > +static inline void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
> > > > > > + const u8 *new, int bytes)
> > > > > > +{
> > > > > > + __kvm_page_track_write(vcpu, gpa, new, bytes);
> > > > > > +
> > > > > > + kvm_mmu_track_write(vcpu, gpa, new, bytes);
> > > > > > +}
> > > > >
> > > > > The kvm_mmu_track_write() is only used for x86, where the incoming parameter
> > > > > "u8 *new" has not been required since 0e0fee5c539b ("kvm: mmu: Fix race in
> > > > > emulated page table writes"), please help confirm if it's still needed ? Thanks.
> > > > > A minor clean up is proposed.
> > > >
> > > > Hmm, unless I'm misreading things, KVMGT ultimately doesn't consume @new either.
> > > > So I think we can remove @new from kvm_page_track_write() entirely.
> > > Sorry for the late reply.
> > > Yes, KVMGT does not consume @new and it reads the guest PTE again in the
> > > page track write handler.
> > >
> > > But I have a couple of questions related to the memtioned commit as
> > > below:
> > >
> > > (1) If "re-reading the current value of the guest PTE after the MMU lock has
> > > been acquired", then should KVMGT also acquire the MMU lock too?
> >
> > No. If applicable, KVMGT should read the new/current value after acquiring
> > whatever lock protects the generation (or update) of the shadow entries. I
> > suspect KVMGT already does this, but I don't have time to confirm that at this
> I think the mutex lock and unlock of info->vgpu_lock you added in
> kvmgt_page_track_write() is the counterpart :)
>
> > exact memory.
> >
> > The race that was fixed in KVM was:
> >
> > vCPU0 vCPU1
> > write X
> > write Y
> > sync SPTE w/ Y
> > sync SPTE w/ X
> >
> > Reading the value after acquiring mmu_lock ensures that both vCPUs will see whatever
> > value "loses" the race, i.e. whatever written value is processed second ('Y' in the
> > above sequence).
> I suspect that vCPU0 may still generate a wrong SPTE if vCPU1 wrote 4
> bytes while vCPU0 wrote 8 bytes, though the chances are very low.
>
This could happen in below sequence:
vCPU0 updates a PTE to AABBCCDD;
vCPU1 updates a PTE to EEFFGGHH in two writes.
(each character stands for a byte)
vCPU0 vCPU1
write AABBCCDD
write GGHH
detect 4 bytes write and hold on sync
sync SPTE w/ AABBGGHH
write EEFF
sync SPTE w/ EEFFGGHH
Do you think it worth below serialization work?
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index a915e23d61fa..51cd0ab73529 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -1445,6 +1445,8 @@ struct kvm_arch {
*/
#define SPLIT_DESC_CACHE_MIN_NR_OBJECTS (SPTE_ENT_PER_PAGE + 1)
struct kvm_mmu_memory_cache split_desc_cache;
+
+ struct xarray track_writing_range;
};
struct kvm_vm_stat {
diff --git a/arch/x86/kvm/mmu/page_track.c b/arch/x86/kvm/mmu/page_track.c
index fd04e618ad2d..4b271701dcf6 100644
--- a/arch/x86/kvm/mmu/page_track.c
+++ b/arch/x86/kvm/mmu/page_track.c
@@ -142,12 +142,14 @@ void kvm_page_track_cleanup(struct kvm *kvm)
head = &kvm->arch.track_notifier_head;
cleanup_srcu_struct(&head->track_srcu);
+ xa_destroy(&kvm->arch.track_writing_range);
}
int kvm_page_track_init(struct kvm *kvm)
{
struct kvm_page_track_notifier_head *head;
+ xa_init(&kvm->arch.track_writing_range);
head = &kvm->arch.track_notifier_head;
INIT_HLIST_HEAD(&head->track_notifier_list);
return init_srcu_struct(&head->track_srcu);
diff --git a/arch/x86/kvm/mmu/page_track.h b/arch/x86/kvm/mmu/page_track.h
index 62f98c6c5af3..1829792b9892 100644
--- a/arch/x86/kvm/mmu/page_track.h
+++ b/arch/x86/kvm/mmu/page_track.h
@@ -47,12 +47,46 @@ static inline bool kvm_page_track_has_external_user(struct kvm *kvm) { return fa
#endif /* CONFIG_KVM_EXTERNAL_WRITE_TRACKING */
-static inline void kvm_page_track_write(struct kvm_vcpu *vcpu, gpa_t gpa,
- const u8 *new, int bytes)
+static inline void kvm_page_track_write_begin(struct kvm_vcpu *vcpu, gpa_t gpa,
+ int bytes)
{
+ struct kvm *kvm = vcpu->kvm;
+ gfn_t gfn = gpa_to_gfn(gpa);
+
+ WARN_ON(gfn != gpa_to_gfn(gpa + bytes - 1));
+
+ if (!kvm_page_track_write_tracking_enabled(kvm))
+ return;
+
+retry:
+ if (xa_insert(&kvm->arch.track_writing_range, gfn, xa_mk_value(gfn),
+ GFP_KERNEL_ACCOUNT)) {
+ cpu_relax();
+ goto retry;
+ }
+ return;
+}
+
+static inline void kvm_page_track_write_abort(struct kvm_vcpu *vcpu, gpa_t gpa,
+ int bytes)
+{
+ if (!kvm_page_track_write_tracking_enabled(vcpu->kvm))
+ return;
+
+ xa_erase(&vcpu->kvm->arch.track_writing_range, gpa_to_gfn(gpa));
+}
+
+static inline void kvm_page_track_write_end(struct kvm_vcpu *vcpu, gpa_t gpa,
+ const u8 *new, int bytes)
+{
+ if (!kvm_page_track_write_tracking_enabled(vcpu->kvm))
+ return;
+
__kvm_page_track_write(vcpu->kvm, gpa, new, bytes);
kvm_mmu_track_write(vcpu, gpa, new, bytes);
+
+ xa_erase(&vcpu->kvm->arch.track_writing_range, gpa_to_gfn(gpa));
}
#endif /* __KVM_X86_PAGE_TRACK_H */
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index 05a68d7d99fe..9b75829d5d7a 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -7544,10 +7544,13 @@ int emulator_write_phys(struct kvm_vcpu *vcpu, gpa_t gpa,
{
int ret;
+ kvm_page_track_write_begin(vcpu, gpa, bytes);
ret = kvm_vcpu_write_guest(vcpu, gpa, val, bytes);
- if (ret < 0)
+ if (ret < 0) {
+ kvm_page_track_write_abort(vcpu, gpa, bytes);
return 0;
- kvm_page_track_write(vcpu, gpa, val, bytes);
+ }
+ kvm_page_track_write_end(vcpu, gpa, val, bytes);
return 1;
}
@@ -7792,6 +7795,7 @@ static int emulator_cmpxchg_emulated(struct x86_emulate_ctxt *ctxt,
hva += offset_in_page(gpa);
+ kvm_page_track_write_begin(vcpu, gpa, bytes);
switch (bytes) {
case 1:
r = emulator_try_cmpxchg_user(u8, hva, old, new);
@@ -7809,12 +7813,16 @@ static int emulator_cmpxchg_emulated(struct x86_emulate_ctxt *ctxt,
BUG();
}
- if (r < 0)
+ if (r < 0) {
+ kvm_page_track_write_abort(vcpu, gpa, bytes);
return X86EMUL_UNHANDLEABLE;
- if (r)
+ }
+ if (r) {
+ kvm_page_track_write_abort(vcpu, gpa, bytes);
return X86EMUL_CMPXCHG_FAILED;
+ }
- kvm_page_track_write(vcpu, gpa, new, bytes);
+ kvm_page_track_write_end(vcpu, gpa, new, bytes);
return X86EMUL_CONTINUE;
On Thu, Aug 10, 2023, Yan Zhao wrote:
> On Thu, Aug 10, 2023 at 07:21:03AM +0800, Yan Zhao wrote:
> > > Reading the value after acquiring mmu_lock ensures that both vCPUs will see whatever
> > > value "loses" the race, i.e. whatever written value is processed second ('Y' in the
> > > above sequence).
> > I suspect that vCPU0 may still generate a wrong SPTE if vCPU1 wrote 4
> > bytes while vCPU0 wrote 8 bytes, though the chances are very low.
> >
> This could happen in below sequence:
> vCPU0 updates a PTE to AABBCCDD;
> vCPU1 updates a PTE to EEFFGGHH in two writes.
> (each character stands for a byte)
>
> vCPU0 vCPU1
> write AABBCCDD
> write GGHH
> detect 4 bytes write and hold on sync
> sync SPTE w/ AABBGGHH
> write EEFF
> sync SPTE w/ EEFFGGHH
>
>
> Do you think it worth below serialization work?
No, because I don't see any KVM bugs with the above sequence. If the guest doesn't
ensure *all* writes from vCPU0 and vCPU1 are fully serialized, then it is completely
legal for hardware (KVM in this case) to consume AABBGGHH as a PTE. The only thing
the guest shouldn't see is EEFFCCDD, but I don't see how that can happen.
On Thu, Aug 10, 2023 at 08:41:14AM -0700, Sean Christopherson wrote:
> On Thu, Aug 10, 2023, Yan Zhao wrote:
> > On Thu, Aug 10, 2023 at 07:21:03AM +0800, Yan Zhao wrote:
> > > > Reading the value after acquiring mmu_lock ensures that both vCPUs will see whatever
> > > > value "loses" the race, i.e. whatever written value is processed second ('Y' in the
> > > > above sequence).
> > > I suspect that vCPU0 may still generate a wrong SPTE if vCPU1 wrote 4
> > > bytes while vCPU0 wrote 8 bytes, though the chances are very low.
> > >
> > This could happen in below sequence:
> > vCPU0 updates a PTE to AABBCCDD;
> > vCPU1 updates a PTE to EEFFGGHH in two writes.
> > (each character stands for a byte)
> >
> > vCPU0 vCPU1
> > write AABBCCDD
> > write GGHH
> > detect 4 bytes write and hold on sync
> > sync SPTE w/ AABBGGHH
> > write EEFF
> > sync SPTE w/ EEFFGGHH
> >
> >
> > Do you think it worth below serialization work?
>
> No, because I don't see any KVM bugs with the above sequence. If the guest doesn't
> ensure *all* writes from vCPU0 and vCPU1 are fully serialized, then it is completely
> legal for hardware (KVM in this case) to consume AABBGGHH as a PTE. The only thing
> the guest shouldn't see is EEFFCCDD, but I don't see how that can happen.
Ok, though still feel it's a little odd when a 1st cmpxch instruction on a GPA is still
under emulation, a 2nd or 3rd... cmpxch instruction to the same GPA may have returned
and they all succeeded :)