diff mbox series

[RFC,4/8] kvm: x86: support walking guest page tables in gmem

Message ID 20240709132041.3625501-5-roypat@amazon.co.uk (mailing list archive)
State New
Headers show
Series Unmapping guest_memfd from Direct Map | expand

Commit Message

Patrick Roy July 9, 2024, 1:20 p.m. UTC
Update the logic in paging_tmpl.h to work with guest_private memory. If
KVM cannot access gmem and the guest's page tables are in gfns marked as
private, then error out.

Let the guest page table walker access gmem by making it use
gfn_to_pfn_caches, which are already gmem aware, and will later also
handle on-demand mapping of gmem once it supports being removed from the
direct map.  We re-use the gfn_to_pfn_cache here to avoid implementing
yet another remapping solution to support the cmpxchg used to set the
"accessed" bit on guest PTEs. The only case that now needs some special
handling is page tables in read-only memslots, as gfn_to_pfn_caches
cannot be used for readonly memory. In this case, use
kvm_vcpu_read_guest (which is also gmem aware), as there is no need to
cache the gfn->pfn translation in this case (there is no need to do a
cmpxchg on the PTE as the walker does not set the accessed bit for
read-only ptes).

Signed-off-by: Patrick Roy <roypat@amazon.co.uk>
---
 arch/x86/kvm/mmu/paging_tmpl.h | 94 ++++++++++++++++++++++++++++------
 1 file changed, 77 insertions(+), 17 deletions(-)
diff mbox series

Patch

diff --git a/arch/x86/kvm/mmu/paging_tmpl.h b/arch/x86/kvm/mmu/paging_tmpl.h
index 69941cebb3a8..ddf3b4bd479e 100644
--- a/arch/x86/kvm/mmu/paging_tmpl.h
+++ b/arch/x86/kvm/mmu/paging_tmpl.h
@@ -84,7 +84,7 @@  struct guest_walker {
 	pt_element_t ptes[PT_MAX_FULL_LEVELS];
 	pt_element_t prefetch_ptes[PTE_PREFETCH_NUM];
 	gpa_t pte_gpa[PT_MAX_FULL_LEVELS];
-	pt_element_t __user *ptep_user[PT_MAX_FULL_LEVELS];
+	struct gfn_to_pfn_cache ptep_caches[PT_MAX_FULL_LEVELS];
 	bool pte_writable[PT_MAX_FULL_LEVELS];
 	unsigned int pt_access[PT_MAX_FULL_LEVELS];
 	unsigned int pte_access;
@@ -201,7 +201,7 @@  static int FNAME(update_accessed_dirty_bits)(struct kvm_vcpu *vcpu,
 {
 	unsigned level, index;
 	pt_element_t pte, orig_pte;
-	pt_element_t __user *ptep_user;
+	struct gfn_to_pfn_cache *pte_cache;
 	gfn_t table_gfn;
 	int ret;
 
@@ -210,10 +210,12 @@  static int FNAME(update_accessed_dirty_bits)(struct kvm_vcpu *vcpu,
 		return 0;
 
 	for (level = walker->max_level; level >= walker->level; --level) {
+		unsigned long flags;
+
 		pte = orig_pte = walker->ptes[level - 1];
 		table_gfn = walker->table_gfn[level - 1];
-		ptep_user = walker->ptep_user[level - 1];
-		index = offset_in_page(ptep_user) / sizeof(pt_element_t);
+		pte_cache = &walker->ptep_caches[level - 1];
+		index = offset_in_page(pte_cache->khva) / sizeof(pt_element_t);
 		if (!(pte & PT_GUEST_ACCESSED_MASK)) {
 			trace_kvm_mmu_set_accessed_bit(table_gfn, index, sizeof(pte));
 			pte |= PT_GUEST_ACCESSED_MASK;
@@ -246,11 +248,26 @@  static int FNAME(update_accessed_dirty_bits)(struct kvm_vcpu *vcpu,
 		if (unlikely(!walker->pte_writable[level - 1]))
 			continue;
 
-		ret = __try_cmpxchg_user(ptep_user, &orig_pte, pte, fault);
+		read_lock_irqsave(&pte_cache->lock, flags);
+		while (!kvm_gpc_check(pte_cache, sizeof(pte))) {
+			read_unlock_irqrestore(&pte_cache->lock, flags);
+
+			ret = kvm_gpc_refresh(pte_cache, sizeof(pte));
+			if (ret)
+				return ret;
+
+			read_lock_irqsave(&pte_cache->lock, flags);
+		}
+		ret = __try_cmpxchg((pt_element_t *)pte_cache->khva, &orig_pte, pte, sizeof(pte));
+
+		if (!ret)
+			kvm_gpc_mark_dirty_in_slot(pte_cache);
+
+		read_unlock_irqrestore(&pte_cache->lock, flags);
+
 		if (ret)
 			return ret;
 
-		kvm_vcpu_mark_page_dirty(vcpu, table_gfn);
 		walker->ptes[level - 1] = pte;
 	}
 	return 0;
@@ -296,6 +313,12 @@  static inline bool FNAME(is_last_gpte)(struct kvm_mmu *mmu,
 
 	return gpte & PT_PAGE_SIZE_MASK;
 }
+
+static void FNAME(walk_deactivate_gpcs)(struct guest_walker *walker) {
+	for (unsigned int level = 0; level < PT_MAX_FULL_LEVELS; ++level)
+		kvm_gpc_deactivate(&walker->ptep_caches[level]);
+}
+
 /*
  * Fetch a guest pte for a guest virtual address, or for an L2's GPA.
  */
@@ -305,7 +328,6 @@  static int FNAME(walk_addr_generic)(struct guest_walker *walker,
 {
 	int ret;
 	pt_element_t pte;
-	pt_element_t __user *ptep_user;
 	gfn_t table_gfn;
 	u64 pt_access, pte_access;
 	unsigned index, accessed_dirty, pte_pkey;
@@ -320,8 +342,17 @@  static int FNAME(walk_addr_generic)(struct guest_walker *walker,
 	u16 errcode = 0;
 	gpa_t real_gpa;
 	gfn_t gfn;
+	struct gfn_to_pfn_cache *pte_cache;
 
 	trace_kvm_mmu_pagetable_walk(addr, access);
+
+	for (unsigned int level = 0; level < PT_MAX_FULL_LEVELS; ++level) {
+		pte_cache = &walker->ptep_caches[level];
+
+		memset(pte_cache, 0, sizeof(*pte_cache));
+		kvm_gpc_init(pte_cache, vcpu->kvm);
+	}
+
 retry_walk:
 	walker->level = mmu->cpu_role.base.level;
 	pte           = kvm_mmu_get_guest_pgd(vcpu, mmu);
@@ -362,11 +393,13 @@  static int FNAME(walk_addr_generic)(struct guest_walker *walker,
 
 	do {
 		struct kvm_memory_slot *slot;
-		unsigned long host_addr;
+		unsigned long flags;
 
 		pt_access = pte_access;
 		--walker->level;
 
+		pte_cache = &walker->ptep_caches[walker->level - 1];
+
 		index = PT_INDEX(addr, walker->level);
 		table_gfn = gpte_to_gfn(pte);
 		offset    = index * sizeof(pt_element_t);
@@ -396,15 +429,36 @@  static int FNAME(walk_addr_generic)(struct guest_walker *walker,
 		if (!kvm_is_visible_memslot(slot))
 			goto error;
 
-		host_addr = gfn_to_hva_memslot_prot(slot, gpa_to_gfn(real_gpa),
-					    &walker->pte_writable[walker->level - 1]);
-		if (unlikely(kvm_is_error_hva(host_addr)))
-			goto error;
+		/*
+		 * gfn_to_pfn_cache expects the memory to be writable. However,
+		 * if the memory is not writable, we do not need caching in the
+		 * first place, as we only need it to later potentially write
+		 * the access bit (which we cannot do anyway if the memory is
+		 * readonly).
+		 */
+		if (slot->flags & KVM_MEM_READONLY) {
+			if (kvm_vcpu_read_guest(vcpu, real_gpa + offset, &pte, sizeof(pte)))
+				goto error;
+		} else {
+			if (kvm_gpc_activate(pte_cache, real_gpa + offset,
+					     sizeof(pte)))
+				goto error;
 
-		ptep_user = (pt_element_t __user *)((void *)host_addr + offset);
-		if (unlikely(__get_user(pte, ptep_user)))
-			goto error;
-		walker->ptep_user[walker->level - 1] = ptep_user;
+			read_lock_irqsave(&pte_cache->lock, flags);
+			while (!kvm_gpc_check(pte_cache, sizeof(pte))) {
+				read_unlock_irqrestore(&pte_cache->lock, flags);
+
+				if (kvm_gpc_refresh(pte_cache, sizeof(pte)))
+					goto error;
+
+				read_lock_irqsave(&pte_cache->lock, flags);
+			}
+
+			pte = *(pt_element_t *)pte_cache->khva;
+			read_unlock_irqrestore(&pte_cache->lock, flags);
+
+			walker->pte_writable[walker->level - 1] = true;
+		}
 
 		trace_kvm_mmu_paging_element(pte, walker->level);
 
@@ -467,13 +521,19 @@  static int FNAME(walk_addr_generic)(struct guest_walker *walker,
 							addr, write_fault);
 		if (unlikely(ret < 0))
 			goto error;
-		else if (ret)
+		else if (ret) {
+			FNAME(walk_deactivate_gpcs)(walker);
 			goto retry_walk;
+		}
 	}
 
+	FNAME(walk_deactivate_gpcs)(walker);
+
 	return 1;
 
 error:
+	FNAME(walk_deactivate_gpcs)(walker);
+
 	errcode |= write_fault | user_fault;
 	if (fetch_fault && (is_efer_nx(mmu) || is_cr4_smep(mmu)))
 		errcode |= PFERR_FETCH_MASK;