1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2020 Google LLC
4 * Author: Quentin Perret <qperret@google.com>
7 #include <linux/kvm_host.h>
8 #include <asm/kvm_hyp.h>
9 #include <asm/kvm_mmu.h>
10 #include <asm/kvm_pgtable.h>
11 #include <asm/kvm_pkvm.h>
12 #include <asm/spectre.h>
14 #include <nvhe/early_alloc.h>
16 #include <nvhe/memory.h>
17 #include <nvhe/mem_protect.h>
19 #include <nvhe/spinlock.h>
21 struct kvm_pgtable pkvm_pgtable;
22 hyp_spinlock_t pkvm_pgd_lock;
24 struct memblock_region hyp_memory[HYP_MEMBLOCK_REGIONS];
25 unsigned int hyp_memblock_nr;
27 static u64 __io_map_base;
29 struct hyp_fixmap_slot {
33 static DEFINE_PER_CPU(struct hyp_fixmap_slot, fixmap_slots);
35 static int __pkvm_create_mappings(unsigned long start, unsigned long size,
36 unsigned long phys, enum kvm_pgtable_prot prot)
40 hyp_spin_lock(&pkvm_pgd_lock);
41 err = kvm_pgtable_hyp_map(&pkvm_pgtable, start, size, phys, prot);
42 hyp_spin_unlock(&pkvm_pgd_lock);
48 * pkvm_alloc_private_va_range - Allocates a private VA range.
49 * @size: The size of the VA range to reserve.
50 * @haddr: The hypervisor virtual start address of the allocation.
52 * The private virtual address (VA) range is allocated above __io_map_base
53 * and aligned based on the order of @size.
55 * Return: 0 on success or negative error code on failure.
57 int pkvm_alloc_private_va_range(size_t size, unsigned long *haddr)
59 unsigned long base, addr;
62 hyp_spin_lock(&pkvm_pgd_lock);
64 /* Align the allocation based on the order of its size */
65 addr = ALIGN(__io_map_base, PAGE_SIZE << get_order(size));
67 /* The allocated size is always a multiple of PAGE_SIZE */
68 base = addr + PAGE_ALIGN(size);
70 /* Are we overflowing on the vmemmap ? */
71 if (!addr || base > __hyp_vmemmap)
78 hyp_spin_unlock(&pkvm_pgd_lock);
83 int __pkvm_create_private_mapping(phys_addr_t phys, size_t size,
84 enum kvm_pgtable_prot prot,
90 size = PAGE_ALIGN(size + offset_in_page(phys));
91 err = pkvm_alloc_private_va_range(size, &addr);
95 err = __pkvm_create_mappings(addr, size, phys, prot);
99 *haddr = addr + offset_in_page(phys);
103 int pkvm_create_mappings_locked(void *from, void *to, enum kvm_pgtable_prot prot)
105 unsigned long start = (unsigned long)from;
106 unsigned long end = (unsigned long)to;
107 unsigned long virt_addr;
110 hyp_assert_lock_held(&pkvm_pgd_lock);
112 start = start & PAGE_MASK;
113 end = PAGE_ALIGN(end);
115 for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
118 phys = hyp_virt_to_phys((void *)virt_addr);
119 err = kvm_pgtable_hyp_map(&pkvm_pgtable, virt_addr, PAGE_SIZE,
128 int pkvm_create_mappings(void *from, void *to, enum kvm_pgtable_prot prot)
132 hyp_spin_lock(&pkvm_pgd_lock);
133 ret = pkvm_create_mappings_locked(from, to, prot);
134 hyp_spin_unlock(&pkvm_pgd_lock);
139 int hyp_back_vmemmap(phys_addr_t back)
141 unsigned long i, start, size, end = 0;
144 for (i = 0; i < hyp_memblock_nr; i++) {
145 start = hyp_memory[i].base;
146 start = ALIGN_DOWN((u64)hyp_phys_to_page(start), PAGE_SIZE);
148 * The begining of the hyp_vmemmap region for the current
149 * memblock may already be backed by the page backing the end
150 * the previous region, so avoid mapping it twice.
152 start = max(start, end);
154 end = hyp_memory[i].base + hyp_memory[i].size;
155 end = PAGE_ALIGN((u64)hyp_phys_to_page(end));
160 ret = __pkvm_create_mappings(start, size, back, PAGE_HYP);
164 memset(hyp_phys_to_virt(back), 0, size);
171 static void *__hyp_bp_vect_base;
172 int pkvm_cpu_set_vector(enum arm64_hyp_spectre_vector slot)
177 case HYP_VECTOR_DIRECT: {
178 vector = __kvm_hyp_vector;
181 case HYP_VECTOR_SPECTRE_DIRECT: {
182 vector = __bp_harden_hyp_vecs;
185 case HYP_VECTOR_INDIRECT:
186 case HYP_VECTOR_SPECTRE_INDIRECT: {
187 vector = (void *)__hyp_bp_vect_base;
194 vector = __kvm_vector_slot2addr(vector, slot);
195 *this_cpu_ptr(&kvm_hyp_vector) = (unsigned long)vector;
200 int hyp_map_vectors(void)
203 unsigned long bp_base;
206 if (!kvm_system_needs_idmapped_vectors()) {
207 __hyp_bp_vect_base = __bp_harden_hyp_vecs;
211 phys = __hyp_pa(__bp_harden_hyp_vecs);
212 ret = __pkvm_create_private_mapping(phys, __BP_HARDEN_HYP_VECS_SZ,
213 PAGE_HYP_EXEC, &bp_base);
217 __hyp_bp_vect_base = (void *)bp_base;
222 void *hyp_fixmap_map(phys_addr_t phys)
224 struct hyp_fixmap_slot *slot = this_cpu_ptr(&fixmap_slots);
225 kvm_pte_t pte, *ptep = slot->ptep;
228 pte &= ~kvm_phys_to_pte(KVM_PHYS_INVALID);
229 pte |= kvm_phys_to_pte(phys) | KVM_PTE_VALID;
230 WRITE_ONCE(*ptep, pte);
233 return (void *)slot->addr;
236 static void fixmap_clear_slot(struct hyp_fixmap_slot *slot)
238 kvm_pte_t *ptep = slot->ptep;
239 u64 addr = slot->addr;
241 WRITE_ONCE(*ptep, *ptep & ~KVM_PTE_VALID);
244 * Irritatingly, the architecture requires that we use inner-shareable
245 * broadcast TLB invalidation here in case another CPU speculates
246 * through our fixmap and decides to create an "amalagamation of the
247 * values held in the TLB" due to the apparent lack of a
248 * break-before-make sequence.
250 * https://lore.kernel.org/kvm/20221017115209.2099-1-will@kernel.org/T/#mf10dfbaf1eaef9274c581b81c53758918c1d0f03
253 __tlbi_level(vale2is, __TLBI_VADDR(addr, 0), (KVM_PGTABLE_MAX_LEVELS - 1));
258 void hyp_fixmap_unmap(void)
260 fixmap_clear_slot(this_cpu_ptr(&fixmap_slots));
263 static int __create_fixmap_slot_cb(const struct kvm_pgtable_visit_ctx *ctx,
264 enum kvm_pgtable_walk_flags visit)
266 struct hyp_fixmap_slot *slot = per_cpu_ptr(&fixmap_slots, (u64)ctx->arg);
268 if (!kvm_pte_valid(ctx->old) || ctx->level != KVM_PGTABLE_MAX_LEVELS - 1)
271 slot->addr = ctx->addr;
272 slot->ptep = ctx->ptep;
275 * Clear the PTE, but keep the page-table page refcount elevated to
276 * prevent it from ever being freed. This lets us manipulate the PTEs
277 * by hand safely without ever needing to allocate memory.
279 fixmap_clear_slot(slot);
284 static int create_fixmap_slot(u64 addr, u64 cpu)
286 struct kvm_pgtable_walker walker = {
287 .cb = __create_fixmap_slot_cb,
288 .flags = KVM_PGTABLE_WALK_LEAF,
292 return kvm_pgtable_walk(&pkvm_pgtable, addr, PAGE_SIZE, &walker);
295 int hyp_create_pcpu_fixmap(void)
297 unsigned long addr, i;
300 for (i = 0; i < hyp_nr_cpus; i++) {
301 ret = pkvm_alloc_private_va_range(PAGE_SIZE, &addr);
305 ret = kvm_pgtable_hyp_map(&pkvm_pgtable, addr, PAGE_SIZE,
306 __hyp_pa(__hyp_bss_start), PAGE_HYP);
310 ret = create_fixmap_slot(addr, i);
318 int hyp_create_idmap(u32 hyp_va_bits)
320 unsigned long start, end;
322 start = hyp_virt_to_phys((void *)__hyp_idmap_text_start);
323 start = ALIGN_DOWN(start, PAGE_SIZE);
325 end = hyp_virt_to_phys((void *)__hyp_idmap_text_end);
326 end = ALIGN(end, PAGE_SIZE);
329 * One half of the VA space is reserved to linearly map portions of
330 * memory -- see va_layout.c for more details. The other half of the VA
331 * space contains the trampoline page, and needs some care. Split that
332 * second half in two and find the quarter of VA space not conflicting
333 * with the idmap to place the IOs and the vmemmap. IOs use the lower
334 * half of the quarter and the vmemmap the upper half.
336 __io_map_base = start & BIT(hyp_va_bits - 2);
337 __io_map_base ^= BIT(hyp_va_bits - 2);
338 __hyp_vmemmap = __io_map_base | BIT(hyp_va_bits - 3);
340 return __pkvm_create_mappings(start, end - start, start, PAGE_HYP_EXEC);
343 static void *admit_host_page(void *arg)
345 struct kvm_hyp_memcache *host_mc = arg;
347 if (!host_mc->nr_pages)
351 * The host still owns the pages in its memcache, so we need to go
352 * through a full host-to-hyp donation cycle to change it. Fortunately,
353 * __pkvm_host_donate_hyp() takes care of races for us, so if it
354 * succeeds we're good to go.
356 if (__pkvm_host_donate_hyp(hyp_phys_to_pfn(host_mc->head), 1))
359 return pop_hyp_memcache(host_mc, hyp_phys_to_virt);
362 /* Refill our local memcache by poping pages from the one provided by the host. */
363 int refill_memcache(struct kvm_hyp_memcache *mc, unsigned long min_pages,
364 struct kvm_hyp_memcache *host_mc)
366 struct kvm_hyp_memcache tmp = *host_mc;
369 ret = __topup_hyp_memcache(mc, min_pages, admit_host_page,
370 hyp_virt_to_phys, &tmp);