arch/arm64/kvm/hyp/nvhe/setup.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2020 Google LLC
   4  * Author: Quentin Perret <qperret@google.com>
   5  */
   6
   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
  12 #include <nvhe/early_alloc.h>
  13 #include <nvhe/gfp.h>
  14 #include <nvhe/memory.h>
  15 #include <nvhe/mem_protect.h>
  16 #include <nvhe/mm.h>
  17 #include <nvhe/trap_handler.h>
  18
  19 struct hyp_pool hpool;
  20 unsigned long hyp_nr_cpus;
  21
  22 #define hyp_percpu_size ((unsigned long)__per_cpu_end - \
  23                          (unsigned long)__per_cpu_start)
  24
  25 static void *vmemmap_base;
  26 static void *hyp_pgt_base;
  27 static void *host_s2_pgt_base;
  28 static struct kvm_pgtable_mm_ops pkvm_pgtable_mm_ops;
  29
  30 static int divide_memory_pool(void *virt, unsigned long size)
  31 {
  32         unsigned long vstart, vend, nr_pages;
  33
  34         hyp_early_alloc_init(virt, size);
  35
  36         hyp_vmemmap_range(__hyp_pa(virt), size, &vstart, &vend);
  37         nr_pages = (vend - vstart) >> PAGE_SHIFT;
  38         vmemmap_base = hyp_early_alloc_contig(nr_pages);
  39         if (!vmemmap_base)
  40                 return -ENOMEM;
  41
  42         nr_pages = hyp_s1_pgtable_pages();
  43         hyp_pgt_base = hyp_early_alloc_contig(nr_pages);
  44         if (!hyp_pgt_base)
  45                 return -ENOMEM;
  46
  47         nr_pages = host_s2_pgtable_pages();
  48         host_s2_pgt_base = hyp_early_alloc_contig(nr_pages);
  49         if (!host_s2_pgt_base)
  50                 return -ENOMEM;
  51
  52         return 0;
  53 }
  54
  55 static int recreate_hyp_mappings(phys_addr_t phys, unsigned long size,
  56                                  unsigned long *per_cpu_base,
  57                                  u32 hyp_va_bits)
  58 {
  59         void *start, *end, *virt = hyp_phys_to_virt(phys);
  60         unsigned long pgt_size = hyp_s1_pgtable_pages() << PAGE_SHIFT;
  61         enum kvm_pgtable_prot prot;
  62         int ret, i;
  63
  64         /* Recreate the hyp page-table using the early page allocator */
  65         hyp_early_alloc_init(hyp_pgt_base, pgt_size);
  66         ret = kvm_pgtable_hyp_init(&pkvm_pgtable, hyp_va_bits,
  67                                    &hyp_early_alloc_mm_ops);
  68         if (ret)
  69                 return ret;
  70
  71         ret = hyp_create_idmap(hyp_va_bits);
  72         if (ret)
  73                 return ret;
  74
  75         ret = hyp_map_vectors();
  76         if (ret)
  77                 return ret;
  78
  79         ret = hyp_back_vmemmap(phys, size, hyp_virt_to_phys(vmemmap_base));
  80         if (ret)
  81                 return ret;
  82
  83         ret = pkvm_create_mappings(__hyp_text_start, __hyp_text_end, PAGE_HYP_EXEC);
  84         if (ret)
  85                 return ret;
  86
  87         ret = pkvm_create_mappings(__hyp_rodata_start, __hyp_rodata_end, PAGE_HYP_RO);
  88         if (ret)
  89                 return ret;
  90
  91         ret = pkvm_create_mappings(__hyp_bss_start, __hyp_bss_end, PAGE_HYP);
  92         if (ret)
  93                 return ret;
  94
  95         ret = pkvm_create_mappings(virt, virt + size, PAGE_HYP);
  96         if (ret)
  97                 return ret;
  98
  99         for (i = 0; i < hyp_nr_cpus; i++) {
 100                 start = (void *)kern_hyp_va(per_cpu_base[i]);
 101                 end = start + PAGE_ALIGN(hyp_percpu_size);
 102                 ret = pkvm_create_mappings(start, end, PAGE_HYP);
 103                 if (ret)
 104                         return ret;
 105
 106                 end = (void *)per_cpu_ptr(&kvm_init_params, i)->stack_hyp_va;
 107                 start = end - PAGE_SIZE;
 108                 ret = pkvm_create_mappings(start, end, PAGE_HYP);
 109                 if (ret)
 110                         return ret;
 111         }
 112
 113         /*
 114          * Map the host's .bss and .rodata sections RO in the hypervisor, but
 115          * transfer the ownership from the host to the hypervisor itself to
 116          * make sure it can't be donated or shared with another entity.
 117          *
 118          * The ownership transition requires matching changes in the host
 119          * stage-2. This will be done later (see finalize_host_mappings()) once
 120          * the hyp_vmemmap is addressable.
 121          */
 122         prot = pkvm_mkstate(PAGE_HYP_RO, PKVM_PAGE_SHARED_OWNED);
 123         ret = pkvm_create_mappings(__start_rodata, __end_rodata, prot);
 124         if (ret)
 125                 return ret;
 126
 127         ret = pkvm_create_mappings(__hyp_bss_end, __bss_stop, prot);
 128         if (ret)
 129                 return ret;
 130
 131         return 0;
 132 }
 133
 134 static void update_nvhe_init_params(void)
 135 {
 136         struct kvm_nvhe_init_params *params;
 137         unsigned long i;
 138
 139         for (i = 0; i < hyp_nr_cpus; i++) {
 140                 params = per_cpu_ptr(&kvm_init_params, i);
 141                 params->pgd_pa = __hyp_pa(pkvm_pgtable.pgd);
 142                 dcache_clean_inval_poc((unsigned long)params,
 143                                     (unsigned long)params + sizeof(*params));
 144         }
 145 }
 146
 147 static void *hyp_zalloc_hyp_page(void *arg)
 148 {
 149         return hyp_alloc_pages(&hpool, 0);
 150 }
 151
 152 static void hpool_get_page(void *addr)
 153 {
 154         hyp_get_page(&hpool, addr);
 155 }
 156
 157 static void hpool_put_page(void *addr)
 158 {
 159         hyp_put_page(&hpool, addr);
 160 }
 161
 162 static int finalize_host_mappings_walker(u64 addr, u64 end, u32 level,
 163                                          kvm_pte_t *ptep,
 164                                          enum kvm_pgtable_walk_flags flag,
 165                                          void * const arg)
 166 {
 167         enum kvm_pgtable_prot prot;
 168         enum pkvm_page_state state;
 169         kvm_pte_t pte = *ptep;
 170         phys_addr_t phys;
 171
 172         if (!kvm_pte_valid(pte))
 173                 return 0;
 174
 175         if (level != (KVM_PGTABLE_MAX_LEVELS - 1))
 176                 return -EINVAL;
 177
 178         phys = kvm_pte_to_phys(pte);
 179         if (!addr_is_memory(phys))
 180                 return 0;
 181
 182         /*
 183          * Adjust the host stage-2 mappings to match the ownership attributes
 184          * configured in the hypervisor stage-1.
 185          */
 186         state = pkvm_getstate(kvm_pgtable_hyp_pte_prot(pte));
 187         switch (state) {
 188         case PKVM_PAGE_OWNED:
 189                 return host_stage2_set_owner_locked(phys, PAGE_SIZE, pkvm_hyp_id);
 190         case PKVM_PAGE_SHARED_OWNED:
 191                 prot = pkvm_mkstate(PKVM_HOST_MEM_PROT, PKVM_PAGE_SHARED_BORROWED);
 192                 break;
 193         case PKVM_PAGE_SHARED_BORROWED:
 194                 prot = pkvm_mkstate(PKVM_HOST_MEM_PROT, PKVM_PAGE_SHARED_OWNED);
 195                 break;
 196         default:
 197                 return -EINVAL;
 198         }
 199
 200         return host_stage2_idmap_locked(phys, PAGE_SIZE, prot);
 201 }
 202
 203 static int finalize_host_mappings(void)
 204 {
 205         struct kvm_pgtable_walker walker = {
 206                 .cb     = finalize_host_mappings_walker,
 207                 .flags  = KVM_PGTABLE_WALK_LEAF,
 208         };
 209
 210         return kvm_pgtable_walk(&pkvm_pgtable, 0, BIT(pkvm_pgtable.ia_bits), &walker);
 211 }
 212
 213 void __noreturn __pkvm_init_finalise(void)
 214 {
 215         struct kvm_host_data *host_data = this_cpu_ptr(&kvm_host_data);
 216         struct kvm_cpu_context *host_ctxt = &host_data->host_ctxt;
 217         unsigned long nr_pages, reserved_pages, pfn;
 218         int ret;
 219
 220         /* Now that the vmemmap is backed, install the full-fledged allocator */
 221         pfn = hyp_virt_to_pfn(hyp_pgt_base);
 222         nr_pages = hyp_s1_pgtable_pages();
 223         reserved_pages = hyp_early_alloc_nr_used_pages();
 224         ret = hyp_pool_init(&hpool, pfn, nr_pages, reserved_pages);
 225         if (ret)
 226                 goto out;
 227
 228         ret = kvm_host_prepare_stage2(host_s2_pgt_base);
 229         if (ret)
 230                 goto out;
 231
 232         ret = finalize_host_mappings();
 233         if (ret)
 234                 goto out;
 235
 236         pkvm_pgtable_mm_ops = (struct kvm_pgtable_mm_ops) {
 237                 .zalloc_page = hyp_zalloc_hyp_page,
 238                 .phys_to_virt = hyp_phys_to_virt,
 239                 .virt_to_phys = hyp_virt_to_phys,
 240                 .get_page = hpool_get_page,
 241                 .put_page = hpool_put_page,
 242         };
 243         pkvm_pgtable.mm_ops = &pkvm_pgtable_mm_ops;
 244
 245 out:
 246         /*
 247          * We tail-called to here from handle___pkvm_init() and will not return,
 248          * so make sure to propagate the return value to the host.
 249          */
 250         cpu_reg(host_ctxt, 1) = ret;
 251
 252         __host_enter(host_ctxt);
 253 }
 254
 255 int __pkvm_init(phys_addr_t phys, unsigned long size, unsigned long nr_cpus,
 256                 unsigned long *per_cpu_base, u32 hyp_va_bits)
 257 {
 258         struct kvm_nvhe_init_params *params;
 259         void *virt = hyp_phys_to_virt(phys);
 260         void (*fn)(phys_addr_t params_pa, void *finalize_fn_va);
 261         int ret;
 262
 263         if (!PAGE_ALIGNED(phys) || !PAGE_ALIGNED(size))
 264                 return -EINVAL;
 265
 266         hyp_spin_lock_init(&pkvm_pgd_lock);
 267         hyp_nr_cpus = nr_cpus;
 268
 269         ret = divide_memory_pool(virt, size);
 270         if (ret)
 271                 return ret;
 272
 273         ret = recreate_hyp_mappings(phys, size, per_cpu_base, hyp_va_bits);
 274         if (ret)
 275                 return ret;
 276
 277         update_nvhe_init_params();
 278
 279         /* Jump in the idmap page to switch to the new page-tables */
 280         params = this_cpu_ptr(&kvm_init_params);
 281         fn = (typeof(fn))__hyp_pa(__pkvm_init_switch_pgd);
 282         fn(__hyp_pa(params), __pkvm_init_finalise);
 283
 284         unreachable();
 285 }