arch/x86/hyperv/mmu.c

   1 #define pr_fmt(fmt)  "Hyper-V: " fmt
   2
   3 #include <linux/hyperv.h>
   4 #include <linux/log2.h>
   5 #include <linux/slab.h>
   6 #include <linux/types.h>
   7
   8 #include <asm/fpu/api.h>
   9 #include <asm/mshyperv.h>
  10 #include <asm/msr.h>
  11 #include <asm/tlbflush.h>
  12 #include <asm/tlb.h>
  13
  14 #define CREATE_TRACE_POINTS
  15 #include <asm/trace/hyperv.h>
  16
  17 /* Each gva in gva_list encodes up to 4096 pages to flush */
  18 #define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE)
  19
  20 static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
  21                                       const struct flush_tlb_info *info);
  22
  23 /*
  24  * Fills in gva_list starting from offset. Returns the number of items added.
  25  */
  26 static inline int fill_gva_list(u64 gva_list[], int offset,
  27                                 unsigned long start, unsigned long end)
  28 {
  29         int gva_n = offset;
  30         unsigned long cur = start, diff;
  31
  32         do {
  33                 diff = end > cur ? end - cur : 0;
  34
  35                 gva_list[gva_n] = cur & PAGE_MASK;
  36                 /*
  37                  * Lower 12 bits encode the number of additional
  38                  * pages to flush (in addition to the 'cur' page).
  39                  */
  40                 if (diff >= HV_TLB_FLUSH_UNIT) {
  41                         gva_list[gva_n] |= ~PAGE_MASK;
  42                         cur += HV_TLB_FLUSH_UNIT;
  43                 }  else if (diff) {
  44                         gva_list[gva_n] |= (diff - 1) >> PAGE_SHIFT;
  45                         cur = end;
  46                 }
  47
  48                 gva_n++;
  49
  50         } while (cur < end);
  51
  52         return gva_n - offset;
  53 }
  54
  55 static void hyperv_flush_tlb_others(const struct cpumask *cpus,
  56                                     const struct flush_tlb_info *info)
  57 {
  58         int cpu, vcpu, gva_n, max_gvas;
  59         struct hv_tlb_flush **flush_pcpu;
  60         struct hv_tlb_flush *flush;
  61         u64 status = U64_MAX;
  62         unsigned long flags;
  63
  64         trace_hyperv_mmu_flush_tlb_others(cpus, info);
  65
  66         if (!hv_hypercall_pg)
  67                 goto do_native;
  68
  69         local_irq_save(flags);
  70
  71         /*
  72          * Only check the mask _after_ interrupt has been disabled to avoid the
  73          * mask changing under our feet.
  74          */
  75         if (cpumask_empty(cpus)) {
  76                 local_irq_restore(flags);
  77                 return;
  78         }
  79
  80         flush_pcpu = (struct hv_tlb_flush **)
  81                      this_cpu_ptr(hyperv_pcpu_input_arg);
  82
  83         flush = *flush_pcpu;
  84
  85         if (unlikely(!flush)) {
  86                 local_irq_restore(flags);
  87                 goto do_native;
  88         }
  89
  90         if (info->mm) {
  91                 /*
  92                  * AddressSpace argument must match the CR3 with PCID bits
  93                  * stripped out.
  94                  */
  95                 flush->address_space = virt_to_phys(info->mm->pgd);
  96                 flush->address_space &= CR3_ADDR_MASK;
  97                 flush->flags = 0;
  98         } else {
  99                 flush->address_space = 0;
 100                 flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
 101         }
 102
 103         flush->processor_mask = 0;
 104         if (cpumask_equal(cpus, cpu_present_mask)) {
 105                 flush->flags |= HV_FLUSH_ALL_PROCESSORS;
 106         } else {
 107                 /*
 108                  * From the supplied CPU set we need to figure out if we can get
 109                  * away with cheaper HVCALL_FLUSH_VIRTUAL_ADDRESS_{LIST,SPACE}
 110                  * hypercalls. This is possible when the highest VP number in
 111                  * the set is < 64. As VP numbers are usually in ascending order
 112                  * and match Linux CPU ids, here is an optimization: we check
 113                  * the VP number for the highest bit in the supplied set first
 114                  * so we can quickly find out if using *_EX hypercalls is a
 115                  * must. We will also check all VP numbers when walking the
 116                  * supplied CPU set to remain correct in all cases.
 117                  */
 118                 if (hv_cpu_number_to_vp_number(cpumask_last(cpus)) >= 64)
 119                         goto do_ex_hypercall;
 120
 121                 for_each_cpu(cpu, cpus) {
 122                         vcpu = hv_cpu_number_to_vp_number(cpu);
 123                         if (vcpu == VP_INVAL) {
 124                                 local_irq_restore(flags);
 125                                 goto do_native;
 126                         }
 127
 128                         if (vcpu >= 64)
 129                                 goto do_ex_hypercall;
 130
 131                         __set_bit(vcpu, (unsigned long *)
 132                                   &flush->processor_mask);
 133                 }
 134         }
 135
 136         /*
 137          * We can flush not more than max_gvas with one hypercall. Flush the
 138          * whole address space if we were asked to do more.
 139          */
 140         max_gvas = (PAGE_SIZE - sizeof(*flush)) / sizeof(flush->gva_list[0]);
 141
 142         if (info->end == TLB_FLUSH_ALL) {
 143                 flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
 144                 status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
 145                                          flush, NULL);
 146         } else if (info->end &&
 147                    ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
 148                 status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
 149                                          flush, NULL);
 150         } else {
 151                 gva_n = fill_gva_list(flush->gva_list, 0,
 152                                       info->start, info->end);
 153                 status = hv_do_rep_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST,
 154                                              gva_n, 0, flush, NULL);
 155         }
 156         goto check_status;
 157
 158 do_ex_hypercall:
 159         status = hyperv_flush_tlb_others_ex(cpus, info);
 160
 161 check_status:
 162         local_irq_restore(flags);
 163
 164         if (!(status & HV_HYPERCALL_RESULT_MASK))
 165                 return;
 166 do_native:
 167         native_flush_tlb_others(cpus, info);
 168 }
 169
 170 static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
 171                                       const struct flush_tlb_info *info)
 172 {
 173         int nr_bank = 0, max_gvas, gva_n;
 174         struct hv_tlb_flush_ex **flush_pcpu;
 175         struct hv_tlb_flush_ex *flush;
 176         u64 status;
 177
 178         if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
 179                 return U64_MAX;
 180
 181         flush_pcpu = (struct hv_tlb_flush_ex **)
 182                      this_cpu_ptr(hyperv_pcpu_input_arg);
 183
 184         flush = *flush_pcpu;
 185
 186         if (info->mm) {
 187                 /*
 188                  * AddressSpace argument must match the CR3 with PCID bits
 189                  * stripped out.
 190                  */
 191                 flush->address_space = virt_to_phys(info->mm->pgd);
 192                 flush->address_space &= CR3_ADDR_MASK;
 193                 flush->flags = 0;
 194         } else {
 195                 flush->address_space = 0;
 196                 flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
 197         }
 198
 199         flush->hv_vp_set.valid_bank_mask = 0;
 200
 201         flush->hv_vp_set.format = HV_GENERIC_SET_SPARSE_4K;
 202         nr_bank = cpumask_to_vpset(&(flush->hv_vp_set), cpus);
 203         if (nr_bank < 0)
 204                 return U64_MAX;
 205
 206         /*
 207          * We can flush not more than max_gvas with one hypercall. Flush the
 208          * whole address space if we were asked to do more.
 209          */
 210         max_gvas =
 211                 (PAGE_SIZE - sizeof(*flush) - nr_bank *
 212                  sizeof(flush->hv_vp_set.bank_contents[0])) /
 213                 sizeof(flush->gva_list[0]);
 214
 215         if (info->end == TLB_FLUSH_ALL) {
 216                 flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
 217                 status = hv_do_rep_hypercall(
 218                         HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
 219                         0, nr_bank, flush, NULL);
 220         } else if (info->end &&
 221                    ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
 222                 status = hv_do_rep_hypercall(
 223                         HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
 224                         0, nr_bank, flush, NULL);
 225         } else {
 226                 gva_n = fill_gva_list(flush->gva_list, nr_bank,
 227                                       info->start, info->end);
 228                 status = hv_do_rep_hypercall(
 229                         HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX,
 230                         gva_n, nr_bank, flush, NULL);
 231         }
 232
 233         return status;
 234 }
 235
 236 void hyperv_setup_mmu_ops(void)
 237 {
 238         if (!(ms_hyperv.hints & HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED))
 239                 return;
 240
 241         pr_info("Using hypercall for remote TLB flush\n");
 242         pv_ops.mmu.flush_tlb_others = hyperv_flush_tlb_others;
 243         pv_ops.mmu.tlb_remove_table = tlb_remove_table;
 244 }