1 #define pr_fmt(fmt) "Hyper-V: " fmt
3 #include <linux/hyperv.h>
4 #include <linux/log2.h>
5 #include <linux/slab.h>
6 #include <linux/types.h>
8 #include <asm/fpu/api.h>
9 #include <asm/mshyperv.h>
11 #include <asm/tlbflush.h>
14 #define CREATE_TRACE_POINTS
15 #include <asm/trace/hyperv.h>
17 /* Each gva in gva_list encodes up to 4096 pages to flush */
18 #define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE)
20 static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
21 const struct flush_tlb_info *info);
24 * Fills in gva_list starting from offset. Returns the number of items added.
26 static inline int fill_gva_list(u64 gva_list[], int offset,
27 unsigned long start, unsigned long end)
30 unsigned long cur = start, diff;
33 diff = end > cur ? end - cur : 0;
35 gva_list[gva_n] = cur & PAGE_MASK;
37 * Lower 12 bits encode the number of additional
38 * pages to flush (in addition to the 'cur' page).
40 if (diff >= HV_TLB_FLUSH_UNIT) {
41 gva_list[gva_n] |= ~PAGE_MASK;
42 cur += HV_TLB_FLUSH_UNIT;
44 gva_list[gva_n] |= (diff - 1) >> PAGE_SHIFT;
52 return gva_n - offset;
55 static bool cpu_is_lazy(int cpu)
57 return per_cpu(cpu_tlbstate_shared.is_lazy, cpu);
60 static void hyperv_flush_tlb_multi(const struct cpumask *cpus,
61 const struct flush_tlb_info *info)
63 int cpu, vcpu, gva_n, max_gvas;
64 struct hv_tlb_flush **flush_pcpu;
65 struct hv_tlb_flush *flush;
68 bool do_lazy = !info->freed_tables;
70 trace_hyperv_mmu_flush_tlb_multi(cpus, info);
75 local_irq_save(flags);
77 flush_pcpu = (struct hv_tlb_flush **)
78 this_cpu_ptr(hyperv_pcpu_input_arg);
82 if (unlikely(!flush)) {
83 local_irq_restore(flags);
89 * AddressSpace argument must match the CR3 with PCID bits
92 flush->address_space = virt_to_phys(info->mm->pgd);
93 flush->address_space &= CR3_ADDR_MASK;
96 flush->address_space = 0;
97 flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
100 flush->processor_mask = 0;
101 if (cpumask_equal(cpus, cpu_present_mask)) {
102 flush->flags |= HV_FLUSH_ALL_PROCESSORS;
105 * From the supplied CPU set we need to figure out if we can get
106 * away with cheaper HVCALL_FLUSH_VIRTUAL_ADDRESS_{LIST,SPACE}
107 * hypercalls. This is possible when the highest VP number in
108 * the set is < 64. As VP numbers are usually in ascending order
109 * and match Linux CPU ids, here is an optimization: we check
110 * the VP number for the highest bit in the supplied set first
111 * so we can quickly find out if using *_EX hypercalls is a
112 * must. We will also check all VP numbers when walking the
113 * supplied CPU set to remain correct in all cases.
115 cpu = cpumask_last(cpus);
117 if (cpu < nr_cpumask_bits && hv_cpu_number_to_vp_number(cpu) >= 64)
118 goto do_ex_hypercall;
120 for_each_cpu(cpu, cpus) {
121 if (do_lazy && cpu_is_lazy(cpu))
123 vcpu = hv_cpu_number_to_vp_number(cpu);
124 if (vcpu == VP_INVAL) {
125 local_irq_restore(flags);
130 goto do_ex_hypercall;
132 __set_bit(vcpu, (unsigned long *)
133 &flush->processor_mask);
136 /* nothing to flush if 'processor_mask' ends up being empty */
137 if (!flush->processor_mask) {
138 local_irq_restore(flags);
144 * We can flush not more than max_gvas with one hypercall. Flush the
145 * whole address space if we were asked to do more.
147 max_gvas = (PAGE_SIZE - sizeof(*flush)) / sizeof(flush->gva_list[0]);
149 if (info->end == TLB_FLUSH_ALL) {
150 flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
151 status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
153 } else if (info->end &&
154 ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
155 status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
158 gva_n = fill_gva_list(flush->gva_list, 0,
159 info->start, info->end);
160 status = hv_do_rep_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST,
161 gva_n, 0, flush, NULL);
166 status = hyperv_flush_tlb_others_ex(cpus, info);
169 local_irq_restore(flags);
171 if (hv_result_success(status))
174 native_flush_tlb_multi(cpus, info);
177 static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
178 const struct flush_tlb_info *info)
180 int nr_bank = 0, max_gvas, gva_n;
181 struct hv_tlb_flush_ex **flush_pcpu;
182 struct hv_tlb_flush_ex *flush;
185 if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
186 return HV_STATUS_INVALID_PARAMETER;
188 flush_pcpu = (struct hv_tlb_flush_ex **)
189 this_cpu_ptr(hyperv_pcpu_input_arg);
195 * AddressSpace argument must match the CR3 with PCID bits
198 flush->address_space = virt_to_phys(info->mm->pgd);
199 flush->address_space &= CR3_ADDR_MASK;
202 flush->address_space = 0;
203 flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
206 flush->hv_vp_set.valid_bank_mask = 0;
208 flush->hv_vp_set.format = HV_GENERIC_SET_SPARSE_4K;
209 nr_bank = cpumask_to_vpset_skip(&flush->hv_vp_set, cpus,
210 info->freed_tables ? NULL : cpu_is_lazy);
212 return HV_STATUS_INVALID_PARAMETER;
215 * We can flush not more than max_gvas with one hypercall. Flush the
216 * whole address space if we were asked to do more.
219 (PAGE_SIZE - sizeof(*flush) - nr_bank *
220 sizeof(flush->hv_vp_set.bank_contents[0])) /
221 sizeof(flush->gva_list[0]);
223 if (info->end == TLB_FLUSH_ALL) {
224 flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
225 status = hv_do_rep_hypercall(
226 HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
227 0, nr_bank, flush, NULL);
228 } else if (info->end &&
229 ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
230 status = hv_do_rep_hypercall(
231 HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
232 0, nr_bank, flush, NULL);
234 gva_n = fill_gva_list(flush->gva_list, nr_bank,
235 info->start, info->end);
236 status = hv_do_rep_hypercall(
237 HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX,
238 gva_n, nr_bank, flush, NULL);
244 void hyperv_setup_mmu_ops(void)
246 if (!(ms_hyperv.hints & HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED))
249 pr_info("Using hypercall for remote TLB flush\n");
250 pv_ops.mmu.flush_tlb_multi = hyperv_flush_tlb_multi;
251 pv_ops.mmu.tlb_remove_table = tlb_remove_table;