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>
13 #define CREATE_TRACE_POINTS
14 #include <asm/trace/hyperv.h>
16 /* HvFlushVirtualAddressSpace, HvFlushVirtualAddressList hypercalls */
17 struct hv_flush_pcpu
{
24 /* HvFlushVirtualAddressSpaceEx, HvFlushVirtualAddressListEx hypercalls */
25 struct hv_flush_pcpu_ex
{
36 /* Each gva in gva_list encodes up to 4096 pages to flush */
37 #define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE)
39 static struct hv_flush_pcpu __percpu
*pcpu_flush
;
41 static struct hv_flush_pcpu_ex __percpu
*pcpu_flush_ex
;
44 * Fills in gva_list starting from offset. Returns the number of items added.
46 static inline int fill_gva_list(u64 gva_list
[], int offset
,
47 unsigned long start
, unsigned long end
)
50 unsigned long cur
= start
, diff
;
53 diff
= end
> cur
? end
- cur
: 0;
55 gva_list
[gva_n
] = cur
& PAGE_MASK
;
57 * Lower 12 bits encode the number of additional
58 * pages to flush (in addition to the 'cur' page).
60 if (diff
>= HV_TLB_FLUSH_UNIT
)
61 gva_list
[gva_n
] |= ~PAGE_MASK
;
63 gva_list
[gva_n
] |= (diff
- 1) >> PAGE_SHIFT
;
65 cur
+= HV_TLB_FLUSH_UNIT
;
70 return gva_n
- offset
;
73 /* Return the number of banks in the resulting vp_set */
74 static inline int cpumask_to_vp_set(struct hv_flush_pcpu_ex
*flush
,
75 const struct cpumask
*cpus
)
77 int cpu
, vcpu
, vcpu_bank
, vcpu_offset
, nr_bank
= 1;
79 /* valid_bank_mask can represent up to 64 banks */
80 if (hv_max_vp_index
/ 64 >= 64)
84 * Clear all banks up to the maximum possible bank as hv_flush_pcpu_ex
85 * structs are not cleared between calls, we risk flushing unneeded
88 for (vcpu_bank
= 0; vcpu_bank
<= hv_max_vp_index
/ 64; vcpu_bank
++)
89 flush
->hv_vp_set
.bank_contents
[vcpu_bank
] = 0;
92 * Some banks may end up being empty but this is acceptable.
94 for_each_cpu(cpu
, cpus
) {
95 vcpu
= hv_cpu_number_to_vp_number(cpu
);
96 vcpu_bank
= vcpu
/ 64;
97 vcpu_offset
= vcpu
% 64;
98 __set_bit(vcpu_offset
, (unsigned long *)
99 &flush
->hv_vp_set
.bank_contents
[vcpu_bank
]);
100 if (vcpu_bank
>= nr_bank
)
101 nr_bank
= vcpu_bank
+ 1;
103 flush
->hv_vp_set
.valid_bank_mask
= GENMASK_ULL(nr_bank
- 1, 0);
108 static void hyperv_flush_tlb_others(const struct cpumask
*cpus
,
109 const struct flush_tlb_info
*info
)
111 int cpu
, vcpu
, gva_n
, max_gvas
;
112 struct hv_flush_pcpu
*flush
;
113 u64 status
= U64_MAX
;
116 trace_hyperv_mmu_flush_tlb_others(cpus
, info
);
118 if (!pcpu_flush
|| !hv_hypercall_pg
)
121 if (cpumask_empty(cpus
))
124 local_irq_save(flags
);
126 flush
= this_cpu_ptr(pcpu_flush
);
129 flush
->address_space
= virt_to_phys(info
->mm
->pgd
);
132 flush
->address_space
= 0;
133 flush
->flags
= HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES
;
136 flush
->processor_mask
= 0;
137 if (cpumask_equal(cpus
, cpu_present_mask
)) {
138 flush
->flags
|= HV_FLUSH_ALL_PROCESSORS
;
140 for_each_cpu(cpu
, cpus
) {
141 vcpu
= hv_cpu_number_to_vp_number(cpu
);
145 __set_bit(vcpu
, (unsigned long *)
146 &flush
->processor_mask
);
151 * We can flush not more than max_gvas with one hypercall. Flush the
152 * whole address space if we were asked to do more.
154 max_gvas
= (PAGE_SIZE
- sizeof(*flush
)) / sizeof(flush
->gva_list
[0]);
156 if (info
->end
== TLB_FLUSH_ALL
) {
157 flush
->flags
|= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY
;
158 status
= hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE
,
160 } else if (info
->end
&&
161 ((info
->end
- info
->start
)/HV_TLB_FLUSH_UNIT
) > max_gvas
) {
162 status
= hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE
,
165 gva_n
= fill_gva_list(flush
->gva_list
, 0,
166 info
->start
, info
->end
);
167 status
= hv_do_rep_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST
,
168 gva_n
, 0, flush
, NULL
);
171 local_irq_restore(flags
);
173 if (!(status
& HV_HYPERCALL_RESULT_MASK
))
176 native_flush_tlb_others(cpus
, info
);
179 static void hyperv_flush_tlb_others_ex(const struct cpumask
*cpus
,
180 const struct flush_tlb_info
*info
)
182 int nr_bank
= 0, max_gvas
, gva_n
;
183 struct hv_flush_pcpu_ex
*flush
;
184 u64 status
= U64_MAX
;
187 trace_hyperv_mmu_flush_tlb_others(cpus
, info
);
189 if (!pcpu_flush_ex
|| !hv_hypercall_pg
)
192 if (cpumask_empty(cpus
))
195 local_irq_save(flags
);
197 flush
= this_cpu_ptr(pcpu_flush_ex
);
200 flush
->address_space
= virt_to_phys(info
->mm
->pgd
);
203 flush
->address_space
= 0;
204 flush
->flags
= HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES
;
207 flush
->hv_vp_set
.valid_bank_mask
= 0;
209 if (!cpumask_equal(cpus
, cpu_present_mask
)) {
210 flush
->hv_vp_set
.format
= HV_GENERIC_SET_SPARCE_4K
;
211 nr_bank
= cpumask_to_vp_set(flush
, cpus
);
215 flush
->hv_vp_set
.format
= HV_GENERIC_SET_ALL
;
216 flush
->flags
|= HV_FLUSH_ALL_PROCESSORS
;
220 * We can flush not more than max_gvas with one hypercall. Flush the
221 * whole address space if we were asked to do more.
224 (PAGE_SIZE
- sizeof(*flush
) - nr_bank
*
225 sizeof(flush
->hv_vp_set
.bank_contents
[0])) /
226 sizeof(flush
->gva_list
[0]);
228 if (info
->end
== TLB_FLUSH_ALL
) {
229 flush
->flags
|= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY
;
230 status
= hv_do_rep_hypercall(
231 HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX
,
232 0, nr_bank
+ 2, flush
, NULL
);
233 } else if (info
->end
&&
234 ((info
->end
- info
->start
)/HV_TLB_FLUSH_UNIT
) > max_gvas
) {
235 status
= hv_do_rep_hypercall(
236 HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX
,
237 0, nr_bank
+ 2, flush
, NULL
);
239 gva_n
= fill_gva_list(flush
->gva_list
, nr_bank
,
240 info
->start
, info
->end
);
241 status
= hv_do_rep_hypercall(
242 HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX
,
243 gva_n
, nr_bank
+ 2, flush
, NULL
);
246 local_irq_restore(flags
);
248 if (!(status
& HV_HYPERCALL_RESULT_MASK
))
251 native_flush_tlb_others(cpus
, info
);
254 void hyperv_setup_mmu_ops(void)
256 if (!(ms_hyperv
.hints
& HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED
))
259 setup_clear_cpu_cap(X86_FEATURE_PCID
);
261 if (!(ms_hyperv
.hints
& HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED
)) {
262 pr_info("Using hypercall for remote TLB flush\n");
263 pv_mmu_ops
.flush_tlb_others
= hyperv_flush_tlb_others
;
265 pr_info("Using ext hypercall for remote TLB flush\n");
266 pv_mmu_ops
.flush_tlb_others
= hyperv_flush_tlb_others_ex
;
270 void hyper_alloc_mmu(void)
272 if (!(ms_hyperv
.hints
& HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED
))
275 if (!(ms_hyperv
.hints
& HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED
))
276 pcpu_flush
= __alloc_percpu(PAGE_SIZE
, PAGE_SIZE
);
278 pcpu_flush_ex
= __alloc_percpu(PAGE_SIZE
, PAGE_SIZE
);