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  */
fill_gva_list(u64 gva_list[],int offset,unsigned long start,unsigned long end)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 
hyperv_flush_tlb_multi(const struct cpumask * cpus,const struct flush_tlb_info * info)55 static void hyperv_flush_tlb_multi(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;
62 	unsigned long flags;
63 
64 	trace_hyperv_mmu_flush_tlb_multi(cpus, info);
65 
66 	if (!hv_hypercall_pg)
67 		goto do_native;
68 
69 	local_irq_save(flags);
70 
71 	flush_pcpu = (struct hv_tlb_flush **)
72 		     this_cpu_ptr(hyperv_pcpu_input_arg);
73 
74 	flush = *flush_pcpu;
75 
76 	if (unlikely(!flush)) {
77 		local_irq_restore(flags);
78 		goto do_native;
79 	}
80 
81 	if (info->mm) {
82 		/*
83 		 * AddressSpace argument must match the CR3 with PCID bits
84 		 * stripped out.
85 		 */
86 		flush->address_space = virt_to_phys(info->mm->pgd);
87 		flush->address_space &= CR3_ADDR_MASK;
88 		flush->flags = 0;
89 	} else {
90 		flush->address_space = 0;
91 		flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
92 	}
93 
94 	flush->processor_mask = 0;
95 	if (cpumask_equal(cpus, cpu_present_mask)) {
96 		flush->flags |= HV_FLUSH_ALL_PROCESSORS;
97 	} else {
98 		/*
99 		 * From the supplied CPU set we need to figure out if we can get
100 		 * away with cheaper HVCALL_FLUSH_VIRTUAL_ADDRESS_{LIST,SPACE}
101 		 * hypercalls. This is possible when the highest VP number in
102 		 * the set is < 64. As VP numbers are usually in ascending order
103 		 * and match Linux CPU ids, here is an optimization: we check
104 		 * the VP number for the highest bit in the supplied set first
105 		 * so we can quickly find out if using *_EX hypercalls is a
106 		 * must. We will also check all VP numbers when walking the
107 		 * supplied CPU set to remain correct in all cases.
108 		 */
109 		cpu = cpumask_last(cpus);
110 
111 		if (cpu < nr_cpumask_bits && hv_cpu_number_to_vp_number(cpu) >= 64)
112 			goto do_ex_hypercall;
113 
114 		for_each_cpu(cpu, cpus) {
115 			vcpu = hv_cpu_number_to_vp_number(cpu);
116 			if (vcpu == VP_INVAL) {
117 				local_irq_restore(flags);
118 				goto do_native;
119 			}
120 
121 			if (vcpu >= 64)
122 				goto do_ex_hypercall;
123 
124 			__set_bit(vcpu, (unsigned long *)
125 				  &flush->processor_mask);
126 		}
127 
128 		/* nothing to flush if 'processor_mask' ends up being empty */
129 		if (!flush->processor_mask) {
130 			local_irq_restore(flags);
131 			return;
132 		}
133 	}
134 
135 	/*
136 	 * We can flush not more than max_gvas with one hypercall. Flush the
137 	 * whole address space if we were asked to do more.
138 	 */
139 	max_gvas = (PAGE_SIZE - sizeof(*flush)) / sizeof(flush->gva_list[0]);
140 
141 	if (info->end == TLB_FLUSH_ALL) {
142 		flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
143 		status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
144 					 flush, NULL);
145 	} else if (info->end &&
146 		   ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
147 		status = hv_do_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE,
148 					 flush, NULL);
149 	} else {
150 		gva_n = fill_gva_list(flush->gva_list, 0,
151 				      info->start, info->end);
152 		status = hv_do_rep_hypercall(HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST,
153 					     gva_n, 0, flush, NULL);
154 	}
155 	goto check_status;
156 
157 do_ex_hypercall:
158 	status = hyperv_flush_tlb_others_ex(cpus, info);
159 
160 check_status:
161 	local_irq_restore(flags);
162 
163 	if (hv_result_success(status))
164 		return;
165 do_native:
166 	native_flush_tlb_multi(cpus, info);
167 }
168 
hyperv_flush_tlb_others_ex(const struct cpumask * cpus,const struct flush_tlb_info * info)169 static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
170 				      const struct flush_tlb_info *info)
171 {
172 	int nr_bank = 0, max_gvas, gva_n;
173 	struct hv_tlb_flush_ex **flush_pcpu;
174 	struct hv_tlb_flush_ex *flush;
175 	u64 status;
176 
177 	if (!(ms_hyperv.hints & HV_X64_EX_PROCESSOR_MASKS_RECOMMENDED))
178 		return HV_STATUS_INVALID_PARAMETER;
179 
180 	flush_pcpu = (struct hv_tlb_flush_ex **)
181 		     this_cpu_ptr(hyperv_pcpu_input_arg);
182 
183 	flush = *flush_pcpu;
184 
185 	if (info->mm) {
186 		/*
187 		 * AddressSpace argument must match the CR3 with PCID bits
188 		 * stripped out.
189 		 */
190 		flush->address_space = virt_to_phys(info->mm->pgd);
191 		flush->address_space &= CR3_ADDR_MASK;
192 		flush->flags = 0;
193 	} else {
194 		flush->address_space = 0;
195 		flush->flags = HV_FLUSH_ALL_VIRTUAL_ADDRESS_SPACES;
196 	}
197 
198 	flush->hv_vp_set.valid_bank_mask = 0;
199 
200 	flush->hv_vp_set.format = HV_GENERIC_SET_SPARSE_4K;
201 	nr_bank = cpumask_to_vpset(&(flush->hv_vp_set), cpus);
202 	if (nr_bank < 0)
203 		return HV_STATUS_INVALID_PARAMETER;
204 
205 	/*
206 	 * We can flush not more than max_gvas with one hypercall. Flush the
207 	 * whole address space if we were asked to do more.
208 	 */
209 	max_gvas =
210 		(PAGE_SIZE - sizeof(*flush) - nr_bank *
211 		 sizeof(flush->hv_vp_set.bank_contents[0])) /
212 		sizeof(flush->gva_list[0]);
213 
214 	if (info->end == TLB_FLUSH_ALL) {
215 		flush->flags |= HV_FLUSH_NON_GLOBAL_MAPPINGS_ONLY;
216 		status = hv_do_rep_hypercall(
217 			HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
218 			0, nr_bank, flush, NULL);
219 	} else if (info->end &&
220 		   ((info->end - info->start)/HV_TLB_FLUSH_UNIT) > max_gvas) {
221 		status = hv_do_rep_hypercall(
222 			HVCALL_FLUSH_VIRTUAL_ADDRESS_SPACE_EX,
223 			0, nr_bank, flush, NULL);
224 	} else {
225 		gva_n = fill_gva_list(flush->gva_list, nr_bank,
226 				      info->start, info->end);
227 		status = hv_do_rep_hypercall(
228 			HVCALL_FLUSH_VIRTUAL_ADDRESS_LIST_EX,
229 			gva_n, nr_bank, flush, NULL);
230 	}
231 
232 	return status;
233 }
234 
hyperv_setup_mmu_ops(void)235 void hyperv_setup_mmu_ops(void)
236 {
237 	if (!(ms_hyperv.hints & HV_X64_REMOTE_TLB_FLUSH_RECOMMENDED))
238 		return;
239 
240 	pr_info("Using hypercall for remote TLB flush\n");
241 	pv_ops.mmu.flush_tlb_multi = hyperv_flush_tlb_multi;
242 	pv_ops.mmu.tlb_remove_table = tlb_remove_table;
243 }
244