arch/x86/include/asm/mshyperv.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _ASM_X86_MSHYPER_H
   3 #define _ASM_X86_MSHYPER_H
   4
   5 #include <linux/types.h>
   6 #include <linux/atomic.h>
   7 #include <linux/nmi.h>
   8 #include <asm/io.h>
   9 #include <asm/hyperv-tlfs.h>
  10 #include <asm/nospec-branch.h>
  11
  12 struct ms_hyperv_info {
  13         u32 features;
  14         u32 misc_features;
  15         u32 hints;
  16         u32 nested_features;
  17         u32 max_vp_index;
  18         u32 max_lp_index;
  19 };
  20
  21 extern struct ms_hyperv_info ms_hyperv;
  22
  23
  24 /*
  25  * Generate the guest ID.
  26  */
  27
  28 static inline  __u64 generate_guest_id(__u64 d_info1, __u64 kernel_version,
  29                                        __u64 d_info2)
  30 {
  31         __u64 guest_id = 0;
  32
  33         guest_id = (((__u64)HV_LINUX_VENDOR_ID) << 48);
  34         guest_id |= (d_info1 << 48);
  35         guest_id |= (kernel_version << 16);
  36         guest_id |= d_info2;
  37
  38         return guest_id;
  39 }
  40
  41
  42 /* Free the message slot and signal end-of-message if required */
  43 static inline void vmbus_signal_eom(struct hv_message *msg, u32 old_msg_type)
  44 {
  45         /*
  46          * On crash we're reading some other CPU's message page and we need
  47          * to be careful: this other CPU may already had cleared the header
  48          * and the host may already had delivered some other message there.
  49          * In case we blindly write msg->header.message_type we're going
  50          * to lose it. We can still lose a message of the same type but
  51          * we count on the fact that there can only be one
  52          * CHANNELMSG_UNLOAD_RESPONSE and we don't care about other messages
  53          * on crash.
  54          */
  55         if (cmpxchg(&msg->header.message_type, old_msg_type,
  56                     HVMSG_NONE) != old_msg_type)
  57                 return;
  58
  59         /*
  60          * Make sure the write to MessageType (ie set to
  61          * HVMSG_NONE) happens before we read the
  62          * MessagePending and EOMing. Otherwise, the EOMing
  63          * will not deliver any more messages since there is
  64          * no empty slot
  65          */
  66         mb();
  67
  68         if (msg->header.message_flags.msg_pending) {
  69                 /*
  70                  * This will cause message queue rescan to
  71                  * possibly deliver another msg from the
  72                  * hypervisor
  73                  */
  74                 wrmsrl(HV_X64_MSR_EOM, 0);
  75         }
  76 }
  77
  78 #define hv_init_timer(timer, tick) wrmsrl(timer, tick)
  79 #define hv_init_timer_config(config, val) wrmsrl(config, val)
  80
  81 #define hv_get_simp(val) rdmsrl(HV_X64_MSR_SIMP, val)
  82 #define hv_set_simp(val) wrmsrl(HV_X64_MSR_SIMP, val)
  83
  84 #define hv_get_siefp(val) rdmsrl(HV_X64_MSR_SIEFP, val)
  85 #define hv_set_siefp(val) wrmsrl(HV_X64_MSR_SIEFP, val)
  86
  87 #define hv_get_synic_state(val) rdmsrl(HV_X64_MSR_SCONTROL, val)
  88 #define hv_set_synic_state(val) wrmsrl(HV_X64_MSR_SCONTROL, val)
  89
  90 #define hv_get_vp_index(index) rdmsrl(HV_X64_MSR_VP_INDEX, index)
  91
  92 #define hv_get_synint_state(int_num, val) rdmsrl(int_num, val)
  93 #define hv_set_synint_state(int_num, val) wrmsrl(int_num, val)
  94
  95 void hyperv_callback_vector(void);
  96 void hyperv_reenlightenment_vector(void);
  97 #ifdef CONFIG_TRACING
  98 #define trace_hyperv_callback_vector hyperv_callback_vector
  99 #endif
 100 void hyperv_vector_handler(struct pt_regs *regs);
 101 void hv_setup_vmbus_irq(void (*handler)(void));
 102 void hv_remove_vmbus_irq(void);
 103
 104 void hv_setup_kexec_handler(void (*handler)(void));
 105 void hv_remove_kexec_handler(void);
 106 void hv_setup_crash_handler(void (*handler)(struct pt_regs *regs));
 107 void hv_remove_crash_handler(void);
 108
 109 /*
 110  * Routines for stimer0 Direct Mode handling.
 111  * On x86/x64, there are no percpu actions to take.
 112  */
 113 void hv_stimer0_vector_handler(struct pt_regs *regs);
 114 void hv_stimer0_callback_vector(void);
 115 int hv_setup_stimer0_irq(int *irq, int *vector, void (*handler)(void));
 116 void hv_remove_stimer0_irq(int irq);
 117
 118 static inline void hv_enable_stimer0_percpu_irq(int irq) {}
 119 static inline void hv_disable_stimer0_percpu_irq(int irq) {}
 120
 121
 122 #if IS_ENABLED(CONFIG_HYPERV)
 123 extern struct clocksource *hyperv_cs;
 124 extern void *hv_hypercall_pg;
 125
 126 static inline u64 hv_do_hypercall(u64 control, void *input, void *output)
 127 {
 128         u64 input_address = input ? virt_to_phys(input) : 0;
 129         u64 output_address = output ? virt_to_phys(output) : 0;
 130         u64 hv_status;
 131
 132 #ifdef CONFIG_X86_64
 133         if (!hv_hypercall_pg)
 134                 return U64_MAX;
 135
 136         __asm__ __volatile__("mov %4, %%r8\n"
 137                              CALL_NOSPEC
 138                              : "=a" (hv_status), ASM_CALL_CONSTRAINT,
 139                                "+c" (control), "+d" (input_address)
 140                              :  "r" (output_address),
 141                                 THUNK_TARGET(hv_hypercall_pg)
 142                              : "cc", "memory", "r8", "r9", "r10", "r11");
 143 #else
 144         u32 input_address_hi = upper_32_bits(input_address);
 145         u32 input_address_lo = lower_32_bits(input_address);
 146         u32 output_address_hi = upper_32_bits(output_address);
 147         u32 output_address_lo = lower_32_bits(output_address);
 148
 149         if (!hv_hypercall_pg)
 150                 return U64_MAX;
 151
 152         __asm__ __volatile__(CALL_NOSPEC
 153                              : "=A" (hv_status),
 154                                "+c" (input_address_lo), ASM_CALL_CONSTRAINT
 155                              : "A" (control),
 156                                "b" (input_address_hi),
 157                                "D"(output_address_hi), "S"(output_address_lo),
 158                                THUNK_TARGET(hv_hypercall_pg)
 159                              : "cc", "memory");
 160 #endif /* !x86_64 */
 161         return hv_status;
 162 }
 163
 164 /* Fast hypercall with 8 bytes of input and no output */
 165 static inline u64 hv_do_fast_hypercall8(u16 code, u64 input1)
 166 {
 167         u64 hv_status, control = (u64)code | HV_HYPERCALL_FAST_BIT;
 168
 169 #ifdef CONFIG_X86_64
 170         {
 171                 __asm__ __volatile__(CALL_NOSPEC
 172                                      : "=a" (hv_status), ASM_CALL_CONSTRAINT,
 173                                        "+c" (control), "+d" (input1)
 174                                      : THUNK_TARGET(hv_hypercall_pg)
 175                                      : "cc", "r8", "r9", "r10", "r11");
 176         }
 177 #else
 178         {
 179                 u32 input1_hi = upper_32_bits(input1);
 180                 u32 input1_lo = lower_32_bits(input1);
 181
 182                 __asm__ __volatile__ (CALL_NOSPEC
 183                                       : "=A"(hv_status),
 184                                         "+c"(input1_lo),
 185                                         ASM_CALL_CONSTRAINT
 186                                       : "A" (control),
 187                                         "b" (input1_hi),
 188                                         THUNK_TARGET(hv_hypercall_pg)
 189                                       : "cc", "edi", "esi");
 190         }
 191 #endif
 192                 return hv_status;
 193 }
 194
 195 /*
 196  * Rep hypercalls. Callers of this functions are supposed to ensure that
 197  * rep_count and varhead_size comply with Hyper-V hypercall definition.
 198  */
 199 static inline u64 hv_do_rep_hypercall(u16 code, u16 rep_count, u16 varhead_size,
 200                                       void *input, void *output)
 201 {
 202         u64 control = code;
 203         u64 status;
 204         u16 rep_comp;
 205
 206         control |= (u64)varhead_size << HV_HYPERCALL_VARHEAD_OFFSET;
 207         control |= (u64)rep_count << HV_HYPERCALL_REP_COMP_OFFSET;
 208
 209         do {
 210                 status = hv_do_hypercall(control, input, output);
 211                 if ((status & HV_HYPERCALL_RESULT_MASK) != HV_STATUS_SUCCESS)
 212                         return status;
 213
 214                 /* Bits 32-43 of status have 'Reps completed' data. */
 215                 rep_comp = (status & HV_HYPERCALL_REP_COMP_MASK) >>
 216                         HV_HYPERCALL_REP_COMP_OFFSET;
 217
 218                 control &= ~HV_HYPERCALL_REP_START_MASK;
 219                 control |= (u64)rep_comp << HV_HYPERCALL_REP_START_OFFSET;
 220
 221                 touch_nmi_watchdog();
 222         } while (rep_comp < rep_count);
 223
 224         return status;
 225 }
 226
 227 /*
 228  * Hypervisor's notion of virtual processor ID is different from
 229  * Linux' notion of CPU ID. This information can only be retrieved
 230  * in the context of the calling CPU. Setup a map for easy access
 231  * to this information.
 232  */
 233 extern u32 *hv_vp_index;
 234 extern u32 hv_max_vp_index;
 235 extern struct hv_vp_assist_page **hv_vp_assist_page;
 236
 237 static inline struct hv_vp_assist_page *hv_get_vp_assist_page(unsigned int cpu)
 238 {
 239         if (!hv_vp_assist_page)
 240                 return NULL;
 241
 242         return hv_vp_assist_page[cpu];
 243 }
 244
 245 /**
 246  * hv_cpu_number_to_vp_number() - Map CPU to VP.
 247  * @cpu_number: CPU number in Linux terms
 248  *
 249  * This function returns the mapping between the Linux processor
 250  * number and the hypervisor's virtual processor number, useful
 251  * in making hypercalls and such that talk about specific
 252  * processors.
 253  *
 254  * Return: Virtual processor number in Hyper-V terms
 255  */
 256 static inline int hv_cpu_number_to_vp_number(int cpu_number)
 257 {
 258         return hv_vp_index[cpu_number];
 259 }
 260
 261 void hyperv_init(void);
 262 void hyperv_setup_mmu_ops(void);
 263 void hyper_alloc_mmu(void);
 264 void hyperv_report_panic(struct pt_regs *regs, long err);
 265 bool hv_is_hyperv_initialized(void);
 266 void hyperv_cleanup(void);
 267
 268 void hyperv_reenlightenment_intr(struct pt_regs *regs);
 269 void set_hv_tscchange_cb(void (*cb)(void));
 270 void clear_hv_tscchange_cb(void);
 271 void hyperv_stop_tsc_emulation(void);
 272 #else /* CONFIG_HYPERV */
 273 static inline void hyperv_init(void) {}
 274 static inline bool hv_is_hyperv_initialized(void) { return false; }
 275 static inline void hyperv_cleanup(void) {}
 276 static inline void hyperv_setup_mmu_ops(void) {}
 277 static inline void set_hv_tscchange_cb(void (*cb)(void)) {}
 278 static inline void clear_hv_tscchange_cb(void) {}
 279 static inline void hyperv_stop_tsc_emulation(void) {};
 280 static inline struct hv_vp_assist_page *hv_get_vp_assist_page(unsigned int cpu)
 281 {
 282         return NULL;
 283 }
 284 #endif /* CONFIG_HYPERV */
 285
 286 #ifdef CONFIG_HYPERV_TSCPAGE
 287 struct ms_hyperv_tsc_page *hv_get_tsc_page(void);
 288 static inline u64 hv_read_tsc_page_tsc(const struct ms_hyperv_tsc_page *tsc_pg,
 289                                        u64 *cur_tsc)
 290 {
 291         u64 scale, offset;
 292         u32 sequence;
 293
 294         /*
 295          * The protocol for reading Hyper-V TSC page is specified in Hypervisor
 296          * Top-Level Functional Specification ver. 3.0 and above. To get the
 297          * reference time we must do the following:
 298          * - READ ReferenceTscSequence
 299          *   A special '0' value indicates the time source is unreliable and we
 300          *   need to use something else. The currently published specification
 301          *   versions (up to 4.0b) contain a mistake and wrongly claim '-1'
 302          *   instead of '0' as the special value, see commit c35b82ef0294.
 303          * - ReferenceTime =
 304          *        ((RDTSC() * ReferenceTscScale) >> 64) + ReferenceTscOffset
 305          * - READ ReferenceTscSequence again. In case its value has changed
 306          *   since our first reading we need to discard ReferenceTime and repeat
 307          *   the whole sequence as the hypervisor was updating the page in
 308          *   between.
 309          */
 310         do {
 311                 sequence = READ_ONCE(tsc_pg->tsc_sequence);
 312                 if (!sequence)
 313                         return U64_MAX;
 314                 /*
 315                  * Make sure we read sequence before we read other values from
 316                  * TSC page.
 317                  */
 318                 smp_rmb();
 319
 320                 scale = READ_ONCE(tsc_pg->tsc_scale);
 321                 offset = READ_ONCE(tsc_pg->tsc_offset);
 322                 *cur_tsc = rdtsc_ordered();
 323
 324                 /*
 325                  * Make sure we read sequence after we read all other values
 326                  * from TSC page.
 327                  */
 328                 smp_rmb();
 329
 330         } while (READ_ONCE(tsc_pg->tsc_sequence) != sequence);
 331
 332         return mul_u64_u64_shr(*cur_tsc, scale, 64) + offset;
 333 }
 334
 335 static inline u64 hv_read_tsc_page(const struct ms_hyperv_tsc_page *tsc_pg)
 336 {
 337         u64 cur_tsc;
 338
 339         return hv_read_tsc_page_tsc(tsc_pg, &cur_tsc);
 340 }
 341
 342 #else
 343 static inline struct ms_hyperv_tsc_page *hv_get_tsc_page(void)
 344 {
 345         return NULL;
 346 }
 347
 348 static inline u64 hv_read_tsc_page_tsc(const struct ms_hyperv_tsc_page *tsc_pg,
 349                                        u64 *cur_tsc)
 350 {
 351         BUG();
 352         return U64_MAX;
 353 }
 354 #endif
 355 #endif