kernel/drivers/hv/hv.c

   1 /*
   2  * Copyright (c) 2009, Microsoft Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
  15  * Place - Suite 330, Boston, MA 02111-1307 USA.
  16  *
  17  * Authors:
  18  *   Haiyang Zhang <haiyangz@microsoft.com>
  19  *   Hank Janssen  <hjanssen@microsoft.com>
  20  *
  21  */
  22 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  23
  24 #include <linux/kernel.h>
  25 #include <linux/mm.h>
  26 #include <linux/slab.h>
  27 #include <linux/vmalloc.h>
  28 #include <linux/hyperv.h>
  29 #include <linux/version.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/clockchips.h>
  32 #include <asm/hyperv.h>
  33 #include <asm/mshyperv.h>
  34 #include "hyperv_vmbus.h"
  35
  36 /* The one and only */
  37 struct hv_context hv_context = {
  38         .synic_initialized      = false,
  39         .hypercall_page         = NULL,
  40 };
  41
  42 #define HV_TIMER_FREQUENCY (10 * 1000 * 1000) /* 100ns period */
  43 #define HV_MAX_MAX_DELTA_TICKS 0xffffffff
  44 #define HV_MIN_DELTA_TICKS 1
  45
  46 /*
  47  * query_hypervisor_info - Get version info of the windows hypervisor
  48  */
  49 unsigned int host_info_eax;
  50 unsigned int host_info_ebx;
  51 unsigned int host_info_ecx;
  52 unsigned int host_info_edx;
  53
  54 static int query_hypervisor_info(void)
  55 {
  56         unsigned int eax;
  57         unsigned int ebx;
  58         unsigned int ecx;
  59         unsigned int edx;
  60         unsigned int max_leaf;
  61         unsigned int op;
  62
  63         /*
  64         * Its assumed that this is called after confirming that Viridian
  65         * is present. Query id and revision.
  66         */
  67         eax = 0;
  68         ebx = 0;
  69         ecx = 0;
  70         edx = 0;
  71         op = HVCPUID_VENDOR_MAXFUNCTION;
  72         cpuid(op, &eax, &ebx, &ecx, &edx);
  73
  74         max_leaf = eax;
  75
  76         if (max_leaf >= HVCPUID_VERSION) {
  77                 eax = 0;
  78                 ebx = 0;
  79                 ecx = 0;
  80                 edx = 0;
  81                 op = HVCPUID_VERSION;
  82                 cpuid(op, &eax, &ebx, &ecx, &edx);
  83                 host_info_eax = eax;
  84                 host_info_ebx = ebx;
  85                 host_info_ecx = ecx;
  86                 host_info_edx = edx;
  87         }
  88         return max_leaf;
  89 }
  90
  91 /*
  92  * do_hypercall- Invoke the specified hypercall
  93  */
  94 static u64 do_hypercall(u64 control, void *input, void *output)
  95 {
  96         u64 input_address = (input) ? virt_to_phys(input) : 0;
  97         u64 output_address = (output) ? virt_to_phys(output) : 0;
  98         void *hypercall_page = hv_context.hypercall_page;
  99 #ifdef CONFIG_X86_64
 100         u64 hv_status = 0;
 101
 102         if (!hypercall_page)
 103                 return (u64)ULLONG_MAX;
 104
 105         __asm__ __volatile__("mov %0, %%r8" : : "r" (output_address) : "r8");
 106         __asm__ __volatile__("call *%3" : "=a" (hv_status) :
 107                              "c" (control), "d" (input_address),
 108                              "m" (hypercall_page));
 109
 110         return hv_status;
 111
 112 #else
 113
 114         u32 control_hi = control >> 32;
 115         u32 control_lo = control & 0xFFFFFFFF;
 116         u32 hv_status_hi = 1;
 117         u32 hv_status_lo = 1;
 118         u32 input_address_hi = input_address >> 32;
 119         u32 input_address_lo = input_address & 0xFFFFFFFF;
 120         u32 output_address_hi = output_address >> 32;
 121         u32 output_address_lo = output_address & 0xFFFFFFFF;
 122
 123         if (!hypercall_page)
 124                 return (u64)ULLONG_MAX;
 125
 126         __asm__ __volatile__ ("call *%8" : "=d"(hv_status_hi),
 127                               "=a"(hv_status_lo) : "d" (control_hi),
 128                               "a" (control_lo), "b" (input_address_hi),
 129                               "c" (input_address_lo), "D"(output_address_hi),
 130                               "S"(output_address_lo), "m" (hypercall_page));
 131
 132         return hv_status_lo | ((u64)hv_status_hi << 32);
 133 #endif /* !x86_64 */
 134 }
 135
 136 #ifdef CONFIG_X86_64
 137 static cycle_t read_hv_clock_tsc(struct clocksource *arg)
 138 {
 139         cycle_t current_tick;
 140         struct ms_hyperv_tsc_page *tsc_pg = hv_context.tsc_page;
 141
 142         if (tsc_pg->tsc_sequence != -1) {
 143                 /*
 144                  * Use the tsc page to compute the value.
 145                  */
 146
 147                 while (1) {
 148                         cycle_t tmp;
 149                         u32 sequence = tsc_pg->tsc_sequence;
 150                         u64 cur_tsc;
 151                         u64 scale = tsc_pg->tsc_scale;
 152                         s64 offset = tsc_pg->tsc_offset;
 153
 154                         rdtscll(cur_tsc);
 155                         /* current_tick = ((cur_tsc *scale) >> 64) + offset */
 156                         asm("mulq %3"
 157                                 : "=d" (current_tick), "=a" (tmp)
 158                                 : "a" (cur_tsc), "r" (scale));
 159
 160                         current_tick += offset;
 161                         if (tsc_pg->tsc_sequence == sequence)
 162                                 return current_tick;
 163
 164                         if (tsc_pg->tsc_sequence != -1)
 165                                 continue;
 166                         /*
 167                          * Fallback using MSR method.
 168                          */
 169                         break;
 170                 }
 171         }
 172         rdmsrl(HV_X64_MSR_TIME_REF_COUNT, current_tick);
 173         return current_tick;
 174 }
 175
 176 static struct clocksource hyperv_cs_tsc = {
 177                 .name           = "hyperv_clocksource_tsc_page",
 178                 .rating         = 425,
 179                 .read           = read_hv_clock_tsc,
 180                 .mask           = CLOCKSOURCE_MASK(64),
 181                 .flags          = CLOCK_SOURCE_IS_CONTINUOUS,
 182 };
 183 #endif
 184
 185
 186 /*
 187  * hv_init - Main initialization routine.
 188  *
 189  * This routine must be called before any other routines in here are called
 190  */
 191 int hv_init(void)
 192 {
 193         int max_leaf;
 194         union hv_x64_msr_hypercall_contents hypercall_msr;
 195         union hv_x64_msr_hypercall_contents tsc_msr;
 196         void *virtaddr = NULL;
 197         void *va_tsc = NULL;
 198
 199         memset(hv_context.synic_event_page, 0, sizeof(void *) * NR_CPUS);
 200         memset(hv_context.synic_message_page, 0,
 201                sizeof(void *) * NR_CPUS);
 202         memset(hv_context.post_msg_page, 0,
 203                sizeof(void *) * NR_CPUS);
 204         memset(hv_context.vp_index, 0,
 205                sizeof(int) * NR_CPUS);
 206         memset(hv_context.event_dpc, 0,
 207                sizeof(void *) * NR_CPUS);
 208         memset(hv_context.clk_evt, 0,
 209                sizeof(void *) * NR_CPUS);
 210
 211         max_leaf = query_hypervisor_info();
 212
 213         /*
 214          * Write our OS ID.
 215          */
 216         hv_context.guestid = generate_guest_id(0, LINUX_VERSION_CODE, 0);
 217         wrmsrl(HV_X64_MSR_GUEST_OS_ID, hv_context.guestid);
 218
 219         /* See if the hypercall page is already set */
 220         rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 221
 222         virtaddr = __vmalloc(PAGE_SIZE, GFP_KERNEL, PAGE_KERNEL_EXEC);
 223
 224         if (!virtaddr)
 225                 goto cleanup;
 226
 227         hypercall_msr.enable = 1;
 228
 229         hypercall_msr.guest_physical_address = vmalloc_to_pfn(virtaddr);
 230         wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 231
 232         /* Confirm that hypercall page did get setup. */
 233         hypercall_msr.as_uint64 = 0;
 234         rdmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 235
 236         if (!hypercall_msr.enable)
 237                 goto cleanup;
 238
 239         hv_context.hypercall_page = virtaddr;
 240
 241 #ifdef CONFIG_X86_64
 242         if (ms_hyperv.features & HV_X64_MSR_REFERENCE_TSC_AVAILABLE) {
 243                 va_tsc = __vmalloc(PAGE_SIZE, GFP_KERNEL, PAGE_KERNEL);
 244                 if (!va_tsc)
 245                         goto cleanup;
 246                 hv_context.tsc_page = va_tsc;
 247
 248                 rdmsrl(HV_X64_MSR_REFERENCE_TSC, tsc_msr.as_uint64);
 249
 250                 tsc_msr.enable = 1;
 251                 tsc_msr.guest_physical_address = vmalloc_to_pfn(va_tsc);
 252
 253                 wrmsrl(HV_X64_MSR_REFERENCE_TSC, tsc_msr.as_uint64);
 254                 clocksource_register_hz(&hyperv_cs_tsc, NSEC_PER_SEC/100);
 255         }
 256 #endif
 257         return 0;
 258
 259 cleanup:
 260         if (virtaddr) {
 261                 if (hypercall_msr.enable) {
 262                         hypercall_msr.as_uint64 = 0;
 263                         wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 264                 }
 265
 266                 vfree(virtaddr);
 267         }
 268
 269         return -ENOTSUPP;
 270 }
 271
 272 /*
 273  * hv_cleanup - Cleanup routine.
 274  *
 275  * This routine is called normally during driver unloading or exiting.
 276  */
 277 void hv_cleanup(void)
 278 {
 279         union hv_x64_msr_hypercall_contents hypercall_msr;
 280
 281         /* Reset our OS id */
 282         wrmsrl(HV_X64_MSR_GUEST_OS_ID, 0);
 283
 284         if (hv_context.hypercall_page) {
 285                 hypercall_msr.as_uint64 = 0;
 286                 wrmsrl(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64);
 287                 vfree(hv_context.hypercall_page);
 288                 hv_context.hypercall_page = NULL;
 289         }
 290
 291 #ifdef CONFIG_X86_64
 292         /*
 293          * Cleanup the TSC page based CS.
 294          */
 295         if (ms_hyperv.features & HV_X64_MSR_REFERENCE_TSC_AVAILABLE) {
 296                 clocksource_change_rating(&hyperv_cs_tsc, 10);
 297                 clocksource_unregister(&hyperv_cs_tsc);
 298
 299                 hypercall_msr.as_uint64 = 0;
 300                 wrmsrl(HV_X64_MSR_REFERENCE_TSC, hypercall_msr.as_uint64);
 301                 vfree(hv_context.tsc_page);
 302                 hv_context.tsc_page = NULL;
 303         }
 304 #endif
 305 }
 306
 307 /*
 308  * hv_post_message - Post a message using the hypervisor message IPC.
 309  *
 310  * This involves a hypercall.
 311  */
 312 int hv_post_message(union hv_connection_id connection_id,
 313                   enum hv_message_type message_type,
 314                   void *payload, size_t payload_size)
 315 {
 316
 317         struct hv_input_post_message *aligned_msg;
 318         u16 status;
 319
 320         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT)
 321                 return -EMSGSIZE;
 322
 323         aligned_msg = (struct hv_input_post_message *)
 324                         hv_context.post_msg_page[get_cpu()];
 325
 326         aligned_msg->connectionid = connection_id;
 327         aligned_msg->reserved = 0;
 328         aligned_msg->message_type = message_type;
 329         aligned_msg->payload_size = payload_size;
 330         memcpy((void *)aligned_msg->payload, payload, payload_size);
 331
 332         status = do_hypercall(HVCALL_POST_MESSAGE, aligned_msg, NULL)
 333                 & 0xFFFF;
 334
 335         put_cpu();
 336         return status;
 337 }
 338
 339
 340 /*
 341  * hv_signal_event -
 342  * Signal an event on the specified connection using the hypervisor event IPC.
 343  *
 344  * This involves a hypercall.
 345  */
 346 u16 hv_signal_event(void *con_id)
 347 {
 348         u16 status;
 349
 350         status = (do_hypercall(HVCALL_SIGNAL_EVENT, con_id, NULL) & 0xFFFF);
 351
 352         return status;
 353 }
 354
 355 static int hv_ce_set_next_event(unsigned long delta,
 356                                 struct clock_event_device *evt)
 357 {
 358         cycle_t current_tick;
 359
 360         WARN_ON(!clockevent_state_oneshot(evt));
 361
 362         rdmsrl(HV_X64_MSR_TIME_REF_COUNT, current_tick);
 363         current_tick += delta;
 364         wrmsrl(HV_X64_MSR_STIMER0_COUNT, current_tick);
 365         return 0;
 366 }
 367
 368 static int hv_ce_shutdown(struct clock_event_device *evt)
 369 {
 370         wrmsrl(HV_X64_MSR_STIMER0_COUNT, 0);
 371         wrmsrl(HV_X64_MSR_STIMER0_CONFIG, 0);
 372
 373         return 0;
 374 }
 375
 376 static int hv_ce_set_oneshot(struct clock_event_device *evt)
 377 {
 378         union hv_timer_config timer_cfg;
 379
 380         timer_cfg.enable = 1;
 381         timer_cfg.auto_enable = 1;
 382         timer_cfg.sintx = VMBUS_MESSAGE_SINT;
 383         wrmsrl(HV_X64_MSR_STIMER0_CONFIG, timer_cfg.as_uint64);
 384
 385         return 0;
 386 }
 387
 388 static void hv_init_clockevent_device(struct clock_event_device *dev, int cpu)
 389 {
 390         dev->name = "Hyper-V clockevent";
 391         dev->features = CLOCK_EVT_FEAT_ONESHOT;
 392         dev->cpumask = cpumask_of(cpu);
 393         dev->rating = 1000;
 394         /*
 395          * Avoid settint dev->owner = THIS_MODULE deliberately as doing so will
 396          * result in clockevents_config_and_register() taking additional
 397          * references to the hv_vmbus module making it impossible to unload.
 398          */
 399
 400         dev->set_state_shutdown = hv_ce_shutdown;
 401         dev->set_state_oneshot = hv_ce_set_oneshot;
 402         dev->set_next_event = hv_ce_set_next_event;
 403 }
 404
 405
 406 int hv_synic_alloc(void)
 407 {
 408         size_t size = sizeof(struct tasklet_struct);
 409         size_t ced_size = sizeof(struct clock_event_device);
 410         int cpu;
 411
 412         hv_context.hv_numa_map = kzalloc(sizeof(struct cpumask) * nr_node_ids,
 413                                          GFP_ATOMIC);
 414         if (hv_context.hv_numa_map == NULL) {
 415                 pr_err("Unable to allocate NUMA map\n");
 416                 goto err;
 417         }
 418
 419         for_each_online_cpu(cpu) {
 420                 hv_context.event_dpc[cpu] = kmalloc(size, GFP_ATOMIC);
 421                 if (hv_context.event_dpc[cpu] == NULL) {
 422                         pr_err("Unable to allocate event dpc\n");
 423                         goto err;
 424                 }
 425                 tasklet_init(hv_context.event_dpc[cpu], vmbus_on_event, cpu);
 426
 427                 hv_context.clk_evt[cpu] = kzalloc(ced_size, GFP_ATOMIC);
 428                 if (hv_context.clk_evt[cpu] == NULL) {
 429                         pr_err("Unable to allocate clock event device\n");
 430                         goto err;
 431                 }
 432
 433                 hv_init_clockevent_device(hv_context.clk_evt[cpu], cpu);
 434
 435                 hv_context.synic_message_page[cpu] =
 436                         (void *)get_zeroed_page(GFP_ATOMIC);
 437
 438                 if (hv_context.synic_message_page[cpu] == NULL) {
 439                         pr_err("Unable to allocate SYNIC message page\n");
 440                         goto err;
 441                 }
 442
 443                 hv_context.synic_event_page[cpu] =
 444                         (void *)get_zeroed_page(GFP_ATOMIC);
 445
 446                 if (hv_context.synic_event_page[cpu] == NULL) {
 447                         pr_err("Unable to allocate SYNIC event page\n");
 448                         goto err;
 449                 }
 450
 451                 hv_context.post_msg_page[cpu] =
 452                         (void *)get_zeroed_page(GFP_ATOMIC);
 453
 454                 if (hv_context.post_msg_page[cpu] == NULL) {
 455                         pr_err("Unable to allocate post msg page\n");
 456                         goto err;
 457                 }
 458         }
 459
 460         return 0;
 461 err:
 462         return -ENOMEM;
 463 }
 464
 465 static void hv_synic_free_cpu(int cpu)
 466 {
 467         kfree(hv_context.event_dpc[cpu]);
 468         kfree(hv_context.clk_evt[cpu]);
 469         if (hv_context.synic_event_page[cpu])
 470                 free_page((unsigned long)hv_context.synic_event_page[cpu]);
 471         if (hv_context.synic_message_page[cpu])
 472                 free_page((unsigned long)hv_context.synic_message_page[cpu]);
 473         if (hv_context.post_msg_page[cpu])
 474                 free_page((unsigned long)hv_context.post_msg_page[cpu]);
 475 }
 476
 477 void hv_synic_free(void)
 478 {
 479         int cpu;
 480
 481         kfree(hv_context.hv_numa_map);
 482         for_each_online_cpu(cpu)
 483                 hv_synic_free_cpu(cpu);
 484 }
 485
 486 /*
 487  * hv_synic_init - Initialize the Synthethic Interrupt Controller.
 488  *
 489  * If it is already initialized by another entity (ie x2v shim), we need to
 490  * retrieve the initialized message and event pages.  Otherwise, we create and
 491  * initialize the message and event pages.
 492  */
 493 void hv_synic_init(void *arg)
 494 {
 495         u64 version;
 496         union hv_synic_simp simp;
 497         union hv_synic_siefp siefp;
 498         union hv_synic_sint shared_sint;
 499         union hv_synic_scontrol sctrl;
 500         u64 vp_index;
 501
 502         int cpu = smp_processor_id();
 503
 504         if (!hv_context.hypercall_page)
 505                 return;
 506
 507         /* Check the version */
 508         rdmsrl(HV_X64_MSR_SVERSION, version);
 509
 510         /* Setup the Synic's message page */
 511         rdmsrl(HV_X64_MSR_SIMP, simp.as_uint64);
 512         simp.simp_enabled = 1;
 513         simp.base_simp_gpa = virt_to_phys(hv_context.synic_message_page[cpu])
 514                 >> PAGE_SHIFT;
 515
 516         wrmsrl(HV_X64_MSR_SIMP, simp.as_uint64);
 517
 518         /* Setup the Synic's event page */
 519         rdmsrl(HV_X64_MSR_SIEFP, siefp.as_uint64);
 520         siefp.siefp_enabled = 1;
 521         siefp.base_siefp_gpa = virt_to_phys(hv_context.synic_event_page[cpu])
 522                 >> PAGE_SHIFT;
 523
 524         wrmsrl(HV_X64_MSR_SIEFP, siefp.as_uint64);
 525
 526         /* Setup the shared SINT. */
 527         rdmsrl(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
 528
 529         shared_sint.as_uint64 = 0;
 530         shared_sint.vector = HYPERVISOR_CALLBACK_VECTOR;
 531         shared_sint.masked = false;
 532         shared_sint.auto_eoi = true;
 533
 534         wrmsrl(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
 535
 536         /* Enable the global synic bit */
 537         rdmsrl(HV_X64_MSR_SCONTROL, sctrl.as_uint64);
 538         sctrl.enable = 1;
 539
 540         wrmsrl(HV_X64_MSR_SCONTROL, sctrl.as_uint64);
 541
 542         hv_context.synic_initialized = true;
 543
 544         /*
 545          * Setup the mapping between Hyper-V's notion
 546          * of cpuid and Linux' notion of cpuid.
 547          * This array will be indexed using Linux cpuid.
 548          */
 549         rdmsrl(HV_X64_MSR_VP_INDEX, vp_index);
 550         hv_context.vp_index[cpu] = (u32)vp_index;
 551
 552         INIT_LIST_HEAD(&hv_context.percpu_list[cpu]);
 553
 554         /*
 555          * Register the per-cpu clockevent source.
 556          */
 557         if (ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE)
 558                 clockevents_config_and_register(hv_context.clk_evt[cpu],
 559                                                 HV_TIMER_FREQUENCY,
 560                                                 HV_MIN_DELTA_TICKS,
 561                                                 HV_MAX_MAX_DELTA_TICKS);
 562         return;
 563 }
 564
 565 /*
 566  * hv_synic_clockevents_cleanup - Cleanup clockevent devices
 567  */
 568 void hv_synic_clockevents_cleanup(void)
 569 {
 570         int cpu;
 571
 572         if (!(ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE))
 573                 return;
 574
 575         for_each_online_cpu(cpu)
 576                 clockevents_unbind_device(hv_context.clk_evt[cpu], cpu);
 577 }
 578
 579 /*
 580  * hv_synic_cleanup - Cleanup routine for hv_synic_init().
 581  */
 582 void hv_synic_cleanup(void *arg)
 583 {
 584         union hv_synic_sint shared_sint;
 585         union hv_synic_simp simp;
 586         union hv_synic_siefp siefp;
 587         union hv_synic_scontrol sctrl;
 588         int cpu = smp_processor_id();
 589
 590         if (!hv_context.synic_initialized)
 591                 return;
 592
 593         /* Turn off clockevent device */
 594         if (ms_hyperv.features & HV_X64_MSR_SYNTIMER_AVAILABLE)
 595                 hv_ce_shutdown(hv_context.clk_evt[cpu]);
 596
 597         rdmsrl(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
 598
 599         shared_sint.masked = 1;
 600
 601         /* Need to correctly cleanup in the case of SMP!!! */
 602         /* Disable the interrupt */
 603         wrmsrl(HV_X64_MSR_SINT0 + VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
 604
 605         rdmsrl(HV_X64_MSR_SIMP, simp.as_uint64);
 606         simp.simp_enabled = 0;
 607         simp.base_simp_gpa = 0;
 608
 609         wrmsrl(HV_X64_MSR_SIMP, simp.as_uint64);
 610
 611         rdmsrl(HV_X64_MSR_SIEFP, siefp.as_uint64);
 612         siefp.siefp_enabled = 0;
 613         siefp.base_siefp_gpa = 0;
 614
 615         wrmsrl(HV_X64_MSR_SIEFP, siefp.as_uint64);
 616
 617         /* Disable the global synic bit */
 618         rdmsrl(HV_X64_MSR_SCONTROL, sctrl.as_uint64);
 619         sctrl.enable = 0;
 620         wrmsrl(HV_X64_MSR_SCONTROL, sctrl.as_uint64);
 621 }