drivers/hv/hv.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  */
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kernel.h>
  12 #include <linux/mm.h>
  13 #include <linux/slab.h>
  14 #include <linux/vmalloc.h>
  15 #include <linux/hyperv.h>
  16 #include <linux/version.h>
  17 #include <linux/random.h>
  18 #include <linux/clockchips.h>
  19 #include <asm/mshyperv.h>
  20 #include "hyperv_vmbus.h"
  21
  22 /* The one and only */
  23 struct hv_context hv_context;
  24
  25 /*
  26  * If false, we're using the old mechanism for stimer0 interrupts
  27  * where it sends a VMbus message when it expires. The old
  28  * mechanism is used when running on older versions of Hyper-V
  29  * that don't support Direct Mode. While Hyper-V provides
  30  * four stimer's per CPU, Linux uses only stimer0.
  31  */
  32 static bool direct_mode_enabled;
  33 static int stimer0_irq;
  34 static int stimer0_vector;
  35
  36 #define HV_TIMER_FREQUENCY (10 * 1000 * 1000) /* 100ns period */
  37 #define HV_MAX_MAX_DELTA_TICKS 0xffffffff
  38 #define HV_MIN_DELTA_TICKS 1
  39
  40 /*
  41  * hv_init - Main initialization routine.
  42  *
  43  * This routine must be called before any other routines in here are called
  44  */
  45 int hv_init(void)
  46 {
  47         hv_context.cpu_context = alloc_percpu(struct hv_per_cpu_context);
  48         if (!hv_context.cpu_context)
  49                 return -ENOMEM;
  50
  51         direct_mode_enabled = ms_hyperv.misc_features &
  52                         HV_STIMER_DIRECT_MODE_AVAILABLE;
  53         return 0;
  54 }
  55
  56 /*
  57  * hv_post_message - Post a message using the hypervisor message IPC.
  58  *
  59  * This involves a hypercall.
  60  */
  61 int hv_post_message(union hv_connection_id connection_id,
  62                   enum hv_message_type message_type,
  63                   void *payload, size_t payload_size)
  64 {
  65         struct hv_input_post_message *aligned_msg;
  66         struct hv_per_cpu_context *hv_cpu;
  67         u64 status;
  68
  69         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT)
  70                 return -EMSGSIZE;
  71
  72         hv_cpu = get_cpu_ptr(hv_context.cpu_context);
  73         aligned_msg = hv_cpu->post_msg_page;
  74         aligned_msg->connectionid = connection_id;
  75         aligned_msg->reserved = 0;
  76         aligned_msg->message_type = message_type;
  77         aligned_msg->payload_size = payload_size;
  78         memcpy((void *)aligned_msg->payload, payload, payload_size);
  79
  80         status = hv_do_hypercall(HVCALL_POST_MESSAGE, aligned_msg, NULL);
  81
  82         /* Preemption must remain disabled until after the hypercall
  83          * so some other thread can't get scheduled onto this cpu and
  84          * corrupt the per-cpu post_msg_page
  85          */
  86         put_cpu_ptr(hv_cpu);
  87
  88         return status & 0xFFFF;
  89 }
  90
  91 /*
  92  * ISR for when stimer0 is operating in Direct Mode.  Direct Mode
  93  * does not use VMbus or any VMbus messages, so process here and not
  94  * in the VMbus driver code.
  95  */
  96
  97 static void hv_stimer0_isr(void)
  98 {
  99         struct hv_per_cpu_context *hv_cpu;
 100
 101         hv_cpu = this_cpu_ptr(hv_context.cpu_context);
 102         hv_cpu->clk_evt->event_handler(hv_cpu->clk_evt);
 103         add_interrupt_randomness(stimer0_vector, 0);
 104 }
 105
 106 static int hv_ce_set_next_event(unsigned long delta,
 107                                 struct clock_event_device *evt)
 108 {
 109         u64 current_tick;
 110
 111         WARN_ON(!clockevent_state_oneshot(evt));
 112
 113         current_tick = hyperv_cs->read(NULL);
 114         current_tick += delta;
 115         hv_init_timer(0, current_tick);
 116         return 0;
 117 }
 118
 119 static int hv_ce_shutdown(struct clock_event_device *evt)
 120 {
 121         hv_init_timer(0, 0);
 122         hv_init_timer_config(0, 0);
 123         if (direct_mode_enabled)
 124                 hv_disable_stimer0_percpu_irq(stimer0_irq);
 125
 126         return 0;
 127 }
 128
 129 static int hv_ce_set_oneshot(struct clock_event_device *evt)
 130 {
 131         union hv_stimer_config timer_cfg;
 132
 133         timer_cfg.as_uint64 = 0;
 134         timer_cfg.enable = 1;
 135         timer_cfg.auto_enable = 1;
 136         if (direct_mode_enabled) {
 137                 /*
 138                  * When it expires, the timer will directly interrupt
 139                  * on the specified hardware vector/IRQ.
 140                  */
 141                 timer_cfg.direct_mode = 1;
 142                 timer_cfg.apic_vector = stimer0_vector;
 143                 hv_enable_stimer0_percpu_irq(stimer0_irq);
 144         } else {
 145                 /*
 146                  * When it expires, the timer will generate a VMbus message,
 147                  * to be handled by the normal VMbus interrupt handler.
 148                  */
 149                 timer_cfg.direct_mode = 0;
 150                 timer_cfg.sintx = VMBUS_MESSAGE_SINT;
 151         }
 152         hv_init_timer_config(0, timer_cfg.as_uint64);
 153         return 0;
 154 }
 155
 156 static void hv_init_clockevent_device(struct clock_event_device *dev, int cpu)
 157 {
 158         dev->name = "Hyper-V clockevent";
 159         dev->features = CLOCK_EVT_FEAT_ONESHOT;
 160         dev->cpumask = cpumask_of(cpu);
 161         dev->rating = 1000;
 162         /*
 163          * Avoid settint dev->owner = THIS_MODULE deliberately as doing so will
 164          * result in clockevents_config_and_register() taking additional
 165          * references to the hv_vmbus module making it impossible to unload.
 166          */
 167
 168         dev->set_state_shutdown = hv_ce_shutdown;
 169         dev->set_state_oneshot = hv_ce_set_oneshot;
 170         dev->set_next_event = hv_ce_set_next_event;
 171 }
 172
 173
 174 int hv_synic_alloc(void)
 175 {
 176         int cpu;
 177         struct hv_per_cpu_context *hv_cpu;
 178
 179         /*
 180          * First, zero all per-cpu memory areas so hv_synic_free() can
 181          * detect what memory has been allocated and cleanup properly
 182          * after any failures.
 183          */
 184         for_each_present_cpu(cpu) {
 185                 hv_cpu = per_cpu_ptr(hv_context.cpu_context, cpu);
 186                 memset(hv_cpu, 0, sizeof(*hv_cpu));
 187         }
 188
 189         hv_context.hv_numa_map = kcalloc(nr_node_ids, sizeof(struct cpumask),
 190                                          GFP_KERNEL);
 191         if (hv_context.hv_numa_map == NULL) {
 192                 pr_err("Unable to allocate NUMA map\n");
 193                 goto err;
 194         }
 195
 196         for_each_present_cpu(cpu) {
 197                 hv_cpu = per_cpu_ptr(hv_context.cpu_context, cpu);
 198
 199                 tasklet_init(&hv_cpu->msg_dpc,
 200                              vmbus_on_msg_dpc, (unsigned long) hv_cpu);
 201
 202                 hv_cpu->clk_evt = kzalloc(sizeof(struct clock_event_device),
 203                                           GFP_KERNEL);
 204                 if (hv_cpu->clk_evt == NULL) {
 205                         pr_err("Unable to allocate clock event device\n");
 206                         goto err;
 207                 }
 208                 hv_init_clockevent_device(hv_cpu->clk_evt, cpu);
 209
 210                 hv_cpu->synic_message_page =
 211                         (void *)get_zeroed_page(GFP_ATOMIC);
 212                 if (hv_cpu->synic_message_page == NULL) {
 213                         pr_err("Unable to allocate SYNIC message page\n");
 214                         goto err;
 215                 }
 216
 217                 hv_cpu->synic_event_page = (void *)get_zeroed_page(GFP_ATOMIC);
 218                 if (hv_cpu->synic_event_page == NULL) {
 219                         pr_err("Unable to allocate SYNIC event page\n");
 220                         goto err;
 221                 }
 222
 223                 hv_cpu->post_msg_page = (void *)get_zeroed_page(GFP_ATOMIC);
 224                 if (hv_cpu->post_msg_page == NULL) {
 225                         pr_err("Unable to allocate post msg page\n");
 226                         goto err;
 227                 }
 228
 229                 INIT_LIST_HEAD(&hv_cpu->chan_list);
 230         }
 231
 232         if (direct_mode_enabled &&
 233             hv_setup_stimer0_irq(&stimer0_irq, &stimer0_vector,
 234                                 hv_stimer0_isr))
 235                 goto err;
 236
 237         return 0;
 238 err:
 239         /*
 240          * Any memory allocations that succeeded will be freed when
 241          * the caller cleans up by calling hv_synic_free()
 242          */
 243         return -ENOMEM;
 244 }
 245
 246
 247 void hv_synic_free(void)
 248 {
 249         int cpu;
 250
 251         for_each_present_cpu(cpu) {
 252                 struct hv_per_cpu_context *hv_cpu
 253                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 254
 255                 kfree(hv_cpu->clk_evt);
 256                 free_page((unsigned long)hv_cpu->synic_event_page);
 257                 free_page((unsigned long)hv_cpu->synic_message_page);
 258                 free_page((unsigned long)hv_cpu->post_msg_page);
 259         }
 260
 261         kfree(hv_context.hv_numa_map);
 262 }
 263
 264 /*
 265  * hv_synic_init - Initialize the Synthetic Interrupt Controller.
 266  *
 267  * If it is already initialized by another entity (ie x2v shim), we need to
 268  * retrieve the initialized message and event pages.  Otherwise, we create and
 269  * initialize the message and event pages.
 270  */
 271 int hv_synic_init(unsigned int cpu)
 272 {
 273         struct hv_per_cpu_context *hv_cpu
 274                 = per_cpu_ptr(hv_context.cpu_context, cpu);
 275         union hv_synic_simp simp;
 276         union hv_synic_siefp siefp;
 277         union hv_synic_sint shared_sint;
 278         union hv_synic_scontrol sctrl;
 279
 280         /* Setup the Synic's message page */
 281         hv_get_simp(simp.as_uint64);
 282         simp.simp_enabled = 1;
 283         simp.base_simp_gpa = virt_to_phys(hv_cpu->synic_message_page)
 284                 >> PAGE_SHIFT;
 285
 286         hv_set_simp(simp.as_uint64);
 287
 288         /* Setup the Synic's event page */
 289         hv_get_siefp(siefp.as_uint64);
 290         siefp.siefp_enabled = 1;
 291         siefp.base_siefp_gpa = virt_to_phys(hv_cpu->synic_event_page)
 292                 >> PAGE_SHIFT;
 293
 294         hv_set_siefp(siefp.as_uint64);
 295
 296         /* Setup the shared SINT. */
 297         hv_get_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
 298
 299         shared_sint.vector = HYPERVISOR_CALLBACK_VECTOR;
 300         shared_sint.masked = false;
 301         if (ms_hyperv.hints & HV_DEPRECATING_AEOI_RECOMMENDED)
 302                 shared_sint.auto_eoi = false;
 303         else
 304                 shared_sint.auto_eoi = true;
 305
 306         hv_set_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
 307
 308         /* Enable the global synic bit */
 309         hv_get_synic_state(sctrl.as_uint64);
 310         sctrl.enable = 1;
 311
 312         hv_set_synic_state(sctrl.as_uint64);
 313
 314         /*
 315          * Register the per-cpu clockevent source.
 316          */
 317         if (ms_hyperv.features & HV_MSR_SYNTIMER_AVAILABLE)
 318                 clockevents_config_and_register(hv_cpu->clk_evt,
 319                                                 HV_TIMER_FREQUENCY,
 320                                                 HV_MIN_DELTA_TICKS,
 321                                                 HV_MAX_MAX_DELTA_TICKS);
 322         return 0;
 323 }
 324
 325 /*
 326  * hv_synic_clockevents_cleanup - Cleanup clockevent devices
 327  */
 328 void hv_synic_clockevents_cleanup(void)
 329 {
 330         int cpu;
 331
 332         if (!(ms_hyperv.features & HV_MSR_SYNTIMER_AVAILABLE))
 333                 return;
 334
 335         if (direct_mode_enabled)
 336                 hv_remove_stimer0_irq(stimer0_irq);
 337
 338         for_each_present_cpu(cpu) {
 339                 struct hv_per_cpu_context *hv_cpu
 340                         = per_cpu_ptr(hv_context.cpu_context, cpu);
 341
 342                 clockevents_unbind_device(hv_cpu->clk_evt, cpu);
 343         }
 344 }
 345
 346 /*
 347  * hv_synic_cleanup - Cleanup routine for hv_synic_init().
 348  */
 349 int hv_synic_cleanup(unsigned int cpu)
 350 {
 351         union hv_synic_sint shared_sint;
 352         union hv_synic_simp simp;
 353         union hv_synic_siefp siefp;
 354         union hv_synic_scontrol sctrl;
 355         struct vmbus_channel *channel, *sc;
 356         bool channel_found = false;
 357         unsigned long flags;
 358
 359         hv_get_synic_state(sctrl.as_uint64);
 360         if (sctrl.enable != 1)
 361                 return -EFAULT;
 362
 363         /*
 364          * Search for channels which are bound to the CPU we're about to
 365          * cleanup. In case we find one and vmbus is still connected we need to
 366          * fail, this will effectively prevent CPU offlining. There is no way
 367          * we can re-bind channels to different CPUs for now.
 368          */
 369         mutex_lock(&vmbus_connection.channel_mutex);
 370         list_for_each_entry(channel, &vmbus_connection.chn_list, listentry) {
 371                 if (channel->target_cpu == cpu) {
 372                         channel_found = true;
 373                         break;
 374                 }
 375                 spin_lock_irqsave(&channel->lock, flags);
 376                 list_for_each_entry(sc, &channel->sc_list, sc_list) {
 377                         if (sc->target_cpu == cpu) {
 378                                 channel_found = true;
 379                                 break;
 380                         }
 381                 }
 382                 spin_unlock_irqrestore(&channel->lock, flags);
 383                 if (channel_found)
 384                         break;
 385         }
 386         mutex_unlock(&vmbus_connection.channel_mutex);
 387
 388         if (channel_found && vmbus_connection.conn_state == CONNECTED)
 389                 return -EBUSY;
 390
 391         /* Turn off clockevent device */
 392         if (ms_hyperv.features & HV_MSR_SYNTIMER_AVAILABLE) {
 393                 struct hv_per_cpu_context *hv_cpu
 394                         = this_cpu_ptr(hv_context.cpu_context);
 395
 396                 clockevents_unbind_device(hv_cpu->clk_evt, cpu);
 397                 hv_ce_shutdown(hv_cpu->clk_evt);
 398         }
 399
 400         hv_get_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
 401
 402         shared_sint.masked = 1;
 403
 404         /* Need to correctly cleanup in the case of SMP!!! */
 405         /* Disable the interrupt */
 406         hv_set_synint_state(VMBUS_MESSAGE_SINT, shared_sint.as_uint64);
 407
 408         hv_get_simp(simp.as_uint64);
 409         simp.simp_enabled = 0;
 410         simp.base_simp_gpa = 0;
 411
 412         hv_set_simp(simp.as_uint64);
 413
 414         hv_get_siefp(siefp.as_uint64);
 415         siefp.siefp_enabled = 0;
 416         siefp.base_siefp_gpa = 0;
 417
 418         hv_set_siefp(siefp.as_uint64);
 419
 420         /* Disable the global synic bit */
 421         sctrl.enable = 0;
 422         hv_set_synic_state(sctrl.as_uint64);
 423
 424         return 0;
 425 }