drivers/xen/events/events_fifo.c

   1 /*
   2  * Xen event channels (FIFO-based ABI)
   3  *
   4  * Copyright (C) 2013 Citrix Systems R&D ltd.
   5  *
   6  * This source code is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU General Public License as
   8  * published by the Free Software Foundation; either version 2 of the
   9  * License, or (at your option) any later version.
  10  *
  11  * Or, when distributed separately from the Linux kernel or
  12  * incorporated into other software packages, subject to the following
  13  * license:
  14  *
  15  * Permission is hereby granted, free of charge, to any person obtaining a copy
  16  * of this source file (the "Software"), to deal in the Software without
  17  * restriction, including without limitation the rights to use, copy, modify,
  18  * merge, publish, distribute, sublicense, and/or sell copies of the Software,
  19  * and to permit persons to whom the Software is furnished to do so, subject to
  20  * the following conditions:
  21  *
  22  * The above copyright notice and this permission notice shall be included in
  23  * all copies or substantial portions of the Software.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  26  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  27  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
  28  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  29  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  30  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  31  * IN THE SOFTWARE.
  32  */
  33
  34 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
  35
  36 #include <linux/linkage.h>
  37 #include <linux/interrupt.h>
  38 #include <linux/irq.h>
  39 #include <linux/smp.h>
  40 #include <linux/percpu.h>
  41 #include <linux/cpu.h>
  42
  43 #include <asm/barrier.h>
  44 #include <asm/sync_bitops.h>
  45 #include <asm/xen/hypercall.h>
  46 #include <asm/xen/hypervisor.h>
  47
  48 #include <xen/xen.h>
  49 #include <xen/xen-ops.h>
  50 #include <xen/events.h>
  51 #include <xen/interface/xen.h>
  52 #include <xen/interface/event_channel.h>
  53 #include <xen/page.h>
  54
  55 #include "events_internal.h"
  56
  57 #define EVENT_WORDS_PER_PAGE (XEN_PAGE_SIZE / sizeof(event_word_t))
  58 #define MAX_EVENT_ARRAY_PAGES (EVTCHN_FIFO_NR_CHANNELS / EVENT_WORDS_PER_PAGE)
  59
  60 struct evtchn_fifo_queue {
  61         uint32_t head[EVTCHN_FIFO_MAX_QUEUES];
  62 };
  63
  64 static DEFINE_PER_CPU(struct evtchn_fifo_control_block *, cpu_control_block);
  65 static DEFINE_PER_CPU(struct evtchn_fifo_queue, cpu_queue);
  66 static event_word_t *event_array[MAX_EVENT_ARRAY_PAGES] __read_mostly;
  67 static unsigned event_array_pages __read_mostly;
  68
  69 /*
  70  * sync_set_bit() and friends must be unsigned long aligned.
  71  */
  72 #if BITS_PER_LONG > 32
  73
  74 #define BM(w) (unsigned long *)((unsigned long)w & ~0x7UL)
  75 #define EVTCHN_FIFO_BIT(b, w) \
  76     (((unsigned long)w & 0x4UL) ? (EVTCHN_FIFO_ ##b + 32) : EVTCHN_FIFO_ ##b)
  77
  78 #else
  79
  80 #define BM(w) ((unsigned long *)(w))
  81 #define EVTCHN_FIFO_BIT(b, w) EVTCHN_FIFO_ ##b
  82
  83 #endif
  84
  85 static inline event_word_t *event_word_from_port(unsigned port)
  86 {
  87         unsigned i = port / EVENT_WORDS_PER_PAGE;
  88
  89         return event_array[i] + port % EVENT_WORDS_PER_PAGE;
  90 }
  91
  92 static unsigned evtchn_fifo_max_channels(void)
  93 {
  94         return EVTCHN_FIFO_NR_CHANNELS;
  95 }
  96
  97 static unsigned evtchn_fifo_nr_channels(void)
  98 {
  99         return event_array_pages * EVENT_WORDS_PER_PAGE;
 100 }
 101
 102 static int init_control_block(int cpu,
 103                               struct evtchn_fifo_control_block *control_block)
 104 {
 105         struct evtchn_fifo_queue *q = &per_cpu(cpu_queue, cpu);
 106         struct evtchn_init_control init_control;
 107         unsigned int i;
 108
 109         /* Reset the control block and the local HEADs. */
 110         clear_page(control_block);
 111         for (i = 0; i < EVTCHN_FIFO_MAX_QUEUES; i++)
 112                 q->head[i] = 0;
 113
 114         init_control.control_gfn = virt_to_gfn(control_block);
 115         init_control.offset      = 0;
 116         init_control.vcpu        = xen_vcpu_nr(cpu);
 117
 118         return HYPERVISOR_event_channel_op(EVTCHNOP_init_control, &init_control);
 119 }
 120
 121 static void free_unused_array_pages(void)
 122 {
 123         unsigned i;
 124
 125         for (i = event_array_pages; i < MAX_EVENT_ARRAY_PAGES; i++) {
 126                 if (!event_array[i])
 127                         break;
 128                 free_page((unsigned long)event_array[i]);
 129                 event_array[i] = NULL;
 130         }
 131 }
 132
 133 static void init_array_page(event_word_t *array_page)
 134 {
 135         unsigned i;
 136
 137         for (i = 0; i < EVENT_WORDS_PER_PAGE; i++)
 138                 array_page[i] = 1 << EVTCHN_FIFO_MASKED;
 139 }
 140
 141 static int evtchn_fifo_setup(struct irq_info *info)
 142 {
 143         unsigned port = info->evtchn;
 144         unsigned new_array_pages;
 145         int ret;
 146
 147         new_array_pages = port / EVENT_WORDS_PER_PAGE + 1;
 148
 149         if (new_array_pages > MAX_EVENT_ARRAY_PAGES)
 150                 return -EINVAL;
 151
 152         while (event_array_pages < new_array_pages) {
 153                 void *array_page;
 154                 struct evtchn_expand_array expand_array;
 155
 156                 /* Might already have a page if we've resumed. */
 157                 array_page = event_array[event_array_pages];
 158                 if (!array_page) {
 159                         array_page = (void *)__get_free_page(GFP_KERNEL);
 160                         if (array_page == NULL) {
 161                                 ret = -ENOMEM;
 162                                 goto error;
 163                         }
 164                         event_array[event_array_pages] = array_page;
 165                 }
 166
 167                 /* Mask all events in this page before adding it. */
 168                 init_array_page(array_page);
 169
 170                 expand_array.array_gfn = virt_to_gfn(array_page);
 171
 172                 ret = HYPERVISOR_event_channel_op(EVTCHNOP_expand_array, &expand_array);
 173                 if (ret < 0)
 174                         goto error;
 175
 176                 event_array_pages++;
 177         }
 178         return 0;
 179
 180   error:
 181         if (event_array_pages == 0)
 182                 panic("xen: unable to expand event array with initial page (%d)\n", ret);
 183         else
 184                 pr_err("unable to expand event array (%d)\n", ret);
 185         free_unused_array_pages();
 186         return ret;
 187 }
 188
 189 static void evtchn_fifo_bind_to_cpu(struct irq_info *info, unsigned cpu)
 190 {
 191         /* no-op */
 192 }
 193
 194 static void evtchn_fifo_clear_pending(unsigned port)
 195 {
 196         event_word_t *word = event_word_from_port(port);
 197         sync_clear_bit(EVTCHN_FIFO_BIT(PENDING, word), BM(word));
 198 }
 199
 200 static void evtchn_fifo_set_pending(unsigned port)
 201 {
 202         event_word_t *word = event_word_from_port(port);
 203         sync_set_bit(EVTCHN_FIFO_BIT(PENDING, word), BM(word));
 204 }
 205
 206 static bool evtchn_fifo_is_pending(unsigned port)
 207 {
 208         event_word_t *word = event_word_from_port(port);
 209         return sync_test_bit(EVTCHN_FIFO_BIT(PENDING, word), BM(word));
 210 }
 211
 212 static void evtchn_fifo_mask(unsigned port)
 213 {
 214         event_word_t *word = event_word_from_port(port);
 215         sync_set_bit(EVTCHN_FIFO_BIT(MASKED, word), BM(word));
 216 }
 217
 218 static bool evtchn_fifo_is_masked(unsigned port)
 219 {
 220         event_word_t *word = event_word_from_port(port);
 221         return sync_test_bit(EVTCHN_FIFO_BIT(MASKED, word), BM(word));
 222 }
 223 /*
 224  * Clear MASKED if not PENDING, spinning if BUSY is set.
 225  * Return true if mask was cleared.
 226  */
 227 static bool clear_masked_cond(volatile event_word_t *word)
 228 {
 229         event_word_t new, old, w;
 230
 231         w = *word;
 232
 233         do {
 234                 if (w & (1 << EVTCHN_FIFO_PENDING))
 235                         return false;
 236
 237                 old = w & ~(1 << EVTCHN_FIFO_BUSY);
 238                 new = old & ~(1 << EVTCHN_FIFO_MASKED);
 239                 w = sync_cmpxchg(word, old, new);
 240         } while (w != old);
 241
 242         return true;
 243 }
 244
 245 static void evtchn_fifo_unmask(unsigned port)
 246 {
 247         event_word_t *word = event_word_from_port(port);
 248
 249         BUG_ON(!irqs_disabled());
 250
 251         if (!clear_masked_cond(word)) {
 252                 struct evtchn_unmask unmask = { .port = port };
 253                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 254         }
 255 }
 256
 257 static uint32_t clear_linked(volatile event_word_t *word)
 258 {
 259         event_word_t new, old, w;
 260
 261         w = *word;
 262
 263         do {
 264                 old = w;
 265                 new = (w & ~((1 << EVTCHN_FIFO_LINKED)
 266                              | EVTCHN_FIFO_LINK_MASK));
 267         } while ((w = sync_cmpxchg(word, old, new)) != old);
 268
 269         return w & EVTCHN_FIFO_LINK_MASK;
 270 }
 271
 272 static void consume_one_event(unsigned cpu, struct evtchn_loop_ctrl *ctrl,
 273                               struct evtchn_fifo_control_block *control_block,
 274                               unsigned priority, unsigned long *ready)
 275 {
 276         struct evtchn_fifo_queue *q = &per_cpu(cpu_queue, cpu);
 277         uint32_t head;
 278         unsigned port;
 279         event_word_t *word;
 280
 281         head = q->head[priority];
 282
 283         /*
 284          * Reached the tail last time?  Read the new HEAD from the
 285          * control block.
 286          */
 287         if (head == 0) {
 288                 virt_rmb(); /* Ensure word is up-to-date before reading head. */
 289                 head = control_block->head[priority];
 290         }
 291
 292         port = head;
 293         word = event_word_from_port(port);
 294         head = clear_linked(word);
 295
 296         /*
 297          * If the link is non-zero, there are more events in the
 298          * queue, otherwise the queue is empty.
 299          *
 300          * If the queue is empty, clear this priority from our local
 301          * copy of the ready word.
 302          */
 303         if (head == 0)
 304                 clear_bit(priority, ready);
 305
 306         if (evtchn_fifo_is_pending(port) && !evtchn_fifo_is_masked(port)) {
 307                 if (unlikely(!ctrl))
 308                         pr_warn("Dropping pending event for port %u\n", port);
 309                 else
 310                         handle_irq_for_port(port, ctrl);
 311         }
 312
 313         q->head[priority] = head;
 314 }
 315
 316 static void __evtchn_fifo_handle_events(unsigned cpu,
 317                                         struct evtchn_loop_ctrl *ctrl)
 318 {
 319         struct evtchn_fifo_control_block *control_block;
 320         unsigned long ready;
 321         unsigned q;
 322
 323         control_block = per_cpu(cpu_control_block, cpu);
 324
 325         ready = xchg(&control_block->ready, 0);
 326
 327         while (ready) {
 328                 q = find_first_bit(&ready, EVTCHN_FIFO_MAX_QUEUES);
 329                 consume_one_event(cpu, ctrl, control_block, q, &ready);
 330                 ready |= xchg(&control_block->ready, 0);
 331         }
 332 }
 333
 334 static void evtchn_fifo_handle_events(unsigned cpu,
 335                                       struct evtchn_loop_ctrl *ctrl)
 336 {
 337         __evtchn_fifo_handle_events(cpu, ctrl);
 338 }
 339
 340 static void evtchn_fifo_resume(void)
 341 {
 342         unsigned cpu;
 343
 344         for_each_possible_cpu(cpu) {
 345                 void *control_block = per_cpu(cpu_control_block, cpu);
 346                 int ret;
 347
 348                 if (!control_block)
 349                         continue;
 350
 351                 /*
 352                  * If this CPU is offline, take the opportunity to
 353                  * free the control block while it is not being
 354                  * used.
 355                  */
 356                 if (!cpu_online(cpu)) {
 357                         free_page((unsigned long)control_block);
 358                         per_cpu(cpu_control_block, cpu) = NULL;
 359                         continue;
 360                 }
 361
 362                 ret = init_control_block(cpu, control_block);
 363                 BUG_ON(ret < 0);
 364         }
 365
 366         /*
 367          * The event array starts out as empty again and is extended
 368          * as normal when events are bound.  The existing pages will
 369          * be reused.
 370          */
 371         event_array_pages = 0;
 372 }
 373
 374 static int evtchn_fifo_alloc_control_block(unsigned cpu)
 375 {
 376         void *control_block = NULL;
 377         int ret = -ENOMEM;
 378
 379         control_block = (void *)__get_free_page(GFP_KERNEL);
 380         if (control_block == NULL)
 381                 goto error;
 382
 383         ret = init_control_block(cpu, control_block);
 384         if (ret < 0)
 385                 goto error;
 386
 387         per_cpu(cpu_control_block, cpu) = control_block;
 388
 389         return 0;
 390
 391   error:
 392         free_page((unsigned long)control_block);
 393         return ret;
 394 }
 395
 396 static int evtchn_fifo_percpu_init(unsigned int cpu)
 397 {
 398         if (!per_cpu(cpu_control_block, cpu))
 399                 return evtchn_fifo_alloc_control_block(cpu);
 400         return 0;
 401 }
 402
 403 static int evtchn_fifo_percpu_deinit(unsigned int cpu)
 404 {
 405         __evtchn_fifo_handle_events(cpu, NULL);
 406         return 0;
 407 }
 408
 409 static const struct evtchn_ops evtchn_ops_fifo = {
 410         .max_channels      = evtchn_fifo_max_channels,
 411         .nr_channels       = evtchn_fifo_nr_channels,
 412         .setup             = evtchn_fifo_setup,
 413         .bind_to_cpu       = evtchn_fifo_bind_to_cpu,
 414         .clear_pending     = evtchn_fifo_clear_pending,
 415         .set_pending       = evtchn_fifo_set_pending,
 416         .is_pending        = evtchn_fifo_is_pending,
 417         .mask              = evtchn_fifo_mask,
 418         .unmask            = evtchn_fifo_unmask,
 419         .handle_events     = evtchn_fifo_handle_events,
 420         .resume            = evtchn_fifo_resume,
 421         .percpu_init       = evtchn_fifo_percpu_init,
 422         .percpu_deinit     = evtchn_fifo_percpu_deinit,
 423 };
 424
 425 int __init xen_evtchn_fifo_init(void)
 426 {
 427         int cpu = smp_processor_id();
 428         int ret;
 429
 430         ret = evtchn_fifo_alloc_control_block(cpu);
 431         if (ret < 0)
 432                 return ret;
 433
 434         pr_info("Using FIFO-based ABI\n");
 435
 436         evtchn_ops = &evtchn_ops_fifo;
 437
 438         return ret;
 439 }