drivers/xen/events/events_2l.c

   1 /*
   2  * Xen event channels (2-level ABI)
   3  *
   4  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
   5  */
   6
   7 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
   8
   9 #include <linux/linkage.h>
  10 #include <linux/interrupt.h>
  11 #include <linux/irq.h>
  12
  13 #include <asm/sync_bitops.h>
  14 #include <asm/xen/hypercall.h>
  15 #include <asm/xen/hypervisor.h>
  16
  17 #include <xen/xen.h>
  18 #include <xen/xen-ops.h>
  19 #include <xen/events.h>
  20 #include <xen/interface/xen.h>
  21 #include <xen/interface/event_channel.h>
  22
  23 #include "events_internal.h"
  24
  25 /*
  26  * Note sizeof(xen_ulong_t) can be more than sizeof(unsigned long). Be
  27  * careful to only use bitops which allow for this (e.g
  28  * test_bit/find_first_bit and friends but not __ffs) and to pass
  29  * BITS_PER_EVTCHN_WORD as the bitmask length.
  30  */
  31 #define BITS_PER_EVTCHN_WORD (sizeof(xen_ulong_t)*8)
  32 /*
  33  * Make a bitmask (i.e. unsigned long *) of a xen_ulong_t
  34  * array. Primarily to avoid long lines (hence the terse name).
  35  */
  36 #define BM(x) (unsigned long *)(x)
  37 /* Find the first set bit in a evtchn mask */
  38 #define EVTCHN_FIRST_BIT(w) find_first_bit(BM(&(w)), BITS_PER_EVTCHN_WORD)
  39
  40 #define EVTCHN_MASK_SIZE (EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD)
  41
  42 static DEFINE_PER_CPU(xen_ulong_t [EVTCHN_MASK_SIZE], cpu_evtchn_mask);
  43
  44 static unsigned evtchn_2l_max_channels(void)
  45 {
  46         return EVTCHN_2L_NR_CHANNELS;
  47 }
  48
  49 static void evtchn_2l_remove(evtchn_port_t evtchn, unsigned int cpu)
  50 {
  51         clear_bit(evtchn, BM(per_cpu(cpu_evtchn_mask, cpu)));
  52 }
  53
  54 static void evtchn_2l_bind_to_cpu(struct irq_info *info, unsigned cpu)
  55 {
  56         clear_bit(info->evtchn, BM(per_cpu(cpu_evtchn_mask, info->cpu)));
  57         set_bit(info->evtchn, BM(per_cpu(cpu_evtchn_mask, cpu)));
  58 }
  59
  60 static void evtchn_2l_clear_pending(unsigned port)
  61 {
  62         struct shared_info *s = HYPERVISOR_shared_info;
  63         sync_clear_bit(port, BM(&s->evtchn_pending[0]));
  64 }
  65
  66 static void evtchn_2l_set_pending(unsigned port)
  67 {
  68         struct shared_info *s = HYPERVISOR_shared_info;
  69         sync_set_bit(port, BM(&s->evtchn_pending[0]));
  70 }
  71
  72 static bool evtchn_2l_is_pending(unsigned port)
  73 {
  74         struct shared_info *s = HYPERVISOR_shared_info;
  75         return sync_test_bit(port, BM(&s->evtchn_pending[0]));
  76 }
  77
  78 static void evtchn_2l_mask(unsigned port)
  79 {
  80         struct shared_info *s = HYPERVISOR_shared_info;
  81         sync_set_bit(port, BM(&s->evtchn_mask[0]));
  82 }
  83
  84 static void evtchn_2l_unmask(unsigned port)
  85 {
  86         struct shared_info *s = HYPERVISOR_shared_info;
  87         unsigned int cpu = get_cpu();
  88         int do_hypercall = 0, evtchn_pending = 0;
  89
  90         BUG_ON(!irqs_disabled());
  91
  92         smp_wmb();      /* All writes before unmask must be visible. */
  93
  94         if (unlikely((cpu != cpu_from_evtchn(port))))
  95                 do_hypercall = 1;
  96         else {
  97                 /*
  98                  * Need to clear the mask before checking pending to
  99                  * avoid a race with an event becoming pending.
 100                  *
 101                  * EVTCHNOP_unmask will only trigger an upcall if the
 102                  * mask bit was set, so if a hypercall is needed
 103                  * remask the event.
 104                  */
 105                 sync_clear_bit(port, BM(&s->evtchn_mask[0]));
 106                 evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0]));
 107
 108                 if (unlikely(evtchn_pending && xen_hvm_domain())) {
 109                         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 110                         do_hypercall = 1;
 111                 }
 112         }
 113
 114         /* Slow path (hypercall) if this is a non-local port or if this is
 115          * an hvm domain and an event is pending (hvm domains don't have
 116          * their own implementation of irq_enable). */
 117         if (do_hypercall) {
 118                 struct evtchn_unmask unmask = { .port = port };
 119                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 120         } else {
 121                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 122
 123                 /*
 124                  * The following is basically the equivalent of
 125                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 126                  * the interrupt edge' if the channel is masked.
 127                  */
 128                 if (evtchn_pending &&
 129                     !sync_test_and_set_bit(port / BITS_PER_EVTCHN_WORD,
 130                                            BM(&vcpu_info->evtchn_pending_sel)))
 131                         vcpu_info->evtchn_upcall_pending = 1;
 132         }
 133
 134         put_cpu();
 135 }
 136
 137 static DEFINE_PER_CPU(unsigned int, current_word_idx);
 138 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
 139
 140 /*
 141  * Mask out the i least significant bits of w
 142  */
 143 #define MASK_LSBS(w, i) (w & ((~((xen_ulong_t)0UL)) << i))
 144
 145 static inline xen_ulong_t active_evtchns(unsigned int cpu,
 146                                          struct shared_info *sh,
 147                                          unsigned int idx)
 148 {
 149         return sh->evtchn_pending[idx] &
 150                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 151                 ~sh->evtchn_mask[idx];
 152 }
 153
 154 /*
 155  * Search the CPU's pending events bitmasks.  For each one found, map
 156  * the event number to an irq, and feed it into do_IRQ() for handling.
 157  *
 158  * Xen uses a two-level bitmap to speed searching.  The first level is
 159  * a bitset of words which contain pending event bits.  The second
 160  * level is a bitset of pending events themselves.
 161  */
 162 static void evtchn_2l_handle_events(unsigned cpu, struct evtchn_loop_ctrl *ctrl)
 163 {
 164         int irq;
 165         xen_ulong_t pending_words;
 166         xen_ulong_t pending_bits;
 167         int start_word_idx, start_bit_idx;
 168         int word_idx, bit_idx;
 169         int i;
 170         struct shared_info *s = HYPERVISOR_shared_info;
 171         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 172
 173         /* Timer interrupt has highest priority. */
 174         irq = irq_from_virq(cpu, VIRQ_TIMER);
 175         if (irq != -1) {
 176                 unsigned int evtchn = evtchn_from_irq(irq);
 177                 word_idx = evtchn / BITS_PER_LONG;
 178                 bit_idx = evtchn % BITS_PER_LONG;
 179                 if (active_evtchns(cpu, s, word_idx) & (1ULL << bit_idx))
 180                         generic_handle_irq(irq);
 181         }
 182
 183         /*
 184          * Master flag must be cleared /before/ clearing
 185          * selector flag. xchg_xen_ulong must contain an
 186          * appropriate barrier.
 187          */
 188         pending_words = xchg_xen_ulong(&vcpu_info->evtchn_pending_sel, 0);
 189
 190         start_word_idx = __this_cpu_read(current_word_idx);
 191         start_bit_idx = __this_cpu_read(current_bit_idx);
 192
 193         word_idx = start_word_idx;
 194
 195         for (i = 0; pending_words != 0; i++) {
 196                 xen_ulong_t words;
 197
 198                 words = MASK_LSBS(pending_words, word_idx);
 199
 200                 /*
 201                  * If we masked out all events, wrap to beginning.
 202                  */
 203                 if (words == 0) {
 204                         word_idx = 0;
 205                         bit_idx = 0;
 206                         continue;
 207                 }
 208                 word_idx = EVTCHN_FIRST_BIT(words);
 209
 210                 pending_bits = active_evtchns(cpu, s, word_idx);
 211                 bit_idx = 0; /* usually scan entire word from start */
 212                 /*
 213                  * We scan the starting word in two parts.
 214                  *
 215                  * 1st time: start in the middle, scanning the
 216                  * upper bits.
 217                  *
 218                  * 2nd time: scan the whole word (not just the
 219                  * parts skipped in the first pass) -- if an
 220                  * event in the previously scanned bits is
 221                  * pending again it would just be scanned on
 222                  * the next loop anyway.
 223                  */
 224                 if (word_idx == start_word_idx) {
 225                         if (i == 0)
 226                                 bit_idx = start_bit_idx;
 227                 }
 228
 229                 do {
 230                         xen_ulong_t bits;
 231                         int port;
 232
 233                         bits = MASK_LSBS(pending_bits, bit_idx);
 234
 235                         /* If we masked out all events, move on. */
 236                         if (bits == 0)
 237                                 break;
 238
 239                         bit_idx = EVTCHN_FIRST_BIT(bits);
 240
 241                         /* Process port. */
 242                         port = (word_idx * BITS_PER_EVTCHN_WORD) + bit_idx;
 243                         handle_irq_for_port(port, ctrl);
 244
 245                         bit_idx = (bit_idx + 1) % BITS_PER_EVTCHN_WORD;
 246
 247                         /* Next caller starts at last processed + 1 */
 248                         __this_cpu_write(current_word_idx,
 249                                          bit_idx ? word_idx :
 250                                          (word_idx+1) % BITS_PER_EVTCHN_WORD);
 251                         __this_cpu_write(current_bit_idx, bit_idx);
 252                 } while (bit_idx != 0);
 253
 254                 /* Scan start_l1i twice; all others once. */
 255                 if ((word_idx != start_word_idx) || (i != 0))
 256                         pending_words &= ~(1UL << word_idx);
 257
 258                 word_idx = (word_idx + 1) % BITS_PER_EVTCHN_WORD;
 259         }
 260 }
 261
 262 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
 263 {
 264         struct shared_info *sh = HYPERVISOR_shared_info;
 265         int cpu = smp_processor_id();
 266         xen_ulong_t *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
 267         int i;
 268         unsigned long flags;
 269         static DEFINE_SPINLOCK(debug_lock);
 270         struct vcpu_info *v;
 271
 272         spin_lock_irqsave(&debug_lock, flags);
 273
 274         printk("\nvcpu %d\n  ", cpu);
 275
 276         for_each_online_cpu(i) {
 277                 int pending;
 278                 v = per_cpu(xen_vcpu, i);
 279                 pending = (get_irq_regs() && i == cpu)
 280                         ? xen_irqs_disabled(get_irq_regs())
 281                         : v->evtchn_upcall_mask;
 282                 printk("%d: masked=%d pending=%d event_sel %0*"PRI_xen_ulong"\n  ", i,
 283                        pending, v->evtchn_upcall_pending,
 284                        (int)(sizeof(v->evtchn_pending_sel)*2),
 285                        v->evtchn_pending_sel);
 286         }
 287         v = per_cpu(xen_vcpu, cpu);
 288
 289         printk("\npending:\n   ");
 290         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
 291                 printk("%0*"PRI_xen_ulong"%s",
 292                        (int)sizeof(sh->evtchn_pending[0])*2,
 293                        sh->evtchn_pending[i],
 294                        i % 8 == 0 ? "\n   " : " ");
 295         printk("\nglobal mask:\n   ");
 296         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 297                 printk("%0*"PRI_xen_ulong"%s",
 298                        (int)(sizeof(sh->evtchn_mask[0])*2),
 299                        sh->evtchn_mask[i],
 300                        i % 8 == 0 ? "\n   " : " ");
 301
 302         printk("\nglobally unmasked:\n   ");
 303         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 304                 printk("%0*"PRI_xen_ulong"%s",
 305                        (int)(sizeof(sh->evtchn_mask[0])*2),
 306                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
 307                        i % 8 == 0 ? "\n   " : " ");
 308
 309         printk("\nlocal cpu%d mask:\n   ", cpu);
 310         for (i = (EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD)-1; i >= 0; i--)
 311                 printk("%0*"PRI_xen_ulong"%s", (int)(sizeof(cpu_evtchn[0])*2),
 312                        cpu_evtchn[i],
 313                        i % 8 == 0 ? "\n   " : " ");
 314
 315         printk("\nlocally unmasked:\n   ");
 316         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
 317                 xen_ulong_t pending = sh->evtchn_pending[i]
 318                         & ~sh->evtchn_mask[i]
 319                         & cpu_evtchn[i];
 320                 printk("%0*"PRI_xen_ulong"%s",
 321                        (int)(sizeof(sh->evtchn_mask[0])*2),
 322                        pending, i % 8 == 0 ? "\n   " : " ");
 323         }
 324
 325         printk("\npending list:\n");
 326         for (i = 0; i < EVTCHN_2L_NR_CHANNELS; i++) {
 327                 if (sync_test_bit(i, BM(sh->evtchn_pending))) {
 328                         int word_idx = i / BITS_PER_EVTCHN_WORD;
 329                         printk("  %d: event %d -> irq %d%s%s%s\n",
 330                                cpu_from_evtchn(i), i,
 331                                get_evtchn_to_irq(i),
 332                                sync_test_bit(word_idx, BM(&v->evtchn_pending_sel))
 333                                ? "" : " l2-clear",
 334                                !sync_test_bit(i, BM(sh->evtchn_mask))
 335                                ? "" : " globally-masked",
 336                                sync_test_bit(i, BM(cpu_evtchn))
 337                                ? "" : " locally-masked");
 338                 }
 339         }
 340
 341         spin_unlock_irqrestore(&debug_lock, flags);
 342
 343         return IRQ_HANDLED;
 344 }
 345
 346 static void evtchn_2l_resume(void)
 347 {
 348         int i;
 349
 350         for_each_online_cpu(i)
 351                 memset(per_cpu(cpu_evtchn_mask, i), 0, sizeof(xen_ulong_t) *
 352                                 EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD);
 353 }
 354
 355 static int evtchn_2l_percpu_deinit(unsigned int cpu)
 356 {
 357         memset(per_cpu(cpu_evtchn_mask, cpu), 0, sizeof(xen_ulong_t) *
 358                         EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD);
 359
 360         return 0;
 361 }
 362
 363 static const struct evtchn_ops evtchn_ops_2l = {
 364         .max_channels      = evtchn_2l_max_channels,
 365         .nr_channels       = evtchn_2l_max_channels,
 366         .remove            = evtchn_2l_remove,
 367         .bind_to_cpu       = evtchn_2l_bind_to_cpu,
 368         .clear_pending     = evtchn_2l_clear_pending,
 369         .set_pending       = evtchn_2l_set_pending,
 370         .is_pending        = evtchn_2l_is_pending,
 371         .mask              = evtchn_2l_mask,
 372         .unmask            = evtchn_2l_unmask,
 373         .handle_events     = evtchn_2l_handle_events,
 374         .resume            = evtchn_2l_resume,
 375         .percpu_deinit     = evtchn_2l_percpu_deinit,
 376 };
 377
 378 void __init xen_evtchn_2l_init(void)
 379 {
 380         pr_info("Using 2-level ABI\n");
 381         evtchn_ops = &evtchn_ops_2l;
 382 }