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 static DEFINE_PER_CPU(xen_ulong_t [EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD],
  41                       cpu_evtchn_mask);
  42
  43 static unsigned evtchn_2l_max_channels(void)
  44 {
  45         return EVTCHN_2L_NR_CHANNELS;
  46 }
  47
  48 static void evtchn_2l_bind_to_cpu(struct irq_info *info, unsigned cpu)
  49 {
  50         clear_bit(info->evtchn, BM(per_cpu(cpu_evtchn_mask, info->cpu)));
  51         set_bit(info->evtchn, BM(per_cpu(cpu_evtchn_mask, cpu)));
  52 }
  53
  54 static void evtchn_2l_clear_pending(unsigned port)
  55 {
  56         struct shared_info *s = HYPERVISOR_shared_info;
  57         sync_clear_bit(port, BM(&s->evtchn_pending[0]));
  58 }
  59
  60 static void evtchn_2l_set_pending(unsigned port)
  61 {
  62         struct shared_info *s = HYPERVISOR_shared_info;
  63         sync_set_bit(port, BM(&s->evtchn_pending[0]));
  64 }
  65
  66 static bool evtchn_2l_is_pending(unsigned port)
  67 {
  68         struct shared_info *s = HYPERVISOR_shared_info;
  69         return sync_test_bit(port, BM(&s->evtchn_pending[0]));
  70 }
  71
  72 static bool evtchn_2l_test_and_set_mask(unsigned port)
  73 {
  74         struct shared_info *s = HYPERVISOR_shared_info;
  75         return sync_test_and_set_bit(port, BM(&s->evtchn_mask[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         if (unlikely((cpu != cpu_from_evtchn(port))))
  93                 do_hypercall = 1;
  94         else {
  95                 /*
  96                  * Need to clear the mask before checking pending to
  97                  * avoid a race with an event becoming pending.
  98                  *
  99                  * EVTCHNOP_unmask will only trigger an upcall if the
 100                  * mask bit was set, so if a hypercall is needed
 101                  * remask the event.
 102                  */
 103                 sync_clear_bit(port, BM(&s->evtchn_mask[0]));
 104                 evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0]));
 105
 106                 if (unlikely(evtchn_pending && xen_hvm_domain())) {
 107                         sync_set_bit(port, BM(&s->evtchn_mask[0]));
 108                         do_hypercall = 1;
 109                 }
 110         }
 111
 112         /* Slow path (hypercall) if this is a non-local port or if this is
 113          * an hvm domain and an event is pending (hvm domains don't have
 114          * their own implementation of irq_enable). */
 115         if (do_hypercall) {
 116                 struct evtchn_unmask unmask = { .port = port };
 117                 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask);
 118         } else {
 119                 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 120
 121                 /*
 122                  * The following is basically the equivalent of
 123                  * 'hw_resend_irq'. Just like a real IO-APIC we 'lose
 124                  * the interrupt edge' if the channel is masked.
 125                  */
 126                 if (evtchn_pending &&
 127                     !sync_test_and_set_bit(port / BITS_PER_EVTCHN_WORD,
 128                                            BM(&vcpu_info->evtchn_pending_sel)))
 129                         vcpu_info->evtchn_upcall_pending = 1;
 130         }
 131
 132         put_cpu();
 133 }
 134
 135 static DEFINE_PER_CPU(unsigned int, current_word_idx);
 136 static DEFINE_PER_CPU(unsigned int, current_bit_idx);
 137
 138 /*
 139  * Mask out the i least significant bits of w
 140  */
 141 #define MASK_LSBS(w, i) (w & ((~((xen_ulong_t)0UL)) << i))
 142
 143 static inline xen_ulong_t active_evtchns(unsigned int cpu,
 144                                          struct shared_info *sh,
 145                                          unsigned int idx)
 146 {
 147         return sh->evtchn_pending[idx] &
 148                 per_cpu(cpu_evtchn_mask, cpu)[idx] &
 149                 ~sh->evtchn_mask[idx];
 150 }
 151
 152 /*
 153  * Search the CPU's pending events bitmasks.  For each one found, map
 154  * the event number to an irq, and feed it into do_IRQ() for handling.
 155  *
 156  * Xen uses a two-level bitmap to speed searching.  The first level is
 157  * a bitset of words which contain pending event bits.  The second
 158  * level is a bitset of pending events themselves.
 159  */
 160 static void evtchn_2l_handle_events(unsigned cpu)
 161 {
 162         int irq;
 163         xen_ulong_t pending_words;
 164         xen_ulong_t pending_bits;
 165         int start_word_idx, start_bit_idx;
 166         int word_idx, bit_idx;
 167         int i;
 168         struct shared_info *s = HYPERVISOR_shared_info;
 169         struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu);
 170
 171         /* Timer interrupt has highest priority. */
 172         irq = irq_from_virq(cpu, VIRQ_TIMER);
 173         if (irq != -1) {
 174                 unsigned int evtchn = evtchn_from_irq(irq);
 175                 word_idx = evtchn / BITS_PER_LONG;
 176                 bit_idx = evtchn % BITS_PER_LONG;
 177                 if (active_evtchns(cpu, s, word_idx) & (1ULL << bit_idx))
 178                         generic_handle_irq(irq);
 179         }
 180
 181         /*
 182          * Master flag must be cleared /before/ clearing
 183          * selector flag. xchg_xen_ulong must contain an
 184          * appropriate barrier.
 185          */
 186         pending_words = xchg_xen_ulong(&vcpu_info->evtchn_pending_sel, 0);
 187
 188         start_word_idx = __this_cpu_read(current_word_idx);
 189         start_bit_idx = __this_cpu_read(current_bit_idx);
 190
 191         word_idx = start_word_idx;
 192
 193         for (i = 0; pending_words != 0; i++) {
 194                 xen_ulong_t words;
 195
 196                 words = MASK_LSBS(pending_words, word_idx);
 197
 198                 /*
 199                  * If we masked out all events, wrap to beginning.
 200                  */
 201                 if (words == 0) {
 202                         word_idx = 0;
 203                         bit_idx = 0;
 204                         continue;
 205                 }
 206                 word_idx = EVTCHN_FIRST_BIT(words);
 207
 208                 pending_bits = active_evtchns(cpu, s, word_idx);
 209                 bit_idx = 0; /* usually scan entire word from start */
 210                 /*
 211                  * We scan the starting word in two parts.
 212                  *
 213                  * 1st time: start in the middle, scanning the
 214                  * upper bits.
 215                  *
 216                  * 2nd time: scan the whole word (not just the
 217                  * parts skipped in the first pass) -- if an
 218                  * event in the previously scanned bits is
 219                  * pending again it would just be scanned on
 220                  * the next loop anyway.
 221                  */
 222                 if (word_idx == start_word_idx) {
 223                         if (i == 0)
 224                                 bit_idx = start_bit_idx;
 225                 }
 226
 227                 do {
 228                         xen_ulong_t bits;
 229                         int port;
 230
 231                         bits = MASK_LSBS(pending_bits, bit_idx);
 232
 233                         /* If we masked out all events, move on. */
 234                         if (bits == 0)
 235                                 break;
 236
 237                         bit_idx = EVTCHN_FIRST_BIT(bits);
 238
 239                         /* Process port. */
 240                         port = (word_idx * BITS_PER_EVTCHN_WORD) + bit_idx;
 241                         irq = get_evtchn_to_irq(port);
 242
 243                         if (irq != -1)
 244                                 generic_handle_irq(irq);
 245
 246                         bit_idx = (bit_idx + 1) % BITS_PER_EVTCHN_WORD;
 247
 248                         /* Next caller starts at last processed + 1 */
 249                         __this_cpu_write(current_word_idx,
 250                                          bit_idx ? word_idx :
 251                                          (word_idx+1) % BITS_PER_EVTCHN_WORD);
 252                         __this_cpu_write(current_bit_idx, bit_idx);
 253                 } while (bit_idx != 0);
 254
 255                 /* Scan start_l1i twice; all others once. */
 256                 if ((word_idx != start_word_idx) || (i != 0))
 257                         pending_words &= ~(1UL << word_idx);
 258
 259                 word_idx = (word_idx + 1) % BITS_PER_EVTCHN_WORD;
 260         }
 261 }
 262
 263 irqreturn_t xen_debug_interrupt(int irq, void *dev_id)
 264 {
 265         struct shared_info *sh = HYPERVISOR_shared_info;
 266         int cpu = smp_processor_id();
 267         xen_ulong_t *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu);
 268         int i;
 269         unsigned long flags;
 270         static DEFINE_SPINLOCK(debug_lock);
 271         struct vcpu_info *v;
 272
 273         spin_lock_irqsave(&debug_lock, flags);
 274
 275         printk("\nvcpu %d\n  ", cpu);
 276
 277         for_each_online_cpu(i) {
 278                 int pending;
 279                 v = per_cpu(xen_vcpu, i);
 280                 pending = (get_irq_regs() && i == cpu)
 281                         ? xen_irqs_disabled(get_irq_regs())
 282                         : v->evtchn_upcall_mask;
 283                 printk("%d: masked=%d pending=%d event_sel %0*"PRI_xen_ulong"\n  ", i,
 284                        pending, v->evtchn_upcall_pending,
 285                        (int)(sizeof(v->evtchn_pending_sel)*2),
 286                        v->evtchn_pending_sel);
 287         }
 288         v = per_cpu(xen_vcpu, cpu);
 289
 290         printk("\npending:\n   ");
 291         for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--)
 292                 printk("%0*"PRI_xen_ulong"%s",
 293                        (int)sizeof(sh->evtchn_pending[0])*2,
 294                        sh->evtchn_pending[i],
 295                        i % 8 == 0 ? "\n   " : " ");
 296         printk("\nglobal mask:\n   ");
 297         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 298                 printk("%0*"PRI_xen_ulong"%s",
 299                        (int)(sizeof(sh->evtchn_mask[0])*2),
 300                        sh->evtchn_mask[i],
 301                        i % 8 == 0 ? "\n   " : " ");
 302
 303         printk("\nglobally unmasked:\n   ");
 304         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--)
 305                 printk("%0*"PRI_xen_ulong"%s",
 306                        (int)(sizeof(sh->evtchn_mask[0])*2),
 307                        sh->evtchn_pending[i] & ~sh->evtchn_mask[i],
 308                        i % 8 == 0 ? "\n   " : " ");
 309
 310         printk("\nlocal cpu%d mask:\n   ", cpu);
 311         for (i = (EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD)-1; i >= 0; i--)
 312                 printk("%0*"PRI_xen_ulong"%s", (int)(sizeof(cpu_evtchn[0])*2),
 313                        cpu_evtchn[i],
 314                        i % 8 == 0 ? "\n   " : " ");
 315
 316         printk("\nlocally unmasked:\n   ");
 317         for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) {
 318                 xen_ulong_t pending = sh->evtchn_pending[i]
 319                         & ~sh->evtchn_mask[i]
 320                         & cpu_evtchn[i];
 321                 printk("%0*"PRI_xen_ulong"%s",
 322                        (int)(sizeof(sh->evtchn_mask[0])*2),
 323                        pending, i % 8 == 0 ? "\n   " : " ");
 324         }
 325
 326         printk("\npending list:\n");
 327         for (i = 0; i < EVTCHN_2L_NR_CHANNELS; i++) {
 328                 if (sync_test_bit(i, BM(sh->evtchn_pending))) {
 329                         int word_idx = i / BITS_PER_EVTCHN_WORD;
 330                         printk("  %d: event %d -> irq %d%s%s%s\n",
 331                                cpu_from_evtchn(i), i,
 332                                get_evtchn_to_irq(i),
 333                                sync_test_bit(word_idx, BM(&v->evtchn_pending_sel))
 334                                ? "" : " l2-clear",
 335                                !sync_test_bit(i, BM(sh->evtchn_mask))
 336                                ? "" : " globally-masked",
 337                                sync_test_bit(i, BM(cpu_evtchn))
 338                                ? "" : " locally-masked");
 339                 }
 340         }
 341
 342         spin_unlock_irqrestore(&debug_lock, flags);
 343
 344         return IRQ_HANDLED;
 345 }
 346
 347 static void evtchn_2l_resume(void)
 348 {
 349         int i;
 350
 351         for_each_online_cpu(i)
 352                 memset(per_cpu(cpu_evtchn_mask, i), 0, sizeof(xen_ulong_t) *
 353                                 EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD);
 354 }
 355
 356 static const struct evtchn_ops evtchn_ops_2l = {
 357         .max_channels      = evtchn_2l_max_channels,
 358         .nr_channels       = evtchn_2l_max_channels,
 359         .bind_to_cpu       = evtchn_2l_bind_to_cpu,
 360         .clear_pending     = evtchn_2l_clear_pending,
 361         .set_pending       = evtchn_2l_set_pending,
 362         .is_pending        = evtchn_2l_is_pending,
 363         .test_and_set_mask = evtchn_2l_test_and_set_mask,
 364         .mask              = evtchn_2l_mask,
 365         .unmask            = evtchn_2l_unmask,
 366         .handle_events     = evtchn_2l_handle_events,
 367         .resume            = evtchn_2l_resume,
 368 };
 369
 370 void __init xen_evtchn_2l_init(void)
 371 {
 372         pr_info("Using 2-level ABI\n");
 373         evtchn_ops = &evtchn_ops_2l;
 374 }