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 #include <linux/module.h> 13 14 #include <asm/sync_bitops.h> 15 #include <asm/xen/hypercall.h> 16 #include <asm/xen/hypervisor.h> 17 18 #include <xen/xen.h> 19 #include <xen/xen-ops.h> 20 #include <xen/events.h> 21 #include <xen/interface/xen.h> 22 #include <xen/interface/event_channel.h> 23 24 #include "events_internal.h" 25 26 /* 27 * Note sizeof(xen_ulong_t) can be more than sizeof(unsigned long). Be 28 * careful to only use bitops which allow for this (e.g 29 * test_bit/find_first_bit and friends but not __ffs) and to pass 30 * BITS_PER_EVTCHN_WORD as the bitmask length. 31 */ 32 #define BITS_PER_EVTCHN_WORD (sizeof(xen_ulong_t)*8) 33 /* 34 * Make a bitmask (i.e. unsigned long *) of a xen_ulong_t 35 * array. Primarily to avoid long lines (hence the terse name). 36 */ 37 #define BM(x) (unsigned long *)(x) 38 /* Find the first set bit in a evtchn mask */ 39 #define EVTCHN_FIRST_BIT(w) find_first_bit(BM(&(w)), BITS_PER_EVTCHN_WORD) 40 41 static DEFINE_PER_CPU(xen_ulong_t [EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD], 42 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_bind_to_cpu(struct irq_info *info, unsigned cpu) 50 { 51 clear_bit(info->evtchn, BM(per_cpu(cpu_evtchn_mask, info->cpu))); 52 set_bit(info->evtchn, BM(per_cpu(cpu_evtchn_mask, cpu))); 53 } 54 55 static void evtchn_2l_clear_pending(unsigned port) 56 { 57 struct shared_info *s = HYPERVISOR_shared_info; 58 sync_clear_bit(port, BM(&s->evtchn_pending[0])); 59 } 60 61 static void evtchn_2l_set_pending(unsigned port) 62 { 63 struct shared_info *s = HYPERVISOR_shared_info; 64 sync_set_bit(port, BM(&s->evtchn_pending[0])); 65 } 66 67 static bool evtchn_2l_is_pending(unsigned port) 68 { 69 struct shared_info *s = HYPERVISOR_shared_info; 70 return sync_test_bit(port, BM(&s->evtchn_pending[0])); 71 } 72 73 static bool evtchn_2l_test_and_set_mask(unsigned port) 74 { 75 struct shared_info *s = HYPERVISOR_shared_info; 76 return sync_test_and_set_bit(port, BM(&s->evtchn_mask[0])); 77 } 78 79 static void evtchn_2l_mask(unsigned port) 80 { 81 struct shared_info *s = HYPERVISOR_shared_info; 82 sync_set_bit(port, BM(&s->evtchn_mask[0])); 83 } 84 85 static void evtchn_2l_unmask(unsigned port) 86 { 87 struct shared_info *s = HYPERVISOR_shared_info; 88 unsigned int cpu = get_cpu(); 89 int do_hypercall = 0, evtchn_pending = 0; 90 91 BUG_ON(!irqs_disabled()); 92 93 if (unlikely((cpu != cpu_from_evtchn(port)))) 94 do_hypercall = 1; 95 else { 96 /* 97 * Need to clear the mask before checking pending to 98 * avoid a race with an event becoming pending. 99 * 100 * EVTCHNOP_unmask will only trigger an upcall if the 101 * mask bit was set, so if a hypercall is needed 102 * remask the event. 103 */ 104 sync_clear_bit(port, BM(&s->evtchn_mask[0])); 105 evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0])); 106 107 if (unlikely(evtchn_pending && xen_hvm_domain())) { 108 sync_set_bit(port, BM(&s->evtchn_mask[0])); 109 do_hypercall = 1; 110 } 111 } 112 113 /* Slow path (hypercall) if this is a non-local port or if this is 114 * an hvm domain and an event is pending (hvm domains don't have 115 * their own implementation of irq_enable). */ 116 if (do_hypercall) { 117 struct evtchn_unmask unmask = { .port = port }; 118 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask); 119 } else { 120 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu); 121 122 /* 123 * The following is basically the equivalent of 124 * 'hw_resend_irq'. Just like a real IO-APIC we 'lose 125 * the interrupt edge' if the channel is masked. 126 */ 127 if (evtchn_pending && 128 !sync_test_and_set_bit(port / BITS_PER_EVTCHN_WORD, 129 BM(&vcpu_info->evtchn_pending_sel))) 130 vcpu_info->evtchn_upcall_pending = 1; 131 } 132 133 put_cpu(); 134 } 135 136 static DEFINE_PER_CPU(unsigned int, current_word_idx); 137 static DEFINE_PER_CPU(unsigned int, current_bit_idx); 138 139 /* 140 * Mask out the i least significant bits of w 141 */ 142 #define MASK_LSBS(w, i) (w & ((~((xen_ulong_t)0UL)) << i)) 143 144 static inline xen_ulong_t active_evtchns(unsigned int cpu, 145 struct shared_info *sh, 146 unsigned int idx) 147 { 148 return sh->evtchn_pending[idx] & 149 per_cpu(cpu_evtchn_mask, cpu)[idx] & 150 ~sh->evtchn_mask[idx]; 151 } 152 153 /* 154 * Search the CPU's pending events bitmasks. For each one found, map 155 * the event number to an irq, and feed it into do_IRQ() for handling. 156 * 157 * Xen uses a two-level bitmap to speed searching. The first level is 158 * a bitset of words which contain pending event bits. The second 159 * level is a bitset of pending events themselves. 160 */ 161 static void evtchn_2l_handle_events(unsigned cpu) 162 { 163 int irq; 164 xen_ulong_t pending_words; 165 xen_ulong_t pending_bits; 166 int start_word_idx, start_bit_idx; 167 int word_idx, bit_idx; 168 int i; 169 struct irq_desc *desc; 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 desc = irq_to_desc(irq); 181 if (desc) 182 generic_handle_irq_desc(irq, desc); 183 } 184 } 185 186 /* 187 * Master flag must be cleared /before/ clearing 188 * selector flag. xchg_xen_ulong must contain an 189 * appropriate barrier. 190 */ 191 pending_words = xchg_xen_ulong(&vcpu_info->evtchn_pending_sel, 0); 192 193 start_word_idx = __this_cpu_read(current_word_idx); 194 start_bit_idx = __this_cpu_read(current_bit_idx); 195 196 word_idx = start_word_idx; 197 198 for (i = 0; pending_words != 0; i++) { 199 xen_ulong_t words; 200 201 words = MASK_LSBS(pending_words, word_idx); 202 203 /* 204 * If we masked out all events, wrap to beginning. 205 */ 206 if (words == 0) { 207 word_idx = 0; 208 bit_idx = 0; 209 continue; 210 } 211 word_idx = EVTCHN_FIRST_BIT(words); 212 213 pending_bits = active_evtchns(cpu, s, word_idx); 214 bit_idx = 0; /* usually scan entire word from start */ 215 /* 216 * We scan the starting word in two parts. 217 * 218 * 1st time: start in the middle, scanning the 219 * upper bits. 220 * 221 * 2nd time: scan the whole word (not just the 222 * parts skipped in the first pass) -- if an 223 * event in the previously scanned bits is 224 * pending again it would just be scanned on 225 * the next loop anyway. 226 */ 227 if (word_idx == start_word_idx) { 228 if (i == 0) 229 bit_idx = start_bit_idx; 230 } 231 232 do { 233 xen_ulong_t bits; 234 int port; 235 236 bits = MASK_LSBS(pending_bits, bit_idx); 237 238 /* If we masked out all events, move on. */ 239 if (bits == 0) 240 break; 241 242 bit_idx = EVTCHN_FIRST_BIT(bits); 243 244 /* Process port. */ 245 port = (word_idx * BITS_PER_EVTCHN_WORD) + bit_idx; 246 irq = get_evtchn_to_irq(port); 247 248 if (irq != -1) { 249 desc = irq_to_desc(irq); 250 if (desc) 251 generic_handle_irq_desc(irq, desc); 252 } 253 254 bit_idx = (bit_idx + 1) % BITS_PER_EVTCHN_WORD; 255 256 /* Next caller starts at last processed + 1 */ 257 __this_cpu_write(current_word_idx, 258 bit_idx ? word_idx : 259 (word_idx+1) % BITS_PER_EVTCHN_WORD); 260 __this_cpu_write(current_bit_idx, bit_idx); 261 } while (bit_idx != 0); 262 263 /* Scan start_l1i twice; all others once. */ 264 if ((word_idx != start_word_idx) || (i != 0)) 265 pending_words &= ~(1UL << word_idx); 266 267 word_idx = (word_idx + 1) % BITS_PER_EVTCHN_WORD; 268 } 269 } 270 271 irqreturn_t xen_debug_interrupt(int irq, void *dev_id) 272 { 273 struct shared_info *sh = HYPERVISOR_shared_info; 274 int cpu = smp_processor_id(); 275 xen_ulong_t *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu); 276 int i; 277 unsigned long flags; 278 static DEFINE_SPINLOCK(debug_lock); 279 struct vcpu_info *v; 280 281 spin_lock_irqsave(&debug_lock, flags); 282 283 printk("\nvcpu %d\n ", cpu); 284 285 for_each_online_cpu(i) { 286 int pending; 287 v = per_cpu(xen_vcpu, i); 288 pending = (get_irq_regs() && i == cpu) 289 ? xen_irqs_disabled(get_irq_regs()) 290 : v->evtchn_upcall_mask; 291 printk("%d: masked=%d pending=%d event_sel %0*"PRI_xen_ulong"\n ", i, 292 pending, v->evtchn_upcall_pending, 293 (int)(sizeof(v->evtchn_pending_sel)*2), 294 v->evtchn_pending_sel); 295 } 296 v = per_cpu(xen_vcpu, cpu); 297 298 printk("\npending:\n "); 299 for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--) 300 printk("%0*"PRI_xen_ulong"%s", 301 (int)sizeof(sh->evtchn_pending[0])*2, 302 sh->evtchn_pending[i], 303 i % 8 == 0 ? "\n " : " "); 304 printk("\nglobal mask:\n "); 305 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) 306 printk("%0*"PRI_xen_ulong"%s", 307 (int)(sizeof(sh->evtchn_mask[0])*2), 308 sh->evtchn_mask[i], 309 i % 8 == 0 ? "\n " : " "); 310 311 printk("\nglobally unmasked:\n "); 312 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) 313 printk("%0*"PRI_xen_ulong"%s", 314 (int)(sizeof(sh->evtchn_mask[0])*2), 315 sh->evtchn_pending[i] & ~sh->evtchn_mask[i], 316 i % 8 == 0 ? "\n " : " "); 317 318 printk("\nlocal cpu%d mask:\n ", cpu); 319 for (i = (EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD)-1; i >= 0; i--) 320 printk("%0*"PRI_xen_ulong"%s", (int)(sizeof(cpu_evtchn[0])*2), 321 cpu_evtchn[i], 322 i % 8 == 0 ? "\n " : " "); 323 324 printk("\nlocally unmasked:\n "); 325 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) { 326 xen_ulong_t pending = sh->evtchn_pending[i] 327 & ~sh->evtchn_mask[i] 328 & cpu_evtchn[i]; 329 printk("%0*"PRI_xen_ulong"%s", 330 (int)(sizeof(sh->evtchn_mask[0])*2), 331 pending, i % 8 == 0 ? "\n " : " "); 332 } 333 334 printk("\npending list:\n"); 335 for (i = 0; i < EVTCHN_2L_NR_CHANNELS; i++) { 336 if (sync_test_bit(i, BM(sh->evtchn_pending))) { 337 int word_idx = i / BITS_PER_EVTCHN_WORD; 338 printk(" %d: event %d -> irq %d%s%s%s\n", 339 cpu_from_evtchn(i), i, 340 get_evtchn_to_irq(i), 341 sync_test_bit(word_idx, BM(&v->evtchn_pending_sel)) 342 ? "" : " l2-clear", 343 !sync_test_bit(i, BM(sh->evtchn_mask)) 344 ? "" : " globally-masked", 345 sync_test_bit(i, BM(cpu_evtchn)) 346 ? "" : " locally-masked"); 347 } 348 } 349 350 spin_unlock_irqrestore(&debug_lock, flags); 351 352 return IRQ_HANDLED; 353 } 354 355 static const struct evtchn_ops evtchn_ops_2l = { 356 .max_channels = evtchn_2l_max_channels, 357 .nr_channels = evtchn_2l_max_channels, 358 .bind_to_cpu = evtchn_2l_bind_to_cpu, 359 .clear_pending = evtchn_2l_clear_pending, 360 .set_pending = evtchn_2l_set_pending, 361 .is_pending = evtchn_2l_is_pending, 362 .test_and_set_mask = evtchn_2l_test_and_set_mask, 363 .mask = evtchn_2l_mask, 364 .unmask = evtchn_2l_unmask, 365 .handle_events = evtchn_2l_handle_events, 366 }; 367 368 void __init xen_evtchn_2l_init(void) 369 { 370 pr_info("Using 2-level ABI\n"); 371 evtchn_ops = &evtchn_ops_2l; 372 } 373