1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Xen event channels (2-level ABI) 4 * 5 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007 6 */ 7 8 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt 9 10 #include <linux/linkage.h> 11 #include <linux/interrupt.h> 12 #include <linux/irq.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 #define EVTCHN_MASK_SIZE (EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD) 42 43 static DEFINE_PER_CPU(xen_ulong_t [EVTCHN_MASK_SIZE], cpu_evtchn_mask); 44 45 static unsigned evtchn_2l_max_channels(void) 46 { 47 return EVTCHN_2L_NR_CHANNELS; 48 } 49 50 static void evtchn_2l_bind_to_cpu(evtchn_port_t evtchn, unsigned int cpu, 51 unsigned int old_cpu) 52 { 53 clear_bit(evtchn, BM(per_cpu(cpu_evtchn_mask, old_cpu))); 54 set_bit(evtchn, BM(per_cpu(cpu_evtchn_mask, cpu))); 55 } 56 57 static void evtchn_2l_clear_pending(evtchn_port_t port) 58 { 59 struct shared_info *s = HYPERVISOR_shared_info; 60 sync_clear_bit(port, BM(&s->evtchn_pending[0])); 61 } 62 63 static void evtchn_2l_set_pending(evtchn_port_t port) 64 { 65 struct shared_info *s = HYPERVISOR_shared_info; 66 sync_set_bit(port, BM(&s->evtchn_pending[0])); 67 } 68 69 static bool evtchn_2l_is_pending(evtchn_port_t port) 70 { 71 struct shared_info *s = HYPERVISOR_shared_info; 72 return sync_test_bit(port, BM(&s->evtchn_pending[0])); 73 } 74 75 static bool evtchn_2l_test_and_set_mask(evtchn_port_t port) 76 { 77 struct shared_info *s = HYPERVISOR_shared_info; 78 return sync_test_and_set_bit(port, BM(&s->evtchn_mask[0])); 79 } 80 81 static void evtchn_2l_mask(evtchn_port_t port) 82 { 83 struct shared_info *s = HYPERVISOR_shared_info; 84 sync_set_bit(port, BM(&s->evtchn_mask[0])); 85 } 86 87 static void evtchn_2l_unmask(evtchn_port_t port) 88 { 89 struct shared_info *s = HYPERVISOR_shared_info; 90 unsigned int cpu = get_cpu(); 91 int do_hypercall = 0, evtchn_pending = 0; 92 93 BUG_ON(!irqs_disabled()); 94 95 smp_wmb(); /* All writes before unmask must be visible. */ 96 97 if (unlikely((cpu != cpu_from_evtchn(port)))) 98 do_hypercall = 1; 99 else { 100 /* 101 * Need to clear the mask before checking pending to 102 * avoid a race with an event becoming pending. 103 * 104 * EVTCHNOP_unmask will only trigger an upcall if the 105 * mask bit was set, so if a hypercall is needed 106 * remask the event. 107 */ 108 sync_clear_bit(port, BM(&s->evtchn_mask[0])); 109 evtchn_pending = sync_test_bit(port, BM(&s->evtchn_pending[0])); 110 111 if (unlikely(evtchn_pending && xen_hvm_domain())) { 112 sync_set_bit(port, BM(&s->evtchn_mask[0])); 113 do_hypercall = 1; 114 } 115 } 116 117 /* Slow path (hypercall) if this is a non-local port or if this is 118 * an hvm domain and an event is pending (hvm domains don't have 119 * their own implementation of irq_enable). */ 120 if (do_hypercall) { 121 struct evtchn_unmask unmask = { .port = port }; 122 (void)HYPERVISOR_event_channel_op(EVTCHNOP_unmask, &unmask); 123 } else { 124 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu); 125 126 /* 127 * The following is basically the equivalent of 128 * 'hw_resend_irq'. Just like a real IO-APIC we 'lose 129 * the interrupt edge' if the channel is masked. 130 */ 131 if (evtchn_pending && 132 !sync_test_and_set_bit(port / BITS_PER_EVTCHN_WORD, 133 BM(&vcpu_info->evtchn_pending_sel))) 134 vcpu_info->evtchn_upcall_pending = 1; 135 } 136 137 put_cpu(); 138 } 139 140 static DEFINE_PER_CPU(unsigned int, current_word_idx); 141 static DEFINE_PER_CPU(unsigned int, current_bit_idx); 142 143 /* 144 * Mask out the i least significant bits of w 145 */ 146 #define MASK_LSBS(w, i) (w & ((~((xen_ulong_t)0UL)) << i)) 147 148 static inline xen_ulong_t active_evtchns(unsigned int cpu, 149 struct shared_info *sh, 150 unsigned int idx) 151 { 152 return sh->evtchn_pending[idx] & 153 per_cpu(cpu_evtchn_mask, cpu)[idx] & 154 ~sh->evtchn_mask[idx]; 155 } 156 157 /* 158 * Search the CPU's pending events bitmasks. For each one found, map 159 * the event number to an irq, and feed it into do_IRQ() for handling. 160 * 161 * Xen uses a two-level bitmap to speed searching. The first level is 162 * a bitset of words which contain pending event bits. The second 163 * level is a bitset of pending events themselves. 164 */ 165 static void evtchn_2l_handle_events(unsigned cpu, struct evtchn_loop_ctrl *ctrl) 166 { 167 int irq; 168 xen_ulong_t pending_words; 169 xen_ulong_t pending_bits; 170 int start_word_idx, start_bit_idx; 171 int word_idx, bit_idx; 172 int i; 173 struct shared_info *s = HYPERVISOR_shared_info; 174 struct vcpu_info *vcpu_info = __this_cpu_read(xen_vcpu); 175 176 /* Timer interrupt has highest priority. */ 177 irq = irq_from_virq(cpu, VIRQ_TIMER); 178 if (irq != -1) { 179 evtchn_port_t evtchn = evtchn_from_irq(irq); 180 word_idx = evtchn / BITS_PER_LONG; 181 bit_idx = evtchn % BITS_PER_LONG; 182 if (active_evtchns(cpu, s, word_idx) & (1ULL << bit_idx)) 183 generic_handle_irq(irq); 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 evtchn_port_t 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 handle_irq_for_port(port, ctrl); 247 248 bit_idx = (bit_idx + 1) % BITS_PER_EVTCHN_WORD; 249 250 /* Next caller starts at last processed + 1 */ 251 __this_cpu_write(current_word_idx, 252 bit_idx ? word_idx : 253 (word_idx+1) % BITS_PER_EVTCHN_WORD); 254 __this_cpu_write(current_bit_idx, bit_idx); 255 } while (bit_idx != 0); 256 257 /* Scan start_l1i twice; all others once. */ 258 if ((word_idx != start_word_idx) || (i != 0)) 259 pending_words &= ~(1UL << word_idx); 260 261 word_idx = (word_idx + 1) % BITS_PER_EVTCHN_WORD; 262 } 263 } 264 265 irqreturn_t xen_debug_interrupt(int irq, void *dev_id) 266 { 267 struct shared_info *sh = HYPERVISOR_shared_info; 268 int cpu = smp_processor_id(); 269 xen_ulong_t *cpu_evtchn = per_cpu(cpu_evtchn_mask, cpu); 270 int i; 271 unsigned long flags; 272 static DEFINE_SPINLOCK(debug_lock); 273 struct vcpu_info *v; 274 275 spin_lock_irqsave(&debug_lock, flags); 276 277 printk("\nvcpu %d\n ", cpu); 278 279 for_each_online_cpu(i) { 280 int pending; 281 v = per_cpu(xen_vcpu, i); 282 pending = (get_irq_regs() && i == cpu) 283 ? xen_irqs_disabled(get_irq_regs()) 284 : v->evtchn_upcall_mask; 285 printk("%d: masked=%d pending=%d event_sel %0*"PRI_xen_ulong"\n ", i, 286 pending, v->evtchn_upcall_pending, 287 (int)(sizeof(v->evtchn_pending_sel)*2), 288 v->evtchn_pending_sel); 289 } 290 v = per_cpu(xen_vcpu, cpu); 291 292 printk("\npending:\n "); 293 for (i = ARRAY_SIZE(sh->evtchn_pending)-1; i >= 0; i--) 294 printk("%0*"PRI_xen_ulong"%s", 295 (int)sizeof(sh->evtchn_pending[0])*2, 296 sh->evtchn_pending[i], 297 i % 8 == 0 ? "\n " : " "); 298 printk("\nglobal mask:\n "); 299 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) 300 printk("%0*"PRI_xen_ulong"%s", 301 (int)(sizeof(sh->evtchn_mask[0])*2), 302 sh->evtchn_mask[i], 303 i % 8 == 0 ? "\n " : " "); 304 305 printk("\nglobally unmasked:\n "); 306 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) 307 printk("%0*"PRI_xen_ulong"%s", 308 (int)(sizeof(sh->evtchn_mask[0])*2), 309 sh->evtchn_pending[i] & ~sh->evtchn_mask[i], 310 i % 8 == 0 ? "\n " : " "); 311 312 printk("\nlocal cpu%d mask:\n ", cpu); 313 for (i = (EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD)-1; i >= 0; i--) 314 printk("%0*"PRI_xen_ulong"%s", (int)(sizeof(cpu_evtchn[0])*2), 315 cpu_evtchn[i], 316 i % 8 == 0 ? "\n " : " "); 317 318 printk("\nlocally unmasked:\n "); 319 for (i = ARRAY_SIZE(sh->evtchn_mask)-1; i >= 0; i--) { 320 xen_ulong_t pending = sh->evtchn_pending[i] 321 & ~sh->evtchn_mask[i] 322 & cpu_evtchn[i]; 323 printk("%0*"PRI_xen_ulong"%s", 324 (int)(sizeof(sh->evtchn_mask[0])*2), 325 pending, i % 8 == 0 ? "\n " : " "); 326 } 327 328 printk("\npending list:\n"); 329 for (i = 0; i < EVTCHN_2L_NR_CHANNELS; i++) { 330 if (sync_test_bit(i, BM(sh->evtchn_pending))) { 331 int word_idx = i / BITS_PER_EVTCHN_WORD; 332 printk(" %d: event %d -> irq %d%s%s%s\n", 333 cpu_from_evtchn(i), i, 334 get_evtchn_to_irq(i), 335 sync_test_bit(word_idx, BM(&v->evtchn_pending_sel)) 336 ? "" : " l2-clear", 337 !sync_test_bit(i, BM(sh->evtchn_mask)) 338 ? "" : " globally-masked", 339 sync_test_bit(i, BM(cpu_evtchn)) 340 ? "" : " locally-masked"); 341 } 342 } 343 344 spin_unlock_irqrestore(&debug_lock, flags); 345 346 return IRQ_HANDLED; 347 } 348 349 static void evtchn_2l_resume(void) 350 { 351 int i; 352 353 for_each_online_cpu(i) 354 memset(per_cpu(cpu_evtchn_mask, i), 0, sizeof(xen_ulong_t) * 355 EVTCHN_2L_NR_CHANNELS/BITS_PER_EVTCHN_WORD); 356 } 357 358 static const struct evtchn_ops evtchn_ops_2l = { 359 .max_channels = evtchn_2l_max_channels, 360 .nr_channels = evtchn_2l_max_channels, 361 .bind_to_cpu = evtchn_2l_bind_to_cpu, 362 .clear_pending = evtchn_2l_clear_pending, 363 .set_pending = evtchn_2l_set_pending, 364 .is_pending = evtchn_2l_is_pending, 365 .test_and_set_mask = evtchn_2l_test_and_set_mask, 366 .mask = evtchn_2l_mask, 367 .unmask = evtchn_2l_unmask, 368 .handle_events = evtchn_2l_handle_events, 369 .resume = evtchn_2l_resume, 370 }; 371 372 void __init xen_evtchn_2l_init(void) 373 { 374 pr_info("Using 2-level ABI\n"); 375 evtchn_ops = &evtchn_ops_2l; 376 } 377