arch/powerpc/kvm/book3s_xics.c

   1 /*
   2  * Copyright 2012 Michael Ellerman, IBM Corporation.
   3  * Copyright 2012 Benjamin Herrenschmidt, IBM Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License, version 2, as
   7  * published by the Free Software Foundation.
   8  */
   9
  10 #include <linux/kernel.h>
  11 #include <linux/kvm_host.h>
  12 #include <linux/err.h>
  13 #include <linux/gfp.h>
  14 #include <linux/anon_inodes.h>
  15
  16 #include <asm/uaccess.h>
  17 #include <asm/kvm_book3s.h>
  18 #include <asm/kvm_ppc.h>
  19 #include <asm/hvcall.h>
  20 #include <asm/xics.h>
  21 #include <asm/debug.h>
  22 #include <asm/time.h>
  23
  24 #include <linux/debugfs.h>
  25 #include <linux/seq_file.h>
  26
  27 #include "book3s_xics.h"
  28
  29 #if 1
  30 #define XICS_DBG(fmt...) do { } while (0)
  31 #else
  32 #define XICS_DBG(fmt...) trace_printk(fmt)
  33 #endif
  34
  35 #define ENABLE_REALMODE true
  36 #define DEBUG_REALMODE  false
  37
  38 /*
  39  * LOCKING
  40  * =======
  41  *
  42  * Each ICS has a mutex protecting the information about the IRQ
  43  * sources and avoiding simultaneous deliveries if the same interrupt.
  44  *
  45  * ICP operations are done via a single compare & swap transaction
  46  * (most ICP state fits in the union kvmppc_icp_state)
  47  */
  48
  49 /*
  50  * TODO
  51  * ====
  52  *
  53  * - To speed up resends, keep a bitmap of "resend" set bits in the
  54  *   ICS
  55  *
  56  * - Speed up server# -> ICP lookup (array ? hash table ?)
  57  *
  58  * - Make ICS lockless as well, or at least a per-interrupt lock or hashed
  59  *   locks array to improve scalability
  60  */
  61
  62 /* -- ICS routines -- */
  63
  64 static void icp_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
  65                             u32 new_irq);
  66
  67 static int ics_deliver_irq(struct kvmppc_xics *xics, u32 irq, u32 level,
  68                            bool report_status)
  69 {
  70         struct ics_irq_state *state;
  71         struct kvmppc_ics *ics;
  72         u16 src;
  73
  74         XICS_DBG("ics deliver %#x (level: %d)\n", irq, level);
  75
  76         ics = kvmppc_xics_find_ics(xics, irq, &src);
  77         if (!ics) {
  78                 XICS_DBG("ics_deliver_irq: IRQ 0x%06x not found !\n", irq);
  79                 return -EINVAL;
  80         }
  81         state = &ics->irq_state[src];
  82         if (!state->exists)
  83                 return -EINVAL;
  84
  85         if (report_status)
  86                 return state->asserted;
  87
  88         /*
  89          * We set state->asserted locklessly. This should be fine as
  90          * we are the only setter, thus concurrent access is undefined
  91          * to begin with.
  92          */
  93         if (level == KVM_INTERRUPT_SET_LEVEL)
  94                 state->asserted = 1;
  95         else if (level == KVM_INTERRUPT_UNSET) {
  96                 state->asserted = 0;
  97                 return 0;
  98         }
  99
 100         /* Attempt delivery */
 101         icp_deliver_irq(xics, NULL, irq);
 102
 103         return state->asserted;
 104 }
 105
 106 static void ics_check_resend(struct kvmppc_xics *xics, struct kvmppc_ics *ics,
 107                              struct kvmppc_icp *icp)
 108 {
 109         int i;
 110
 111         mutex_lock(&ics->lock);
 112
 113         for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
 114                 struct ics_irq_state *state = &ics->irq_state[i];
 115
 116                 if (!state->resend)
 117                         continue;
 118
 119                 XICS_DBG("resend %#x prio %#x\n", state->number,
 120                               state->priority);
 121
 122                 mutex_unlock(&ics->lock);
 123                 icp_deliver_irq(xics, icp, state->number);
 124                 mutex_lock(&ics->lock);
 125         }
 126
 127         mutex_unlock(&ics->lock);
 128 }
 129
 130 static bool write_xive(struct kvmppc_xics *xics, struct kvmppc_ics *ics,
 131                        struct ics_irq_state *state,
 132                        u32 server, u32 priority, u32 saved_priority)
 133 {
 134         bool deliver;
 135
 136         mutex_lock(&ics->lock);
 137
 138         state->server = server;
 139         state->priority = priority;
 140         state->saved_priority = saved_priority;
 141         deliver = false;
 142         if ((state->masked_pending || state->resend) && priority != MASKED) {
 143                 state->masked_pending = 0;
 144                 deliver = true;
 145         }
 146
 147         mutex_unlock(&ics->lock);
 148
 149         return deliver;
 150 }
 151
 152 int kvmppc_xics_set_xive(struct kvm *kvm, u32 irq, u32 server, u32 priority)
 153 {
 154         struct kvmppc_xics *xics = kvm->arch.xics;
 155         struct kvmppc_icp *icp;
 156         struct kvmppc_ics *ics;
 157         struct ics_irq_state *state;
 158         u16 src;
 159
 160         if (!xics)
 161                 return -ENODEV;
 162
 163         ics = kvmppc_xics_find_ics(xics, irq, &src);
 164         if (!ics)
 165                 return -EINVAL;
 166         state = &ics->irq_state[src];
 167
 168         icp = kvmppc_xics_find_server(kvm, server);
 169         if (!icp)
 170                 return -EINVAL;
 171
 172         XICS_DBG("set_xive %#x server %#x prio %#x MP:%d RS:%d\n",
 173                  irq, server, priority,
 174                  state->masked_pending, state->resend);
 175
 176         if (write_xive(xics, ics, state, server, priority, priority))
 177                 icp_deliver_irq(xics, icp, irq);
 178
 179         return 0;
 180 }
 181
 182 int kvmppc_xics_get_xive(struct kvm *kvm, u32 irq, u32 *server, u32 *priority)
 183 {
 184         struct kvmppc_xics *xics = kvm->arch.xics;
 185         struct kvmppc_ics *ics;
 186         struct ics_irq_state *state;
 187         u16 src;
 188
 189         if (!xics)
 190                 return -ENODEV;
 191
 192         ics = kvmppc_xics_find_ics(xics, irq, &src);
 193         if (!ics)
 194                 return -EINVAL;
 195         state = &ics->irq_state[src];
 196
 197         mutex_lock(&ics->lock);
 198         *server = state->server;
 199         *priority = state->priority;
 200         mutex_unlock(&ics->lock);
 201
 202         return 0;
 203 }
 204
 205 int kvmppc_xics_int_on(struct kvm *kvm, u32 irq)
 206 {
 207         struct kvmppc_xics *xics = kvm->arch.xics;
 208         struct kvmppc_icp *icp;
 209         struct kvmppc_ics *ics;
 210         struct ics_irq_state *state;
 211         u16 src;
 212
 213         if (!xics)
 214                 return -ENODEV;
 215
 216         ics = kvmppc_xics_find_ics(xics, irq, &src);
 217         if (!ics)
 218                 return -EINVAL;
 219         state = &ics->irq_state[src];
 220
 221         icp = kvmppc_xics_find_server(kvm, state->server);
 222         if (!icp)
 223                 return -EINVAL;
 224
 225         if (write_xive(xics, ics, state, state->server, state->saved_priority,
 226                        state->saved_priority))
 227                 icp_deliver_irq(xics, icp, irq);
 228
 229         return 0;
 230 }
 231
 232 int kvmppc_xics_int_off(struct kvm *kvm, u32 irq)
 233 {
 234         struct kvmppc_xics *xics = kvm->arch.xics;
 235         struct kvmppc_ics *ics;
 236         struct ics_irq_state *state;
 237         u16 src;
 238
 239         if (!xics)
 240                 return -ENODEV;
 241
 242         ics = kvmppc_xics_find_ics(xics, irq, &src);
 243         if (!ics)
 244                 return -EINVAL;
 245         state = &ics->irq_state[src];
 246
 247         write_xive(xics, ics, state, state->server, MASKED, state->priority);
 248
 249         return 0;
 250 }
 251
 252 /* -- ICP routines, including hcalls -- */
 253
 254 static inline bool icp_try_update(struct kvmppc_icp *icp,
 255                                   union kvmppc_icp_state old,
 256                                   union kvmppc_icp_state new,
 257                                   bool change_self)
 258 {
 259         bool success;
 260
 261         /* Calculate new output value */
 262         new.out_ee = (new.xisr && (new.pending_pri < new.cppr));
 263
 264         /* Attempt atomic update */
 265         success = cmpxchg64(&icp->state.raw, old.raw, new.raw) == old.raw;
 266         if (!success)
 267                 goto bail;
 268
 269         XICS_DBG("UPD [%04x] - C:%02x M:%02x PP: %02x PI:%06x R:%d O:%d\n",
 270                  icp->server_num,
 271                  old.cppr, old.mfrr, old.pending_pri, old.xisr,
 272                  old.need_resend, old.out_ee);
 273         XICS_DBG("UPD        - C:%02x M:%02x PP: %02x PI:%06x R:%d O:%d\n",
 274                  new.cppr, new.mfrr, new.pending_pri, new.xisr,
 275                  new.need_resend, new.out_ee);
 276         /*
 277          * Check for output state update
 278          *
 279          * Note that this is racy since another processor could be updating
 280          * the state already. This is why we never clear the interrupt output
 281          * here, we only ever set it. The clear only happens prior to doing
 282          * an update and only by the processor itself. Currently we do it
 283          * in Accept (H_XIRR) and Up_Cppr (H_XPPR).
 284          *
 285          * We also do not try to figure out whether the EE state has changed,
 286          * we unconditionally set it if the new state calls for it. The reason
 287          * for that is that we opportunistically remove the pending interrupt
 288          * flag when raising CPPR, so we need to set it back here if an
 289          * interrupt is still pending.
 290          */
 291         if (new.out_ee) {
 292                 kvmppc_book3s_queue_irqprio(icp->vcpu,
 293                                             BOOK3S_INTERRUPT_EXTERNAL_LEVEL);
 294                 if (!change_self)
 295                         kvmppc_fast_vcpu_kick(icp->vcpu);
 296         }
 297  bail:
 298         return success;
 299 }
 300
 301 static void icp_check_resend(struct kvmppc_xics *xics,
 302                              struct kvmppc_icp *icp)
 303 {
 304         u32 icsid;
 305
 306         /* Order this load with the test for need_resend in the caller */
 307         smp_rmb();
 308         for_each_set_bit(icsid, icp->resend_map, xics->max_icsid + 1) {
 309                 struct kvmppc_ics *ics = xics->ics[icsid];
 310
 311                 if (!test_and_clear_bit(icsid, icp->resend_map))
 312                         continue;
 313                 if (!ics)
 314                         continue;
 315                 ics_check_resend(xics, ics, icp);
 316         }
 317 }
 318
 319 static bool icp_try_to_deliver(struct kvmppc_icp *icp, u32 irq, u8 priority,
 320                                u32 *reject)
 321 {
 322         union kvmppc_icp_state old_state, new_state;
 323         bool success;
 324
 325         XICS_DBG("try deliver %#x(P:%#x) to server %#x\n", irq, priority,
 326                  icp->server_num);
 327
 328         do {
 329                 old_state = new_state = ACCESS_ONCE(icp->state);
 330
 331                 *reject = 0;
 332
 333                 /* See if we can deliver */
 334                 success = new_state.cppr > priority &&
 335                         new_state.mfrr > priority &&
 336                         new_state.pending_pri > priority;
 337
 338                 /*
 339                  * If we can, check for a rejection and perform the
 340                  * delivery
 341                  */
 342                 if (success) {
 343                         *reject = new_state.xisr;
 344                         new_state.xisr = irq;
 345                         new_state.pending_pri = priority;
 346                 } else {
 347                         /*
 348                          * If we failed to deliver we set need_resend
 349                          * so a subsequent CPPR state change causes us
 350                          * to try a new delivery.
 351                          */
 352                         new_state.need_resend = true;
 353                 }
 354
 355         } while (!icp_try_update(icp, old_state, new_state, false));
 356
 357         return success;
 358 }
 359
 360 static void icp_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
 361                             u32 new_irq)
 362 {
 363         struct ics_irq_state *state;
 364         struct kvmppc_ics *ics;
 365         u32 reject;
 366         u16 src;
 367
 368         /*
 369          * This is used both for initial delivery of an interrupt and
 370          * for subsequent rejection.
 371          *
 372          * Rejection can be racy vs. resends. We have evaluated the
 373          * rejection in an atomic ICP transaction which is now complete,
 374          * so potentially the ICP can already accept the interrupt again.
 375          *
 376          * So we need to retry the delivery. Essentially the reject path
 377          * boils down to a failed delivery. Always.
 378          *
 379          * Now the interrupt could also have moved to a different target,
 380          * thus we may need to re-do the ICP lookup as well
 381          */
 382
 383  again:
 384         /* Get the ICS state and lock it */
 385         ics = kvmppc_xics_find_ics(xics, new_irq, &src);
 386         if (!ics) {
 387                 XICS_DBG("icp_deliver_irq: IRQ 0x%06x not found !\n", new_irq);
 388                 return;
 389         }
 390         state = &ics->irq_state[src];
 391
 392         /* Get a lock on the ICS */
 393         mutex_lock(&ics->lock);
 394
 395         /* Get our server */
 396         if (!icp || state->server != icp->server_num) {
 397                 icp = kvmppc_xics_find_server(xics->kvm, state->server);
 398                 if (!icp) {
 399                         pr_warn("icp_deliver_irq: IRQ 0x%06x server 0x%x not found !\n",
 400                                 new_irq, state->server);
 401                         goto out;
 402                 }
 403         }
 404
 405         /* Clear the resend bit of that interrupt */
 406         state->resend = 0;
 407
 408         /*
 409          * If masked, bail out
 410          *
 411          * Note: PAPR doesn't mention anything about masked pending
 412          * when doing a resend, only when doing a delivery.
 413          *
 414          * However that would have the effect of losing a masked
 415          * interrupt that was rejected and isn't consistent with
 416          * the whole masked_pending business which is about not
 417          * losing interrupts that occur while masked.
 418          *
 419          * I don't differenciate normal deliveries and resends, this
 420          * implementation will differ from PAPR and not lose such
 421          * interrupts.
 422          */
 423         if (state->priority == MASKED) {
 424                 XICS_DBG("irq %#x masked pending\n", new_irq);
 425                 state->masked_pending = 1;
 426                 goto out;
 427         }
 428
 429         /*
 430          * Try the delivery, this will set the need_resend flag
 431          * in the ICP as part of the atomic transaction if the
 432          * delivery is not possible.
 433          *
 434          * Note that if successful, the new delivery might have itself
 435          * rejected an interrupt that was "delivered" before we took the
 436          * icp mutex.
 437          *
 438          * In this case we do the whole sequence all over again for the
 439          * new guy. We cannot assume that the rejected interrupt is less
 440          * favored than the new one, and thus doesn't need to be delivered,
 441          * because by the time we exit icp_try_to_deliver() the target
 442          * processor may well have alrady consumed & completed it, and thus
 443          * the rejected interrupt might actually be already acceptable.
 444          */
 445         if (icp_try_to_deliver(icp, new_irq, state->priority, &reject)) {
 446                 /*
 447                  * Delivery was successful, did we reject somebody else ?
 448                  */
 449                 if (reject && reject != XICS_IPI) {
 450                         mutex_unlock(&ics->lock);
 451                         new_irq = reject;
 452                         goto again;
 453                 }
 454         } else {
 455                 /*
 456                  * We failed to deliver the interrupt we need to set the
 457                  * resend map bit and mark the ICS state as needing a resend
 458                  */
 459                 set_bit(ics->icsid, icp->resend_map);
 460                 state->resend = 1;
 461
 462                 /*
 463                  * If the need_resend flag got cleared in the ICP some time
 464                  * between icp_try_to_deliver() atomic update and now, then
 465                  * we know it might have missed the resend_map bit. So we
 466                  * retry
 467                  */
 468                 smp_mb();
 469                 if (!icp->state.need_resend) {
 470                         mutex_unlock(&ics->lock);
 471                         goto again;
 472                 }
 473         }
 474  out:
 475         mutex_unlock(&ics->lock);
 476 }
 477
 478 static void icp_down_cppr(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
 479                           u8 new_cppr)
 480 {
 481         union kvmppc_icp_state old_state, new_state;
 482         bool resend;
 483
 484         /*
 485          * This handles several related states in one operation:
 486          *
 487          * ICP State: Down_CPPR
 488          *
 489          * Load CPPR with new value and if the XISR is 0
 490          * then check for resends:
 491          *
 492          * ICP State: Resend
 493          *
 494          * If MFRR is more favored than CPPR, check for IPIs
 495          * and notify ICS of a potential resend. This is done
 496          * asynchronously (when used in real mode, we will have
 497          * to exit here).
 498          *
 499          * We do not handle the complete Check_IPI as documented
 500          * here. In the PAPR, this state will be used for both
 501          * Set_MFRR and Down_CPPR. However, we know that we aren't
 502          * changing the MFRR state here so we don't need to handle
 503          * the case of an MFRR causing a reject of a pending irq,
 504          * this will have been handled when the MFRR was set in the
 505          * first place.
 506          *
 507          * Thus we don't have to handle rejects, only resends.
 508          *
 509          * When implementing real mode for HV KVM, resend will lead to
 510          * a H_TOO_HARD return and the whole transaction will be handled
 511          * in virtual mode.
 512          */
 513         do {
 514                 old_state = new_state = ACCESS_ONCE(icp->state);
 515
 516                 /* Down_CPPR */
 517                 new_state.cppr = new_cppr;
 518
 519                 /*
 520                  * Cut down Resend / Check_IPI / IPI
 521                  *
 522                  * The logic is that we cannot have a pending interrupt
 523                  * trumped by an IPI at this point (see above), so we
 524                  * know that either the pending interrupt is already an
 525                  * IPI (in which case we don't care to override it) or
 526                  * it's either more favored than us or non existent
 527                  */
 528                 if (new_state.mfrr < new_cppr &&
 529                     new_state.mfrr <= new_state.pending_pri) {
 530                         WARN_ON(new_state.xisr != XICS_IPI &&
 531                                 new_state.xisr != 0);
 532                         new_state.pending_pri = new_state.mfrr;
 533                         new_state.xisr = XICS_IPI;
 534                 }
 535
 536                 /* Latch/clear resend bit */
 537                 resend = new_state.need_resend;
 538                 new_state.need_resend = 0;
 539
 540         } while (!icp_try_update(icp, old_state, new_state, true));
 541
 542         /*
 543          * Now handle resend checks. Those are asynchronous to the ICP
 544          * state update in HW (ie bus transactions) so we can handle them
 545          * separately here too
 546          */
 547         if (resend)
 548                 icp_check_resend(xics, icp);
 549 }
 550
 551 static noinline unsigned long kvmppc_h_xirr(struct kvm_vcpu *vcpu)
 552 {
 553         union kvmppc_icp_state old_state, new_state;
 554         struct kvmppc_icp *icp = vcpu->arch.icp;
 555         u32 xirr;
 556
 557         /* First, remove EE from the processor */
 558         kvmppc_book3s_dequeue_irqprio(icp->vcpu,
 559                                       BOOK3S_INTERRUPT_EXTERNAL_LEVEL);
 560
 561         /*
 562          * ICP State: Accept_Interrupt
 563          *
 564          * Return the pending interrupt (if any) along with the
 565          * current CPPR, then clear the XISR & set CPPR to the
 566          * pending priority
 567          */
 568         do {
 569                 old_state = new_state = ACCESS_ONCE(icp->state);
 570
 571                 xirr = old_state.xisr | (((u32)old_state.cppr) << 24);
 572                 if (!old_state.xisr)
 573                         break;
 574                 new_state.cppr = new_state.pending_pri;
 575                 new_state.pending_pri = 0xff;
 576                 new_state.xisr = 0;
 577
 578         } while (!icp_try_update(icp, old_state, new_state, true));
 579
 580         XICS_DBG("h_xirr vcpu %d xirr %#x\n", vcpu->vcpu_id, xirr);
 581
 582         return xirr;
 583 }
 584
 585 static noinline int kvmppc_h_ipi(struct kvm_vcpu *vcpu, unsigned long server,
 586                                  unsigned long mfrr)
 587 {
 588         union kvmppc_icp_state old_state, new_state;
 589         struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
 590         struct kvmppc_icp *icp;
 591         u32 reject;
 592         bool resend;
 593         bool local;
 594
 595         XICS_DBG("h_ipi vcpu %d to server %lu mfrr %#lx\n",
 596                  vcpu->vcpu_id, server, mfrr);
 597
 598         icp = vcpu->arch.icp;
 599         local = icp->server_num == server;
 600         if (!local) {
 601                 icp = kvmppc_xics_find_server(vcpu->kvm, server);
 602                 if (!icp)
 603                         return H_PARAMETER;
 604         }
 605
 606         /*
 607          * ICP state: Set_MFRR
 608          *
 609          * If the CPPR is more favored than the new MFRR, then
 610          * nothing needs to be rejected as there can be no XISR to
 611          * reject.  If the MFRR is being made less favored then
 612          * there might be a previously-rejected interrupt needing
 613          * to be resent.
 614          *
 615          * If the CPPR is less favored, then we might be replacing
 616          * an interrupt, and thus need to possibly reject it as in
 617          *
 618          * ICP state: Check_IPI
 619          */
 620         do {
 621                 old_state = new_state = ACCESS_ONCE(icp->state);
 622
 623                 /* Set_MFRR */
 624                 new_state.mfrr = mfrr;
 625
 626                 /* Check_IPI */
 627                 reject = 0;
 628                 resend = false;
 629                 if (mfrr < new_state.cppr) {
 630                         /* Reject a pending interrupt if not an IPI */
 631                         if (mfrr <= new_state.pending_pri)
 632                                 reject = new_state.xisr;
 633                         new_state.pending_pri = mfrr;
 634                         new_state.xisr = XICS_IPI;
 635                 }
 636
 637                 if (mfrr > old_state.mfrr && mfrr > new_state.cppr) {
 638                         resend = new_state.need_resend;
 639                         new_state.need_resend = 0;
 640                 }
 641         } while (!icp_try_update(icp, old_state, new_state, local));
 642
 643         /* Handle reject */
 644         if (reject && reject != XICS_IPI)
 645                 icp_deliver_irq(xics, icp, reject);
 646
 647         /* Handle resend */
 648         if (resend)
 649                 icp_check_resend(xics, icp);
 650
 651         return H_SUCCESS;
 652 }
 653
 654 static int kvmppc_h_ipoll(struct kvm_vcpu *vcpu, unsigned long server)
 655 {
 656         union kvmppc_icp_state state;
 657         struct kvmppc_icp *icp;
 658
 659         icp = vcpu->arch.icp;
 660         if (icp->server_num != server) {
 661                 icp = kvmppc_xics_find_server(vcpu->kvm, server);
 662                 if (!icp)
 663                         return H_PARAMETER;
 664         }
 665         state = ACCESS_ONCE(icp->state);
 666         kvmppc_set_gpr(vcpu, 4, ((u32)state.cppr << 24) | state.xisr);
 667         kvmppc_set_gpr(vcpu, 5, state.mfrr);
 668         return H_SUCCESS;
 669 }
 670
 671 static noinline void kvmppc_h_cppr(struct kvm_vcpu *vcpu, unsigned long cppr)
 672 {
 673         union kvmppc_icp_state old_state, new_state;
 674         struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
 675         struct kvmppc_icp *icp = vcpu->arch.icp;
 676         u32 reject;
 677
 678         XICS_DBG("h_cppr vcpu %d cppr %#lx\n", vcpu->vcpu_id, cppr);
 679
 680         /*
 681          * ICP State: Set_CPPR
 682          *
 683          * We can safely compare the new value with the current
 684          * value outside of the transaction as the CPPR is only
 685          * ever changed by the processor on itself
 686          */
 687         if (cppr > icp->state.cppr)
 688                 icp_down_cppr(xics, icp, cppr);
 689         else if (cppr == icp->state.cppr)
 690                 return;
 691
 692         /*
 693          * ICP State: Up_CPPR
 694          *
 695          * The processor is raising its priority, this can result
 696          * in a rejection of a pending interrupt:
 697          *
 698          * ICP State: Reject_Current
 699          *
 700          * We can remove EE from the current processor, the update
 701          * transaction will set it again if needed
 702          */
 703         kvmppc_book3s_dequeue_irqprio(icp->vcpu,
 704                                       BOOK3S_INTERRUPT_EXTERNAL_LEVEL);
 705
 706         do {
 707                 old_state = new_state = ACCESS_ONCE(icp->state);
 708
 709                 reject = 0;
 710                 new_state.cppr = cppr;
 711
 712                 if (cppr <= new_state.pending_pri) {
 713                         reject = new_state.xisr;
 714                         new_state.xisr = 0;
 715                         new_state.pending_pri = 0xff;
 716                 }
 717
 718         } while (!icp_try_update(icp, old_state, new_state, true));
 719
 720         /*
 721          * Check for rejects. They are handled by doing a new delivery
 722          * attempt (see comments in icp_deliver_irq).
 723          */
 724         if (reject && reject != XICS_IPI)
 725                 icp_deliver_irq(xics, icp, reject);
 726 }
 727
 728 static noinline int kvmppc_h_eoi(struct kvm_vcpu *vcpu, unsigned long xirr)
 729 {
 730         struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
 731         struct kvmppc_icp *icp = vcpu->arch.icp;
 732         struct kvmppc_ics *ics;
 733         struct ics_irq_state *state;
 734         u32 irq = xirr & 0x00ffffff;
 735         u16 src;
 736
 737         XICS_DBG("h_eoi vcpu %d eoi %#lx\n", vcpu->vcpu_id, xirr);
 738
 739         /*
 740          * ICP State: EOI
 741          *
 742          * Note: If EOI is incorrectly used by SW to lower the CPPR
 743          * value (ie more favored), we do not check for rejection of
 744          * a pending interrupt, this is a SW error and PAPR sepcifies
 745          * that we don't have to deal with it.
 746          *
 747          * The sending of an EOI to the ICS is handled after the
 748          * CPPR update
 749          *
 750          * ICP State: Down_CPPR which we handle
 751          * in a separate function as it's shared with H_CPPR.
 752          */
 753         icp_down_cppr(xics, icp, xirr >> 24);
 754
 755         /* IPIs have no EOI */
 756         if (irq == XICS_IPI)
 757                 return H_SUCCESS;
 758         /*
 759          * EOI handling: If the interrupt is still asserted, we need to
 760          * resend it. We can take a lockless "peek" at the ICS state here.
 761          *
 762          * "Message" interrupts will never have "asserted" set
 763          */
 764         ics = kvmppc_xics_find_ics(xics, irq, &src);
 765         if (!ics) {
 766                 XICS_DBG("h_eoi: IRQ 0x%06x not found !\n", irq);
 767                 return H_PARAMETER;
 768         }
 769         state = &ics->irq_state[src];
 770
 771         /* Still asserted, resend it */
 772         if (state->asserted)
 773                 icp_deliver_irq(xics, icp, irq);
 774
 775         return H_SUCCESS;
 776 }
 777
 778 static noinline int kvmppc_xics_rm_complete(struct kvm_vcpu *vcpu, u32 hcall)
 779 {
 780         struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
 781         struct kvmppc_icp *icp = vcpu->arch.icp;
 782
 783         XICS_DBG("XICS_RM: H_%x completing, act: %x state: %lx tgt: %p\n",
 784                  hcall, icp->rm_action, icp->rm_dbgstate.raw, icp->rm_dbgtgt);
 785
 786         if (icp->rm_action & XICS_RM_KICK_VCPU)
 787                 kvmppc_fast_vcpu_kick(icp->rm_kick_target);
 788         if (icp->rm_action & XICS_RM_CHECK_RESEND)
 789                 icp_check_resend(xics, icp);
 790         if (icp->rm_action & XICS_RM_REJECT)
 791                 icp_deliver_irq(xics, icp, icp->rm_reject);
 792
 793         icp->rm_action = 0;
 794
 795         return H_SUCCESS;
 796 }
 797
 798 int kvmppc_xics_hcall(struct kvm_vcpu *vcpu, u32 req)
 799 {
 800         struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
 801         unsigned long res;
 802         int rc = H_SUCCESS;
 803
 804         /* Check if we have an ICP */
 805         if (!xics || !vcpu->arch.icp)
 806                 return H_HARDWARE;
 807
 808         /* These requests don't have real-mode implementations at present */
 809         switch (req) {
 810         case H_XIRR_X:
 811                 res = kvmppc_h_xirr(vcpu);
 812                 kvmppc_set_gpr(vcpu, 4, res);
 813                 kvmppc_set_gpr(vcpu, 5, get_tb());
 814                 return rc;
 815         case H_IPOLL:
 816                 rc = kvmppc_h_ipoll(vcpu, kvmppc_get_gpr(vcpu, 4));
 817                 return rc;
 818         }
 819
 820         /* Check for real mode returning too hard */
 821         if (xics->real_mode)
 822                 return kvmppc_xics_rm_complete(vcpu, req);
 823
 824         switch (req) {
 825         case H_XIRR:
 826                 res = kvmppc_h_xirr(vcpu);
 827                 kvmppc_set_gpr(vcpu, 4, res);
 828                 break;
 829         case H_CPPR:
 830                 kvmppc_h_cppr(vcpu, kvmppc_get_gpr(vcpu, 4));
 831                 break;
 832         case H_EOI:
 833                 rc = kvmppc_h_eoi(vcpu, kvmppc_get_gpr(vcpu, 4));
 834                 break;
 835         case H_IPI:
 836                 rc = kvmppc_h_ipi(vcpu, kvmppc_get_gpr(vcpu, 4),
 837                                   kvmppc_get_gpr(vcpu, 5));
 838                 break;
 839         }
 840
 841         return rc;
 842 }
 843
 844
 845 /* -- Initialisation code etc. -- */
 846
 847 static int xics_debug_show(struct seq_file *m, void *private)
 848 {
 849         struct kvmppc_xics *xics = m->private;
 850         struct kvm *kvm = xics->kvm;
 851         struct kvm_vcpu *vcpu;
 852         int icsid, i;
 853
 854         if (!kvm)
 855                 return 0;
 856
 857         seq_printf(m, "=========\nICP state\n=========\n");
 858
 859         kvm_for_each_vcpu(i, vcpu, kvm) {
 860                 struct kvmppc_icp *icp = vcpu->arch.icp;
 861                 union kvmppc_icp_state state;
 862
 863                 if (!icp)
 864                         continue;
 865
 866                 state.raw = ACCESS_ONCE(icp->state.raw);
 867                 seq_printf(m, "cpu server %#lx XIRR:%#x PPRI:%#x CPPR:%#x MFRR:%#x OUT:%d NR:%d\n",
 868                            icp->server_num, state.xisr,
 869                            state.pending_pri, state.cppr, state.mfrr,
 870                            state.out_ee, state.need_resend);
 871         }
 872
 873         for (icsid = 0; icsid <= KVMPPC_XICS_MAX_ICS_ID; icsid++) {
 874                 struct kvmppc_ics *ics = xics->ics[icsid];
 875
 876                 if (!ics)
 877                         continue;
 878
 879                 seq_printf(m, "=========\nICS state for ICS 0x%x\n=========\n",
 880                            icsid);
 881
 882                 mutex_lock(&ics->lock);
 883
 884                 for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
 885                         struct ics_irq_state *irq = &ics->irq_state[i];
 886
 887                         seq_printf(m, "irq 0x%06x: server %#x prio %#x save prio %#x asserted %d resend %d masked pending %d\n",
 888                                    irq->number, irq->server, irq->priority,
 889                                    irq->saved_priority, irq->asserted,
 890                                    irq->resend, irq->masked_pending);
 891
 892                 }
 893                 mutex_unlock(&ics->lock);
 894         }
 895         return 0;
 896 }
 897
 898 static int xics_debug_open(struct inode *inode, struct file *file)
 899 {
 900         return single_open(file, xics_debug_show, inode->i_private);
 901 }
 902
 903 static const struct file_operations xics_debug_fops = {
 904         .open = xics_debug_open,
 905         .read = seq_read,
 906         .llseek = seq_lseek,
 907         .release = single_release,
 908 };
 909
 910 static void xics_debugfs_init(struct kvmppc_xics *xics)
 911 {
 912         char *name;
 913
 914         name = kasprintf(GFP_KERNEL, "kvm-xics-%p", xics);
 915         if (!name) {
 916                 pr_err("%s: no memory for name\n", __func__);
 917                 return;
 918         }
 919
 920         xics->dentry = debugfs_create_file(name, S_IRUGO, powerpc_debugfs_root,
 921                                            xics, &xics_debug_fops);
 922
 923         pr_debug("%s: created %s\n", __func__, name);
 924         kfree(name);
 925 }
 926
 927 static struct kvmppc_ics *kvmppc_xics_create_ics(struct kvm *kvm,
 928                                         struct kvmppc_xics *xics, int irq)
 929 {
 930         struct kvmppc_ics *ics;
 931         int i, icsid;
 932
 933         icsid = irq >> KVMPPC_XICS_ICS_SHIFT;
 934
 935         mutex_lock(&kvm->lock);
 936
 937         /* ICS already exists - somebody else got here first */
 938         if (xics->ics[icsid])
 939                 goto out;
 940
 941         /* Create the ICS */
 942         ics = kzalloc(sizeof(struct kvmppc_ics), GFP_KERNEL);
 943         if (!ics)
 944                 goto out;
 945
 946         mutex_init(&ics->lock);
 947         ics->icsid = icsid;
 948
 949         for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
 950                 ics->irq_state[i].number = (icsid << KVMPPC_XICS_ICS_SHIFT) | i;
 951                 ics->irq_state[i].priority = MASKED;
 952                 ics->irq_state[i].saved_priority = MASKED;
 953         }
 954         smp_wmb();
 955         xics->ics[icsid] = ics;
 956
 957         if (icsid > xics->max_icsid)
 958                 xics->max_icsid = icsid;
 959
 960  out:
 961         mutex_unlock(&kvm->lock);
 962         return xics->ics[icsid];
 963 }
 964
 965 int kvmppc_xics_create_icp(struct kvm_vcpu *vcpu, unsigned long server_num)
 966 {
 967         struct kvmppc_icp *icp;
 968
 969         if (!vcpu->kvm->arch.xics)
 970                 return -ENODEV;
 971
 972         if (kvmppc_xics_find_server(vcpu->kvm, server_num))
 973                 return -EEXIST;
 974
 975         icp = kzalloc(sizeof(struct kvmppc_icp), GFP_KERNEL);
 976         if (!icp)
 977                 return -ENOMEM;
 978
 979         icp->vcpu = vcpu;
 980         icp->server_num = server_num;
 981         icp->state.mfrr = MASKED;
 982         icp->state.pending_pri = MASKED;
 983         vcpu->arch.icp = icp;
 984
 985         XICS_DBG("created server for vcpu %d\n", vcpu->vcpu_id);
 986
 987         return 0;
 988 }
 989
 990 u64 kvmppc_xics_get_icp(struct kvm_vcpu *vcpu)
 991 {
 992         struct kvmppc_icp *icp = vcpu->arch.icp;
 993         union kvmppc_icp_state state;
 994
 995         if (!icp)
 996                 return 0;
 997         state = icp->state;
 998         return ((u64)state.cppr << KVM_REG_PPC_ICP_CPPR_SHIFT) |
 999                 ((u64)state.xisr << KVM_REG_PPC_ICP_XISR_SHIFT) |
1000                 ((u64)state.mfrr << KVM_REG_PPC_ICP_MFRR_SHIFT) |
1001                 ((u64)state.pending_pri << KVM_REG_PPC_ICP_PPRI_SHIFT);
1002 }
1003
1004 int kvmppc_xics_set_icp(struct kvm_vcpu *vcpu, u64 icpval)
1005 {
1006         struct kvmppc_icp *icp = vcpu->arch.icp;
1007         struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
1008         union kvmppc_icp_state old_state, new_state;
1009         struct kvmppc_ics *ics;
1010         u8 cppr, mfrr, pending_pri;
1011         u32 xisr;
1012         u16 src;
1013         bool resend;
1014
1015         if (!icp || !xics)
1016                 return -ENOENT;
1017
1018         cppr = icpval >> KVM_REG_PPC_ICP_CPPR_SHIFT;
1019         xisr = (icpval >> KVM_REG_PPC_ICP_XISR_SHIFT) &
1020                 KVM_REG_PPC_ICP_XISR_MASK;
1021         mfrr = icpval >> KVM_REG_PPC_ICP_MFRR_SHIFT;
1022         pending_pri = icpval >> KVM_REG_PPC_ICP_PPRI_SHIFT;
1023
1024         /* Require the new state to be internally consistent */
1025         if (xisr == 0) {
1026                 if (pending_pri != 0xff)
1027                         return -EINVAL;
1028         } else if (xisr == XICS_IPI) {
1029                 if (pending_pri != mfrr || pending_pri >= cppr)
1030                         return -EINVAL;
1031         } else {
1032                 if (pending_pri >= mfrr || pending_pri >= cppr)
1033                         return -EINVAL;
1034                 ics = kvmppc_xics_find_ics(xics, xisr, &src);
1035                 if (!ics)
1036                         return -EINVAL;
1037         }
1038
1039         new_state.raw = 0;
1040         new_state.cppr = cppr;
1041         new_state.xisr = xisr;
1042         new_state.mfrr = mfrr;
1043         new_state.pending_pri = pending_pri;
1044
1045         /*
1046          * Deassert the CPU interrupt request.
1047          * icp_try_update will reassert it if necessary.
1048          */
1049         kvmppc_book3s_dequeue_irqprio(icp->vcpu,
1050                                       BOOK3S_INTERRUPT_EXTERNAL_LEVEL);
1051
1052         /*
1053          * Note that if we displace an interrupt from old_state.xisr,
1054          * we don't mark it as rejected.  We expect userspace to set
1055          * the state of the interrupt sources to be consistent with
1056          * the ICP states (either before or afterwards, which doesn't
1057          * matter).  We do handle resends due to CPPR becoming less
1058          * favoured because that is necessary to end up with a
1059          * consistent state in the situation where userspace restores
1060          * the ICS states before the ICP states.
1061          */
1062         do {
1063                 old_state = ACCESS_ONCE(icp->state);
1064
1065                 if (new_state.mfrr <= old_state.mfrr) {
1066                         resend = false;
1067                         new_state.need_resend = old_state.need_resend;
1068                 } else {
1069                         resend = old_state.need_resend;
1070                         new_state.need_resend = 0;
1071                 }
1072         } while (!icp_try_update(icp, old_state, new_state, false));
1073
1074         if (resend)
1075                 icp_check_resend(xics, icp);
1076
1077         return 0;
1078 }
1079
1080 static int xics_get_source(struct kvmppc_xics *xics, long irq, u64 addr)
1081 {
1082         int ret;
1083         struct kvmppc_ics *ics;
1084         struct ics_irq_state *irqp;
1085         u64 __user *ubufp = (u64 __user *) addr;
1086         u16 idx;
1087         u64 val, prio;
1088
1089         ics = kvmppc_xics_find_ics(xics, irq, &idx);
1090         if (!ics)
1091                 return -ENOENT;
1092
1093         irqp = &ics->irq_state[idx];
1094         mutex_lock(&ics->lock);
1095         ret = -ENOENT;
1096         if (irqp->exists) {
1097                 val = irqp->server;
1098                 prio = irqp->priority;
1099                 if (prio == MASKED) {
1100                         val |= KVM_XICS_MASKED;
1101                         prio = irqp->saved_priority;
1102                 }
1103                 val |= prio << KVM_XICS_PRIORITY_SHIFT;
1104                 if (irqp->asserted)
1105                         val |= KVM_XICS_LEVEL_SENSITIVE | KVM_XICS_PENDING;
1106                 else if (irqp->masked_pending || irqp->resend)
1107                         val |= KVM_XICS_PENDING;
1108                 ret = 0;
1109         }
1110         mutex_unlock(&ics->lock);
1111
1112         if (!ret && put_user(val, ubufp))
1113                 ret = -EFAULT;
1114
1115         return ret;
1116 }
1117
1118 static int xics_set_source(struct kvmppc_xics *xics, long irq, u64 addr)
1119 {
1120         struct kvmppc_ics *ics;
1121         struct ics_irq_state *irqp;
1122         u64 __user *ubufp = (u64 __user *) addr;
1123         u16 idx;
1124         u64 val;
1125         u8 prio;
1126         u32 server;
1127
1128         if (irq < KVMPPC_XICS_FIRST_IRQ || irq >= KVMPPC_XICS_NR_IRQS)
1129                 return -ENOENT;
1130
1131         ics = kvmppc_xics_find_ics(xics, irq, &idx);
1132         if (!ics) {
1133                 ics = kvmppc_xics_create_ics(xics->kvm, xics, irq);
1134                 if (!ics)
1135                         return -ENOMEM;
1136         }
1137         irqp = &ics->irq_state[idx];
1138         if (get_user(val, ubufp))
1139                 return -EFAULT;
1140
1141         server = val & KVM_XICS_DESTINATION_MASK;
1142         prio = val >> KVM_XICS_PRIORITY_SHIFT;
1143         if (prio != MASKED &&
1144             kvmppc_xics_find_server(xics->kvm, server) == NULL)
1145                 return -EINVAL;
1146
1147         mutex_lock(&ics->lock);
1148         irqp->server = server;
1149         irqp->saved_priority = prio;
1150         if (val & KVM_XICS_MASKED)
1151                 prio = MASKED;
1152         irqp->priority = prio;
1153         irqp->resend = 0;
1154         irqp->masked_pending = 0;
1155         irqp->asserted = 0;
1156         if ((val & KVM_XICS_PENDING) && (val & KVM_XICS_LEVEL_SENSITIVE))
1157                 irqp->asserted = 1;
1158         irqp->exists = 1;
1159         mutex_unlock(&ics->lock);
1160
1161         if (val & KVM_XICS_PENDING)
1162                 icp_deliver_irq(xics, NULL, irqp->number);
1163
1164         return 0;
1165 }
1166
1167 int kvm_set_irq(struct kvm *kvm, int irq_source_id, u32 irq, int level,
1168                 bool line_status)
1169 {
1170         struct kvmppc_xics *xics = kvm->arch.xics;
1171
1172         return ics_deliver_irq(xics, irq, level, line_status);
1173 }
1174
1175 static int xics_set_attr(struct kvm_device *dev, struct kvm_device_attr *attr)
1176 {
1177         struct kvmppc_xics *xics = dev->private;
1178
1179         switch (attr->group) {
1180         case KVM_DEV_XICS_GRP_SOURCES:
1181                 return xics_set_source(xics, attr->attr, attr->addr);
1182         }
1183         return -ENXIO;
1184 }
1185
1186 static int xics_get_attr(struct kvm_device *dev, struct kvm_device_attr *attr)
1187 {
1188         struct kvmppc_xics *xics = dev->private;
1189
1190         switch (attr->group) {
1191         case KVM_DEV_XICS_GRP_SOURCES:
1192                 return xics_get_source(xics, attr->attr, attr->addr);
1193         }
1194         return -ENXIO;
1195 }
1196
1197 static int xics_has_attr(struct kvm_device *dev, struct kvm_device_attr *attr)
1198 {
1199         switch (attr->group) {
1200         case KVM_DEV_XICS_GRP_SOURCES:
1201                 if (attr->attr >= KVMPPC_XICS_FIRST_IRQ &&
1202                     attr->attr < KVMPPC_XICS_NR_IRQS)
1203                         return 0;
1204                 break;
1205         }
1206         return -ENXIO;
1207 }
1208
1209 static void kvmppc_xics_free(struct kvm_device *dev)
1210 {
1211         struct kvmppc_xics *xics = dev->private;
1212         int i;
1213         struct kvm *kvm = xics->kvm;
1214
1215         debugfs_remove(xics->dentry);
1216
1217         if (kvm)
1218                 kvm->arch.xics = NULL;
1219
1220         for (i = 0; i <= xics->max_icsid; i++)
1221                 kfree(xics->ics[i]);
1222         kfree(xics);
1223         kfree(dev);
1224 }
1225
1226 static int kvmppc_xics_create(struct kvm_device *dev, u32 type)
1227 {
1228         struct kvmppc_xics *xics;
1229         struct kvm *kvm = dev->kvm;
1230         int ret = 0;
1231
1232         xics = kzalloc(sizeof(*xics), GFP_KERNEL);
1233         if (!xics)
1234                 return -ENOMEM;
1235
1236         dev->private = xics;
1237         xics->dev = dev;
1238         xics->kvm = kvm;
1239
1240         /* Already there ? */
1241         mutex_lock(&kvm->lock);
1242         if (kvm->arch.xics)
1243                 ret = -EEXIST;
1244         else
1245                 kvm->arch.xics = xics;
1246         mutex_unlock(&kvm->lock);
1247
1248         if (ret)
1249                 return ret;
1250
1251         xics_debugfs_init(xics);
1252
1253 #ifdef CONFIG_KVM_BOOK3S_64_HV
1254         if (cpu_has_feature(CPU_FTR_ARCH_206)) {
1255                 /* Enable real mode support */
1256                 xics->real_mode = ENABLE_REALMODE;
1257                 xics->real_mode_dbg = DEBUG_REALMODE;
1258         }
1259 #endif /* CONFIG_KVM_BOOK3S_64_HV */
1260
1261         return 0;
1262 }
1263
1264 struct kvm_device_ops kvm_xics_ops = {
1265         .name = "kvm-xics",
1266         .create = kvmppc_xics_create,
1267         .destroy = kvmppc_xics_free,
1268         .set_attr = xics_set_attr,
1269         .get_attr = xics_get_attr,
1270         .has_attr = xics_has_attr,
1271 };
1272
1273 int kvmppc_xics_connect_vcpu(struct kvm_device *dev, struct kvm_vcpu *vcpu,
1274                              u32 xcpu)
1275 {
1276         struct kvmppc_xics *xics = dev->private;
1277         int r = -EBUSY;
1278
1279         if (dev->ops != &kvm_xics_ops)
1280                 return -EPERM;
1281         if (xics->kvm != vcpu->kvm)
1282                 return -EPERM;
1283         if (vcpu->arch.irq_type)
1284                 return -EBUSY;
1285
1286         r = kvmppc_xics_create_icp(vcpu, xcpu);
1287         if (!r)
1288                 vcpu->arch.irq_type = KVMPPC_IRQ_XICS;
1289
1290         return r;
1291 }
1292
1293 void kvmppc_xics_free_icp(struct kvm_vcpu *vcpu)
1294 {
1295         if (!vcpu->arch.icp)
1296                 return;
1297         kfree(vcpu->arch.icp);
1298         vcpu->arch.icp = NULL;
1299         vcpu->arch.irq_type = KVMPPC_IRQ_DEFAULT;
1300 }