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/kernel_stat.h>
14 
15 #include <asm/kvm_book3s.h>
16 #include <asm/kvm_ppc.h>
17 #include <asm/hvcall.h>
18 #include <asm/xics.h>
19 #include <asm/synch.h>
20 #include <asm/cputhreads.h>
21 #include <asm/pgtable.h>
22 #include <asm/ppc-opcode.h>
23 #include <asm/pnv-pci.h>
24 #include <asm/opal.h>
25 #include <asm/smp.h>
26 
27 #include "book3s_xics.h"
28 
29 #define DEBUG_PASSUP
30 
31 int h_ipi_redirect = 1;
32 EXPORT_SYMBOL(h_ipi_redirect);
33 int kvm_irq_bypass = 1;
34 EXPORT_SYMBOL(kvm_irq_bypass);
35 
36 static void icp_rm_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
37 			    u32 new_irq, bool check_resend);
38 static int xics_opal_set_server(unsigned int hw_irq, int server_cpu);
39 
40 /* -- ICS routines -- */
41 static void ics_rm_check_resend(struct kvmppc_xics *xics,
42 				struct kvmppc_ics *ics, struct kvmppc_icp *icp)
43 {
44 	int i;
45 
46 	for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
47 		struct ics_irq_state *state = &ics->irq_state[i];
48 		if (state->resend)
49 			icp_rm_deliver_irq(xics, icp, state->number, true);
50 	}
51 
52 }
53 
54 /* -- ICP routines -- */
55 
56 #ifdef CONFIG_SMP
57 static inline void icp_send_hcore_msg(int hcore, struct kvm_vcpu *vcpu)
58 {
59 	int hcpu;
60 
61 	hcpu = hcore << threads_shift;
62 	kvmppc_host_rm_ops_hv->rm_core[hcore].rm_data = vcpu;
63 	smp_muxed_ipi_set_message(hcpu, PPC_MSG_RM_HOST_ACTION);
64 	kvmppc_set_host_ipi(hcpu, 1);
65 	smp_mb();
66 	kvmhv_rm_send_ipi(hcpu);
67 }
68 #else
69 static inline void icp_send_hcore_msg(int hcore, struct kvm_vcpu *vcpu) { }
70 #endif
71 
72 /*
73  * We start the search from our current CPU Id in the core map
74  * and go in a circle until we get back to our ID looking for a
75  * core that is running in host context and that hasn't already
76  * been targeted for another rm_host_ops.
77  *
78  * In the future, could consider using a fairer algorithm (one
79  * that distributes the IPIs better)
80  *
81  * Returns -1, if no CPU could be found in the host
82  * Else, returns a CPU Id which has been reserved for use
83  */
84 static inline int grab_next_hostcore(int start,
85 		struct kvmppc_host_rm_core *rm_core, int max, int action)
86 {
87 	bool success;
88 	int core;
89 	union kvmppc_rm_state old, new;
90 
91 	for (core = start + 1; core < max; core++)  {
92 		old = new = READ_ONCE(rm_core[core].rm_state);
93 
94 		if (!old.in_host || old.rm_action)
95 			continue;
96 
97 		/* Try to grab this host core if not taken already. */
98 		new.rm_action = action;
99 
100 		success = cmpxchg64(&rm_core[core].rm_state.raw,
101 						old.raw, new.raw) == old.raw;
102 		if (success) {
103 			/*
104 			 * Make sure that the store to the rm_action is made
105 			 * visible before we return to caller (and the
106 			 * subsequent store to rm_data) to synchronize with
107 			 * the IPI handler.
108 			 */
109 			smp_wmb();
110 			return core;
111 		}
112 	}
113 
114 	return -1;
115 }
116 
117 static inline int find_available_hostcore(int action)
118 {
119 	int core;
120 	int my_core = smp_processor_id() >> threads_shift;
121 	struct kvmppc_host_rm_core *rm_core = kvmppc_host_rm_ops_hv->rm_core;
122 
123 	core = grab_next_hostcore(my_core, rm_core, cpu_nr_cores(), action);
124 	if (core == -1)
125 		core = grab_next_hostcore(core, rm_core, my_core, action);
126 
127 	return core;
128 }
129 
130 static void icp_rm_set_vcpu_irq(struct kvm_vcpu *vcpu,
131 				struct kvm_vcpu *this_vcpu)
132 {
133 	struct kvmppc_icp *this_icp = this_vcpu->arch.icp;
134 	int cpu;
135 	int hcore;
136 
137 	/* Mark the target VCPU as having an interrupt pending */
138 	vcpu->stat.queue_intr++;
139 	set_bit(BOOK3S_IRQPRIO_EXTERNAL, &vcpu->arch.pending_exceptions);
140 
141 	/* Kick self ? Just set MER and return */
142 	if (vcpu == this_vcpu) {
143 		mtspr(SPRN_LPCR, mfspr(SPRN_LPCR) | LPCR_MER);
144 		return;
145 	}
146 
147 	if (xive_enabled() && kvmhv_on_pseries()) {
148 		/* No XICS access or hypercalls available, too hard */
149 		this_icp->rm_action |= XICS_RM_KICK_VCPU;
150 		this_icp->rm_kick_target = vcpu;
151 		return;
152 	}
153 
154 	/*
155 	 * Check if the core is loaded,
156 	 * if not, find an available host core to post to wake the VCPU,
157 	 * if we can't find one, set up state to eventually return too hard.
158 	 */
159 	cpu = vcpu->arch.thread_cpu;
160 	if (cpu < 0 || cpu >= nr_cpu_ids) {
161 		hcore = -1;
162 		if (kvmppc_host_rm_ops_hv && h_ipi_redirect)
163 			hcore = find_available_hostcore(XICS_RM_KICK_VCPU);
164 		if (hcore != -1) {
165 			icp_send_hcore_msg(hcore, vcpu);
166 		} else {
167 			this_icp->rm_action |= XICS_RM_KICK_VCPU;
168 			this_icp->rm_kick_target = vcpu;
169 		}
170 		return;
171 	}
172 
173 	smp_mb();
174 	kvmhv_rm_send_ipi(cpu);
175 }
176 
177 static void icp_rm_clr_vcpu_irq(struct kvm_vcpu *vcpu)
178 {
179 	/* Note: Only called on self ! */
180 	clear_bit(BOOK3S_IRQPRIO_EXTERNAL, &vcpu->arch.pending_exceptions);
181 	mtspr(SPRN_LPCR, mfspr(SPRN_LPCR) & ~LPCR_MER);
182 }
183 
184 static inline bool icp_rm_try_update(struct kvmppc_icp *icp,
185 				     union kvmppc_icp_state old,
186 				     union kvmppc_icp_state new)
187 {
188 	struct kvm_vcpu *this_vcpu = local_paca->kvm_hstate.kvm_vcpu;
189 	bool success;
190 
191 	/* Calculate new output value */
192 	new.out_ee = (new.xisr && (new.pending_pri < new.cppr));
193 
194 	/* Attempt atomic update */
195 	success = cmpxchg64(&icp->state.raw, old.raw, new.raw) == old.raw;
196 	if (!success)
197 		goto bail;
198 
199 	/*
200 	 * Check for output state update
201 	 *
202 	 * Note that this is racy since another processor could be updating
203 	 * the state already. This is why we never clear the interrupt output
204 	 * here, we only ever set it. The clear only happens prior to doing
205 	 * an update and only by the processor itself. Currently we do it
206 	 * in Accept (H_XIRR) and Up_Cppr (H_XPPR).
207 	 *
208 	 * We also do not try to figure out whether the EE state has changed,
209 	 * we unconditionally set it if the new state calls for it. The reason
210 	 * for that is that we opportunistically remove the pending interrupt
211 	 * flag when raising CPPR, so we need to set it back here if an
212 	 * interrupt is still pending.
213 	 */
214 	if (new.out_ee)
215 		icp_rm_set_vcpu_irq(icp->vcpu, this_vcpu);
216 
217 	/* Expose the state change for debug purposes */
218 	this_vcpu->arch.icp->rm_dbgstate = new;
219 	this_vcpu->arch.icp->rm_dbgtgt = icp->vcpu;
220 
221  bail:
222 	return success;
223 }
224 
225 static inline int check_too_hard(struct kvmppc_xics *xics,
226 				 struct kvmppc_icp *icp)
227 {
228 	return (xics->real_mode_dbg || icp->rm_action) ? H_TOO_HARD : H_SUCCESS;
229 }
230 
231 static void icp_rm_check_resend(struct kvmppc_xics *xics,
232 			     struct kvmppc_icp *icp)
233 {
234 	u32 icsid;
235 
236 	/* Order this load with the test for need_resend in the caller */
237 	smp_rmb();
238 	for_each_set_bit(icsid, icp->resend_map, xics->max_icsid + 1) {
239 		struct kvmppc_ics *ics = xics->ics[icsid];
240 
241 		if (!test_and_clear_bit(icsid, icp->resend_map))
242 			continue;
243 		if (!ics)
244 			continue;
245 		ics_rm_check_resend(xics, ics, icp);
246 	}
247 }
248 
249 static bool icp_rm_try_to_deliver(struct kvmppc_icp *icp, u32 irq, u8 priority,
250 			       u32 *reject)
251 {
252 	union kvmppc_icp_state old_state, new_state;
253 	bool success;
254 
255 	do {
256 		old_state = new_state = READ_ONCE(icp->state);
257 
258 		*reject = 0;
259 
260 		/* See if we can deliver */
261 		success = new_state.cppr > priority &&
262 			new_state.mfrr > priority &&
263 			new_state.pending_pri > priority;
264 
265 		/*
266 		 * If we can, check for a rejection and perform the
267 		 * delivery
268 		 */
269 		if (success) {
270 			*reject = new_state.xisr;
271 			new_state.xisr = irq;
272 			new_state.pending_pri = priority;
273 		} else {
274 			/*
275 			 * If we failed to deliver we set need_resend
276 			 * so a subsequent CPPR state change causes us
277 			 * to try a new delivery.
278 			 */
279 			new_state.need_resend = true;
280 		}
281 
282 	} while (!icp_rm_try_update(icp, old_state, new_state));
283 
284 	return success;
285 }
286 
287 static void icp_rm_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
288 			    u32 new_irq, bool check_resend)
289 {
290 	struct ics_irq_state *state;
291 	struct kvmppc_ics *ics;
292 	u32 reject;
293 	u16 src;
294 
295 	/*
296 	 * This is used both for initial delivery of an interrupt and
297 	 * for subsequent rejection.
298 	 *
299 	 * Rejection can be racy vs. resends. We have evaluated the
300 	 * rejection in an atomic ICP transaction which is now complete,
301 	 * so potentially the ICP can already accept the interrupt again.
302 	 *
303 	 * So we need to retry the delivery. Essentially the reject path
304 	 * boils down to a failed delivery. Always.
305 	 *
306 	 * Now the interrupt could also have moved to a different target,
307 	 * thus we may need to re-do the ICP lookup as well
308 	 */
309 
310  again:
311 	/* Get the ICS state and lock it */
312 	ics = kvmppc_xics_find_ics(xics, new_irq, &src);
313 	if (!ics) {
314 		/* Unsafe increment, but this does not need to be accurate */
315 		xics->err_noics++;
316 		return;
317 	}
318 	state = &ics->irq_state[src];
319 
320 	/* Get a lock on the ICS */
321 	arch_spin_lock(&ics->lock);
322 
323 	/* Get our server */
324 	if (!icp || state->server != icp->server_num) {
325 		icp = kvmppc_xics_find_server(xics->kvm, state->server);
326 		if (!icp) {
327 			/* Unsafe increment again*/
328 			xics->err_noicp++;
329 			goto out;
330 		}
331 	}
332 
333 	if (check_resend)
334 		if (!state->resend)
335 			goto out;
336 
337 	/* Clear the resend bit of that interrupt */
338 	state->resend = 0;
339 
340 	/*
341 	 * If masked, bail out
342 	 *
343 	 * Note: PAPR doesn't mention anything about masked pending
344 	 * when doing a resend, only when doing a delivery.
345 	 *
346 	 * However that would have the effect of losing a masked
347 	 * interrupt that was rejected and isn't consistent with
348 	 * the whole masked_pending business which is about not
349 	 * losing interrupts that occur while masked.
350 	 *
351 	 * I don't differentiate normal deliveries and resends, this
352 	 * implementation will differ from PAPR and not lose such
353 	 * interrupts.
354 	 */
355 	if (state->priority == MASKED) {
356 		state->masked_pending = 1;
357 		goto out;
358 	}
359 
360 	/*
361 	 * Try the delivery, this will set the need_resend flag
362 	 * in the ICP as part of the atomic transaction if the
363 	 * delivery is not possible.
364 	 *
365 	 * Note that if successful, the new delivery might have itself
366 	 * rejected an interrupt that was "delivered" before we took the
367 	 * ics spin lock.
368 	 *
369 	 * In this case we do the whole sequence all over again for the
370 	 * new guy. We cannot assume that the rejected interrupt is less
371 	 * favored than the new one, and thus doesn't need to be delivered,
372 	 * because by the time we exit icp_rm_try_to_deliver() the target
373 	 * processor may well have already consumed & completed it, and thus
374 	 * the rejected interrupt might actually be already acceptable.
375 	 */
376 	if (icp_rm_try_to_deliver(icp, new_irq, state->priority, &reject)) {
377 		/*
378 		 * Delivery was successful, did we reject somebody else ?
379 		 */
380 		if (reject && reject != XICS_IPI) {
381 			arch_spin_unlock(&ics->lock);
382 			icp->n_reject++;
383 			new_irq = reject;
384 			check_resend = 0;
385 			goto again;
386 		}
387 	} else {
388 		/*
389 		 * We failed to deliver the interrupt we need to set the
390 		 * resend map bit and mark the ICS state as needing a resend
391 		 */
392 		state->resend = 1;
393 
394 		/*
395 		 * Make sure when checking resend, we don't miss the resend
396 		 * if resend_map bit is seen and cleared.
397 		 */
398 		smp_wmb();
399 		set_bit(ics->icsid, icp->resend_map);
400 
401 		/*
402 		 * If the need_resend flag got cleared in the ICP some time
403 		 * between icp_rm_try_to_deliver() atomic update and now, then
404 		 * we know it might have missed the resend_map bit. So we
405 		 * retry
406 		 */
407 		smp_mb();
408 		if (!icp->state.need_resend) {
409 			state->resend = 0;
410 			arch_spin_unlock(&ics->lock);
411 			check_resend = 0;
412 			goto again;
413 		}
414 	}
415  out:
416 	arch_spin_unlock(&ics->lock);
417 }
418 
419 static void icp_rm_down_cppr(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
420 			     u8 new_cppr)
421 {
422 	union kvmppc_icp_state old_state, new_state;
423 	bool resend;
424 
425 	/*
426 	 * This handles several related states in one operation:
427 	 *
428 	 * ICP State: Down_CPPR
429 	 *
430 	 * Load CPPR with new value and if the XISR is 0
431 	 * then check for resends:
432 	 *
433 	 * ICP State: Resend
434 	 *
435 	 * If MFRR is more favored than CPPR, check for IPIs
436 	 * and notify ICS of a potential resend. This is done
437 	 * asynchronously (when used in real mode, we will have
438 	 * to exit here).
439 	 *
440 	 * We do not handle the complete Check_IPI as documented
441 	 * here. In the PAPR, this state will be used for both
442 	 * Set_MFRR and Down_CPPR. However, we know that we aren't
443 	 * changing the MFRR state here so we don't need to handle
444 	 * the case of an MFRR causing a reject of a pending irq,
445 	 * this will have been handled when the MFRR was set in the
446 	 * first place.
447 	 *
448 	 * Thus we don't have to handle rejects, only resends.
449 	 *
450 	 * When implementing real mode for HV KVM, resend will lead to
451 	 * a H_TOO_HARD return and the whole transaction will be handled
452 	 * in virtual mode.
453 	 */
454 	do {
455 		old_state = new_state = READ_ONCE(icp->state);
456 
457 		/* Down_CPPR */
458 		new_state.cppr = new_cppr;
459 
460 		/*
461 		 * Cut down Resend / Check_IPI / IPI
462 		 *
463 		 * The logic is that we cannot have a pending interrupt
464 		 * trumped by an IPI at this point (see above), so we
465 		 * know that either the pending interrupt is already an
466 		 * IPI (in which case we don't care to override it) or
467 		 * it's either more favored than us or non existent
468 		 */
469 		if (new_state.mfrr < new_cppr &&
470 		    new_state.mfrr <= new_state.pending_pri) {
471 			new_state.pending_pri = new_state.mfrr;
472 			new_state.xisr = XICS_IPI;
473 		}
474 
475 		/* Latch/clear resend bit */
476 		resend = new_state.need_resend;
477 		new_state.need_resend = 0;
478 
479 	} while (!icp_rm_try_update(icp, old_state, new_state));
480 
481 	/*
482 	 * Now handle resend checks. Those are asynchronous to the ICP
483 	 * state update in HW (ie bus transactions) so we can handle them
484 	 * separately here as well.
485 	 */
486 	if (resend) {
487 		icp->n_check_resend++;
488 		icp_rm_check_resend(xics, icp);
489 	}
490 }
491 
492 
493 unsigned long xics_rm_h_xirr(struct kvm_vcpu *vcpu)
494 {
495 	union kvmppc_icp_state old_state, new_state;
496 	struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
497 	struct kvmppc_icp *icp = vcpu->arch.icp;
498 	u32 xirr;
499 
500 	if (!xics || !xics->real_mode)
501 		return H_TOO_HARD;
502 
503 	/* First clear the interrupt */
504 	icp_rm_clr_vcpu_irq(icp->vcpu);
505 
506 	/*
507 	 * ICP State: Accept_Interrupt
508 	 *
509 	 * Return the pending interrupt (if any) along with the
510 	 * current CPPR, then clear the XISR & set CPPR to the
511 	 * pending priority
512 	 */
513 	do {
514 		old_state = new_state = READ_ONCE(icp->state);
515 
516 		xirr = old_state.xisr | (((u32)old_state.cppr) << 24);
517 		if (!old_state.xisr)
518 			break;
519 		new_state.cppr = new_state.pending_pri;
520 		new_state.pending_pri = 0xff;
521 		new_state.xisr = 0;
522 
523 	} while (!icp_rm_try_update(icp, old_state, new_state));
524 
525 	/* Return the result in GPR4 */
526 	vcpu->arch.regs.gpr[4] = xirr;
527 
528 	return check_too_hard(xics, icp);
529 }
530 
531 int xics_rm_h_ipi(struct kvm_vcpu *vcpu, unsigned long server,
532 		  unsigned long mfrr)
533 {
534 	union kvmppc_icp_state old_state, new_state;
535 	struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
536 	struct kvmppc_icp *icp, *this_icp = vcpu->arch.icp;
537 	u32 reject;
538 	bool resend;
539 	bool local;
540 
541 	if (!xics || !xics->real_mode)
542 		return H_TOO_HARD;
543 
544 	local = this_icp->server_num == server;
545 	if (local)
546 		icp = this_icp;
547 	else
548 		icp = kvmppc_xics_find_server(vcpu->kvm, server);
549 	if (!icp)
550 		return H_PARAMETER;
551 
552 	/*
553 	 * ICP state: Set_MFRR
554 	 *
555 	 * If the CPPR is more favored than the new MFRR, then
556 	 * nothing needs to be done as there can be no XISR to
557 	 * reject.
558 	 *
559 	 * ICP state: Check_IPI
560 	 *
561 	 * If the CPPR is less favored, then we might be replacing
562 	 * an interrupt, and thus need to possibly reject it.
563 	 *
564 	 * ICP State: IPI
565 	 *
566 	 * Besides rejecting any pending interrupts, we also
567 	 * update XISR and pending_pri to mark IPI as pending.
568 	 *
569 	 * PAPR does not describe this state, but if the MFRR is being
570 	 * made less favored than its earlier value, there might be
571 	 * a previously-rejected interrupt needing to be resent.
572 	 * Ideally, we would want to resend only if
573 	 *	prio(pending_interrupt) < mfrr &&
574 	 *	prio(pending_interrupt) < cppr
575 	 * where pending interrupt is the one that was rejected. But
576 	 * we don't have that state, so we simply trigger a resend
577 	 * whenever the MFRR is made less favored.
578 	 */
579 	do {
580 		old_state = new_state = READ_ONCE(icp->state);
581 
582 		/* Set_MFRR */
583 		new_state.mfrr = mfrr;
584 
585 		/* Check_IPI */
586 		reject = 0;
587 		resend = false;
588 		if (mfrr < new_state.cppr) {
589 			/* Reject a pending interrupt if not an IPI */
590 			if (mfrr <= new_state.pending_pri) {
591 				reject = new_state.xisr;
592 				new_state.pending_pri = mfrr;
593 				new_state.xisr = XICS_IPI;
594 			}
595 		}
596 
597 		if (mfrr > old_state.mfrr) {
598 			resend = new_state.need_resend;
599 			new_state.need_resend = 0;
600 		}
601 	} while (!icp_rm_try_update(icp, old_state, new_state));
602 
603 	/* Handle reject in real mode */
604 	if (reject && reject != XICS_IPI) {
605 		this_icp->n_reject++;
606 		icp_rm_deliver_irq(xics, icp, reject, false);
607 	}
608 
609 	/* Handle resends in real mode */
610 	if (resend) {
611 		this_icp->n_check_resend++;
612 		icp_rm_check_resend(xics, icp);
613 	}
614 
615 	return check_too_hard(xics, this_icp);
616 }
617 
618 int xics_rm_h_cppr(struct kvm_vcpu *vcpu, unsigned long cppr)
619 {
620 	union kvmppc_icp_state old_state, new_state;
621 	struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
622 	struct kvmppc_icp *icp = vcpu->arch.icp;
623 	u32 reject;
624 
625 	if (!xics || !xics->real_mode)
626 		return H_TOO_HARD;
627 
628 	/*
629 	 * ICP State: Set_CPPR
630 	 *
631 	 * We can safely compare the new value with the current
632 	 * value outside of the transaction as the CPPR is only
633 	 * ever changed by the processor on itself
634 	 */
635 	if (cppr > icp->state.cppr) {
636 		icp_rm_down_cppr(xics, icp, cppr);
637 		goto bail;
638 	} else if (cppr == icp->state.cppr)
639 		return H_SUCCESS;
640 
641 	/*
642 	 * ICP State: Up_CPPR
643 	 *
644 	 * The processor is raising its priority, this can result
645 	 * in a rejection of a pending interrupt:
646 	 *
647 	 * ICP State: Reject_Current
648 	 *
649 	 * We can remove EE from the current processor, the update
650 	 * transaction will set it again if needed
651 	 */
652 	icp_rm_clr_vcpu_irq(icp->vcpu);
653 
654 	do {
655 		old_state = new_state = READ_ONCE(icp->state);
656 
657 		reject = 0;
658 		new_state.cppr = cppr;
659 
660 		if (cppr <= new_state.pending_pri) {
661 			reject = new_state.xisr;
662 			new_state.xisr = 0;
663 			new_state.pending_pri = 0xff;
664 		}
665 
666 	} while (!icp_rm_try_update(icp, old_state, new_state));
667 
668 	/*
669 	 * Check for rejects. They are handled by doing a new delivery
670 	 * attempt (see comments in icp_rm_deliver_irq).
671 	 */
672 	if (reject && reject != XICS_IPI) {
673 		icp->n_reject++;
674 		icp_rm_deliver_irq(xics, icp, reject, false);
675 	}
676  bail:
677 	return check_too_hard(xics, icp);
678 }
679 
680 static int ics_rm_eoi(struct kvm_vcpu *vcpu, u32 irq)
681 {
682 	struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
683 	struct kvmppc_icp *icp = vcpu->arch.icp;
684 	struct kvmppc_ics *ics;
685 	struct ics_irq_state *state;
686 	u16 src;
687 	u32 pq_old, pq_new;
688 
689 	/*
690 	 * ICS EOI handling: For LSI, if P bit is still set, we need to
691 	 * resend it.
692 	 *
693 	 * For MSI, we move Q bit into P (and clear Q). If it is set,
694 	 * resend it.
695 	 */
696 
697 	ics = kvmppc_xics_find_ics(xics, irq, &src);
698 	if (!ics)
699 		goto bail;
700 
701 	state = &ics->irq_state[src];
702 
703 	if (state->lsi)
704 		pq_new = state->pq_state;
705 	else
706 		do {
707 			pq_old = state->pq_state;
708 			pq_new = pq_old >> 1;
709 		} while (cmpxchg(&state->pq_state, pq_old, pq_new) != pq_old);
710 
711 	if (pq_new & PQ_PRESENTED)
712 		icp_rm_deliver_irq(xics, NULL, irq, false);
713 
714 	if (!hlist_empty(&vcpu->kvm->irq_ack_notifier_list)) {
715 		icp->rm_action |= XICS_RM_NOTIFY_EOI;
716 		icp->rm_eoied_irq = irq;
717 	}
718 
719 	if (state->host_irq) {
720 		++vcpu->stat.pthru_all;
721 		if (state->intr_cpu != -1) {
722 			int pcpu = raw_smp_processor_id();
723 
724 			pcpu = cpu_first_thread_sibling(pcpu);
725 			++vcpu->stat.pthru_host;
726 			if (state->intr_cpu != pcpu) {
727 				++vcpu->stat.pthru_bad_aff;
728 				xics_opal_set_server(state->host_irq, pcpu);
729 			}
730 			state->intr_cpu = -1;
731 		}
732 	}
733 
734  bail:
735 	return check_too_hard(xics, icp);
736 }
737 
738 int xics_rm_h_eoi(struct kvm_vcpu *vcpu, unsigned long xirr)
739 {
740 	struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
741 	struct kvmppc_icp *icp = vcpu->arch.icp;
742 	u32 irq = xirr & 0x00ffffff;
743 
744 	if (!xics || !xics->real_mode)
745 		return H_TOO_HARD;
746 
747 	/*
748 	 * ICP State: EOI
749 	 *
750 	 * Note: If EOI is incorrectly used by SW to lower the CPPR
751 	 * value (ie more favored), we do not check for rejection of
752 	 * a pending interrupt, this is a SW error and PAPR specifies
753 	 * that we don't have to deal with it.
754 	 *
755 	 * The sending of an EOI to the ICS is handled after the
756 	 * CPPR update
757 	 *
758 	 * ICP State: Down_CPPR which we handle
759 	 * in a separate function as it's shared with H_CPPR.
760 	 */
761 	icp_rm_down_cppr(xics, icp, xirr >> 24);
762 
763 	/* IPIs have no EOI */
764 	if (irq == XICS_IPI)
765 		return check_too_hard(xics, icp);
766 
767 	return ics_rm_eoi(vcpu, irq);
768 }
769 
770 unsigned long eoi_rc;
771 
772 static void icp_eoi(struct irq_chip *c, u32 hwirq, __be32 xirr, bool *again)
773 {
774 	void __iomem *xics_phys;
775 	int64_t rc;
776 
777 	if (kvmhv_on_pseries()) {
778 		unsigned long retbuf[PLPAR_HCALL_BUFSIZE];
779 
780 		iosync();
781 		plpar_hcall_raw(H_EOI, retbuf, hwirq);
782 		return;
783 	}
784 
785 	rc = pnv_opal_pci_msi_eoi(c, hwirq);
786 
787 	if (rc)
788 		eoi_rc = rc;
789 
790 	iosync();
791 
792 	/* EOI it */
793 	xics_phys = local_paca->kvm_hstate.xics_phys;
794 	if (xics_phys) {
795 		__raw_rm_writel(xirr, xics_phys + XICS_XIRR);
796 	} else {
797 		rc = opal_int_eoi(be32_to_cpu(xirr));
798 		*again = rc > 0;
799 	}
800 }
801 
802 static int xics_opal_set_server(unsigned int hw_irq, int server_cpu)
803 {
804 	unsigned int mangle_cpu = get_hard_smp_processor_id(server_cpu) << 2;
805 
806 	return opal_set_xive(hw_irq, mangle_cpu, DEFAULT_PRIORITY);
807 }
808 
809 /*
810  * Increment a per-CPU 32-bit unsigned integer variable.
811  * Safe to call in real-mode. Handles vmalloc'ed addresses
812  *
813  * ToDo: Make this work for any integral type
814  */
815 
816 static inline void this_cpu_inc_rm(unsigned int __percpu *addr)
817 {
818 	unsigned long l;
819 	unsigned int *raddr;
820 	int cpu = smp_processor_id();
821 
822 	raddr = per_cpu_ptr(addr, cpu);
823 	l = (unsigned long)raddr;
824 
825 	if (REGION_ID(l) == VMALLOC_REGION_ID) {
826 		l = vmalloc_to_phys(raddr);
827 		raddr = (unsigned int *)l;
828 	}
829 	++*raddr;
830 }
831 
832 /*
833  * We don't try to update the flags in the irq_desc 'istate' field in
834  * here as would happen in the normal IRQ handling path for several reasons:
835  *  - state flags represent internal IRQ state and are not expected to be
836  *    updated outside the IRQ subsystem
837  *  - more importantly, these are useful for edge triggered interrupts,
838  *    IRQ probing, etc., but we are only handling MSI/MSIx interrupts here
839  *    and these states shouldn't apply to us.
840  *
841  * However, we do update irq_stats - we somewhat duplicate the code in
842  * kstat_incr_irqs_this_cpu() for this since this function is defined
843  * in irq/internal.h which we don't want to include here.
844  * The only difference is that desc->kstat_irqs is an allocated per CPU
845  * variable and could have been vmalloc'ed, so we can't directly
846  * call __this_cpu_inc() on it. The kstat structure is a static
847  * per CPU variable and it should be accessible by real-mode KVM.
848  *
849  */
850 static void kvmppc_rm_handle_irq_desc(struct irq_desc *desc)
851 {
852 	this_cpu_inc_rm(desc->kstat_irqs);
853 	__this_cpu_inc(kstat.irqs_sum);
854 }
855 
856 long kvmppc_deliver_irq_passthru(struct kvm_vcpu *vcpu,
857 				 __be32 xirr,
858 				 struct kvmppc_irq_map *irq_map,
859 				 struct kvmppc_passthru_irqmap *pimap,
860 				 bool *again)
861 {
862 	struct kvmppc_xics *xics;
863 	struct kvmppc_icp *icp;
864 	struct kvmppc_ics *ics;
865 	struct ics_irq_state *state;
866 	u32 irq;
867 	u16 src;
868 	u32 pq_old, pq_new;
869 
870 	irq = irq_map->v_hwirq;
871 	xics = vcpu->kvm->arch.xics;
872 	icp = vcpu->arch.icp;
873 
874 	kvmppc_rm_handle_irq_desc(irq_map->desc);
875 
876 	ics = kvmppc_xics_find_ics(xics, irq, &src);
877 	if (!ics)
878 		return 2;
879 
880 	state = &ics->irq_state[src];
881 
882 	/* only MSIs register bypass producers, so it must be MSI here */
883 	do {
884 		pq_old = state->pq_state;
885 		pq_new = ((pq_old << 1) & 3) | PQ_PRESENTED;
886 	} while (cmpxchg(&state->pq_state, pq_old, pq_new) != pq_old);
887 
888 	/* Test P=1, Q=0, this is the only case where we present */
889 	if (pq_new == PQ_PRESENTED)
890 		icp_rm_deliver_irq(xics, icp, irq, false);
891 
892 	/* EOI the interrupt */
893 	icp_eoi(irq_desc_get_chip(irq_map->desc), irq_map->r_hwirq, xirr,
894 		again);
895 
896 	if (check_too_hard(xics, icp) == H_TOO_HARD)
897 		return 2;
898 	else
899 		return -2;
900 }
901 
902 /*  --- Non-real mode XICS-related built-in routines ---  */
903 
904 /**
905  * Host Operations poked by RM KVM
906  */
907 static void rm_host_ipi_action(int action, void *data)
908 {
909 	switch (action) {
910 	case XICS_RM_KICK_VCPU:
911 		kvmppc_host_rm_ops_hv->vcpu_kick(data);
912 		break;
913 	default:
914 		WARN(1, "Unexpected rm_action=%d data=%p\n", action, data);
915 		break;
916 	}
917 
918 }
919 
920 void kvmppc_xics_ipi_action(void)
921 {
922 	int core;
923 	unsigned int cpu = smp_processor_id();
924 	struct kvmppc_host_rm_core *rm_corep;
925 
926 	core = cpu >> threads_shift;
927 	rm_corep = &kvmppc_host_rm_ops_hv->rm_core[core];
928 
929 	if (rm_corep->rm_data) {
930 		rm_host_ipi_action(rm_corep->rm_state.rm_action,
931 							rm_corep->rm_data);
932 		/* Order these stores against the real mode KVM */
933 		rm_corep->rm_data = NULL;
934 		smp_wmb();
935 		rm_corep->rm_state.rm_action = 0;
936 	}
937 }
938