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/debug.h>
20 #include <asm/synch.h>
21 #include <asm/cputhreads.h>
22 #include <asm/pgtable.h>
23 #include <asm/ppc-opcode.h>
24 #include <asm/pnv-pci.h>
25 #include <asm/opal.h>
26 #include <asm/smp.h>
27 
28 #include "book3s_xics.h"
29 
30 #define DEBUG_PASSUP
31 
32 int h_ipi_redirect = 1;
33 EXPORT_SYMBOL(h_ipi_redirect);
34 int kvm_irq_bypass = 1;
35 EXPORT_SYMBOL(kvm_irq_bypass);
36 
37 static void icp_rm_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
38 			    u32 new_irq);
39 static int xics_opal_rm_set_server(unsigned int hw_irq, int server_cpu);
40 
41 /* -- ICS routines -- */
42 static void ics_rm_check_resend(struct kvmppc_xics *xics,
43 				struct kvmppc_ics *ics, struct kvmppc_icp *icp)
44 {
45 	int i;
46 
47 	arch_spin_lock(&ics->lock);
48 
49 	for (i = 0; i < KVMPPC_XICS_IRQ_PER_ICS; i++) {
50 		struct ics_irq_state *state = &ics->irq_state[i];
51 
52 		if (!state->resend)
53 			continue;
54 
55 		arch_spin_unlock(&ics->lock);
56 		icp_rm_deliver_irq(xics, icp, state->number);
57 		arch_spin_lock(&ics->lock);
58 	}
59 
60 	arch_spin_unlock(&ics->lock);
61 }
62 
63 /* -- ICP routines -- */
64 
65 #ifdef CONFIG_SMP
66 static inline void icp_send_hcore_msg(int hcore, struct kvm_vcpu *vcpu)
67 {
68 	int hcpu;
69 
70 	hcpu = hcore << threads_shift;
71 	kvmppc_host_rm_ops_hv->rm_core[hcore].rm_data = vcpu;
72 	smp_muxed_ipi_set_message(hcpu, PPC_MSG_RM_HOST_ACTION);
73 	if (paca[hcpu].kvm_hstate.xics_phys)
74 		icp_native_cause_ipi_rm(hcpu);
75 	else
76 		opal_rm_int_set_mfrr(get_hard_smp_processor_id(hcpu),
77 				     IPI_PRIORITY);
78 }
79 #else
80 static inline void icp_send_hcore_msg(int hcore, struct kvm_vcpu *vcpu) { }
81 #endif
82 
83 /*
84  * We start the search from our current CPU Id in the core map
85  * and go in a circle until we get back to our ID looking for a
86  * core that is running in host context and that hasn't already
87  * been targeted for another rm_host_ops.
88  *
89  * In the future, could consider using a fairer algorithm (one
90  * that distributes the IPIs better)
91  *
92  * Returns -1, if no CPU could be found in the host
93  * Else, returns a CPU Id which has been reserved for use
94  */
95 static inline int grab_next_hostcore(int start,
96 		struct kvmppc_host_rm_core *rm_core, int max, int action)
97 {
98 	bool success;
99 	int core;
100 	union kvmppc_rm_state old, new;
101 
102 	for (core = start + 1; core < max; core++)  {
103 		old = new = READ_ONCE(rm_core[core].rm_state);
104 
105 		if (!old.in_host || old.rm_action)
106 			continue;
107 
108 		/* Try to grab this host core if not taken already. */
109 		new.rm_action = action;
110 
111 		success = cmpxchg64(&rm_core[core].rm_state.raw,
112 						old.raw, new.raw) == old.raw;
113 		if (success) {
114 			/*
115 			 * Make sure that the store to the rm_action is made
116 			 * visible before we return to caller (and the
117 			 * subsequent store to rm_data) to synchronize with
118 			 * the IPI handler.
119 			 */
120 			smp_wmb();
121 			return core;
122 		}
123 	}
124 
125 	return -1;
126 }
127 
128 static inline int find_available_hostcore(int action)
129 {
130 	int core;
131 	int my_core = smp_processor_id() >> threads_shift;
132 	struct kvmppc_host_rm_core *rm_core = kvmppc_host_rm_ops_hv->rm_core;
133 
134 	core = grab_next_hostcore(my_core, rm_core, cpu_nr_cores(), action);
135 	if (core == -1)
136 		core = grab_next_hostcore(core, rm_core, my_core, action);
137 
138 	return core;
139 }
140 
141 static void icp_rm_set_vcpu_irq(struct kvm_vcpu *vcpu,
142 				struct kvm_vcpu *this_vcpu)
143 {
144 	struct kvmppc_icp *this_icp = this_vcpu->arch.icp;
145 	int cpu;
146 	int hcore;
147 
148 	/* Mark the target VCPU as having an interrupt pending */
149 	vcpu->stat.queue_intr++;
150 	set_bit(BOOK3S_IRQPRIO_EXTERNAL_LEVEL, &vcpu->arch.pending_exceptions);
151 
152 	/* Kick self ? Just set MER and return */
153 	if (vcpu == this_vcpu) {
154 		mtspr(SPRN_LPCR, mfspr(SPRN_LPCR) | LPCR_MER);
155 		return;
156 	}
157 
158 	/*
159 	 * Check if the core is loaded,
160 	 * if not, find an available host core to post to wake the VCPU,
161 	 * if we can't find one, set up state to eventually return too hard.
162 	 */
163 	cpu = vcpu->arch.thread_cpu;
164 	if (cpu < 0 || cpu >= nr_cpu_ids) {
165 		hcore = -1;
166 		if (kvmppc_host_rm_ops_hv && h_ipi_redirect)
167 			hcore = find_available_hostcore(XICS_RM_KICK_VCPU);
168 		if (hcore != -1) {
169 			icp_send_hcore_msg(hcore, vcpu);
170 		} else {
171 			this_icp->rm_action |= XICS_RM_KICK_VCPU;
172 			this_icp->rm_kick_target = vcpu;
173 		}
174 		return;
175 	}
176 
177 	smp_mb();
178 	kvmhv_rm_send_ipi(cpu);
179 }
180 
181 static void icp_rm_clr_vcpu_irq(struct kvm_vcpu *vcpu)
182 {
183 	/* Note: Only called on self ! */
184 	clear_bit(BOOK3S_IRQPRIO_EXTERNAL_LEVEL,
185 		  &vcpu->arch.pending_exceptions);
186 	mtspr(SPRN_LPCR, mfspr(SPRN_LPCR) & ~LPCR_MER);
187 }
188 
189 static inline bool icp_rm_try_update(struct kvmppc_icp *icp,
190 				     union kvmppc_icp_state old,
191 				     union kvmppc_icp_state new)
192 {
193 	struct kvm_vcpu *this_vcpu = local_paca->kvm_hstate.kvm_vcpu;
194 	bool success;
195 
196 	/* Calculate new output value */
197 	new.out_ee = (new.xisr && (new.pending_pri < new.cppr));
198 
199 	/* Attempt atomic update */
200 	success = cmpxchg64(&icp->state.raw, old.raw, new.raw) == old.raw;
201 	if (!success)
202 		goto bail;
203 
204 	/*
205 	 * Check for output state update
206 	 *
207 	 * Note that this is racy since another processor could be updating
208 	 * the state already. This is why we never clear the interrupt output
209 	 * here, we only ever set it. The clear only happens prior to doing
210 	 * an update and only by the processor itself. Currently we do it
211 	 * in Accept (H_XIRR) and Up_Cppr (H_XPPR).
212 	 *
213 	 * We also do not try to figure out whether the EE state has changed,
214 	 * we unconditionally set it if the new state calls for it. The reason
215 	 * for that is that we opportunistically remove the pending interrupt
216 	 * flag when raising CPPR, so we need to set it back here if an
217 	 * interrupt is still pending.
218 	 */
219 	if (new.out_ee)
220 		icp_rm_set_vcpu_irq(icp->vcpu, this_vcpu);
221 
222 	/* Expose the state change for debug purposes */
223 	this_vcpu->arch.icp->rm_dbgstate = new;
224 	this_vcpu->arch.icp->rm_dbgtgt = icp->vcpu;
225 
226  bail:
227 	return success;
228 }
229 
230 static inline int check_too_hard(struct kvmppc_xics *xics,
231 				 struct kvmppc_icp *icp)
232 {
233 	return (xics->real_mode_dbg || icp->rm_action) ? H_TOO_HARD : H_SUCCESS;
234 }
235 
236 static void icp_rm_check_resend(struct kvmppc_xics *xics,
237 			     struct kvmppc_icp *icp)
238 {
239 	u32 icsid;
240 
241 	/* Order this load with the test for need_resend in the caller */
242 	smp_rmb();
243 	for_each_set_bit(icsid, icp->resend_map, xics->max_icsid + 1) {
244 		struct kvmppc_ics *ics = xics->ics[icsid];
245 
246 		if (!test_and_clear_bit(icsid, icp->resend_map))
247 			continue;
248 		if (!ics)
249 			continue;
250 		ics_rm_check_resend(xics, ics, icp);
251 	}
252 }
253 
254 static bool icp_rm_try_to_deliver(struct kvmppc_icp *icp, u32 irq, u8 priority,
255 			       u32 *reject)
256 {
257 	union kvmppc_icp_state old_state, new_state;
258 	bool success;
259 
260 	do {
261 		old_state = new_state = READ_ONCE(icp->state);
262 
263 		*reject = 0;
264 
265 		/* See if we can deliver */
266 		success = new_state.cppr > priority &&
267 			new_state.mfrr > priority &&
268 			new_state.pending_pri > priority;
269 
270 		/*
271 		 * If we can, check for a rejection and perform the
272 		 * delivery
273 		 */
274 		if (success) {
275 			*reject = new_state.xisr;
276 			new_state.xisr = irq;
277 			new_state.pending_pri = priority;
278 		} else {
279 			/*
280 			 * If we failed to deliver we set need_resend
281 			 * so a subsequent CPPR state change causes us
282 			 * to try a new delivery.
283 			 */
284 			new_state.need_resend = true;
285 		}
286 
287 	} while (!icp_rm_try_update(icp, old_state, new_state));
288 
289 	return success;
290 }
291 
292 static void icp_rm_deliver_irq(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
293 			    u32 new_irq)
294 {
295 	struct ics_irq_state *state;
296 	struct kvmppc_ics *ics;
297 	u32 reject;
298 	u16 src;
299 
300 	/*
301 	 * This is used both for initial delivery of an interrupt and
302 	 * for subsequent rejection.
303 	 *
304 	 * Rejection can be racy vs. resends. We have evaluated the
305 	 * rejection in an atomic ICP transaction which is now complete,
306 	 * so potentially the ICP can already accept the interrupt again.
307 	 *
308 	 * So we need to retry the delivery. Essentially the reject path
309 	 * boils down to a failed delivery. Always.
310 	 *
311 	 * Now the interrupt could also have moved to a different target,
312 	 * thus we may need to re-do the ICP lookup as well
313 	 */
314 
315  again:
316 	/* Get the ICS state and lock it */
317 	ics = kvmppc_xics_find_ics(xics, new_irq, &src);
318 	if (!ics) {
319 		/* Unsafe increment, but this does not need to be accurate */
320 		xics->err_noics++;
321 		return;
322 	}
323 	state = &ics->irq_state[src];
324 
325 	/* Get a lock on the ICS */
326 	arch_spin_lock(&ics->lock);
327 
328 	/* Get our server */
329 	if (!icp || state->server != icp->server_num) {
330 		icp = kvmppc_xics_find_server(xics->kvm, state->server);
331 		if (!icp) {
332 			/* Unsafe increment again*/
333 			xics->err_noicp++;
334 			goto out;
335 		}
336 	}
337 
338 	/* Clear the resend bit of that interrupt */
339 	state->resend = 0;
340 
341 	/*
342 	 * If masked, bail out
343 	 *
344 	 * Note: PAPR doesn't mention anything about masked pending
345 	 * when doing a resend, only when doing a delivery.
346 	 *
347 	 * However that would have the effect of losing a masked
348 	 * interrupt that was rejected and isn't consistent with
349 	 * the whole masked_pending business which is about not
350 	 * losing interrupts that occur while masked.
351 	 *
352 	 * I don't differentiate normal deliveries and resends, this
353 	 * implementation will differ from PAPR and not lose such
354 	 * interrupts.
355 	 */
356 	if (state->priority == MASKED) {
357 		state->masked_pending = 1;
358 		goto out;
359 	}
360 
361 	/*
362 	 * Try the delivery, this will set the need_resend flag
363 	 * in the ICP as part of the atomic transaction if the
364 	 * delivery is not possible.
365 	 *
366 	 * Note that if successful, the new delivery might have itself
367 	 * rejected an interrupt that was "delivered" before we took the
368 	 * ics spin lock.
369 	 *
370 	 * In this case we do the whole sequence all over again for the
371 	 * new guy. We cannot assume that the rejected interrupt is less
372 	 * favored than the new one, and thus doesn't need to be delivered,
373 	 * because by the time we exit icp_rm_try_to_deliver() the target
374 	 * processor may well have already consumed & completed it, and thus
375 	 * the rejected interrupt might actually be already acceptable.
376 	 */
377 	if (icp_rm_try_to_deliver(icp, new_irq, state->priority, &reject)) {
378 		/*
379 		 * Delivery was successful, did we reject somebody else ?
380 		 */
381 		if (reject && reject != XICS_IPI) {
382 			arch_spin_unlock(&ics->lock);
383 			new_irq = reject;
384 			goto again;
385 		}
386 	} else {
387 		/*
388 		 * We failed to deliver the interrupt we need to set the
389 		 * resend map bit and mark the ICS state as needing a resend
390 		 */
391 		set_bit(ics->icsid, icp->resend_map);
392 		state->resend = 1;
393 
394 		/*
395 		 * If the need_resend flag got cleared in the ICP some time
396 		 * between icp_rm_try_to_deliver() atomic update and now, then
397 		 * we know it might have missed the resend_map bit. So we
398 		 * retry
399 		 */
400 		smp_mb();
401 		if (!icp->state.need_resend) {
402 			arch_spin_unlock(&ics->lock);
403 			goto again;
404 		}
405 	}
406  out:
407 	arch_spin_unlock(&ics->lock);
408 }
409 
410 static void icp_rm_down_cppr(struct kvmppc_xics *xics, struct kvmppc_icp *icp,
411 			     u8 new_cppr)
412 {
413 	union kvmppc_icp_state old_state, new_state;
414 	bool resend;
415 
416 	/*
417 	 * This handles several related states in one operation:
418 	 *
419 	 * ICP State: Down_CPPR
420 	 *
421 	 * Load CPPR with new value and if the XISR is 0
422 	 * then check for resends:
423 	 *
424 	 * ICP State: Resend
425 	 *
426 	 * If MFRR is more favored than CPPR, check for IPIs
427 	 * and notify ICS of a potential resend. This is done
428 	 * asynchronously (when used in real mode, we will have
429 	 * to exit here).
430 	 *
431 	 * We do not handle the complete Check_IPI as documented
432 	 * here. In the PAPR, this state will be used for both
433 	 * Set_MFRR and Down_CPPR. However, we know that we aren't
434 	 * changing the MFRR state here so we don't need to handle
435 	 * the case of an MFRR causing a reject of a pending irq,
436 	 * this will have been handled when the MFRR was set in the
437 	 * first place.
438 	 *
439 	 * Thus we don't have to handle rejects, only resends.
440 	 *
441 	 * When implementing real mode for HV KVM, resend will lead to
442 	 * a H_TOO_HARD return and the whole transaction will be handled
443 	 * in virtual mode.
444 	 */
445 	do {
446 		old_state = new_state = READ_ONCE(icp->state);
447 
448 		/* Down_CPPR */
449 		new_state.cppr = new_cppr;
450 
451 		/*
452 		 * Cut down Resend / Check_IPI / IPI
453 		 *
454 		 * The logic is that we cannot have a pending interrupt
455 		 * trumped by an IPI at this point (see above), so we
456 		 * know that either the pending interrupt is already an
457 		 * IPI (in which case we don't care to override it) or
458 		 * it's either more favored than us or non existent
459 		 */
460 		if (new_state.mfrr < new_cppr &&
461 		    new_state.mfrr <= new_state.pending_pri) {
462 			new_state.pending_pri = new_state.mfrr;
463 			new_state.xisr = XICS_IPI;
464 		}
465 
466 		/* Latch/clear resend bit */
467 		resend = new_state.need_resend;
468 		new_state.need_resend = 0;
469 
470 	} while (!icp_rm_try_update(icp, old_state, new_state));
471 
472 	/*
473 	 * Now handle resend checks. Those are asynchronous to the ICP
474 	 * state update in HW (ie bus transactions) so we can handle them
475 	 * separately here as well.
476 	 */
477 	if (resend) {
478 		icp->n_check_resend++;
479 		icp_rm_check_resend(xics, icp);
480 	}
481 }
482 
483 
484 unsigned long kvmppc_rm_h_xirr(struct kvm_vcpu *vcpu)
485 {
486 	union kvmppc_icp_state old_state, new_state;
487 	struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
488 	struct kvmppc_icp *icp = vcpu->arch.icp;
489 	u32 xirr;
490 
491 	if (!xics || !xics->real_mode)
492 		return H_TOO_HARD;
493 
494 	/* First clear the interrupt */
495 	icp_rm_clr_vcpu_irq(icp->vcpu);
496 
497 	/*
498 	 * ICP State: Accept_Interrupt
499 	 *
500 	 * Return the pending interrupt (if any) along with the
501 	 * current CPPR, then clear the XISR & set CPPR to the
502 	 * pending priority
503 	 */
504 	do {
505 		old_state = new_state = READ_ONCE(icp->state);
506 
507 		xirr = old_state.xisr | (((u32)old_state.cppr) << 24);
508 		if (!old_state.xisr)
509 			break;
510 		new_state.cppr = new_state.pending_pri;
511 		new_state.pending_pri = 0xff;
512 		new_state.xisr = 0;
513 
514 	} while (!icp_rm_try_update(icp, old_state, new_state));
515 
516 	/* Return the result in GPR4 */
517 	vcpu->arch.gpr[4] = xirr;
518 
519 	return check_too_hard(xics, icp);
520 }
521 
522 int kvmppc_rm_h_ipi(struct kvm_vcpu *vcpu, unsigned long server,
523 		    unsigned long mfrr)
524 {
525 	union kvmppc_icp_state old_state, new_state;
526 	struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
527 	struct kvmppc_icp *icp, *this_icp = vcpu->arch.icp;
528 	u32 reject;
529 	bool resend;
530 	bool local;
531 
532 	if (!xics || !xics->real_mode)
533 		return H_TOO_HARD;
534 
535 	local = this_icp->server_num == server;
536 	if (local)
537 		icp = this_icp;
538 	else
539 		icp = kvmppc_xics_find_server(vcpu->kvm, server);
540 	if (!icp)
541 		return H_PARAMETER;
542 
543 	/*
544 	 * ICP state: Set_MFRR
545 	 *
546 	 * If the CPPR is more favored than the new MFRR, then
547 	 * nothing needs to be done as there can be no XISR to
548 	 * reject.
549 	 *
550 	 * ICP state: Check_IPI
551 	 *
552 	 * If the CPPR is less favored, then we might be replacing
553 	 * an interrupt, and thus need to possibly reject it.
554 	 *
555 	 * ICP State: IPI
556 	 *
557 	 * Besides rejecting any pending interrupts, we also
558 	 * update XISR and pending_pri to mark IPI as pending.
559 	 *
560 	 * PAPR does not describe this state, but if the MFRR is being
561 	 * made less favored than its earlier value, there might be
562 	 * a previously-rejected interrupt needing to be resent.
563 	 * Ideally, we would want to resend only if
564 	 *	prio(pending_interrupt) < mfrr &&
565 	 *	prio(pending_interrupt) < cppr
566 	 * where pending interrupt is the one that was rejected. But
567 	 * we don't have that state, so we simply trigger a resend
568 	 * whenever the MFRR is made less favored.
569 	 */
570 	do {
571 		old_state = new_state = READ_ONCE(icp->state);
572 
573 		/* Set_MFRR */
574 		new_state.mfrr = mfrr;
575 
576 		/* Check_IPI */
577 		reject = 0;
578 		resend = false;
579 		if (mfrr < new_state.cppr) {
580 			/* Reject a pending interrupt if not an IPI */
581 			if (mfrr <= new_state.pending_pri) {
582 				reject = new_state.xisr;
583 				new_state.pending_pri = mfrr;
584 				new_state.xisr = XICS_IPI;
585 			}
586 		}
587 
588 		if (mfrr > old_state.mfrr) {
589 			resend = new_state.need_resend;
590 			new_state.need_resend = 0;
591 		}
592 	} while (!icp_rm_try_update(icp, old_state, new_state));
593 
594 	/* Handle reject in real mode */
595 	if (reject && reject != XICS_IPI) {
596 		this_icp->n_reject++;
597 		icp_rm_deliver_irq(xics, icp, reject);
598 	}
599 
600 	/* Handle resends in real mode */
601 	if (resend) {
602 		this_icp->n_check_resend++;
603 		icp_rm_check_resend(xics, icp);
604 	}
605 
606 	return check_too_hard(xics, this_icp);
607 }
608 
609 int kvmppc_rm_h_cppr(struct kvm_vcpu *vcpu, unsigned long cppr)
610 {
611 	union kvmppc_icp_state old_state, new_state;
612 	struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
613 	struct kvmppc_icp *icp = vcpu->arch.icp;
614 	u32 reject;
615 
616 	if (!xics || !xics->real_mode)
617 		return H_TOO_HARD;
618 
619 	/*
620 	 * ICP State: Set_CPPR
621 	 *
622 	 * We can safely compare the new value with the current
623 	 * value outside of the transaction as the CPPR is only
624 	 * ever changed by the processor on itself
625 	 */
626 	if (cppr > icp->state.cppr) {
627 		icp_rm_down_cppr(xics, icp, cppr);
628 		goto bail;
629 	} else if (cppr == icp->state.cppr)
630 		return H_SUCCESS;
631 
632 	/*
633 	 * ICP State: Up_CPPR
634 	 *
635 	 * The processor is raising its priority, this can result
636 	 * in a rejection of a pending interrupt:
637 	 *
638 	 * ICP State: Reject_Current
639 	 *
640 	 * We can remove EE from the current processor, the update
641 	 * transaction will set it again if needed
642 	 */
643 	icp_rm_clr_vcpu_irq(icp->vcpu);
644 
645 	do {
646 		old_state = new_state = READ_ONCE(icp->state);
647 
648 		reject = 0;
649 		new_state.cppr = cppr;
650 
651 		if (cppr <= new_state.pending_pri) {
652 			reject = new_state.xisr;
653 			new_state.xisr = 0;
654 			new_state.pending_pri = 0xff;
655 		}
656 
657 	} while (!icp_rm_try_update(icp, old_state, new_state));
658 
659 	/*
660 	 * Check for rejects. They are handled by doing a new delivery
661 	 * attempt (see comments in icp_rm_deliver_irq).
662 	 */
663 	if (reject && reject != XICS_IPI) {
664 		icp->n_reject++;
665 		icp_rm_deliver_irq(xics, icp, reject);
666 	}
667  bail:
668 	return check_too_hard(xics, icp);
669 }
670 
671 int kvmppc_rm_h_eoi(struct kvm_vcpu *vcpu, unsigned long xirr)
672 {
673 	struct kvmppc_xics *xics = vcpu->kvm->arch.xics;
674 	struct kvmppc_icp *icp = vcpu->arch.icp;
675 	struct kvmppc_ics *ics;
676 	struct ics_irq_state *state;
677 	u32 irq = xirr & 0x00ffffff;
678 	u16 src;
679 
680 	if (!xics || !xics->real_mode)
681 		return H_TOO_HARD;
682 
683 	/*
684 	 * ICP State: EOI
685 	 *
686 	 * Note: If EOI is incorrectly used by SW to lower the CPPR
687 	 * value (ie more favored), we do not check for rejection of
688 	 * a pending interrupt, this is a SW error and PAPR sepcifies
689 	 * that we don't have to deal with it.
690 	 *
691 	 * The sending of an EOI to the ICS is handled after the
692 	 * CPPR update
693 	 *
694 	 * ICP State: Down_CPPR which we handle
695 	 * in a separate function as it's shared with H_CPPR.
696 	 */
697 	icp_rm_down_cppr(xics, icp, xirr >> 24);
698 
699 	/* IPIs have no EOI */
700 	if (irq == XICS_IPI)
701 		goto bail;
702 	/*
703 	 * EOI handling: If the interrupt is still asserted, we need to
704 	 * resend it. We can take a lockless "peek" at the ICS state here.
705 	 *
706 	 * "Message" interrupts will never have "asserted" set
707 	 */
708 	ics = kvmppc_xics_find_ics(xics, irq, &src);
709 	if (!ics)
710 		goto bail;
711 	state = &ics->irq_state[src];
712 
713 	/* Still asserted, resend it */
714 	if (state->asserted) {
715 		icp->n_reject++;
716 		icp_rm_deliver_irq(xics, icp, irq);
717 	}
718 
719 	if (!hlist_empty(&vcpu->kvm->irq_ack_notifier_list)) {
720 		icp->rm_action |= XICS_RM_NOTIFY_EOI;
721 		icp->rm_eoied_irq = irq;
722 	}
723 
724 	if (state->host_irq) {
725 		++vcpu->stat.pthru_all;
726 		if (state->intr_cpu != -1) {
727 			int pcpu = raw_smp_processor_id();
728 
729 			pcpu = cpu_first_thread_sibling(pcpu);
730 			++vcpu->stat.pthru_host;
731 			if (state->intr_cpu != pcpu) {
732 				++vcpu->stat.pthru_bad_aff;
733 				xics_opal_rm_set_server(state->host_irq, pcpu);
734 			}
735 			state->intr_cpu = -1;
736 		}
737 	}
738  bail:
739 	return check_too_hard(xics, icp);
740 }
741 
742 unsigned long eoi_rc;
743 
744 static void icp_eoi(struct irq_chip *c, u32 hwirq, __be32 xirr, bool *again)
745 {
746 	unsigned long xics_phys;
747 	int64_t rc;
748 
749 	rc = pnv_opal_pci_msi_eoi(c, hwirq);
750 
751 	if (rc)
752 		eoi_rc = rc;
753 
754 	iosync();
755 
756 	/* EOI it */
757 	xics_phys = local_paca->kvm_hstate.xics_phys;
758 	if (xics_phys) {
759 		_stwcix(xics_phys + XICS_XIRR, xirr);
760 	} else {
761 		rc = opal_rm_int_eoi(be32_to_cpu(xirr));
762 		*again = rc > 0;
763 	}
764 }
765 
766 static int xics_opal_rm_set_server(unsigned int hw_irq, int server_cpu)
767 {
768 	unsigned int mangle_cpu = get_hard_smp_processor_id(server_cpu) << 2;
769 
770 	return opal_rm_set_xive(hw_irq, mangle_cpu, DEFAULT_PRIORITY);
771 }
772 
773 /*
774  * Increment a per-CPU 32-bit unsigned integer variable.
775  * Safe to call in real-mode. Handles vmalloc'ed addresses
776  *
777  * ToDo: Make this work for any integral type
778  */
779 
780 static inline void this_cpu_inc_rm(unsigned int __percpu *addr)
781 {
782 	unsigned long l;
783 	unsigned int *raddr;
784 	int cpu = smp_processor_id();
785 
786 	raddr = per_cpu_ptr(addr, cpu);
787 	l = (unsigned long)raddr;
788 
789 	if (REGION_ID(l) == VMALLOC_REGION_ID) {
790 		l = vmalloc_to_phys(raddr);
791 		raddr = (unsigned int *)l;
792 	}
793 	++*raddr;
794 }
795 
796 /*
797  * We don't try to update the flags in the irq_desc 'istate' field in
798  * here as would happen in the normal IRQ handling path for several reasons:
799  *  - state flags represent internal IRQ state and are not expected to be
800  *    updated outside the IRQ subsystem
801  *  - more importantly, these are useful for edge triggered interrupts,
802  *    IRQ probing, etc., but we are only handling MSI/MSIx interrupts here
803  *    and these states shouldn't apply to us.
804  *
805  * However, we do update irq_stats - we somewhat duplicate the code in
806  * kstat_incr_irqs_this_cpu() for this since this function is defined
807  * in irq/internal.h which we don't want to include here.
808  * The only difference is that desc->kstat_irqs is an allocated per CPU
809  * variable and could have been vmalloc'ed, so we can't directly
810  * call __this_cpu_inc() on it. The kstat structure is a static
811  * per CPU variable and it should be accessible by real-mode KVM.
812  *
813  */
814 static void kvmppc_rm_handle_irq_desc(struct irq_desc *desc)
815 {
816 	this_cpu_inc_rm(desc->kstat_irqs);
817 	__this_cpu_inc(kstat.irqs_sum);
818 }
819 
820 long kvmppc_deliver_irq_passthru(struct kvm_vcpu *vcpu,
821 				 __be32 xirr,
822 				 struct kvmppc_irq_map *irq_map,
823 				 struct kvmppc_passthru_irqmap *pimap,
824 				 bool *again)
825 {
826 	struct kvmppc_xics *xics;
827 	struct kvmppc_icp *icp;
828 	u32 irq;
829 
830 	irq = irq_map->v_hwirq;
831 	xics = vcpu->kvm->arch.xics;
832 	icp = vcpu->arch.icp;
833 
834 	kvmppc_rm_handle_irq_desc(irq_map->desc);
835 	icp_rm_deliver_irq(xics, icp, irq);
836 
837 	/* EOI the interrupt */
838 	icp_eoi(irq_desc_get_chip(irq_map->desc), irq_map->r_hwirq, xirr,
839 		again);
840 
841 	if (check_too_hard(xics, icp) == H_TOO_HARD)
842 		return 2;
843 	else
844 		return -2;
845 }
846 
847 /*  --- Non-real mode XICS-related built-in routines ---  */
848 
849 /**
850  * Host Operations poked by RM KVM
851  */
852 static void rm_host_ipi_action(int action, void *data)
853 {
854 	switch (action) {
855 	case XICS_RM_KICK_VCPU:
856 		kvmppc_host_rm_ops_hv->vcpu_kick(data);
857 		break;
858 	default:
859 		WARN(1, "Unexpected rm_action=%d data=%p\n", action, data);
860 		break;
861 	}
862 
863 }
864 
865 void kvmppc_xics_ipi_action(void)
866 {
867 	int core;
868 	unsigned int cpu = smp_processor_id();
869 	struct kvmppc_host_rm_core *rm_corep;
870 
871 	core = cpu >> threads_shift;
872 	rm_corep = &kvmppc_host_rm_ops_hv->rm_core[core];
873 
874 	if (rm_corep->rm_data) {
875 		rm_host_ipi_action(rm_corep->rm_state.rm_action,
876 							rm_corep->rm_data);
877 		/* Order these stores against the real mode KVM */
878 		rm_corep->rm_data = NULL;
879 		smp_wmb();
880 		rm_corep->rm_state.rm_action = 0;
881 	}
882 }
883