1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * Copyright 2017 Benjamin Herrenschmidt, IBM Corporation
4 */
5
6 /* File to be included by other .c files */
7
8 #define XGLUE(a,b) a##b
9 #define GLUE(a,b) XGLUE(a,b)
10
11 /* Dummy interrupt used when taking interrupts out of a queue in H_CPPR */
12 #define XICS_DUMMY 1
13
GLUE(X_PFX,ack_pending)14 static void GLUE(X_PFX,ack_pending)(struct kvmppc_xive_vcpu *xc)
15 {
16 u8 cppr;
17 u16 ack;
18
19 /*
20 * Ensure any previous store to CPPR is ordered vs.
21 * the subsequent loads from PIPR or ACK.
22 */
23 eieio();
24
25 /* Perform the acknowledge OS to register cycle. */
26 ack = be16_to_cpu(__x_readw(__x_tima + TM_SPC_ACK_OS_REG));
27
28 /* Synchronize subsequent queue accesses */
29 mb();
30
31 /* XXX Check grouping level */
32
33 /* Anything ? */
34 if (!((ack >> 8) & TM_QW1_NSR_EO))
35 return;
36
37 /* Grab CPPR of the most favored pending interrupt */
38 cppr = ack & 0xff;
39 if (cppr < 8)
40 xc->pending |= 1 << cppr;
41
42 #ifdef XIVE_RUNTIME_CHECKS
43 /* Check consistency */
44 if (cppr >= xc->hw_cppr)
45 pr_warn("KVM-XIVE: CPU %d odd ack CPPR, got %d at %d\n",
46 smp_processor_id(), cppr, xc->hw_cppr);
47 #endif
48
49 /*
50 * Update our image of the HW CPPR. We don't yet modify
51 * xc->cppr, this will be done as we scan for interrupts
52 * in the queues.
53 */
54 xc->hw_cppr = cppr;
55 }
56
GLUE(X_PFX,esb_load)57 static u8 GLUE(X_PFX,esb_load)(struct xive_irq_data *xd, u32 offset)
58 {
59 u64 val;
60
61 if (xd->flags & XIVE_IRQ_FLAG_SHIFT_BUG)
62 offset |= offset << 4;
63
64 val =__x_readq(__x_eoi_page(xd) + offset);
65 #ifdef __LITTLE_ENDIAN__
66 val >>= 64-8;
67 #endif
68 return (u8)val;
69 }
70
71
GLUE(X_PFX,source_eoi)72 static void GLUE(X_PFX,source_eoi)(u32 hw_irq, struct xive_irq_data *xd)
73 {
74 /* If the XIVE supports the new "store EOI facility, use it */
75 if (xd->flags & XIVE_IRQ_FLAG_STORE_EOI)
76 __x_writeq(0, __x_eoi_page(xd) + XIVE_ESB_STORE_EOI);
77 else if (hw_irq && xd->flags & XIVE_IRQ_FLAG_EOI_FW)
78 opal_int_eoi(hw_irq);
79 else if (xd->flags & XIVE_IRQ_FLAG_LSI) {
80 /*
81 * For LSIs the HW EOI cycle is used rather than PQ bits,
82 * as they are automatically re-triggred in HW when still
83 * pending.
84 */
85 __x_readq(__x_eoi_page(xd) + XIVE_ESB_LOAD_EOI);
86 } else {
87 uint64_t eoi_val;
88
89 /*
90 * Otherwise for EOI, we use the special MMIO that does
91 * a clear of both P and Q and returns the old Q,
92 * except for LSIs where we use the "EOI cycle" special
93 * load.
94 *
95 * This allows us to then do a re-trigger if Q was set
96 * rather than synthetizing an interrupt in software
97 */
98 eoi_val = GLUE(X_PFX,esb_load)(xd, XIVE_ESB_SET_PQ_00);
99
100 /* Re-trigger if needed */
101 if ((eoi_val & 1) && __x_trig_page(xd))
102 __x_writeq(0, __x_trig_page(xd));
103 }
104 }
105
106 enum {
107 scan_fetch,
108 scan_poll,
109 scan_eoi,
110 };
111
GLUE(X_PFX,scan_interrupts)112 static u32 GLUE(X_PFX,scan_interrupts)(struct kvmppc_xive_vcpu *xc,
113 u8 pending, int scan_type)
114 {
115 u32 hirq = 0;
116 u8 prio = 0xff;
117
118 /* Find highest pending priority */
119 while ((xc->mfrr != 0xff || pending != 0) && hirq == 0) {
120 struct xive_q *q;
121 u32 idx, toggle;
122 __be32 *qpage;
123
124 /*
125 * If pending is 0 this will return 0xff which is what
126 * we want
127 */
128 prio = ffs(pending) - 1;
129
130 /* Don't scan past the guest cppr */
131 if (prio >= xc->cppr || prio > 7) {
132 if (xc->mfrr < xc->cppr) {
133 prio = xc->mfrr;
134 hirq = XICS_IPI;
135 }
136 break;
137 }
138
139 /* Grab queue and pointers */
140 q = &xc->queues[prio];
141 idx = q->idx;
142 toggle = q->toggle;
143
144 /*
145 * Snapshot the queue page. The test further down for EOI
146 * must use the same "copy" that was used by __xive_read_eq
147 * since qpage can be set concurrently and we don't want
148 * to miss an EOI.
149 */
150 qpage = READ_ONCE(q->qpage);
151
152 skip_ipi:
153 /*
154 * Try to fetch from the queue. Will return 0 for a
155 * non-queueing priority (ie, qpage = 0).
156 */
157 hirq = __xive_read_eq(qpage, q->msk, &idx, &toggle);
158
159 /*
160 * If this was a signal for an MFFR change done by
161 * H_IPI we skip it. Additionally, if we were fetching
162 * we EOI it now, thus re-enabling reception of a new
163 * such signal.
164 *
165 * We also need to do that if prio is 0 and we had no
166 * page for the queue. In this case, we have non-queued
167 * IPI that needs to be EOId.
168 *
169 * This is safe because if we have another pending MFRR
170 * change that wasn't observed above, the Q bit will have
171 * been set and another occurrence of the IPI will trigger.
172 */
173 if (hirq == XICS_IPI || (prio == 0 && !qpage)) {
174 if (scan_type == scan_fetch) {
175 GLUE(X_PFX,source_eoi)(xc->vp_ipi,
176 &xc->vp_ipi_data);
177 q->idx = idx;
178 q->toggle = toggle;
179 }
180 /* Loop back on same queue with updated idx/toggle */
181 #ifdef XIVE_RUNTIME_CHECKS
182 WARN_ON(hirq && hirq != XICS_IPI);
183 #endif
184 if (hirq)
185 goto skip_ipi;
186 }
187
188 /* If it's the dummy interrupt, continue searching */
189 if (hirq == XICS_DUMMY)
190 goto skip_ipi;
191
192 /* Clear the pending bit if the queue is now empty */
193 if (!hirq) {
194 pending &= ~(1 << prio);
195
196 /*
197 * Check if the queue count needs adjusting due to
198 * interrupts being moved away.
199 */
200 if (atomic_read(&q->pending_count)) {
201 int p = atomic_xchg(&q->pending_count, 0);
202 if (p) {
203 #ifdef XIVE_RUNTIME_CHECKS
204 WARN_ON(p > atomic_read(&q->count));
205 #endif
206 atomic_sub(p, &q->count);
207 }
208 }
209 }
210
211 /*
212 * If the most favoured prio we found pending is less
213 * favored (or equal) than a pending IPI, we return
214 * the IPI instead.
215 */
216 if (prio >= xc->mfrr && xc->mfrr < xc->cppr) {
217 prio = xc->mfrr;
218 hirq = XICS_IPI;
219 break;
220 }
221
222 /* If fetching, update queue pointers */
223 if (scan_type == scan_fetch) {
224 q->idx = idx;
225 q->toggle = toggle;
226 }
227 }
228
229 /* If we are just taking a "peek", do nothing else */
230 if (scan_type == scan_poll)
231 return hirq;
232
233 /* Update the pending bits */
234 xc->pending = pending;
235
236 /*
237 * If this is an EOI that's it, no CPPR adjustment done here,
238 * all we needed was cleanup the stale pending bits and check
239 * if there's anything left.
240 */
241 if (scan_type == scan_eoi)
242 return hirq;
243
244 /*
245 * If we found an interrupt, adjust what the guest CPPR should
246 * be as if we had just fetched that interrupt from HW.
247 *
248 * Note: This can only make xc->cppr smaller as the previous
249 * loop will only exit with hirq != 0 if prio is lower than
250 * the current xc->cppr. Thus we don't need to re-check xc->mfrr
251 * for pending IPIs.
252 */
253 if (hirq)
254 xc->cppr = prio;
255 /*
256 * If it was an IPI the HW CPPR might have been lowered too much
257 * as the HW interrupt we use for IPIs is routed to priority 0.
258 *
259 * We re-sync it here.
260 */
261 if (xc->cppr != xc->hw_cppr) {
262 xc->hw_cppr = xc->cppr;
263 __x_writeb(xc->cppr, __x_tima + TM_QW1_OS + TM_CPPR);
264 }
265
266 return hirq;
267 }
268
GLUE(X_PFX,h_xirr)269 X_STATIC unsigned long GLUE(X_PFX,h_xirr)(struct kvm_vcpu *vcpu)
270 {
271 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
272 u8 old_cppr;
273 u32 hirq;
274
275 pr_devel("H_XIRR\n");
276
277 xc->GLUE(X_STAT_PFX,h_xirr)++;
278
279 /* First collect pending bits from HW */
280 GLUE(X_PFX,ack_pending)(xc);
281
282 pr_devel(" new pending=0x%02x hw_cppr=%d cppr=%d\n",
283 xc->pending, xc->hw_cppr, xc->cppr);
284
285 /* Grab previous CPPR and reverse map it */
286 old_cppr = xive_prio_to_guest(xc->cppr);
287
288 /* Scan for actual interrupts */
289 hirq = GLUE(X_PFX,scan_interrupts)(xc, xc->pending, scan_fetch);
290
291 pr_devel(" got hirq=0x%x hw_cppr=%d cppr=%d\n",
292 hirq, xc->hw_cppr, xc->cppr);
293
294 #ifdef XIVE_RUNTIME_CHECKS
295 /* That should never hit */
296 if (hirq & 0xff000000)
297 pr_warn("XIVE: Weird guest interrupt number 0x%08x\n", hirq);
298 #endif
299
300 /*
301 * XXX We could check if the interrupt is masked here and
302 * filter it. If we chose to do so, we would need to do:
303 *
304 * if (masked) {
305 * lock();
306 * if (masked) {
307 * old_Q = true;
308 * hirq = 0;
309 * }
310 * unlock();
311 * }
312 */
313
314 /* Return interrupt and old CPPR in GPR4 */
315 vcpu->arch.regs.gpr[4] = hirq | (old_cppr << 24);
316
317 return H_SUCCESS;
318 }
319
GLUE(X_PFX,h_ipoll)320 X_STATIC unsigned long GLUE(X_PFX,h_ipoll)(struct kvm_vcpu *vcpu, unsigned long server)
321 {
322 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
323 u8 pending = xc->pending;
324 u32 hirq;
325
326 pr_devel("H_IPOLL(server=%ld)\n", server);
327
328 xc->GLUE(X_STAT_PFX,h_ipoll)++;
329
330 /* Grab the target VCPU if not the current one */
331 if (xc->server_num != server) {
332 vcpu = kvmppc_xive_find_server(vcpu->kvm, server);
333 if (!vcpu)
334 return H_PARAMETER;
335 xc = vcpu->arch.xive_vcpu;
336
337 /* Scan all priorities */
338 pending = 0xff;
339 } else {
340 /* Grab pending interrupt if any */
341 __be64 qw1 = __x_readq(__x_tima + TM_QW1_OS);
342 u8 pipr = be64_to_cpu(qw1) & 0xff;
343 if (pipr < 8)
344 pending |= 1 << pipr;
345 }
346
347 hirq = GLUE(X_PFX,scan_interrupts)(xc, pending, scan_poll);
348
349 /* Return interrupt and old CPPR in GPR4 */
350 vcpu->arch.regs.gpr[4] = hirq | (xc->cppr << 24);
351
352 return H_SUCCESS;
353 }
354
GLUE(X_PFX,push_pending_to_hw)355 static void GLUE(X_PFX,push_pending_to_hw)(struct kvmppc_xive_vcpu *xc)
356 {
357 u8 pending, prio;
358
359 pending = xc->pending;
360 if (xc->mfrr != 0xff) {
361 if (xc->mfrr < 8)
362 pending |= 1 << xc->mfrr;
363 else
364 pending |= 0x80;
365 }
366 if (!pending)
367 return;
368 prio = ffs(pending) - 1;
369
370 __x_writeb(prio, __x_tima + TM_SPC_SET_OS_PENDING);
371 }
372
GLUE(X_PFX,scan_for_rerouted_irqs)373 static void GLUE(X_PFX,scan_for_rerouted_irqs)(struct kvmppc_xive *xive,
374 struct kvmppc_xive_vcpu *xc)
375 {
376 unsigned int prio;
377
378 /* For each priority that is now masked */
379 for (prio = xc->cppr; prio < KVMPPC_XIVE_Q_COUNT; prio++) {
380 struct xive_q *q = &xc->queues[prio];
381 struct kvmppc_xive_irq_state *state;
382 struct kvmppc_xive_src_block *sb;
383 u32 idx, toggle, entry, irq, hw_num;
384 struct xive_irq_data *xd;
385 __be32 *qpage;
386 u16 src;
387
388 idx = q->idx;
389 toggle = q->toggle;
390 qpage = READ_ONCE(q->qpage);
391 if (!qpage)
392 continue;
393
394 /* For each interrupt in the queue */
395 for (;;) {
396 entry = be32_to_cpup(qpage + idx);
397
398 /* No more ? */
399 if ((entry >> 31) == toggle)
400 break;
401 irq = entry & 0x7fffffff;
402
403 /* Skip dummies and IPIs */
404 if (irq == XICS_DUMMY || irq == XICS_IPI)
405 goto next;
406 sb = kvmppc_xive_find_source(xive, irq, &src);
407 if (!sb)
408 goto next;
409 state = &sb->irq_state[src];
410
411 /* Has it been rerouted ? */
412 if (xc->server_num == state->act_server)
413 goto next;
414
415 /*
416 * Allright, it *has* been re-routed, kill it from
417 * the queue.
418 */
419 qpage[idx] = cpu_to_be32((entry & 0x80000000) | XICS_DUMMY);
420
421 /* Find the HW interrupt */
422 kvmppc_xive_select_irq(state, &hw_num, &xd);
423
424 /* If it's not an LSI, set PQ to 11 the EOI will force a resend */
425 if (!(xd->flags & XIVE_IRQ_FLAG_LSI))
426 GLUE(X_PFX,esb_load)(xd, XIVE_ESB_SET_PQ_11);
427
428 /* EOI the source */
429 GLUE(X_PFX,source_eoi)(hw_num, xd);
430
431 next:
432 idx = (idx + 1) & q->msk;
433 if (idx == 0)
434 toggle ^= 1;
435 }
436 }
437 }
438
GLUE(X_PFX,h_cppr)439 X_STATIC int GLUE(X_PFX,h_cppr)(struct kvm_vcpu *vcpu, unsigned long cppr)
440 {
441 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
442 struct kvmppc_xive *xive = vcpu->kvm->arch.xive;
443 u8 old_cppr;
444
445 pr_devel("H_CPPR(cppr=%ld)\n", cppr);
446
447 xc->GLUE(X_STAT_PFX,h_cppr)++;
448
449 /* Map CPPR */
450 cppr = xive_prio_from_guest(cppr);
451
452 /* Remember old and update SW state */
453 old_cppr = xc->cppr;
454 xc->cppr = cppr;
455
456 /*
457 * Order the above update of xc->cppr with the subsequent
458 * read of xc->mfrr inside push_pending_to_hw()
459 */
460 smp_mb();
461
462 if (cppr > old_cppr) {
463 /*
464 * We are masking less, we need to look for pending things
465 * to deliver and set VP pending bits accordingly to trigger
466 * a new interrupt otherwise we might miss MFRR changes for
467 * which we have optimized out sending an IPI signal.
468 */
469 GLUE(X_PFX,push_pending_to_hw)(xc);
470 } else {
471 /*
472 * We are masking more, we need to check the queue for any
473 * interrupt that has been routed to another CPU, take
474 * it out (replace it with the dummy) and retrigger it.
475 *
476 * This is necessary since those interrupts may otherwise
477 * never be processed, at least not until this CPU restores
478 * its CPPR.
479 *
480 * This is in theory racy vs. HW adding new interrupts to
481 * the queue. In practice this works because the interesting
482 * cases are when the guest has done a set_xive() to move the
483 * interrupt away, which flushes the xive, followed by the
484 * target CPU doing a H_CPPR. So any new interrupt coming into
485 * the queue must still be routed to us and isn't a source
486 * of concern.
487 */
488 GLUE(X_PFX,scan_for_rerouted_irqs)(xive, xc);
489 }
490
491 /* Apply new CPPR */
492 xc->hw_cppr = cppr;
493 __x_writeb(cppr, __x_tima + TM_QW1_OS + TM_CPPR);
494
495 return H_SUCCESS;
496 }
497
GLUE(X_PFX,h_eoi)498 X_STATIC int GLUE(X_PFX,h_eoi)(struct kvm_vcpu *vcpu, unsigned long xirr)
499 {
500 struct kvmppc_xive *xive = vcpu->kvm->arch.xive;
501 struct kvmppc_xive_src_block *sb;
502 struct kvmppc_xive_irq_state *state;
503 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
504 struct xive_irq_data *xd;
505 u8 new_cppr = xirr >> 24;
506 u32 irq = xirr & 0x00ffffff, hw_num;
507 u16 src;
508 int rc = 0;
509
510 pr_devel("H_EOI(xirr=%08lx)\n", xirr);
511
512 xc->GLUE(X_STAT_PFX,h_eoi)++;
513
514 xc->cppr = xive_prio_from_guest(new_cppr);
515
516 /*
517 * IPIs are synthetized from MFRR and thus don't need
518 * any special EOI handling. The underlying interrupt
519 * used to signal MFRR changes is EOId when fetched from
520 * the queue.
521 */
522 if (irq == XICS_IPI || irq == 0) {
523 /*
524 * This barrier orders the setting of xc->cppr vs.
525 * subsquent test of xc->mfrr done inside
526 * scan_interrupts and push_pending_to_hw
527 */
528 smp_mb();
529 goto bail;
530 }
531
532 /* Find interrupt source */
533 sb = kvmppc_xive_find_source(xive, irq, &src);
534 if (!sb) {
535 pr_devel(" source not found !\n");
536 rc = H_PARAMETER;
537 /* Same as above */
538 smp_mb();
539 goto bail;
540 }
541 state = &sb->irq_state[src];
542 kvmppc_xive_select_irq(state, &hw_num, &xd);
543
544 state->in_eoi = true;
545
546 /*
547 * This barrier orders both setting of in_eoi above vs,
548 * subsequent test of guest_priority, and the setting
549 * of xc->cppr vs. subsquent test of xc->mfrr done inside
550 * scan_interrupts and push_pending_to_hw
551 */
552 smp_mb();
553
554 again:
555 if (state->guest_priority == MASKED) {
556 arch_spin_lock(&sb->lock);
557 if (state->guest_priority != MASKED) {
558 arch_spin_unlock(&sb->lock);
559 goto again;
560 }
561 pr_devel(" EOI on saved P...\n");
562
563 /* Clear old_p, that will cause unmask to perform an EOI */
564 state->old_p = false;
565
566 arch_spin_unlock(&sb->lock);
567 } else {
568 pr_devel(" EOI on source...\n");
569
570 /* Perform EOI on the source */
571 GLUE(X_PFX,source_eoi)(hw_num, xd);
572
573 /* If it's an emulated LSI, check level and resend */
574 if (state->lsi && state->asserted)
575 __x_writeq(0, __x_trig_page(xd));
576
577 }
578
579 /*
580 * This barrier orders the above guest_priority check
581 * and spin_lock/unlock with clearing in_eoi below.
582 *
583 * It also has to be a full mb() as it must ensure
584 * the MMIOs done in source_eoi() are completed before
585 * state->in_eoi is visible.
586 */
587 mb();
588 state->in_eoi = false;
589 bail:
590
591 /* Re-evaluate pending IRQs and update HW */
592 GLUE(X_PFX,scan_interrupts)(xc, xc->pending, scan_eoi);
593 GLUE(X_PFX,push_pending_to_hw)(xc);
594 pr_devel(" after scan pending=%02x\n", xc->pending);
595
596 /* Apply new CPPR */
597 xc->hw_cppr = xc->cppr;
598 __x_writeb(xc->cppr, __x_tima + TM_QW1_OS + TM_CPPR);
599
600 return rc;
601 }
602
GLUE(X_PFX,h_ipi)603 X_STATIC int GLUE(X_PFX,h_ipi)(struct kvm_vcpu *vcpu, unsigned long server,
604 unsigned long mfrr)
605 {
606 struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
607
608 pr_devel("H_IPI(server=%08lx,mfrr=%ld)\n", server, mfrr);
609
610 xc->GLUE(X_STAT_PFX,h_ipi)++;
611
612 /* Find target */
613 vcpu = kvmppc_xive_find_server(vcpu->kvm, server);
614 if (!vcpu)
615 return H_PARAMETER;
616 xc = vcpu->arch.xive_vcpu;
617
618 /* Locklessly write over MFRR */
619 xc->mfrr = mfrr;
620
621 /*
622 * The load of xc->cppr below and the subsequent MMIO store
623 * to the IPI must happen after the above mfrr update is
624 * globally visible so that:
625 *
626 * - Synchronize with another CPU doing an H_EOI or a H_CPPR
627 * updating xc->cppr then reading xc->mfrr.
628 *
629 * - The target of the IPI sees the xc->mfrr update
630 */
631 mb();
632
633 /* Shoot the IPI if most favored than target cppr */
634 if (mfrr < xc->cppr)
635 __x_writeq(0, __x_trig_page(&xc->vp_ipi_data));
636
637 return H_SUCCESS;
638 }
639
