1// SPDX-License-Identifier: GPL-2.0-or-later
2/*
3 * Copyright 2016,2017 IBM Corporation.
4 */
5
6#define pr_fmt(fmt) "xive: " fmt
7
8#include <linux/types.h>
9#include <linux/threads.h>
10#include <linux/kernel.h>
11#include <linux/irq.h>
12#include <linux/debugfs.h>
13#include <linux/smp.h>
14#include <linux/interrupt.h>
15#include <linux/seq_file.h>
16#include <linux/init.h>
17#include <linux/cpu.h>
18#include <linux/of.h>
19#include <linux/slab.h>
20#include <linux/spinlock.h>
21#include <linux/msi.h>
22#include <linux/vmalloc.h>
23
24#include <asm/debugfs.h>
25#include <asm/prom.h>
26#include <asm/io.h>
27#include <asm/smp.h>
28#include <asm/machdep.h>
29#include <asm/irq.h>
30#include <asm/errno.h>
31#include <asm/xive.h>
32#include <asm/xive-regs.h>
33#include <asm/xmon.h>
34
35#include "xive-internal.h"
36
37#undef DEBUG_FLUSH
38#undef DEBUG_ALL
39
40#ifdef DEBUG_ALL
41#define DBG_VERBOSE(fmt, ...)	pr_devel("cpu %d - " fmt, \
42					 smp_processor_id(), ## __VA_ARGS__)
43#else
44#define DBG_VERBOSE(fmt...)	do { } while(0)
45#endif
46
47bool __xive_enabled;
48EXPORT_SYMBOL_GPL(__xive_enabled);
49bool xive_cmdline_disabled;
50
51/* We use only one priority for now */
52static u8 xive_irq_priority;
53
54/* TIMA exported to KVM */
55void __iomem *xive_tima;
56EXPORT_SYMBOL_GPL(xive_tima);
57u32 xive_tima_offset;
58
59/* Backend ops */
60static const struct xive_ops *xive_ops;
61
62/* Our global interrupt domain */
63static struct irq_domain *xive_irq_domain;
64
65#ifdef CONFIG_SMP
66/* The IPIs all use the same logical irq number */
67static u32 xive_ipi_irq;
68#endif
69
70/* Xive state for each CPU */
71static DEFINE_PER_CPU(struct xive_cpu *, xive_cpu);
72
73/* An invalid CPU target */
74#define XIVE_INVALID_TARGET	(-1)
75
76/*
77 * Read the next entry in a queue, return its content if it's valid
78 * or 0 if there is no new entry.
79 *
80 * The queue pointer is moved forward unless "just_peek" is set
81 */
82static u32 xive_read_eq(struct xive_q *q, bool just_peek)
83{
84	u32 cur;
85
86	if (!q->qpage)
87		return 0;
88	cur = be32_to_cpup(q->qpage + q->idx);
89
90	/* Check valid bit (31) vs current toggle polarity */
91	if ((cur >> 31) == q->toggle)
92		return 0;
93
94	/* If consuming from the queue ... */
95	if (!just_peek) {
96		/* Next entry */
97		q->idx = (q->idx + 1) & q->msk;
98
99		/* Wrap around: flip valid toggle */
100		if (q->idx == 0)
101			q->toggle ^= 1;
102	}
103	/* Mask out the valid bit (31) */
104	return cur & 0x7fffffff;
105}
106
107/*
108 * Scans all the queue that may have interrupts in them
109 * (based on "pending_prio") in priority order until an
110 * interrupt is found or all the queues are empty.
111 *
112 * Then updates the CPPR (Current Processor Priority
113 * Register) based on the most favored interrupt found
114 * (0xff if none) and return what was found (0 if none).
115 *
116 * If just_peek is set, return the most favored pending
117 * interrupt if any but don't update the queue pointers.
118 *
119 * Note: This function can operate generically on any number
120 * of queues (up to 8). The current implementation of the XIVE
121 * driver only uses a single queue however.
122 *
123 * Note2: This will also "flush" "the pending_count" of a queue
124 * into the "count" when that queue is observed to be empty.
125 * This is used to keep track of the amount of interrupts
126 * targetting a queue. When an interrupt is moved away from
127 * a queue, we only decrement that queue count once the queue
128 * has been observed empty to avoid races.
129 */
130static u32 xive_scan_interrupts(struct xive_cpu *xc, bool just_peek)
131{
132	u32 irq = 0;
133	u8 prio = 0;
134
135	/* Find highest pending priority */
136	while (xc->pending_prio != 0) {
137		struct xive_q *q;
138
139		prio = ffs(xc->pending_prio) - 1;
140		DBG_VERBOSE("scan_irq: trying prio %d\n", prio);
141
142		/* Try to fetch */
143		irq = xive_read_eq(&xc->queue[prio], just_peek);
144
145		/* Found something ? That's it */
146		if (irq) {
147			if (just_peek || irq_to_desc(irq))
148				break;
149			/*
150			 * We should never get here; if we do then we must
151			 * have failed to synchronize the interrupt properly
152			 * when shutting it down.
153			 */
154			pr_crit("xive: got interrupt %d without descriptor, dropping\n",
155				irq);
156			WARN_ON(1);
157			continue;
158		}
159
160		/* Clear pending bits */
161		xc->pending_prio &= ~(1 << prio);
162
163		/*
164		 * Check if the queue count needs adjusting due to
165		 * interrupts being moved away. See description of
166		 * xive_dec_target_count()
167		 */
168		q = &xc->queue[prio];
169		if (atomic_read(&q->pending_count)) {
170			int p = atomic_xchg(&q->pending_count, 0);
171			if (p) {
172				WARN_ON(p > atomic_read(&q->count));
173				atomic_sub(p, &q->count);
174			}
175		}
176	}
177
178	/* If nothing was found, set CPPR to 0xff */
179	if (irq == 0)
180		prio = 0xff;
181
182	/* Update HW CPPR to match if necessary */
183	if (prio != xc->cppr) {
184		DBG_VERBOSE("scan_irq: adjusting CPPR to %d\n", prio);
185		xc->cppr = prio;
186		out_8(xive_tima + xive_tima_offset + TM_CPPR, prio);
187	}
188
189	return irq;
190}
191
192/*
193 * This is used to perform the magic loads from an ESB
194 * described in xive-regs.h
195 */
196static notrace u8 xive_esb_read(struct xive_irq_data *xd, u32 offset)
197{
198	u64 val;
199
200	if (offset == XIVE_ESB_SET_PQ_10 && xd->flags & XIVE_IRQ_FLAG_STORE_EOI)
201		offset |= XIVE_ESB_LD_ST_MO;
202
203	/* Handle HW errata */
204	if (xd->flags & XIVE_IRQ_FLAG_SHIFT_BUG)
205		offset |= offset << 4;
206
207	if ((xd->flags & XIVE_IRQ_FLAG_H_INT_ESB) && xive_ops->esb_rw)
208		val = xive_ops->esb_rw(xd->hw_irq, offset, 0, 0);
209	else
210		val = in_be64(xd->eoi_mmio + offset);
211
212	return (u8)val;
213}
214
215static void xive_esb_write(struct xive_irq_data *xd, u32 offset, u64 data)
216{
217	/* Handle HW errata */
218	if (xd->flags & XIVE_IRQ_FLAG_SHIFT_BUG)
219		offset |= offset << 4;
220
221	if ((xd->flags & XIVE_IRQ_FLAG_H_INT_ESB) && xive_ops->esb_rw)
222		xive_ops->esb_rw(xd->hw_irq, offset, data, 1);
223	else
224		out_be64(xd->eoi_mmio + offset, data);
225}
226
227#ifdef CONFIG_XMON
228static notrace void xive_dump_eq(const char *name, struct xive_q *q)
229{
230	u32 i0, i1, idx;
231
232	if (!q->qpage)
233		return;
234	idx = q->idx;
235	i0 = be32_to_cpup(q->qpage + idx);
236	idx = (idx + 1) & q->msk;
237	i1 = be32_to_cpup(q->qpage + idx);
238	xmon_printf("%s idx=%d T=%d %08x %08x ...", name,
239		     q->idx, q->toggle, i0, i1);
240}
241
242notrace void xmon_xive_do_dump(int cpu)
243{
244	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
245
246	xmon_printf("CPU %d:", cpu);
247	if (xc) {
248		xmon_printf("pp=%02x CPPR=%02x ", xc->pending_prio, xc->cppr);
249
250#ifdef CONFIG_SMP
251		{
252			u64 val = xive_esb_read(&xc->ipi_data, XIVE_ESB_GET);
253
254			xmon_printf("IPI=0x%08x PQ=%c%c ", xc->hw_ipi,
255				    val & XIVE_ESB_VAL_P ? 'P' : '-',
256				    val & XIVE_ESB_VAL_Q ? 'Q' : '-');
257		}
258#endif
259		xive_dump_eq("EQ", &xc->queue[xive_irq_priority]);
260	}
261	xmon_printf("\n");
262}
263
264static struct irq_data *xive_get_irq_data(u32 hw_irq)
265{
266	unsigned int irq = irq_find_mapping(xive_irq_domain, hw_irq);
267
268	return irq ? irq_get_irq_data(irq) : NULL;
269}
270
271int xmon_xive_get_irq_config(u32 hw_irq, struct irq_data *d)
272{
273	int rc;
274	u32 target;
275	u8 prio;
276	u32 lirq;
277
278	rc = xive_ops->get_irq_config(hw_irq, &target, &prio, &lirq);
279	if (rc) {
280		xmon_printf("IRQ 0x%08x : no config rc=%d\n", hw_irq, rc);
281		return rc;
282	}
283
284	xmon_printf("IRQ 0x%08x : target=0x%x prio=%02x lirq=0x%x ",
285		    hw_irq, target, prio, lirq);
286
287	if (!d)
288		d = xive_get_irq_data(hw_irq);
289
290	if (d) {
291		struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
292		u64 val = xive_esb_read(xd, XIVE_ESB_GET);
293
294		xmon_printf("flags=%c%c%c PQ=%c%c",
295			    xd->flags & XIVE_IRQ_FLAG_STORE_EOI ? 'S' : ' ',
296			    xd->flags & XIVE_IRQ_FLAG_LSI ? 'L' : ' ',
297			    xd->flags & XIVE_IRQ_FLAG_H_INT_ESB ? 'H' : ' ',
298			    val & XIVE_ESB_VAL_P ? 'P' : '-',
299			    val & XIVE_ESB_VAL_Q ? 'Q' : '-');
300	}
301
302	xmon_printf("\n");
303	return 0;
304}
305
306#endif /* CONFIG_XMON */
307
308static unsigned int xive_get_irq(void)
309{
310	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
311	u32 irq;
312
313	/*
314	 * This can be called either as a result of a HW interrupt or
315	 * as a "replay" because EOI decided there was still something
316	 * in one of the queues.
317	 *
318	 * First we perform an ACK cycle in order to update our mask
319	 * of pending priorities. This will also have the effect of
320	 * updating the CPPR to the most favored pending interrupts.
321	 *
322	 * In the future, if we have a way to differentiate a first
323	 * entry (on HW interrupt) from a replay triggered by EOI,
324	 * we could skip this on replays unless we soft-mask tells us
325	 * that a new HW interrupt occurred.
326	 */
327	xive_ops->update_pending(xc);
328
329	DBG_VERBOSE("get_irq: pending=%02x\n", xc->pending_prio);
330
331	/* Scan our queue(s) for interrupts */
332	irq = xive_scan_interrupts(xc, false);
333
334	DBG_VERBOSE("get_irq: got irq 0x%x, new pending=0x%02x\n",
335	    irq, xc->pending_prio);
336
337	/* Return pending interrupt if any */
338	if (irq == XIVE_BAD_IRQ)
339		return 0;
340	return irq;
341}
342
343/*
344 * After EOI'ing an interrupt, we need to re-check the queue
345 * to see if another interrupt is pending since multiple
346 * interrupts can coalesce into a single notification to the
347 * CPU.
348 *
349 * If we find that there is indeed more in there, we call
350 * force_external_irq_replay() to make Linux synthetize an
351 * external interrupt on the next call to local_irq_restore().
352 */
353static void xive_do_queue_eoi(struct xive_cpu *xc)
354{
355	if (xive_scan_interrupts(xc, true) != 0) {
356		DBG_VERBOSE("eoi: pending=0x%02x\n", xc->pending_prio);
357		force_external_irq_replay();
358	}
359}
360
361/*
362 * EOI an interrupt at the source. There are several methods
363 * to do this depending on the HW version and source type
364 */
365static void xive_do_source_eoi(u32 hw_irq, struct xive_irq_data *xd)
366{
367	xd->stale_p = false;
368	/* If the XIVE supports the new "store EOI facility, use it */
369	if (xd->flags & XIVE_IRQ_FLAG_STORE_EOI)
370		xive_esb_write(xd, XIVE_ESB_STORE_EOI, 0);
371	else if (hw_irq && xd->flags & XIVE_IRQ_FLAG_EOI_FW) {
372		/*
373		 * The FW told us to call it. This happens for some
374		 * interrupt sources that need additional HW whacking
375		 * beyond the ESB manipulation. For example LPC interrupts
376		 * on P9 DD1.0 needed a latch to be clared in the LPC bridge
377		 * itself. The Firmware will take care of it.
378		 */
379		if (WARN_ON_ONCE(!xive_ops->eoi))
380			return;
381		xive_ops->eoi(hw_irq);
382	} else {
383		u8 eoi_val;
384
385		/*
386		 * Otherwise for EOI, we use the special MMIO that does
387		 * a clear of both P and Q and returns the old Q,
388		 * except for LSIs where we use the "EOI cycle" special
389		 * load.
390		 *
391		 * This allows us to then do a re-trigger if Q was set
392		 * rather than synthesizing an interrupt in software
393		 *
394		 * For LSIs the HW EOI cycle is used rather than PQ bits,
395		 * as they are automatically re-triggred in HW when still
396		 * pending.
397		 */
398		if (xd->flags & XIVE_IRQ_FLAG_LSI)
399			xive_esb_read(xd, XIVE_ESB_LOAD_EOI);
400		else {
401			eoi_val = xive_esb_read(xd, XIVE_ESB_SET_PQ_00);
402			DBG_VERBOSE("eoi_val=%x\n", eoi_val);
403
404			/* Re-trigger if needed */
405			if ((eoi_val & XIVE_ESB_VAL_Q) && xd->trig_mmio)
406				out_be64(xd->trig_mmio, 0);
407		}
408	}
409}
410
411/* irq_chip eoi callback, called with irq descriptor lock held */
412static void xive_irq_eoi(struct irq_data *d)
413{
414	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
415	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
416
417	DBG_VERBOSE("eoi_irq: irq=%d [0x%lx] pending=%02x\n",
418		    d->irq, irqd_to_hwirq(d), xc->pending_prio);
419
420	/*
421	 * EOI the source if it hasn't been disabled and hasn't
422	 * been passed-through to a KVM guest
423	 */
424	if (!irqd_irq_disabled(d) && !irqd_is_forwarded_to_vcpu(d) &&
425	    !(xd->flags & XIVE_IRQ_NO_EOI))
426		xive_do_source_eoi(irqd_to_hwirq(d), xd);
427	else
428		xd->stale_p = true;
429
430	/*
431	 * Clear saved_p to indicate that it's no longer occupying
432	 * a queue slot on the target queue
433	 */
434	xd->saved_p = false;
435
436	/* Check for more work in the queue */
437	xive_do_queue_eoi(xc);
438}
439
440/*
441 * Helper used to mask and unmask an interrupt source. This
442 * is only called for normal interrupts that do not require
443 * masking/unmasking via firmware.
444 */
445static void xive_do_source_set_mask(struct xive_irq_data *xd,
446				    bool mask)
447{
448	u64 val;
449
450	/*
451	 * If the interrupt had P set, it may be in a queue.
452	 *
453	 * We need to make sure we don't re-enable it until it
454	 * has been fetched from that queue and EOId. We keep
455	 * a copy of that P state and use it to restore the
456	 * ESB accordingly on unmask.
457	 */
458	if (mask) {
459		val = xive_esb_read(xd, XIVE_ESB_SET_PQ_01);
460		if (!xd->stale_p && !!(val & XIVE_ESB_VAL_P))
461			xd->saved_p = true;
462		xd->stale_p = false;
463	} else if (xd->saved_p) {
464		xive_esb_read(xd, XIVE_ESB_SET_PQ_10);
465		xd->saved_p = false;
466	} else {
467		xive_esb_read(xd, XIVE_ESB_SET_PQ_00);
468		xd->stale_p = false;
469	}
470}
471
472/*
473 * Try to chose "cpu" as a new interrupt target. Increments
474 * the queue accounting for that target if it's not already
475 * full.
476 */
477static bool xive_try_pick_target(int cpu)
478{
479	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
480	struct xive_q *q = &xc->queue[xive_irq_priority];
481	int max;
482
483	/*
484	 * Calculate max number of interrupts in that queue.
485	 *
486	 * We leave a gap of 1 just in case...
487	 */
488	max = (q->msk + 1) - 1;
489	return !!atomic_add_unless(&q->count, 1, max);
490}
491
492/*
493 * Un-account an interrupt for a target CPU. We don't directly
494 * decrement q->count since the interrupt might still be present
495 * in the queue.
496 *
497 * Instead increment a separate counter "pending_count" which
498 * will be substracted from "count" later when that CPU observes
499 * the queue to be empty.
500 */
501static void xive_dec_target_count(int cpu)
502{
503	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
504	struct xive_q *q = &xc->queue[xive_irq_priority];
505
506	if (WARN_ON(cpu < 0 || !xc)) {
507		pr_err("%s: cpu=%d xc=%p\n", __func__, cpu, xc);
508		return;
509	}
510
511	/*
512	 * We increment the "pending count" which will be used
513	 * to decrement the target queue count whenever it's next
514	 * processed and found empty. This ensure that we don't
515	 * decrement while we still have the interrupt there
516	 * occupying a slot.
517	 */
518	atomic_inc(&q->pending_count);
519}
520
521/* Find a tentative CPU target in a CPU mask */
522static int xive_find_target_in_mask(const struct cpumask *mask,
523				    unsigned int fuzz)
524{
525	int cpu, first, num, i;
526
527	/* Pick up a starting point CPU in the mask based on  fuzz */
528	num = min_t(int, cpumask_weight(mask), nr_cpu_ids);
529	first = fuzz % num;
530
531	/* Locate it */
532	cpu = cpumask_first(mask);
533	for (i = 0; i < first && cpu < nr_cpu_ids; i++)
534		cpu = cpumask_next(cpu, mask);
535
536	/* Sanity check */
537	if (WARN_ON(cpu >= nr_cpu_ids))
538		cpu = cpumask_first(cpu_online_mask);
539
540	/* Remember first one to handle wrap-around */
541	first = cpu;
542
543	/*
544	 * Now go through the entire mask until we find a valid
545	 * target.
546	 */
547	do {
548		/*
549		 * We re-check online as the fallback case passes us
550		 * an untested affinity mask
551		 */
552		if (cpu_online(cpu) && xive_try_pick_target(cpu))
553			return cpu;
554		cpu = cpumask_next(cpu, mask);
555		/* Wrap around */
556		if (cpu >= nr_cpu_ids)
557			cpu = cpumask_first(mask);
558	} while (cpu != first);
559
560	return -1;
561}
562
563/*
564 * Pick a target CPU for an interrupt. This is done at
565 * startup or if the affinity is changed in a way that
566 * invalidates the current target.
567 */
568static int xive_pick_irq_target(struct irq_data *d,
569				const struct cpumask *affinity)
570{
571	static unsigned int fuzz;
572	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
573	cpumask_var_t mask;
574	int cpu = -1;
575
576	/*
577	 * If we have chip IDs, first we try to build a mask of
578	 * CPUs matching the CPU and find a target in there
579	 */
580	if (xd->src_chip != XIVE_INVALID_CHIP_ID &&
581		zalloc_cpumask_var(&mask, GFP_ATOMIC)) {
582		/* Build a mask of matching chip IDs */
583		for_each_cpu_and(cpu, affinity, cpu_online_mask) {
584			struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
585			if (xc->chip_id == xd->src_chip)
586				cpumask_set_cpu(cpu, mask);
587		}
588		/* Try to find a target */
589		if (cpumask_empty(mask))
590			cpu = -1;
591		else
592			cpu = xive_find_target_in_mask(mask, fuzz++);
593		free_cpumask_var(mask);
594		if (cpu >= 0)
595			return cpu;
596		fuzz--;
597	}
598
599	/* No chip IDs, fallback to using the affinity mask */
600	return xive_find_target_in_mask(affinity, fuzz++);
601}
602
603static unsigned int xive_irq_startup(struct irq_data *d)
604{
605	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
606	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
607	int target, rc;
608
609	xd->saved_p = false;
610	xd->stale_p = false;
611	pr_devel("xive_irq_startup: irq %d [0x%x] data @%p\n",
612		 d->irq, hw_irq, d);
613
614#ifdef CONFIG_PCI_MSI
615	/*
616	 * The generic MSI code returns with the interrupt disabled on the
617	 * card, using the MSI mask bits. Firmware doesn't appear to unmask
618	 * at that level, so we do it here by hand.
619	 */
620	if (irq_data_get_msi_desc(d))
621		pci_msi_unmask_irq(d);
622#endif
623
624	/* Pick a target */
625	target = xive_pick_irq_target(d, irq_data_get_affinity_mask(d));
626	if (target == XIVE_INVALID_TARGET) {
627		/* Try again breaking affinity */
628		target = xive_pick_irq_target(d, cpu_online_mask);
629		if (target == XIVE_INVALID_TARGET)
630			return -ENXIO;
631		pr_warn("irq %d started with broken affinity\n", d->irq);
632	}
633
634	/* Sanity check */
635	if (WARN_ON(target == XIVE_INVALID_TARGET ||
636		    target >= nr_cpu_ids))
637		target = smp_processor_id();
638
639	xd->target = target;
640
641	/*
642	 * Configure the logical number to be the Linux IRQ number
643	 * and set the target queue
644	 */
645	rc = xive_ops->configure_irq(hw_irq,
646				     get_hard_smp_processor_id(target),
647				     xive_irq_priority, d->irq);
648	if (rc)
649		return rc;
650
651	/* Unmask the ESB */
652	xive_do_source_set_mask(xd, false);
653
654	return 0;
655}
656
657/* called with irq descriptor lock held */
658static void xive_irq_shutdown(struct irq_data *d)
659{
660	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
661	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
662
663	pr_devel("xive_irq_shutdown: irq %d [0x%x] data @%p\n",
664		 d->irq, hw_irq, d);
665
666	if (WARN_ON(xd->target == XIVE_INVALID_TARGET))
667		return;
668
669	/* Mask the interrupt at the source */
670	xive_do_source_set_mask(xd, true);
671
672	/*
673	 * Mask the interrupt in HW in the IVT/EAS and set the number
674	 * to be the "bad" IRQ number
675	 */
676	xive_ops->configure_irq(hw_irq,
677				get_hard_smp_processor_id(xd->target),
678				0xff, XIVE_BAD_IRQ);
679
680	xive_dec_target_count(xd->target);
681	xd->target = XIVE_INVALID_TARGET;
682}
683
684static void xive_irq_unmask(struct irq_data *d)
685{
686	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
687
688	pr_devel("xive_irq_unmask: irq %d data @%p\n", d->irq, xd);
689
690	/*
691	 * This is a workaround for PCI LSI problems on P9, for
692	 * these, we call FW to set the mask. The problems might
693	 * be fixed by P9 DD2.0, if that is the case, firmware
694	 * will no longer set that flag.
695	 */
696	if (xd->flags & XIVE_IRQ_FLAG_MASK_FW) {
697		unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
698		xive_ops->configure_irq(hw_irq,
699					get_hard_smp_processor_id(xd->target),
700					xive_irq_priority, d->irq);
701		return;
702	}
703
704	xive_do_source_set_mask(xd, false);
705}
706
707static void xive_irq_mask(struct irq_data *d)
708{
709	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
710
711	pr_devel("xive_irq_mask: irq %d data @%p\n", d->irq, xd);
712
713	/*
714	 * This is a workaround for PCI LSI problems on P9, for
715	 * these, we call OPAL to set the mask. The problems might
716	 * be fixed by P9 DD2.0, if that is the case, firmware
717	 * will no longer set that flag.
718	 */
719	if (xd->flags & XIVE_IRQ_FLAG_MASK_FW) {
720		unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
721		xive_ops->configure_irq(hw_irq,
722					get_hard_smp_processor_id(xd->target),
723					0xff, d->irq);
724		return;
725	}
726
727	xive_do_source_set_mask(xd, true);
728}
729
730static int xive_irq_set_affinity(struct irq_data *d,
731				 const struct cpumask *cpumask,
732				 bool force)
733{
734	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
735	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
736	u32 target, old_target;
737	int rc = 0;
738
739	pr_devel("xive_irq_set_affinity: irq %d\n", d->irq);
740
741	/* Is this valid ? */
742	if (cpumask_any_and(cpumask, cpu_online_mask) >= nr_cpu_ids)
743		return -EINVAL;
744
745	/* Don't do anything if the interrupt isn't started */
746	if (!irqd_is_started(d))
747		return IRQ_SET_MASK_OK;
748
749	/*
750	 * If existing target is already in the new mask, and is
751	 * online then do nothing.
752	 */
753	if (xd->target != XIVE_INVALID_TARGET &&
754	    cpu_online(xd->target) &&
755	    cpumask_test_cpu(xd->target, cpumask))
756		return IRQ_SET_MASK_OK;
757
758	/* Pick a new target */
759	target = xive_pick_irq_target(d, cpumask);
760
761	/* No target found */
762	if (target == XIVE_INVALID_TARGET)
763		return -ENXIO;
764
765	/* Sanity check */
766	if (WARN_ON(target >= nr_cpu_ids))
767		target = smp_processor_id();
768
769	old_target = xd->target;
770
771	/*
772	 * Only configure the irq if it's not currently passed-through to
773	 * a KVM guest
774	 */
775	if (!irqd_is_forwarded_to_vcpu(d))
776		rc = xive_ops->configure_irq(hw_irq,
777					     get_hard_smp_processor_id(target),
778					     xive_irq_priority, d->irq);
779	if (rc < 0) {
780		pr_err("Error %d reconfiguring irq %d\n", rc, d->irq);
781		return rc;
782	}
783
784	pr_devel("  target: 0x%x\n", target);
785	xd->target = target;
786
787	/* Give up previous target */
788	if (old_target != XIVE_INVALID_TARGET)
789	    xive_dec_target_count(old_target);
790
791	return IRQ_SET_MASK_OK;
792}
793
794static int xive_irq_set_type(struct irq_data *d, unsigned int flow_type)
795{
796	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
797
798	/*
799	 * We only support these. This has really no effect other than setting
800	 * the corresponding descriptor bits mind you but those will in turn
801	 * affect the resend function when re-enabling an edge interrupt.
802	 *
803	 * Set set the default to edge as explained in map().
804	 */
805	if (flow_type == IRQ_TYPE_DEFAULT || flow_type == IRQ_TYPE_NONE)
806		flow_type = IRQ_TYPE_EDGE_RISING;
807
808	if (flow_type != IRQ_TYPE_EDGE_RISING &&
809	    flow_type != IRQ_TYPE_LEVEL_LOW)
810		return -EINVAL;
811
812	irqd_set_trigger_type(d, flow_type);
813
814	/*
815	 * Double check it matches what the FW thinks
816	 *
817	 * NOTE: We don't know yet if the PAPR interface will provide
818	 * the LSI vs MSI information apart from the device-tree so
819	 * this check might have to move into an optional backend call
820	 * that is specific to the native backend
821	 */
822	if ((flow_type == IRQ_TYPE_LEVEL_LOW) !=
823	    !!(xd->flags & XIVE_IRQ_FLAG_LSI)) {
824		pr_warn("Interrupt %d (HW 0x%x) type mismatch, Linux says %s, FW says %s\n",
825			d->irq, (u32)irqd_to_hwirq(d),
826			(flow_type == IRQ_TYPE_LEVEL_LOW) ? "Level" : "Edge",
827			(xd->flags & XIVE_IRQ_FLAG_LSI) ? "Level" : "Edge");
828	}
829
830	return IRQ_SET_MASK_OK_NOCOPY;
831}
832
833static int xive_irq_retrigger(struct irq_data *d)
834{
835	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
836
837	/* This should be only for MSIs */
838	if (WARN_ON(xd->flags & XIVE_IRQ_FLAG_LSI))
839		return 0;
840
841	/*
842	 * To perform a retrigger, we first set the PQ bits to
843	 * 11, then perform an EOI.
844	 */
845	xive_esb_read(xd, XIVE_ESB_SET_PQ_11);
846
847	/*
848	 * Note: We pass "0" to the hw_irq argument in order to
849	 * avoid calling into the backend EOI code which we don't
850	 * want to do in the case of a re-trigger. Backends typically
851	 * only do EOI for LSIs anyway.
852	 */
853	xive_do_source_eoi(0, xd);
854
855	return 1;
856}
857
858/*
859 * Caller holds the irq descriptor lock, so this won't be called
860 * concurrently with xive_get_irqchip_state on the same interrupt.
861 */
862static int xive_irq_set_vcpu_affinity(struct irq_data *d, void *state)
863{
864	struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
865	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
866	int rc;
867	u8 pq;
868
869	/*
870	 * We only support this on interrupts that do not require
871	 * firmware calls for masking and unmasking
872	 */
873	if (xd->flags & XIVE_IRQ_FLAG_MASK_FW)
874		return -EIO;
875
876	/*
877	 * This is called by KVM with state non-NULL for enabling
878	 * pass-through or NULL for disabling it
879	 */
880	if (state) {
881		irqd_set_forwarded_to_vcpu(d);
882
883		/* Set it to PQ=10 state to prevent further sends */
884		pq = xive_esb_read(xd, XIVE_ESB_SET_PQ_10);
885		if (!xd->stale_p) {
886			xd->saved_p = !!(pq & XIVE_ESB_VAL_P);
887			xd->stale_p = !xd->saved_p;
888		}
889
890		/* No target ? nothing to do */
891		if (xd->target == XIVE_INVALID_TARGET) {
892			/*
893			 * An untargetted interrupt should have been
894			 * also masked at the source
895			 */
896			WARN_ON(xd->saved_p);
897
898			return 0;
899		}
900
901		/*
902		 * If P was set, adjust state to PQ=11 to indicate
903		 * that a resend is needed for the interrupt to reach
904		 * the guest. Also remember the value of P.
905		 *
906		 * This also tells us that it's in flight to a host queue
907		 * or has already been fetched but hasn't been EOIed yet
908		 * by the host. This it's potentially using up a host
909		 * queue slot. This is important to know because as long
910		 * as this is the case, we must not hard-unmask it when
911		 * "returning" that interrupt to the host.
912		 *
913		 * This saved_p is cleared by the host EOI, when we know
914		 * for sure the queue slot is no longer in use.
915		 */
916		if (xd->saved_p) {
917			xive_esb_read(xd, XIVE_ESB_SET_PQ_11);
918
919			/*
920			 * Sync the XIVE source HW to ensure the interrupt
921			 * has gone through the EAS before we change its
922			 * target to the guest. That should guarantee us
923			 * that we *will* eventually get an EOI for it on
924			 * the host. Otherwise there would be a small window
925			 * for P to be seen here but the interrupt going
926			 * to the guest queue.
927			 */
928			if (xive_ops->sync_source)
929				xive_ops->sync_source(hw_irq);
930		}
931	} else {
932		irqd_clr_forwarded_to_vcpu(d);
933
934		/* No host target ? hard mask and return */
935		if (xd->target == XIVE_INVALID_TARGET) {
936			xive_do_source_set_mask(xd, true);
937			return 0;
938		}
939
940		/*
941		 * Sync the XIVE source HW to ensure the interrupt
942		 * has gone through the EAS before we change its
943		 * target to the host.
944		 */
945		if (xive_ops->sync_source)
946			xive_ops->sync_source(hw_irq);
947
948		/*
949		 * By convention we are called with the interrupt in
950		 * a PQ=10 or PQ=11 state, ie, it won't fire and will
951		 * have latched in Q whether there's a pending HW
952		 * interrupt or not.
953		 *
954		 * First reconfigure the target.
955		 */
956		rc = xive_ops->configure_irq(hw_irq,
957					     get_hard_smp_processor_id(xd->target),
958					     xive_irq_priority, d->irq);
959		if (rc)
960			return rc;
961
962		/*
963		 * Then if saved_p is not set, effectively re-enable the
964		 * interrupt with an EOI. If it is set, we know there is
965		 * still a message in a host queue somewhere that will be
966		 * EOId eventually.
967		 *
968		 * Note: We don't check irqd_irq_disabled(). Effectively,
969		 * we *will* let the irq get through even if masked if the
970		 * HW is still firing it in order to deal with the whole
971		 * saved_p business properly. If the interrupt triggers
972		 * while masked, the generic code will re-mask it anyway.
973		 */
974		if (!xd->saved_p)
975			xive_do_source_eoi(hw_irq, xd);
976
977	}
978	return 0;
979}
980
981/* Called with irq descriptor lock held. */
982static int xive_get_irqchip_state(struct irq_data *data,
983				  enum irqchip_irq_state which, bool *state)
984{
985	struct xive_irq_data *xd = irq_data_get_irq_handler_data(data);
986	u8 pq;
987
988	switch (which) {
989	case IRQCHIP_STATE_ACTIVE:
990		pq = xive_esb_read(xd, XIVE_ESB_GET);
991
992		/*
993		 * The esb value being all 1's means we couldn't get
994		 * the PQ state of the interrupt through mmio. It may
995		 * happen, for example when querying a PHB interrupt
996		 * while the PHB is in an error state. We consider the
997		 * interrupt to be inactive in that case.
998		 */
999		*state = (pq != XIVE_ESB_INVALID) && !xd->stale_p &&
1000			(xd->saved_p || (!!(pq & XIVE_ESB_VAL_P) &&
1001			 !irqd_irq_disabled(data)));
1002		return 0;
1003	default:
1004		return -EINVAL;
1005	}
1006}
1007
1008static struct irq_chip xive_irq_chip = {
1009	.name = "XIVE-IRQ",
1010	.irq_startup = xive_irq_startup,
1011	.irq_shutdown = xive_irq_shutdown,
1012	.irq_eoi = xive_irq_eoi,
1013	.irq_mask = xive_irq_mask,
1014	.irq_unmask = xive_irq_unmask,
1015	.irq_set_affinity = xive_irq_set_affinity,
1016	.irq_set_type = xive_irq_set_type,
1017	.irq_retrigger = xive_irq_retrigger,
1018	.irq_set_vcpu_affinity = xive_irq_set_vcpu_affinity,
1019	.irq_get_irqchip_state = xive_get_irqchip_state,
1020};
1021
1022bool is_xive_irq(struct irq_chip *chip)
1023{
1024	return chip == &xive_irq_chip;
1025}
1026EXPORT_SYMBOL_GPL(is_xive_irq);
1027
1028void xive_cleanup_irq_data(struct xive_irq_data *xd)
1029{
1030	if (xd->eoi_mmio) {
1031		unmap_kernel_range((unsigned long)xd->eoi_mmio,
1032				   1u << xd->esb_shift);
1033		iounmap(xd->eoi_mmio);
1034		if (xd->eoi_mmio == xd->trig_mmio)
1035			xd->trig_mmio = NULL;
1036		xd->eoi_mmio = NULL;
1037	}
1038	if (xd->trig_mmio) {
1039		unmap_kernel_range((unsigned long)xd->trig_mmio,
1040				   1u << xd->esb_shift);
1041		iounmap(xd->trig_mmio);
1042		xd->trig_mmio = NULL;
1043	}
1044}
1045EXPORT_SYMBOL_GPL(xive_cleanup_irq_data);
1046
1047static int xive_irq_alloc_data(unsigned int virq, irq_hw_number_t hw)
1048{
1049	struct xive_irq_data *xd;
1050	int rc;
1051
1052	xd = kzalloc(sizeof(struct xive_irq_data), GFP_KERNEL);
1053	if (!xd)
1054		return -ENOMEM;
1055	rc = xive_ops->populate_irq_data(hw, xd);
1056	if (rc) {
1057		kfree(xd);
1058		return rc;
1059	}
1060	xd->target = XIVE_INVALID_TARGET;
1061	irq_set_handler_data(virq, xd);
1062
1063	/*
1064	 * Turn OFF by default the interrupt being mapped. A side
1065	 * effect of this check is the mapping the ESB page of the
1066	 * interrupt in the Linux address space. This prevents page
1067	 * fault issues in the crash handler which masks all
1068	 * interrupts.
1069	 */
1070	xive_esb_read(xd, XIVE_ESB_SET_PQ_01);
1071
1072	return 0;
1073}
1074
1075static void xive_irq_free_data(unsigned int virq)
1076{
1077	struct xive_irq_data *xd = irq_get_handler_data(virq);
1078
1079	if (!xd)
1080		return;
1081	irq_set_handler_data(virq, NULL);
1082	xive_cleanup_irq_data(xd);
1083	kfree(xd);
1084}
1085
1086#ifdef CONFIG_SMP
1087
1088static void xive_cause_ipi(int cpu)
1089{
1090	struct xive_cpu *xc;
1091	struct xive_irq_data *xd;
1092
1093	xc = per_cpu(xive_cpu, cpu);
1094
1095	DBG_VERBOSE("IPI CPU %d -> %d (HW IRQ 0x%x)\n",
1096		    smp_processor_id(), cpu, xc->hw_ipi);
1097
1098	xd = &xc->ipi_data;
1099	if (WARN_ON(!xd->trig_mmio))
1100		return;
1101	out_be64(xd->trig_mmio, 0);
1102}
1103
1104static irqreturn_t xive_muxed_ipi_action(int irq, void *dev_id)
1105{
1106	return smp_ipi_demux();
1107}
1108
1109static void xive_ipi_eoi(struct irq_data *d)
1110{
1111	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1112
1113	/* Handle possible race with unplug and drop stale IPIs */
1114	if (!xc)
1115		return;
1116
1117	DBG_VERBOSE("IPI eoi: irq=%d [0x%lx] (HW IRQ 0x%x) pending=%02x\n",
1118		    d->irq, irqd_to_hwirq(d), xc->hw_ipi, xc->pending_prio);
1119
1120	xive_do_source_eoi(xc->hw_ipi, &xc->ipi_data);
1121	xive_do_queue_eoi(xc);
1122}
1123
1124static void xive_ipi_do_nothing(struct irq_data *d)
1125{
1126	/*
1127	 * Nothing to do, we never mask/unmask IPIs, but the callback
1128	 * has to exist for the struct irq_chip.
1129	 */
1130}
1131
1132static struct irq_chip xive_ipi_chip = {
1133	.name = "XIVE-IPI",
1134	.irq_eoi = xive_ipi_eoi,
1135	.irq_mask = xive_ipi_do_nothing,
1136	.irq_unmask = xive_ipi_do_nothing,
1137};
1138
1139static void __init xive_request_ipi(void)
1140{
1141	unsigned int virq;
1142
1143	/*
1144	 * Initialization failed, move on, we might manage to
1145	 * reach the point where we display our errors before
1146	 * the system falls appart
1147	 */
1148	if (!xive_irq_domain)
1149		return;
1150
1151	/* Initialize it */
1152	virq = irq_create_mapping(xive_irq_domain, 0);
1153	xive_ipi_irq = virq;
1154
1155	WARN_ON(request_irq(virq, xive_muxed_ipi_action,
1156			    IRQF_PERCPU | IRQF_NO_THREAD, "IPI", NULL));
1157}
1158
1159static int xive_setup_cpu_ipi(unsigned int cpu)
1160{
1161	struct xive_cpu *xc;
1162	int rc;
1163
1164	pr_debug("Setting up IPI for CPU %d\n", cpu);
1165
1166	xc = per_cpu(xive_cpu, cpu);
1167
1168	/* Check if we are already setup */
1169	if (xc->hw_ipi != XIVE_BAD_IRQ)
1170		return 0;
1171
1172	/* Grab an IPI from the backend, this will populate xc->hw_ipi */
1173	if (xive_ops->get_ipi(cpu, xc))
1174		return -EIO;
1175
1176	/*
1177	 * Populate the IRQ data in the xive_cpu structure and
1178	 * configure the HW / enable the IPIs.
1179	 */
1180	rc = xive_ops->populate_irq_data(xc->hw_ipi, &xc->ipi_data);
1181	if (rc) {
1182		pr_err("Failed to populate IPI data on CPU %d\n", cpu);
1183		return -EIO;
1184	}
1185	rc = xive_ops->configure_irq(xc->hw_ipi,
1186				     get_hard_smp_processor_id(cpu),
1187				     xive_irq_priority, xive_ipi_irq);
1188	if (rc) {
1189		pr_err("Failed to map IPI CPU %d\n", cpu);
1190		return -EIO;
1191	}
1192	pr_devel("CPU %d HW IPI %x, virq %d, trig_mmio=%p\n", cpu,
1193	    xc->hw_ipi, xive_ipi_irq, xc->ipi_data.trig_mmio);
1194
1195	/* Unmask it */
1196	xive_do_source_set_mask(&xc->ipi_data, false);
1197
1198	return 0;
1199}
1200
1201static void xive_cleanup_cpu_ipi(unsigned int cpu, struct xive_cpu *xc)
1202{
1203	/* Disable the IPI and free the IRQ data */
1204
1205	/* Already cleaned up ? */
1206	if (xc->hw_ipi == XIVE_BAD_IRQ)
1207		return;
1208
1209	/* Mask the IPI */
1210	xive_do_source_set_mask(&xc->ipi_data, true);
1211
1212	/*
1213	 * Note: We don't call xive_cleanup_irq_data() to free
1214	 * the mappings as this is called from an IPI on kexec
1215	 * which is not a safe environment to call iounmap()
1216	 */
1217
1218	/* Deconfigure/mask in the backend */
1219	xive_ops->configure_irq(xc->hw_ipi, hard_smp_processor_id(),
1220				0xff, xive_ipi_irq);
1221
1222	/* Free the IPIs in the backend */
1223	xive_ops->put_ipi(cpu, xc);
1224}
1225
1226void __init xive_smp_probe(void)
1227{
1228	smp_ops->cause_ipi = xive_cause_ipi;
1229
1230	/* Register the IPI */
1231	xive_request_ipi();
1232
1233	/* Allocate and setup IPI for the boot CPU */
1234	xive_setup_cpu_ipi(smp_processor_id());
1235}
1236
1237#endif /* CONFIG_SMP */
1238
1239static int xive_irq_domain_map(struct irq_domain *h, unsigned int virq,
1240			       irq_hw_number_t hw)
1241{
1242	int rc;
1243
1244	/*
1245	 * Mark interrupts as edge sensitive by default so that resend
1246	 * actually works. Will fix that up below if needed.
1247	 */
1248	irq_clear_status_flags(virq, IRQ_LEVEL);
1249
1250#ifdef CONFIG_SMP
1251	/* IPIs are special and come up with HW number 0 */
1252	if (hw == 0) {
1253		/*
1254		 * IPIs are marked per-cpu. We use separate HW interrupts under
1255		 * the hood but associated with the same "linux" interrupt
1256		 */
1257		irq_set_chip_and_handler(virq, &xive_ipi_chip,
1258					 handle_percpu_irq);
1259		return 0;
1260	}
1261#endif
1262
1263	rc = xive_irq_alloc_data(virq, hw);
1264	if (rc)
1265		return rc;
1266
1267	irq_set_chip_and_handler(virq, &xive_irq_chip, handle_fasteoi_irq);
1268
1269	return 0;
1270}
1271
1272static void xive_irq_domain_unmap(struct irq_domain *d, unsigned int virq)
1273{
1274	struct irq_data *data = irq_get_irq_data(virq);
1275	unsigned int hw_irq;
1276
1277	/* XXX Assign BAD number */
1278	if (!data)
1279		return;
1280	hw_irq = (unsigned int)irqd_to_hwirq(data);
1281	if (hw_irq)
1282		xive_irq_free_data(virq);
1283}
1284
1285static int xive_irq_domain_xlate(struct irq_domain *h, struct device_node *ct,
1286				 const u32 *intspec, unsigned int intsize,
1287				 irq_hw_number_t *out_hwirq, unsigned int *out_flags)
1288
1289{
1290	*out_hwirq = intspec[0];
1291
1292	/*
1293	 * If intsize is at least 2, we look for the type in the second cell,
1294	 * we assume the LSB indicates a level interrupt.
1295	 */
1296	if (intsize > 1) {
1297		if (intspec[1] & 1)
1298			*out_flags = IRQ_TYPE_LEVEL_LOW;
1299		else
1300			*out_flags = IRQ_TYPE_EDGE_RISING;
1301	} else
1302		*out_flags = IRQ_TYPE_LEVEL_LOW;
1303
1304	return 0;
1305}
1306
1307static int xive_irq_domain_match(struct irq_domain *h, struct device_node *node,
1308				 enum irq_domain_bus_token bus_token)
1309{
1310	return xive_ops->match(node);
1311}
1312
1313static const struct irq_domain_ops xive_irq_domain_ops = {
1314	.match = xive_irq_domain_match,
1315	.map = xive_irq_domain_map,
1316	.unmap = xive_irq_domain_unmap,
1317	.xlate = xive_irq_domain_xlate,
1318};
1319
1320static void __init xive_init_host(void)
1321{
1322	xive_irq_domain = irq_domain_add_nomap(NULL, XIVE_MAX_IRQ,
1323					       &xive_irq_domain_ops, NULL);
1324	if (WARN_ON(xive_irq_domain == NULL))
1325		return;
1326	irq_set_default_host(xive_irq_domain);
1327}
1328
1329static void xive_cleanup_cpu_queues(unsigned int cpu, struct xive_cpu *xc)
1330{
1331	if (xc->queue[xive_irq_priority].qpage)
1332		xive_ops->cleanup_queue(cpu, xc, xive_irq_priority);
1333}
1334
1335static int xive_setup_cpu_queues(unsigned int cpu, struct xive_cpu *xc)
1336{
1337	int rc = 0;
1338
1339	/* We setup 1 queues for now with a 64k page */
1340	if (!xc->queue[xive_irq_priority].qpage)
1341		rc = xive_ops->setup_queue(cpu, xc, xive_irq_priority);
1342
1343	return rc;
1344}
1345
1346static int xive_prepare_cpu(unsigned int cpu)
1347{
1348	struct xive_cpu *xc;
1349
1350	xc = per_cpu(xive_cpu, cpu);
1351	if (!xc) {
1352		struct device_node *np;
1353
1354		xc = kzalloc_node(sizeof(struct xive_cpu),
1355				  GFP_KERNEL, cpu_to_node(cpu));
1356		if (!xc)
1357			return -ENOMEM;
1358		np = of_get_cpu_node(cpu, NULL);
1359		if (np)
1360			xc->chip_id = of_get_ibm_chip_id(np);
1361		of_node_put(np);
1362		xc->hw_ipi = XIVE_BAD_IRQ;
1363
1364		per_cpu(xive_cpu, cpu) = xc;
1365	}
1366
1367	/* Setup EQs if not already */
1368	return xive_setup_cpu_queues(cpu, xc);
1369}
1370
1371static void xive_setup_cpu(void)
1372{
1373	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1374
1375	/* The backend might have additional things to do */
1376	if (xive_ops->setup_cpu)
1377		xive_ops->setup_cpu(smp_processor_id(), xc);
1378
1379	/* Set CPPR to 0xff to enable flow of interrupts */
1380	xc->cppr = 0xff;
1381	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0xff);
1382}
1383
1384#ifdef CONFIG_SMP
1385void xive_smp_setup_cpu(void)
1386{
1387	pr_devel("SMP setup CPU %d\n", smp_processor_id());
1388
1389	/* This will have already been done on the boot CPU */
1390	if (smp_processor_id() != boot_cpuid)
1391		xive_setup_cpu();
1392
1393}
1394
1395int xive_smp_prepare_cpu(unsigned int cpu)
1396{
1397	int rc;
1398
1399	/* Allocate per-CPU data and queues */
1400	rc = xive_prepare_cpu(cpu);
1401	if (rc)
1402		return rc;
1403
1404	/* Allocate and setup IPI for the new CPU */
1405	return xive_setup_cpu_ipi(cpu);
1406}
1407
1408#ifdef CONFIG_HOTPLUG_CPU
1409static void xive_flush_cpu_queue(unsigned int cpu, struct xive_cpu *xc)
1410{
1411	u32 irq;
1412
1413	/* We assume local irqs are disabled */
1414	WARN_ON(!irqs_disabled());
1415
1416	/* Check what's already in the CPU queue */
1417	while ((irq = xive_scan_interrupts(xc, false)) != 0) {
1418		/*
1419		 * We need to re-route that interrupt to its new destination.
1420		 * First get and lock the descriptor
1421		 */
1422		struct irq_desc *desc = irq_to_desc(irq);
1423		struct irq_data *d = irq_desc_get_irq_data(desc);
1424		struct xive_irq_data *xd;
1425		unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
1426
1427		/*
1428		 * Ignore anything that isn't a XIVE irq and ignore
1429		 * IPIs, so can just be dropped.
1430		 */
1431		if (d->domain != xive_irq_domain || hw_irq == 0)
1432			continue;
1433
1434		/*
1435		 * The IRQ should have already been re-routed, it's just a
1436		 * stale in the old queue, so re-trigger it in order to make
1437		 * it reach is new destination.
1438		 */
1439#ifdef DEBUG_FLUSH
1440		pr_info("CPU %d: Got irq %d while offline, re-sending...\n",
1441			cpu, irq);
1442#endif
1443		raw_spin_lock(&desc->lock);
1444		xd = irq_desc_get_handler_data(desc);
1445
1446		/*
1447		 * Clear saved_p to indicate that it's no longer pending
1448		 */
1449		xd->saved_p = false;
1450
1451		/*
1452		 * For LSIs, we EOI, this will cause a resend if it's
1453		 * still asserted. Otherwise do an MSI retrigger.
1454		 */
1455		if (xd->flags & XIVE_IRQ_FLAG_LSI)
1456			xive_do_source_eoi(irqd_to_hwirq(d), xd);
1457		else
1458			xive_irq_retrigger(d);
1459
1460		raw_spin_unlock(&desc->lock);
1461	}
1462}
1463
1464void xive_smp_disable_cpu(void)
1465{
1466	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1467	unsigned int cpu = smp_processor_id();
1468
1469	/* Migrate interrupts away from the CPU */
1470	irq_migrate_all_off_this_cpu();
1471
1472	/* Set CPPR to 0 to disable flow of interrupts */
1473	xc->cppr = 0;
1474	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0);
1475
1476	/* Flush everything still in the queue */
1477	xive_flush_cpu_queue(cpu, xc);
1478
1479	/* Re-enable CPPR  */
1480	xc->cppr = 0xff;
1481	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0xff);
1482}
1483
1484void xive_flush_interrupt(void)
1485{
1486	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1487	unsigned int cpu = smp_processor_id();
1488
1489	/* Called if an interrupt occurs while the CPU is hot unplugged */
1490	xive_flush_cpu_queue(cpu, xc);
1491}
1492
1493#endif /* CONFIG_HOTPLUG_CPU */
1494
1495#endif /* CONFIG_SMP */
1496
1497void xive_teardown_cpu(void)
1498{
1499	struct xive_cpu *xc = __this_cpu_read(xive_cpu);
1500	unsigned int cpu = smp_processor_id();
1501
1502	/* Set CPPR to 0 to disable flow of interrupts */
1503	xc->cppr = 0;
1504	out_8(xive_tima + xive_tima_offset + TM_CPPR, 0);
1505
1506	if (xive_ops->teardown_cpu)
1507		xive_ops->teardown_cpu(cpu, xc);
1508
1509#ifdef CONFIG_SMP
1510	/* Get rid of IPI */
1511	xive_cleanup_cpu_ipi(cpu, xc);
1512#endif
1513
1514	/* Disable and free the queues */
1515	xive_cleanup_cpu_queues(cpu, xc);
1516}
1517
1518void xive_shutdown(void)
1519{
1520	xive_ops->shutdown();
1521}
1522
1523bool __init xive_core_init(const struct xive_ops *ops, void __iomem *area, u32 offset,
1524			   u8 max_prio)
1525{
1526	xive_tima = area;
1527	xive_tima_offset = offset;
1528	xive_ops = ops;
1529	xive_irq_priority = max_prio;
1530
1531	ppc_md.get_irq = xive_get_irq;
1532	__xive_enabled = true;
1533
1534	pr_devel("Initializing host..\n");
1535	xive_init_host();
1536
1537	pr_devel("Initializing boot CPU..\n");
1538
1539	/* Allocate per-CPU data and queues */
1540	xive_prepare_cpu(smp_processor_id());
1541
1542	/* Get ready for interrupts */
1543	xive_setup_cpu();
1544
1545	pr_info("Interrupt handling initialized with %s backend\n",
1546		xive_ops->name);
1547	pr_info("Using priority %d for all interrupts\n", max_prio);
1548
1549	return true;
1550}
1551
1552__be32 *xive_queue_page_alloc(unsigned int cpu, u32 queue_shift)
1553{
1554	unsigned int alloc_order;
1555	struct page *pages;
1556	__be32 *qpage;
1557
1558	alloc_order = xive_alloc_order(queue_shift);
1559	pages = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, alloc_order);
1560	if (!pages)
1561		return ERR_PTR(-ENOMEM);
1562	qpage = (__be32 *)page_address(pages);
1563	memset(qpage, 0, 1 << queue_shift);
1564
1565	return qpage;
1566}
1567
1568static int __init xive_off(char *arg)
1569{
1570	xive_cmdline_disabled = true;
1571	return 0;
1572}
1573__setup("xive=off", xive_off);
1574
1575static void xive_debug_show_cpu(struct seq_file *m, int cpu)
1576{
1577	struct xive_cpu *xc = per_cpu(xive_cpu, cpu);
1578
1579	seq_printf(m, "CPU %d:", cpu);
1580	if (xc) {
1581		seq_printf(m, "pp=%02x CPPR=%02x ", xc->pending_prio, xc->cppr);
1582
1583#ifdef CONFIG_SMP
1584		{
1585			u64 val = xive_esb_read(&xc->ipi_data, XIVE_ESB_GET);
1586
1587			seq_printf(m, "IPI=0x%08x PQ=%c%c ", xc->hw_ipi,
1588				   val & XIVE_ESB_VAL_P ? 'P' : '-',
1589				   val & XIVE_ESB_VAL_Q ? 'Q' : '-');
1590		}
1591#endif
1592		{
1593			struct xive_q *q = &xc->queue[xive_irq_priority];
1594			u32 i0, i1, idx;
1595
1596			if (q->qpage) {
1597				idx = q->idx;
1598				i0 = be32_to_cpup(q->qpage + idx);
1599				idx = (idx + 1) & q->msk;
1600				i1 = be32_to_cpup(q->qpage + idx);
1601				seq_printf(m, "EQ idx=%d T=%d %08x %08x ...",
1602					   q->idx, q->toggle, i0, i1);
1603			}
1604		}
1605	}
1606	seq_puts(m, "\n");
1607}
1608
1609static void xive_debug_show_irq(struct seq_file *m, u32 hw_irq, struct irq_data *d)
1610{
1611	struct irq_chip *chip = irq_data_get_irq_chip(d);
1612	int rc;
1613	u32 target;
1614	u8 prio;
1615	u32 lirq;
1616	struct xive_irq_data *xd;
1617	u64 val;
1618
1619	if (!is_xive_irq(chip))
1620		return;
1621
1622	rc = xive_ops->get_irq_config(hw_irq, &target, &prio, &lirq);
1623	if (rc) {
1624		seq_printf(m, "IRQ 0x%08x : no config rc=%d\n", hw_irq, rc);
1625		return;
1626	}
1627
1628	seq_printf(m, "IRQ 0x%08x : target=0x%x prio=%02x lirq=0x%x ",
1629		   hw_irq, target, prio, lirq);
1630
1631	xd = irq_data_get_irq_handler_data(d);
1632	val = xive_esb_read(xd, XIVE_ESB_GET);
1633	seq_printf(m, "flags=%c%c%c PQ=%c%c",
1634		   xd->flags & XIVE_IRQ_FLAG_STORE_EOI ? 'S' : ' ',
1635		   xd->flags & XIVE_IRQ_FLAG_LSI ? 'L' : ' ',
1636		   xd->flags & XIVE_IRQ_FLAG_H_INT_ESB ? 'H' : ' ',
1637		   val & XIVE_ESB_VAL_P ? 'P' : '-',
1638		   val & XIVE_ESB_VAL_Q ? 'Q' : '-');
1639	seq_puts(m, "\n");
1640}
1641
1642static int xive_core_debug_show(struct seq_file *m, void *private)
1643{
1644	unsigned int i;
1645	struct irq_desc *desc;
1646	int cpu;
1647
1648	if (xive_ops->debug_show)
1649		xive_ops->debug_show(m, private);
1650
1651	for_each_possible_cpu(cpu)
1652		xive_debug_show_cpu(m, cpu);
1653
1654	for_each_irq_desc(i, desc) {
1655		struct irq_data *d = irq_desc_get_irq_data(desc);
1656		unsigned int hw_irq;
1657
1658		if (!d)
1659			continue;
1660
1661		hw_irq = (unsigned int)irqd_to_hwirq(d);
1662
1663		/* IPIs are special (HW number 0) */
1664		if (hw_irq)
1665			xive_debug_show_irq(m, hw_irq, d);
1666	}
1667	return 0;
1668}
1669DEFINE_SHOW_ATTRIBUTE(xive_core_debug);
1670
1671int xive_core_debug_init(void)
1672{
1673	if (xive_enabled())
1674		debugfs_create_file("xive", 0400, powerpc_debugfs_root,
1675				    NULL, &xive_core_debug_fops);
1676	return 0;
1677}
1678