1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2015 Broadcom Corporation
4  */
5 
6 #include <linux/interrupt.h>
7 #include <linux/irqchip/chained_irq.h>
8 #include <linux/irqdomain.h>
9 #include <linux/msi.h>
10 #include <linux/of_irq.h>
11 #include <linux/of_pci.h>
12 #include <linux/pci.h>
13 
14 #include "pcie-iproc.h"
15 
16 #define IPROC_MSI_INTR_EN_SHIFT        11
17 #define IPROC_MSI_INTR_EN              BIT(IPROC_MSI_INTR_EN_SHIFT)
18 #define IPROC_MSI_INT_N_EVENT_SHIFT    1
19 #define IPROC_MSI_INT_N_EVENT          BIT(IPROC_MSI_INT_N_EVENT_SHIFT)
20 #define IPROC_MSI_EQ_EN_SHIFT          0
21 #define IPROC_MSI_EQ_EN                BIT(IPROC_MSI_EQ_EN_SHIFT)
22 
23 #define IPROC_MSI_EQ_MASK              0x3f
24 
25 /* Max number of GIC interrupts */
26 #define NR_HW_IRQS                     6
27 
28 /* Number of entries in each event queue */
29 #define EQ_LEN                         64
30 
31 /* Size of each event queue memory region */
32 #define EQ_MEM_REGION_SIZE             SZ_4K
33 
34 /* Size of each MSI address region */
35 #define MSI_MEM_REGION_SIZE            SZ_4K
36 
37 enum iproc_msi_reg {
38 	IPROC_MSI_EQ_PAGE = 0,
39 	IPROC_MSI_EQ_PAGE_UPPER,
40 	IPROC_MSI_PAGE,
41 	IPROC_MSI_PAGE_UPPER,
42 	IPROC_MSI_CTRL,
43 	IPROC_MSI_EQ_HEAD,
44 	IPROC_MSI_EQ_TAIL,
45 	IPROC_MSI_INTS_EN,
46 	IPROC_MSI_REG_SIZE,
47 };
48 
49 struct iproc_msi;
50 
51 /**
52  * struct iproc_msi_grp - iProc MSI group
53  *
54  * One MSI group is allocated per GIC interrupt, serviced by one iProc MSI
55  * event queue.
56  *
57  * @msi: pointer to iProc MSI data
58  * @gic_irq: GIC interrupt
59  * @eq: Event queue number
60  */
61 struct iproc_msi_grp {
62 	struct iproc_msi *msi;
63 	int gic_irq;
64 	unsigned int eq;
65 };
66 
67 /**
68  * struct iproc_msi - iProc event queue based MSI
69  *
70  * Only meant to be used on platforms without MSI support integrated into the
71  * GIC.
72  *
73  * @pcie: pointer to iProc PCIe data
74  * @reg_offsets: MSI register offsets
75  * @grps: MSI groups
76  * @nr_irqs: number of total interrupts connected to GIC
77  * @nr_cpus: number of toal CPUs
78  * @has_inten_reg: indicates the MSI interrupt enable register needs to be
79  * set explicitly (required for some legacy platforms)
80  * @bitmap: MSI vector bitmap
81  * @bitmap_lock: lock to protect access to the MSI bitmap
82  * @nr_msi_vecs: total number of MSI vectors
83  * @inner_domain: inner IRQ domain
84  * @msi_domain: MSI IRQ domain
85  * @nr_eq_region: required number of 4K aligned memory region for MSI event
86  * queues
87  * @nr_msi_region: required number of 4K aligned address region for MSI posted
88  * writes
89  * @eq_cpu: pointer to allocated memory region for MSI event queues
90  * @eq_dma: DMA address of MSI event queues
91  * @msi_addr: MSI address
92  */
93 struct iproc_msi {
94 	struct iproc_pcie *pcie;
95 	const u16 (*reg_offsets)[IPROC_MSI_REG_SIZE];
96 	struct iproc_msi_grp *grps;
97 	int nr_irqs;
98 	int nr_cpus;
99 	bool has_inten_reg;
100 	unsigned long *bitmap;
101 	struct mutex bitmap_lock;
102 	unsigned int nr_msi_vecs;
103 	struct irq_domain *inner_domain;
104 	struct irq_domain *msi_domain;
105 	unsigned int nr_eq_region;
106 	unsigned int nr_msi_region;
107 	void *eq_cpu;
108 	dma_addr_t eq_dma;
109 	phys_addr_t msi_addr;
110 };
111 
112 static const u16 iproc_msi_reg_paxb[NR_HW_IRQS][IPROC_MSI_REG_SIZE] = {
113 	{ 0x200, 0x2c0, 0x204, 0x2c4, 0x210, 0x250, 0x254, 0x208 },
114 	{ 0x200, 0x2c0, 0x204, 0x2c4, 0x214, 0x258, 0x25c, 0x208 },
115 	{ 0x200, 0x2c0, 0x204, 0x2c4, 0x218, 0x260, 0x264, 0x208 },
116 	{ 0x200, 0x2c0, 0x204, 0x2c4, 0x21c, 0x268, 0x26c, 0x208 },
117 	{ 0x200, 0x2c0, 0x204, 0x2c4, 0x220, 0x270, 0x274, 0x208 },
118 	{ 0x200, 0x2c0, 0x204, 0x2c4, 0x224, 0x278, 0x27c, 0x208 },
119 };
120 
121 static const u16 iproc_msi_reg_paxc[NR_HW_IRQS][IPROC_MSI_REG_SIZE] = {
122 	{ 0xc00, 0xc04, 0xc08, 0xc0c, 0xc40, 0xc50, 0xc60 },
123 	{ 0xc10, 0xc14, 0xc18, 0xc1c, 0xc44, 0xc54, 0xc64 },
124 	{ 0xc20, 0xc24, 0xc28, 0xc2c, 0xc48, 0xc58, 0xc68 },
125 	{ 0xc30, 0xc34, 0xc38, 0xc3c, 0xc4c, 0xc5c, 0xc6c },
126 };
127 
iproc_msi_read_reg(struct iproc_msi * msi,enum iproc_msi_reg reg,unsigned int eq)128 static inline u32 iproc_msi_read_reg(struct iproc_msi *msi,
129 				     enum iproc_msi_reg reg,
130 				     unsigned int eq)
131 {
132 	struct iproc_pcie *pcie = msi->pcie;
133 
134 	return readl_relaxed(pcie->base + msi->reg_offsets[eq][reg]);
135 }
136 
iproc_msi_write_reg(struct iproc_msi * msi,enum iproc_msi_reg reg,int eq,u32 val)137 static inline void iproc_msi_write_reg(struct iproc_msi *msi,
138 				       enum iproc_msi_reg reg,
139 				       int eq, u32 val)
140 {
141 	struct iproc_pcie *pcie = msi->pcie;
142 
143 	writel_relaxed(val, pcie->base + msi->reg_offsets[eq][reg]);
144 }
145 
hwirq_to_group(struct iproc_msi * msi,unsigned long hwirq)146 static inline u32 hwirq_to_group(struct iproc_msi *msi, unsigned long hwirq)
147 {
148 	return (hwirq % msi->nr_irqs);
149 }
150 
iproc_msi_addr_offset(struct iproc_msi * msi,unsigned long hwirq)151 static inline unsigned int iproc_msi_addr_offset(struct iproc_msi *msi,
152 						 unsigned long hwirq)
153 {
154 	if (msi->nr_msi_region > 1)
155 		return hwirq_to_group(msi, hwirq) * MSI_MEM_REGION_SIZE;
156 	else
157 		return hwirq_to_group(msi, hwirq) * sizeof(u32);
158 }
159 
iproc_msi_eq_offset(struct iproc_msi * msi,u32 eq)160 static inline unsigned int iproc_msi_eq_offset(struct iproc_msi *msi, u32 eq)
161 {
162 	if (msi->nr_eq_region > 1)
163 		return eq * EQ_MEM_REGION_SIZE;
164 	else
165 		return eq * EQ_LEN * sizeof(u32);
166 }
167 
168 static struct irq_chip iproc_msi_irq_chip = {
169 	.name = "iProc-MSI",
170 };
171 
172 static struct msi_domain_info iproc_msi_domain_info = {
173 	.flags = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
174 		MSI_FLAG_PCI_MSIX,
175 	.chip = &iproc_msi_irq_chip,
176 };
177 
178 /*
179  * In iProc PCIe core, each MSI group is serviced by a GIC interrupt and a
180  * dedicated event queue.  Each MSI group can support up to 64 MSI vectors.
181  *
182  * The number of MSI groups varies between different iProc SoCs.  The total
183  * number of CPU cores also varies.  To support MSI IRQ affinity, we
184  * distribute GIC interrupts across all available CPUs.  MSI vector is moved
185  * from one GIC interrupt to another to steer to the target CPU.
186  *
187  * Assuming:
188  * - the number of MSI groups is M
189  * - the number of CPU cores is N
190  * - M is always a multiple of N
191  *
192  * Total number of raw MSI vectors = M * 64
193  * Total number of supported MSI vectors = (M * 64) / N
194  */
hwirq_to_cpu(struct iproc_msi * msi,unsigned long hwirq)195 static inline int hwirq_to_cpu(struct iproc_msi *msi, unsigned long hwirq)
196 {
197 	return (hwirq % msi->nr_cpus);
198 }
199 
hwirq_to_canonical_hwirq(struct iproc_msi * msi,unsigned long hwirq)200 static inline unsigned long hwirq_to_canonical_hwirq(struct iproc_msi *msi,
201 						     unsigned long hwirq)
202 {
203 	return (hwirq - hwirq_to_cpu(msi, hwirq));
204 }
205 
iproc_msi_irq_set_affinity(struct irq_data * data,const struct cpumask * mask,bool force)206 static int iproc_msi_irq_set_affinity(struct irq_data *data,
207 				      const struct cpumask *mask, bool force)
208 {
209 	struct iproc_msi *msi = irq_data_get_irq_chip_data(data);
210 	int target_cpu = cpumask_first(mask);
211 	int curr_cpu;
212 	int ret;
213 
214 	curr_cpu = hwirq_to_cpu(msi, data->hwirq);
215 	if (curr_cpu == target_cpu)
216 		ret = IRQ_SET_MASK_OK_DONE;
217 	else {
218 		/* steer MSI to the target CPU */
219 		data->hwirq = hwirq_to_canonical_hwirq(msi, data->hwirq) + target_cpu;
220 		ret = IRQ_SET_MASK_OK;
221 	}
222 
223 	irq_data_update_effective_affinity(data, cpumask_of(target_cpu));
224 
225 	return ret;
226 }
227 
iproc_msi_irq_compose_msi_msg(struct irq_data * data,struct msi_msg * msg)228 static void iproc_msi_irq_compose_msi_msg(struct irq_data *data,
229 					  struct msi_msg *msg)
230 {
231 	struct iproc_msi *msi = irq_data_get_irq_chip_data(data);
232 	dma_addr_t addr;
233 
234 	addr = msi->msi_addr + iproc_msi_addr_offset(msi, data->hwirq);
235 	msg->address_lo = lower_32_bits(addr);
236 	msg->address_hi = upper_32_bits(addr);
237 	msg->data = data->hwirq << 5;
238 }
239 
240 static struct irq_chip iproc_msi_bottom_irq_chip = {
241 	.name = "MSI",
242 	.irq_set_affinity = iproc_msi_irq_set_affinity,
243 	.irq_compose_msi_msg = iproc_msi_irq_compose_msi_msg,
244 };
245 
iproc_msi_irq_domain_alloc(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs,void * args)246 static int iproc_msi_irq_domain_alloc(struct irq_domain *domain,
247 				      unsigned int virq, unsigned int nr_irqs,
248 				      void *args)
249 {
250 	struct iproc_msi *msi = domain->host_data;
251 	int hwirq, i;
252 
253 	if (msi->nr_cpus > 1 && nr_irqs > 1)
254 		return -EINVAL;
255 
256 	mutex_lock(&msi->bitmap_lock);
257 
258 	/*
259 	 * Allocate 'nr_irqs' multiplied by 'nr_cpus' number of MSI vectors
260 	 * each time
261 	 */
262 	hwirq = bitmap_find_free_region(msi->bitmap, msi->nr_msi_vecs,
263 					order_base_2(msi->nr_cpus * nr_irqs));
264 
265 	mutex_unlock(&msi->bitmap_lock);
266 
267 	if (hwirq < 0)
268 		return -ENOSPC;
269 
270 	for (i = 0; i < nr_irqs; i++) {
271 		irq_domain_set_info(domain, virq + i, hwirq + i,
272 				    &iproc_msi_bottom_irq_chip,
273 				    domain->host_data, handle_simple_irq,
274 				    NULL, NULL);
275 	}
276 
277 	return 0;
278 }
279 
iproc_msi_irq_domain_free(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs)280 static void iproc_msi_irq_domain_free(struct irq_domain *domain,
281 				      unsigned int virq, unsigned int nr_irqs)
282 {
283 	struct irq_data *data = irq_domain_get_irq_data(domain, virq);
284 	struct iproc_msi *msi = irq_data_get_irq_chip_data(data);
285 	unsigned int hwirq;
286 
287 	mutex_lock(&msi->bitmap_lock);
288 
289 	hwirq = hwirq_to_canonical_hwirq(msi, data->hwirq);
290 	bitmap_release_region(msi->bitmap, hwirq,
291 			      order_base_2(msi->nr_cpus * nr_irqs));
292 
293 	mutex_unlock(&msi->bitmap_lock);
294 
295 	irq_domain_free_irqs_parent(domain, virq, nr_irqs);
296 }
297 
298 static const struct irq_domain_ops msi_domain_ops = {
299 	.alloc = iproc_msi_irq_domain_alloc,
300 	.free = iproc_msi_irq_domain_free,
301 };
302 
decode_msi_hwirq(struct iproc_msi * msi,u32 eq,u32 head)303 static inline u32 decode_msi_hwirq(struct iproc_msi *msi, u32 eq, u32 head)
304 {
305 	u32 __iomem *msg;
306 	u32 hwirq;
307 	unsigned int offs;
308 
309 	offs = iproc_msi_eq_offset(msi, eq) + head * sizeof(u32);
310 	msg = (u32 __iomem *)(msi->eq_cpu + offs);
311 	hwirq = readl(msg);
312 	hwirq = (hwirq >> 5) + (hwirq & 0x1f);
313 
314 	/*
315 	 * Since we have multiple hwirq mapped to a single MSI vector,
316 	 * now we need to derive the hwirq at CPU0.  It can then be used to
317 	 * mapped back to virq.
318 	 */
319 	return hwirq_to_canonical_hwirq(msi, hwirq);
320 }
321 
iproc_msi_handler(struct irq_desc * desc)322 static void iproc_msi_handler(struct irq_desc *desc)
323 {
324 	struct irq_chip *chip = irq_desc_get_chip(desc);
325 	struct iproc_msi_grp *grp;
326 	struct iproc_msi *msi;
327 	u32 eq, head, tail, nr_events;
328 	unsigned long hwirq;
329 
330 	chained_irq_enter(chip, desc);
331 
332 	grp = irq_desc_get_handler_data(desc);
333 	msi = grp->msi;
334 	eq = grp->eq;
335 
336 	/*
337 	 * iProc MSI event queue is tracked by head and tail pointers.  Head
338 	 * pointer indicates the next entry (MSI data) to be consumed by SW in
339 	 * the queue and needs to be updated by SW.  iProc MSI core uses the
340 	 * tail pointer as the next data insertion point.
341 	 *
342 	 * Entries between head and tail pointers contain valid MSI data.  MSI
343 	 * data is guaranteed to be in the event queue memory before the tail
344 	 * pointer is updated by the iProc MSI core.
345 	 */
346 	head = iproc_msi_read_reg(msi, IPROC_MSI_EQ_HEAD,
347 				  eq) & IPROC_MSI_EQ_MASK;
348 	do {
349 		tail = iproc_msi_read_reg(msi, IPROC_MSI_EQ_TAIL,
350 					  eq) & IPROC_MSI_EQ_MASK;
351 
352 		/*
353 		 * Figure out total number of events (MSI data) to be
354 		 * processed.
355 		 */
356 		nr_events = (tail < head) ?
357 			(EQ_LEN - (head - tail)) : (tail - head);
358 		if (!nr_events)
359 			break;
360 
361 		/* process all outstanding events */
362 		while (nr_events--) {
363 			hwirq = decode_msi_hwirq(msi, eq, head);
364 			generic_handle_domain_irq(msi->inner_domain, hwirq);
365 
366 			head++;
367 			head %= EQ_LEN;
368 		}
369 
370 		/*
371 		 * Now all outstanding events have been processed.  Update the
372 		 * head pointer.
373 		 */
374 		iproc_msi_write_reg(msi, IPROC_MSI_EQ_HEAD, eq, head);
375 
376 		/*
377 		 * Now go read the tail pointer again to see if there are new
378 		 * outstanding events that came in during the above window.
379 		 */
380 	} while (true);
381 
382 	chained_irq_exit(chip, desc);
383 }
384 
iproc_msi_enable(struct iproc_msi * msi)385 static void iproc_msi_enable(struct iproc_msi *msi)
386 {
387 	int i, eq;
388 	u32 val;
389 
390 	/* Program memory region for each event queue */
391 	for (i = 0; i < msi->nr_eq_region; i++) {
392 		dma_addr_t addr = msi->eq_dma + (i * EQ_MEM_REGION_SIZE);
393 
394 		iproc_msi_write_reg(msi, IPROC_MSI_EQ_PAGE, i,
395 				    lower_32_bits(addr));
396 		iproc_msi_write_reg(msi, IPROC_MSI_EQ_PAGE_UPPER, i,
397 				    upper_32_bits(addr));
398 	}
399 
400 	/* Program address region for MSI posted writes */
401 	for (i = 0; i < msi->nr_msi_region; i++) {
402 		phys_addr_t addr = msi->msi_addr + (i * MSI_MEM_REGION_SIZE);
403 
404 		iproc_msi_write_reg(msi, IPROC_MSI_PAGE, i,
405 				    lower_32_bits(addr));
406 		iproc_msi_write_reg(msi, IPROC_MSI_PAGE_UPPER, i,
407 				    upper_32_bits(addr));
408 	}
409 
410 	for (eq = 0; eq < msi->nr_irqs; eq++) {
411 		/* Enable MSI event queue */
412 		val = IPROC_MSI_INTR_EN | IPROC_MSI_INT_N_EVENT |
413 			IPROC_MSI_EQ_EN;
414 		iproc_msi_write_reg(msi, IPROC_MSI_CTRL, eq, val);
415 
416 		/*
417 		 * Some legacy platforms require the MSI interrupt enable
418 		 * register to be set explicitly.
419 		 */
420 		if (msi->has_inten_reg) {
421 			val = iproc_msi_read_reg(msi, IPROC_MSI_INTS_EN, eq);
422 			val |= BIT(eq);
423 			iproc_msi_write_reg(msi, IPROC_MSI_INTS_EN, eq, val);
424 		}
425 	}
426 }
427 
iproc_msi_disable(struct iproc_msi * msi)428 static void iproc_msi_disable(struct iproc_msi *msi)
429 {
430 	u32 eq, val;
431 
432 	for (eq = 0; eq < msi->nr_irqs; eq++) {
433 		if (msi->has_inten_reg) {
434 			val = iproc_msi_read_reg(msi, IPROC_MSI_INTS_EN, eq);
435 			val &= ~BIT(eq);
436 			iproc_msi_write_reg(msi, IPROC_MSI_INTS_EN, eq, val);
437 		}
438 
439 		val = iproc_msi_read_reg(msi, IPROC_MSI_CTRL, eq);
440 		val &= ~(IPROC_MSI_INTR_EN | IPROC_MSI_INT_N_EVENT |
441 			 IPROC_MSI_EQ_EN);
442 		iproc_msi_write_reg(msi, IPROC_MSI_CTRL, eq, val);
443 	}
444 }
445 
iproc_msi_alloc_domains(struct device_node * node,struct iproc_msi * msi)446 static int iproc_msi_alloc_domains(struct device_node *node,
447 				   struct iproc_msi *msi)
448 {
449 	msi->inner_domain = irq_domain_add_linear(NULL, msi->nr_msi_vecs,
450 						  &msi_domain_ops, msi);
451 	if (!msi->inner_domain)
452 		return -ENOMEM;
453 
454 	msi->msi_domain = pci_msi_create_irq_domain(of_node_to_fwnode(node),
455 						    &iproc_msi_domain_info,
456 						    msi->inner_domain);
457 	if (!msi->msi_domain) {
458 		irq_domain_remove(msi->inner_domain);
459 		return -ENOMEM;
460 	}
461 
462 	return 0;
463 }
464 
iproc_msi_free_domains(struct iproc_msi * msi)465 static void iproc_msi_free_domains(struct iproc_msi *msi)
466 {
467 	if (msi->msi_domain)
468 		irq_domain_remove(msi->msi_domain);
469 
470 	if (msi->inner_domain)
471 		irq_domain_remove(msi->inner_domain);
472 }
473 
iproc_msi_irq_free(struct iproc_msi * msi,unsigned int cpu)474 static void iproc_msi_irq_free(struct iproc_msi *msi, unsigned int cpu)
475 {
476 	int i;
477 
478 	for (i = cpu; i < msi->nr_irqs; i += msi->nr_cpus) {
479 		irq_set_chained_handler_and_data(msi->grps[i].gic_irq,
480 						 NULL, NULL);
481 	}
482 }
483 
iproc_msi_irq_setup(struct iproc_msi * msi,unsigned int cpu)484 static int iproc_msi_irq_setup(struct iproc_msi *msi, unsigned int cpu)
485 {
486 	int i, ret;
487 	cpumask_var_t mask;
488 	struct iproc_pcie *pcie = msi->pcie;
489 
490 	for (i = cpu; i < msi->nr_irqs; i += msi->nr_cpus) {
491 		irq_set_chained_handler_and_data(msi->grps[i].gic_irq,
492 						 iproc_msi_handler,
493 						 &msi->grps[i]);
494 		/* Dedicate GIC interrupt to each CPU core */
495 		if (alloc_cpumask_var(&mask, GFP_KERNEL)) {
496 			cpumask_clear(mask);
497 			cpumask_set_cpu(cpu, mask);
498 			ret = irq_set_affinity(msi->grps[i].gic_irq, mask);
499 			if (ret)
500 				dev_err(pcie->dev,
501 					"failed to set affinity for IRQ%d\n",
502 					msi->grps[i].gic_irq);
503 			free_cpumask_var(mask);
504 		} else {
505 			dev_err(pcie->dev, "failed to alloc CPU mask\n");
506 			ret = -EINVAL;
507 		}
508 
509 		if (ret) {
510 			/* Free all configured/unconfigured IRQs */
511 			iproc_msi_irq_free(msi, cpu);
512 			return ret;
513 		}
514 	}
515 
516 	return 0;
517 }
518 
iproc_msi_init(struct iproc_pcie * pcie,struct device_node * node)519 int iproc_msi_init(struct iproc_pcie *pcie, struct device_node *node)
520 {
521 	struct iproc_msi *msi;
522 	int i, ret;
523 	unsigned int cpu;
524 
525 	if (!of_device_is_compatible(node, "brcm,iproc-msi"))
526 		return -ENODEV;
527 
528 	if (!of_property_read_bool(node, "msi-controller"))
529 		return -ENODEV;
530 
531 	if (pcie->msi)
532 		return -EBUSY;
533 
534 	msi = devm_kzalloc(pcie->dev, sizeof(*msi), GFP_KERNEL);
535 	if (!msi)
536 		return -ENOMEM;
537 
538 	msi->pcie = pcie;
539 	pcie->msi = msi;
540 	msi->msi_addr = pcie->base_addr;
541 	mutex_init(&msi->bitmap_lock);
542 	msi->nr_cpus = num_possible_cpus();
543 
544 	if (msi->nr_cpus == 1)
545 		iproc_msi_domain_info.flags |=  MSI_FLAG_MULTI_PCI_MSI;
546 
547 	msi->nr_irqs = of_irq_count(node);
548 	if (!msi->nr_irqs) {
549 		dev_err(pcie->dev, "found no MSI GIC interrupt\n");
550 		return -ENODEV;
551 	}
552 
553 	if (msi->nr_irqs > NR_HW_IRQS) {
554 		dev_warn(pcie->dev, "too many MSI GIC interrupts defined %d\n",
555 			 msi->nr_irqs);
556 		msi->nr_irqs = NR_HW_IRQS;
557 	}
558 
559 	if (msi->nr_irqs < msi->nr_cpus) {
560 		dev_err(pcie->dev,
561 			"not enough GIC interrupts for MSI affinity\n");
562 		return -EINVAL;
563 	}
564 
565 	if (msi->nr_irqs % msi->nr_cpus != 0) {
566 		msi->nr_irqs -= msi->nr_irqs % msi->nr_cpus;
567 		dev_warn(pcie->dev, "Reducing number of interrupts to %d\n",
568 			 msi->nr_irqs);
569 	}
570 
571 	switch (pcie->type) {
572 	case IPROC_PCIE_PAXB_BCMA:
573 	case IPROC_PCIE_PAXB:
574 		msi->reg_offsets = iproc_msi_reg_paxb;
575 		msi->nr_eq_region = 1;
576 		msi->nr_msi_region = 1;
577 		break;
578 	case IPROC_PCIE_PAXC:
579 		msi->reg_offsets = iproc_msi_reg_paxc;
580 		msi->nr_eq_region = msi->nr_irqs;
581 		msi->nr_msi_region = msi->nr_irqs;
582 		break;
583 	default:
584 		dev_err(pcie->dev, "incompatible iProc PCIe interface\n");
585 		return -EINVAL;
586 	}
587 
588 	msi->has_inten_reg = of_property_read_bool(node, "brcm,pcie-msi-inten");
589 
590 	msi->nr_msi_vecs = msi->nr_irqs * EQ_LEN;
591 	msi->bitmap = devm_bitmap_zalloc(pcie->dev, msi->nr_msi_vecs,
592 					 GFP_KERNEL);
593 	if (!msi->bitmap)
594 		return -ENOMEM;
595 
596 	msi->grps = devm_kcalloc(pcie->dev, msi->nr_irqs, sizeof(*msi->grps),
597 				 GFP_KERNEL);
598 	if (!msi->grps)
599 		return -ENOMEM;
600 
601 	for (i = 0; i < msi->nr_irqs; i++) {
602 		unsigned int irq = irq_of_parse_and_map(node, i);
603 
604 		if (!irq) {
605 			dev_err(pcie->dev, "unable to parse/map interrupt\n");
606 			ret = -ENODEV;
607 			goto free_irqs;
608 		}
609 		msi->grps[i].gic_irq = irq;
610 		msi->grps[i].msi = msi;
611 		msi->grps[i].eq = i;
612 	}
613 
614 	/* Reserve memory for event queue and make sure memories are zeroed */
615 	msi->eq_cpu = dma_alloc_coherent(pcie->dev,
616 					 msi->nr_eq_region * EQ_MEM_REGION_SIZE,
617 					 &msi->eq_dma, GFP_KERNEL);
618 	if (!msi->eq_cpu) {
619 		ret = -ENOMEM;
620 		goto free_irqs;
621 	}
622 
623 	ret = iproc_msi_alloc_domains(node, msi);
624 	if (ret) {
625 		dev_err(pcie->dev, "failed to create MSI domains\n");
626 		goto free_eq_dma;
627 	}
628 
629 	for_each_online_cpu(cpu) {
630 		ret = iproc_msi_irq_setup(msi, cpu);
631 		if (ret)
632 			goto free_msi_irq;
633 	}
634 
635 	iproc_msi_enable(msi);
636 
637 	return 0;
638 
639 free_msi_irq:
640 	for_each_online_cpu(cpu)
641 		iproc_msi_irq_free(msi, cpu);
642 	iproc_msi_free_domains(msi);
643 
644 free_eq_dma:
645 	dma_free_coherent(pcie->dev, msi->nr_eq_region * EQ_MEM_REGION_SIZE,
646 			  msi->eq_cpu, msi->eq_dma);
647 
648 free_irqs:
649 	for (i = 0; i < msi->nr_irqs; i++) {
650 		if (msi->grps[i].gic_irq)
651 			irq_dispose_mapping(msi->grps[i].gic_irq);
652 	}
653 	pcie->msi = NULL;
654 	return ret;
655 }
656 EXPORT_SYMBOL(iproc_msi_init);
657 
iproc_msi_exit(struct iproc_pcie * pcie)658 void iproc_msi_exit(struct iproc_pcie *pcie)
659 {
660 	struct iproc_msi *msi = pcie->msi;
661 	unsigned int i, cpu;
662 
663 	if (!msi)
664 		return;
665 
666 	iproc_msi_disable(msi);
667 
668 	for_each_online_cpu(cpu)
669 		iproc_msi_irq_free(msi, cpu);
670 
671 	iproc_msi_free_domains(msi);
672 
673 	dma_free_coherent(pcie->dev, msi->nr_eq_region * EQ_MEM_REGION_SIZE,
674 			  msi->eq_cpu, msi->eq_dma);
675 
676 	for (i = 0; i < msi->nr_irqs; i++) {
677 		if (msi->grps[i].gic_irq)
678 			irq_dispose_mapping(msi->grps[i].gic_irq);
679 	}
680 }
681 EXPORT_SYMBOL(iproc_msi_exit);
682