1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * APM X-Gene MSI Driver
4  *
5  * Copyright (c) 2014, Applied Micro Circuits Corporation
6  * Author: Tanmay Inamdar <tinamdar@apm.com>
7  *	   Duc Dang <dhdang@apm.com>
8  */
9 #include <linux/cpu.h>
10 #include <linux/interrupt.h>
11 #include <linux/module.h>
12 #include <linux/msi.h>
13 #include <linux/of_irq.h>
14 #include <linux/irqchip/chained_irq.h>
15 #include <linux/pci.h>
16 #include <linux/platform_device.h>
17 #include <linux/of_pci.h>
18 
19 #define MSI_IR0			0x000000
20 #define MSI_INT0		0x800000
21 #define IDX_PER_GROUP		8
22 #define IRQS_PER_IDX		16
23 #define NR_HW_IRQS		16
24 #define NR_MSI_VEC		(IDX_PER_GROUP * IRQS_PER_IDX * NR_HW_IRQS)
25 
26 struct xgene_msi_group {
27 	struct xgene_msi	*msi;
28 	int			gic_irq;
29 	u32			msi_grp;
30 };
31 
32 struct xgene_msi {
33 	struct device_node	*node;
34 	struct irq_domain	*inner_domain;
35 	struct irq_domain	*msi_domain;
36 	u64			msi_addr;
37 	void __iomem		*msi_regs;
38 	unsigned long		*bitmap;
39 	struct mutex		bitmap_lock;
40 	struct xgene_msi_group	*msi_groups;
41 	int			num_cpus;
42 };
43 
44 /* Global data */
45 static struct xgene_msi xgene_msi_ctrl;
46 
47 static struct irq_chip xgene_msi_top_irq_chip = {
48 	.name		= "X-Gene1 MSI",
49 	.irq_enable	= pci_msi_unmask_irq,
50 	.irq_disable	= pci_msi_mask_irq,
51 	.irq_mask	= pci_msi_mask_irq,
52 	.irq_unmask	= pci_msi_unmask_irq,
53 };
54 
55 static struct  msi_domain_info xgene_msi_domain_info = {
56 	.flags	= (MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
57 		  MSI_FLAG_PCI_MSIX),
58 	.chip	= &xgene_msi_top_irq_chip,
59 };
60 
61 /*
62  * X-Gene v1 has 16 groups of MSI termination registers MSInIRx, where
63  * n is group number (0..F), x is index of registers in each group (0..7)
64  * The register layout is as follows:
65  * MSI0IR0			base_addr
66  * MSI0IR1			base_addr +  0x10000
67  * ...				...
68  * MSI0IR6			base_addr +  0x60000
69  * MSI0IR7			base_addr +  0x70000
70  * MSI1IR0			base_addr +  0x80000
71  * MSI1IR1			base_addr +  0x90000
72  * ...				...
73  * MSI1IR7			base_addr +  0xF0000
74  * MSI2IR0			base_addr + 0x100000
75  * ...				...
76  * MSIFIR0			base_addr + 0x780000
77  * MSIFIR1			base_addr + 0x790000
78  * ...				...
79  * MSIFIR7			base_addr + 0x7F0000
80  * MSIINT0			base_addr + 0x800000
81  * MSIINT1			base_addr + 0x810000
82  * ...				...
83  * MSIINTF			base_addr + 0x8F0000
84  *
85  * Each index register supports 16 MSI vectors (0..15) to generate interrupt.
86  * There are total 16 GIC IRQs assigned for these 16 groups of MSI termination
87  * registers.
88  *
89  * Each MSI termination group has 1 MSIINTn register (n is 0..15) to indicate
90  * the MSI pending status caused by 1 of its 8 index registers.
91  */
92 
93 /* MSInIRx read helper */
94 static u32 xgene_msi_ir_read(struct xgene_msi *msi,
95 				    u32 msi_grp, u32 msir_idx)
96 {
97 	return readl_relaxed(msi->msi_regs + MSI_IR0 +
98 			      (msi_grp << 19) + (msir_idx << 16));
99 }
100 
101 /* MSIINTn read helper */
102 static u32 xgene_msi_int_read(struct xgene_msi *msi, u32 msi_grp)
103 {
104 	return readl_relaxed(msi->msi_regs + MSI_INT0 + (msi_grp << 16));
105 }
106 
107 /*
108  * With 2048 MSI vectors supported, the MSI message can be constructed using
109  * following scheme:
110  * - Divide into 8 256-vector groups
111  *		Group 0: 0-255
112  *		Group 1: 256-511
113  *		Group 2: 512-767
114  *		...
115  *		Group 7: 1792-2047
116  * - Each 256-vector group is divided into 16 16-vector groups
117  *	As an example: 16 16-vector groups for 256-vector group 0-255 is
118  *		Group 0: 0-15
119  *		Group 1: 16-32
120  *		...
121  *		Group 15: 240-255
122  * - The termination address of MSI vector in 256-vector group n and 16-vector
123  *   group x is the address of MSIxIRn
124  * - The data for MSI vector in 16-vector group x is x
125  */
126 static u32 hwirq_to_reg_set(unsigned long hwirq)
127 {
128 	return (hwirq / (NR_HW_IRQS * IRQS_PER_IDX));
129 }
130 
131 static u32 hwirq_to_group(unsigned long hwirq)
132 {
133 	return (hwirq % NR_HW_IRQS);
134 }
135 
136 static u32 hwirq_to_msi_data(unsigned long hwirq)
137 {
138 	return ((hwirq / NR_HW_IRQS) % IRQS_PER_IDX);
139 }
140 
141 static void xgene_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
142 {
143 	struct xgene_msi *msi = irq_data_get_irq_chip_data(data);
144 	u32 reg_set = hwirq_to_reg_set(data->hwirq);
145 	u32 group = hwirq_to_group(data->hwirq);
146 	u64 target_addr = msi->msi_addr + (((8 * group) + reg_set) << 16);
147 
148 	msg->address_hi = upper_32_bits(target_addr);
149 	msg->address_lo = lower_32_bits(target_addr);
150 	msg->data = hwirq_to_msi_data(data->hwirq);
151 }
152 
153 /*
154  * X-Gene v1 only has 16 MSI GIC IRQs for 2048 MSI vectors.  To maintain
155  * the expected behaviour of .set_affinity for each MSI interrupt, the 16
156  * MSI GIC IRQs are statically allocated to 8 X-Gene v1 cores (2 GIC IRQs
157  * for each core).  The MSI vector is moved fom 1 MSI GIC IRQ to another
158  * MSI GIC IRQ to steer its MSI interrupt to correct X-Gene v1 core.  As a
159  * consequence, the total MSI vectors that X-Gene v1 supports will be
160  * reduced to 256 (2048/8) vectors.
161  */
162 static int hwirq_to_cpu(unsigned long hwirq)
163 {
164 	return (hwirq % xgene_msi_ctrl.num_cpus);
165 }
166 
167 static unsigned long hwirq_to_canonical_hwirq(unsigned long hwirq)
168 {
169 	return (hwirq - hwirq_to_cpu(hwirq));
170 }
171 
172 static int xgene_msi_set_affinity(struct irq_data *irqdata,
173 				  const struct cpumask *mask, bool force)
174 {
175 	int target_cpu = cpumask_first(mask);
176 	int curr_cpu;
177 
178 	curr_cpu = hwirq_to_cpu(irqdata->hwirq);
179 	if (curr_cpu == target_cpu)
180 		return IRQ_SET_MASK_OK_DONE;
181 
182 	/* Update MSI number to target the new CPU */
183 	irqdata->hwirq = hwirq_to_canonical_hwirq(irqdata->hwirq) + target_cpu;
184 
185 	return IRQ_SET_MASK_OK;
186 }
187 
188 static struct irq_chip xgene_msi_bottom_irq_chip = {
189 	.name			= "MSI",
190 	.irq_set_affinity       = xgene_msi_set_affinity,
191 	.irq_compose_msi_msg	= xgene_compose_msi_msg,
192 };
193 
194 static int xgene_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
195 				  unsigned int nr_irqs, void *args)
196 {
197 	struct xgene_msi *msi = domain->host_data;
198 	int msi_irq;
199 
200 	mutex_lock(&msi->bitmap_lock);
201 
202 	msi_irq = bitmap_find_next_zero_area(msi->bitmap, NR_MSI_VEC, 0,
203 					     msi->num_cpus, 0);
204 	if (msi_irq < NR_MSI_VEC)
205 		bitmap_set(msi->bitmap, msi_irq, msi->num_cpus);
206 	else
207 		msi_irq = -ENOSPC;
208 
209 	mutex_unlock(&msi->bitmap_lock);
210 
211 	if (msi_irq < 0)
212 		return msi_irq;
213 
214 	irq_domain_set_info(domain, virq, msi_irq,
215 			    &xgene_msi_bottom_irq_chip, domain->host_data,
216 			    handle_simple_irq, NULL, NULL);
217 
218 	return 0;
219 }
220 
221 static void xgene_irq_domain_free(struct irq_domain *domain,
222 				  unsigned int virq, unsigned int nr_irqs)
223 {
224 	struct irq_data *d = irq_domain_get_irq_data(domain, virq);
225 	struct xgene_msi *msi = irq_data_get_irq_chip_data(d);
226 	u32 hwirq;
227 
228 	mutex_lock(&msi->bitmap_lock);
229 
230 	hwirq = hwirq_to_canonical_hwirq(d->hwirq);
231 	bitmap_clear(msi->bitmap, hwirq, msi->num_cpus);
232 
233 	mutex_unlock(&msi->bitmap_lock);
234 
235 	irq_domain_free_irqs_parent(domain, virq, nr_irqs);
236 }
237 
238 static const struct irq_domain_ops msi_domain_ops = {
239 	.alloc  = xgene_irq_domain_alloc,
240 	.free   = xgene_irq_domain_free,
241 };
242 
243 static int xgene_allocate_domains(struct xgene_msi *msi)
244 {
245 	msi->inner_domain = irq_domain_add_linear(NULL, NR_MSI_VEC,
246 						  &msi_domain_ops, msi);
247 	if (!msi->inner_domain)
248 		return -ENOMEM;
249 
250 	msi->msi_domain = pci_msi_create_irq_domain(of_node_to_fwnode(msi->node),
251 						    &xgene_msi_domain_info,
252 						    msi->inner_domain);
253 
254 	if (!msi->msi_domain) {
255 		irq_domain_remove(msi->inner_domain);
256 		return -ENOMEM;
257 	}
258 
259 	return 0;
260 }
261 
262 static void xgene_free_domains(struct xgene_msi *msi)
263 {
264 	if (msi->msi_domain)
265 		irq_domain_remove(msi->msi_domain);
266 	if (msi->inner_domain)
267 		irq_domain_remove(msi->inner_domain);
268 }
269 
270 static int xgene_msi_init_allocator(struct xgene_msi *xgene_msi)
271 {
272 	int size = BITS_TO_LONGS(NR_MSI_VEC) * sizeof(long);
273 
274 	xgene_msi->bitmap = kzalloc(size, GFP_KERNEL);
275 	if (!xgene_msi->bitmap)
276 		return -ENOMEM;
277 
278 	mutex_init(&xgene_msi->bitmap_lock);
279 
280 	xgene_msi->msi_groups = kcalloc(NR_HW_IRQS,
281 					sizeof(struct xgene_msi_group),
282 					GFP_KERNEL);
283 	if (!xgene_msi->msi_groups)
284 		return -ENOMEM;
285 
286 	return 0;
287 }
288 
289 static void xgene_msi_isr(struct irq_desc *desc)
290 {
291 	struct irq_chip *chip = irq_desc_get_chip(desc);
292 	struct xgene_msi_group *msi_groups;
293 	struct xgene_msi *xgene_msi;
294 	unsigned int virq;
295 	int msir_index, msir_val, hw_irq;
296 	u32 intr_index, grp_select, msi_grp;
297 
298 	chained_irq_enter(chip, desc);
299 
300 	msi_groups = irq_desc_get_handler_data(desc);
301 	xgene_msi = msi_groups->msi;
302 	msi_grp = msi_groups->msi_grp;
303 
304 	/*
305 	 * MSIINTn (n is 0..F) indicates if there is a pending MSI interrupt
306 	 * If bit x of this register is set (x is 0..7), one or more interupts
307 	 * corresponding to MSInIRx is set.
308 	 */
309 	grp_select = xgene_msi_int_read(xgene_msi, msi_grp);
310 	while (grp_select) {
311 		msir_index = ffs(grp_select) - 1;
312 		/*
313 		 * Calculate MSInIRx address to read to check for interrupts
314 		 * (refer to termination address and data assignment
315 		 * described in xgene_compose_msi_msg() )
316 		 */
317 		msir_val = xgene_msi_ir_read(xgene_msi, msi_grp, msir_index);
318 		while (msir_val) {
319 			intr_index = ffs(msir_val) - 1;
320 			/*
321 			 * Calculate MSI vector number (refer to the termination
322 			 * address and data assignment described in
323 			 * xgene_compose_msi_msg function)
324 			 */
325 			hw_irq = (((msir_index * IRQS_PER_IDX) + intr_index) *
326 				 NR_HW_IRQS) + msi_grp;
327 			/*
328 			 * As we have multiple hw_irq that maps to single MSI,
329 			 * always look up the virq using the hw_irq as seen from
330 			 * CPU0
331 			 */
332 			hw_irq = hwirq_to_canonical_hwirq(hw_irq);
333 			virq = irq_find_mapping(xgene_msi->inner_domain, hw_irq);
334 			WARN_ON(!virq);
335 			if (virq != 0)
336 				generic_handle_irq(virq);
337 			msir_val &= ~(1 << intr_index);
338 		}
339 		grp_select &= ~(1 << msir_index);
340 
341 		if (!grp_select) {
342 			/*
343 			 * We handled all interrupts happened in this group,
344 			 * resample this group MSI_INTx register in case
345 			 * something else has been made pending in the meantime
346 			 */
347 			grp_select = xgene_msi_int_read(xgene_msi, msi_grp);
348 		}
349 	}
350 
351 	chained_irq_exit(chip, desc);
352 }
353 
354 static enum cpuhp_state pci_xgene_online;
355 
356 static int xgene_msi_remove(struct platform_device *pdev)
357 {
358 	struct xgene_msi *msi = platform_get_drvdata(pdev);
359 
360 	if (pci_xgene_online)
361 		cpuhp_remove_state(pci_xgene_online);
362 	cpuhp_remove_state(CPUHP_PCI_XGENE_DEAD);
363 
364 	kfree(msi->msi_groups);
365 
366 	kfree(msi->bitmap);
367 	msi->bitmap = NULL;
368 
369 	xgene_free_domains(msi);
370 
371 	return 0;
372 }
373 
374 static int xgene_msi_hwirq_alloc(unsigned int cpu)
375 {
376 	struct xgene_msi *msi = &xgene_msi_ctrl;
377 	struct xgene_msi_group *msi_group;
378 	cpumask_var_t mask;
379 	int i;
380 	int err;
381 
382 	for (i = cpu; i < NR_HW_IRQS; i += msi->num_cpus) {
383 		msi_group = &msi->msi_groups[i];
384 		if (!msi_group->gic_irq)
385 			continue;
386 
387 		irq_set_chained_handler(msi_group->gic_irq,
388 					xgene_msi_isr);
389 		err = irq_set_handler_data(msi_group->gic_irq, msi_group);
390 		if (err) {
391 			pr_err("failed to register GIC IRQ handler\n");
392 			return -EINVAL;
393 		}
394 		/*
395 		 * Statically allocate MSI GIC IRQs to each CPU core.
396 		 * With 8-core X-Gene v1, 2 MSI GIC IRQs are allocated
397 		 * to each core.
398 		 */
399 		if (alloc_cpumask_var(&mask, GFP_KERNEL)) {
400 			cpumask_clear(mask);
401 			cpumask_set_cpu(cpu, mask);
402 			err = irq_set_affinity(msi_group->gic_irq, mask);
403 			if (err)
404 				pr_err("failed to set affinity for GIC IRQ");
405 			free_cpumask_var(mask);
406 		} else {
407 			pr_err("failed to alloc CPU mask for affinity\n");
408 			err = -EINVAL;
409 		}
410 
411 		if (err) {
412 			irq_set_chained_handler_and_data(msi_group->gic_irq,
413 							 NULL, NULL);
414 			return err;
415 		}
416 	}
417 
418 	return 0;
419 }
420 
421 static int xgene_msi_hwirq_free(unsigned int cpu)
422 {
423 	struct xgene_msi *msi = &xgene_msi_ctrl;
424 	struct xgene_msi_group *msi_group;
425 	int i;
426 
427 	for (i = cpu; i < NR_HW_IRQS; i += msi->num_cpus) {
428 		msi_group = &msi->msi_groups[i];
429 		if (!msi_group->gic_irq)
430 			continue;
431 
432 		irq_set_chained_handler_and_data(msi_group->gic_irq, NULL,
433 						 NULL);
434 	}
435 	return 0;
436 }
437 
438 static const struct of_device_id xgene_msi_match_table[] = {
439 	{.compatible = "apm,xgene1-msi"},
440 	{},
441 };
442 
443 static int xgene_msi_probe(struct platform_device *pdev)
444 {
445 	struct resource *res;
446 	int rc, irq_index;
447 	struct xgene_msi *xgene_msi;
448 	int virt_msir;
449 	u32 msi_val, msi_idx;
450 
451 	xgene_msi = &xgene_msi_ctrl;
452 
453 	platform_set_drvdata(pdev, xgene_msi);
454 
455 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
456 	xgene_msi->msi_regs = devm_ioremap_resource(&pdev->dev, res);
457 	if (IS_ERR(xgene_msi->msi_regs)) {
458 		dev_err(&pdev->dev, "no reg space\n");
459 		rc = PTR_ERR(xgene_msi->msi_regs);
460 		goto error;
461 	}
462 	xgene_msi->msi_addr = res->start;
463 	xgene_msi->node = pdev->dev.of_node;
464 	xgene_msi->num_cpus = num_possible_cpus();
465 
466 	rc = xgene_msi_init_allocator(xgene_msi);
467 	if (rc) {
468 		dev_err(&pdev->dev, "Error allocating MSI bitmap\n");
469 		goto error;
470 	}
471 
472 	rc = xgene_allocate_domains(xgene_msi);
473 	if (rc) {
474 		dev_err(&pdev->dev, "Failed to allocate MSI domain\n");
475 		goto error;
476 	}
477 
478 	for (irq_index = 0; irq_index < NR_HW_IRQS; irq_index++) {
479 		virt_msir = platform_get_irq(pdev, irq_index);
480 		if (virt_msir < 0) {
481 			dev_err(&pdev->dev, "Cannot translate IRQ index %d\n",
482 				irq_index);
483 			rc = virt_msir;
484 			goto error;
485 		}
486 		xgene_msi->msi_groups[irq_index].gic_irq = virt_msir;
487 		xgene_msi->msi_groups[irq_index].msi_grp = irq_index;
488 		xgene_msi->msi_groups[irq_index].msi = xgene_msi;
489 	}
490 
491 	/*
492 	 * MSInIRx registers are read-to-clear; before registering
493 	 * interrupt handlers, read all of them to clear spurious
494 	 * interrupts that may occur before the driver is probed.
495 	 */
496 	for (irq_index = 0; irq_index < NR_HW_IRQS; irq_index++) {
497 		for (msi_idx = 0; msi_idx < IDX_PER_GROUP; msi_idx++)
498 			msi_val = xgene_msi_ir_read(xgene_msi, irq_index,
499 						    msi_idx);
500 		/* Read MSIINTn to confirm */
501 		msi_val = xgene_msi_int_read(xgene_msi, irq_index);
502 		if (msi_val) {
503 			dev_err(&pdev->dev, "Failed to clear spurious IRQ\n");
504 			rc = -EINVAL;
505 			goto error;
506 		}
507 	}
508 
509 	rc = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "pci/xgene:online",
510 			       xgene_msi_hwirq_alloc, NULL);
511 	if (rc < 0)
512 		goto err_cpuhp;
513 	pci_xgene_online = rc;
514 	rc = cpuhp_setup_state(CPUHP_PCI_XGENE_DEAD, "pci/xgene:dead", NULL,
515 			       xgene_msi_hwirq_free);
516 	if (rc)
517 		goto err_cpuhp;
518 
519 	dev_info(&pdev->dev, "APM X-Gene PCIe MSI driver loaded\n");
520 
521 	return 0;
522 
523 err_cpuhp:
524 	dev_err(&pdev->dev, "failed to add CPU MSI notifier\n");
525 error:
526 	xgene_msi_remove(pdev);
527 	return rc;
528 }
529 
530 static struct platform_driver xgene_msi_driver = {
531 	.driver = {
532 		.name = "xgene-msi",
533 		.of_match_table = xgene_msi_match_table,
534 	},
535 	.probe = xgene_msi_probe,
536 	.remove = xgene_msi_remove,
537 };
538 
539 static int __init xgene_pcie_msi_init(void)
540 {
541 	return platform_driver_register(&xgene_msi_driver);
542 }
543 subsys_initcall(xgene_pcie_msi_init);
544