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 */
xgene_msi_ir_read(struct xgene_msi * msi,u32 msi_grp,u32 msir_idx)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 */
xgene_msi_int_read(struct xgene_msi * msi,u32 msi_grp)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 */
hwirq_to_reg_set(unsigned long hwirq)126 static u32 hwirq_to_reg_set(unsigned long hwirq)
127 {
128 return (hwirq / (NR_HW_IRQS * IRQS_PER_IDX));
129 }
130
hwirq_to_group(unsigned long hwirq)131 static u32 hwirq_to_group(unsigned long hwirq)
132 {
133 return (hwirq % NR_HW_IRQS);
134 }
135
hwirq_to_msi_data(unsigned long hwirq)136 static u32 hwirq_to_msi_data(unsigned long hwirq)
137 {
138 return ((hwirq / NR_HW_IRQS) % IRQS_PER_IDX);
139 }
140
xgene_compose_msi_msg(struct irq_data * data,struct msi_msg * msg)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 */
hwirq_to_cpu(unsigned long hwirq)162 static int hwirq_to_cpu(unsigned long hwirq)
163 {
164 return (hwirq % xgene_msi_ctrl.num_cpus);
165 }
166
hwirq_to_canonical_hwirq(unsigned long hwirq)167 static unsigned long hwirq_to_canonical_hwirq(unsigned long hwirq)
168 {
169 return (hwirq - hwirq_to_cpu(hwirq));
170 }
171
xgene_msi_set_affinity(struct irq_data * irqdata,const struct cpumask * mask,bool force)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
xgene_irq_domain_alloc(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs,void * args)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
xgene_irq_domain_free(struct irq_domain * domain,unsigned int virq,unsigned int nr_irqs)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
xgene_allocate_domains(struct xgene_msi * msi)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
xgene_free_domains(struct xgene_msi * msi)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
xgene_msi_init_allocator(struct xgene_msi * xgene_msi)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
xgene_msi_isr(struct irq_desc * desc)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 int msir_index, msir_val, hw_irq, ret;
295 u32 intr_index, grp_select, msi_grp;
296
297 chained_irq_enter(chip, desc);
298
299 msi_groups = irq_desc_get_handler_data(desc);
300 xgene_msi = msi_groups->msi;
301 msi_grp = msi_groups->msi_grp;
302
303 /*
304 * MSIINTn (n is 0..F) indicates if there is a pending MSI interrupt
305 * If bit x of this register is set (x is 0..7), one or more interrupts
306 * corresponding to MSInIRx is set.
307 */
308 grp_select = xgene_msi_int_read(xgene_msi, msi_grp);
309 while (grp_select) {
310 msir_index = ffs(grp_select) - 1;
311 /*
312 * Calculate MSInIRx address to read to check for interrupts
313 * (refer to termination address and data assignment
314 * described in xgene_compose_msi_msg() )
315 */
316 msir_val = xgene_msi_ir_read(xgene_msi, msi_grp, msir_index);
317 while (msir_val) {
318 intr_index = ffs(msir_val) - 1;
319 /*
320 * Calculate MSI vector number (refer to the termination
321 * address and data assignment described in
322 * xgene_compose_msi_msg function)
323 */
324 hw_irq = (((msir_index * IRQS_PER_IDX) + intr_index) *
325 NR_HW_IRQS) + msi_grp;
326 /*
327 * As we have multiple hw_irq that maps to single MSI,
328 * always look up the virq using the hw_irq as seen from
329 * CPU0
330 */
331 hw_irq = hwirq_to_canonical_hwirq(hw_irq);
332 ret = generic_handle_domain_irq(xgene_msi->inner_domain, hw_irq);
333 WARN_ON_ONCE(ret);
334 msir_val &= ~(1 << intr_index);
335 }
336 grp_select &= ~(1 << msir_index);
337
338 if (!grp_select) {
339 /*
340 * We handled all interrupts happened in this group,
341 * resample this group MSI_INTx register in case
342 * something else has been made pending in the meantime
343 */
344 grp_select = xgene_msi_int_read(xgene_msi, msi_grp);
345 }
346 }
347
348 chained_irq_exit(chip, desc);
349 }
350
351 static enum cpuhp_state pci_xgene_online;
352
xgene_msi_remove(struct platform_device * pdev)353 static int xgene_msi_remove(struct platform_device *pdev)
354 {
355 struct xgene_msi *msi = platform_get_drvdata(pdev);
356
357 if (pci_xgene_online)
358 cpuhp_remove_state(pci_xgene_online);
359 cpuhp_remove_state(CPUHP_PCI_XGENE_DEAD);
360
361 kfree(msi->msi_groups);
362
363 kfree(msi->bitmap);
364 msi->bitmap = NULL;
365
366 xgene_free_domains(msi);
367
368 return 0;
369 }
370
xgene_msi_hwirq_alloc(unsigned int cpu)371 static int xgene_msi_hwirq_alloc(unsigned int cpu)
372 {
373 struct xgene_msi *msi = &xgene_msi_ctrl;
374 struct xgene_msi_group *msi_group;
375 cpumask_var_t mask;
376 int i;
377 int err;
378
379 for (i = cpu; i < NR_HW_IRQS; i += msi->num_cpus) {
380 msi_group = &msi->msi_groups[i];
381 if (!msi_group->gic_irq)
382 continue;
383
384 irq_set_chained_handler_and_data(msi_group->gic_irq,
385 xgene_msi_isr, msi_group);
386
387 /*
388 * Statically allocate MSI GIC IRQs to each CPU core.
389 * With 8-core X-Gene v1, 2 MSI GIC IRQs are allocated
390 * to each core.
391 */
392 if (alloc_cpumask_var(&mask, GFP_KERNEL)) {
393 cpumask_clear(mask);
394 cpumask_set_cpu(cpu, mask);
395 err = irq_set_affinity(msi_group->gic_irq, mask);
396 if (err)
397 pr_err("failed to set affinity for GIC IRQ");
398 free_cpumask_var(mask);
399 } else {
400 pr_err("failed to alloc CPU mask for affinity\n");
401 err = -EINVAL;
402 }
403
404 if (err) {
405 irq_set_chained_handler_and_data(msi_group->gic_irq,
406 NULL, NULL);
407 return err;
408 }
409 }
410
411 return 0;
412 }
413
xgene_msi_hwirq_free(unsigned int cpu)414 static int xgene_msi_hwirq_free(unsigned int cpu)
415 {
416 struct xgene_msi *msi = &xgene_msi_ctrl;
417 struct xgene_msi_group *msi_group;
418 int i;
419
420 for (i = cpu; i < NR_HW_IRQS; i += msi->num_cpus) {
421 msi_group = &msi->msi_groups[i];
422 if (!msi_group->gic_irq)
423 continue;
424
425 irq_set_chained_handler_and_data(msi_group->gic_irq, NULL,
426 NULL);
427 }
428 return 0;
429 }
430
431 static const struct of_device_id xgene_msi_match_table[] = {
432 {.compatible = "apm,xgene1-msi"},
433 {},
434 };
435
xgene_msi_probe(struct platform_device * pdev)436 static int xgene_msi_probe(struct platform_device *pdev)
437 {
438 struct resource *res;
439 int rc, irq_index;
440 struct xgene_msi *xgene_msi;
441 int virt_msir;
442 u32 msi_val, msi_idx;
443
444 xgene_msi = &xgene_msi_ctrl;
445
446 platform_set_drvdata(pdev, xgene_msi);
447
448 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
449 xgene_msi->msi_regs = devm_ioremap_resource(&pdev->dev, res);
450 if (IS_ERR(xgene_msi->msi_regs)) {
451 rc = PTR_ERR(xgene_msi->msi_regs);
452 goto error;
453 }
454 xgene_msi->msi_addr = res->start;
455 xgene_msi->node = pdev->dev.of_node;
456 xgene_msi->num_cpus = num_possible_cpus();
457
458 rc = xgene_msi_init_allocator(xgene_msi);
459 if (rc) {
460 dev_err(&pdev->dev, "Error allocating MSI bitmap\n");
461 goto error;
462 }
463
464 rc = xgene_allocate_domains(xgene_msi);
465 if (rc) {
466 dev_err(&pdev->dev, "Failed to allocate MSI domain\n");
467 goto error;
468 }
469
470 for (irq_index = 0; irq_index < NR_HW_IRQS; irq_index++) {
471 virt_msir = platform_get_irq(pdev, irq_index);
472 if (virt_msir < 0) {
473 rc = virt_msir;
474 goto error;
475 }
476 xgene_msi->msi_groups[irq_index].gic_irq = virt_msir;
477 xgene_msi->msi_groups[irq_index].msi_grp = irq_index;
478 xgene_msi->msi_groups[irq_index].msi = xgene_msi;
479 }
480
481 /*
482 * MSInIRx registers are read-to-clear; before registering
483 * interrupt handlers, read all of them to clear spurious
484 * interrupts that may occur before the driver is probed.
485 */
486 for (irq_index = 0; irq_index < NR_HW_IRQS; irq_index++) {
487 for (msi_idx = 0; msi_idx < IDX_PER_GROUP; msi_idx++)
488 xgene_msi_ir_read(xgene_msi, irq_index, msi_idx);
489
490 /* Read MSIINTn to confirm */
491 msi_val = xgene_msi_int_read(xgene_msi, irq_index);
492 if (msi_val) {
493 dev_err(&pdev->dev, "Failed to clear spurious IRQ\n");
494 rc = -EINVAL;
495 goto error;
496 }
497 }
498
499 rc = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "pci/xgene:online",
500 xgene_msi_hwirq_alloc, NULL);
501 if (rc < 0)
502 goto err_cpuhp;
503 pci_xgene_online = rc;
504 rc = cpuhp_setup_state(CPUHP_PCI_XGENE_DEAD, "pci/xgene:dead", NULL,
505 xgene_msi_hwirq_free);
506 if (rc)
507 goto err_cpuhp;
508
509 dev_info(&pdev->dev, "APM X-Gene PCIe MSI driver loaded\n");
510
511 return 0;
512
513 err_cpuhp:
514 dev_err(&pdev->dev, "failed to add CPU MSI notifier\n");
515 error:
516 xgene_msi_remove(pdev);
517 return rc;
518 }
519
520 static struct platform_driver xgene_msi_driver = {
521 .driver = {
522 .name = "xgene-msi",
523 .of_match_table = xgene_msi_match_table,
524 },
525 .probe = xgene_msi_probe,
526 .remove = xgene_msi_remove,
527 };
528
xgene_pcie_msi_init(void)529 static int __init xgene_pcie_msi_init(void)
530 {
531 return platform_driver_register(&xgene_msi_driver);
532 }
533 subsys_initcall(xgene_pcie_msi_init);
534