2 * Copyright (C) 2015 Broadcom Corporation
4 * This program is free software; you can redistribute it and/or
5 * modify it under the terms of the GNU General Public License as
6 * published by the Free Software Foundation version 2.
8 * This program is distributed "as is" WITHOUT ANY WARRANTY of any
9 * kind, whether express or implied; without even the implied warranty
10 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11 * GNU General Public License for more details.
14 #include <linux/interrupt.h>
15 #include <linux/irqchip/chained_irq.h>
16 #include <linux/irqdomain.h>
17 #include <linux/msi.h>
18 #include <linux/of_irq.h>
19 #include <linux/of_pci.h>
20 #include <linux/pci.h>
22 #include "pcie-iproc.h"
24 #define IPROC_MSI_INTR_EN_SHIFT 11
25 #define IPROC_MSI_INTR_EN BIT(IPROC_MSI_INTR_EN_SHIFT)
26 #define IPROC_MSI_INT_N_EVENT_SHIFT 1
27 #define IPROC_MSI_INT_N_EVENT BIT(IPROC_MSI_INT_N_EVENT_SHIFT)
28 #define IPROC_MSI_EQ_EN_SHIFT 0
29 #define IPROC_MSI_EQ_EN BIT(IPROC_MSI_EQ_EN_SHIFT)
31 #define IPROC_MSI_EQ_MASK 0x3f
33 /* Max number of GIC interrupts */
36 /* Number of entries in each event queue */
39 /* Size of each event queue memory region */
40 #define EQ_MEM_REGION_SIZE SZ_4K
42 /* Size of each MSI address region */
43 #define MSI_MEM_REGION_SIZE SZ_4K
46 IPROC_MSI_EQ_PAGE = 0,
47 IPROC_MSI_EQ_PAGE_UPPER,
62 * One MSI group is allocated per GIC interrupt, serviced by one iProc MSI
65 * @msi: pointer to iProc MSI data
66 * @gic_irq: GIC interrupt
67 * @eq: Event queue number
69 struct iproc_msi_grp {
70 struct iproc_msi *msi;
76 * iProc event queue based MSI
78 * Only meant to be used on platforms without MSI support integrated into the
81 * @pcie: pointer to iProc PCIe data
82 * @reg_offsets: MSI register offsets
84 * @nr_irqs: number of total interrupts connected to GIC
85 * @nr_cpus: number of toal CPUs
86 * @has_inten_reg: indicates the MSI interrupt enable register needs to be
87 * set explicitly (required for some legacy platforms)
88 * @bitmap: MSI vector bitmap
89 * @bitmap_lock: lock to protect access to the MSI bitmap
90 * @nr_msi_vecs: total number of MSI vectors
91 * @inner_domain: inner IRQ domain
92 * @msi_domain: MSI IRQ domain
93 * @nr_eq_region: required number of 4K aligned memory region for MSI event
95 * @nr_msi_region: required number of 4K aligned address region for MSI posted
97 * @eq_cpu: pointer to allocated memory region for MSI event queues
98 * @eq_dma: DMA address of MSI event queues
99 * @msi_addr: MSI address
102 struct iproc_pcie *pcie;
103 const u16 (*reg_offsets)[IPROC_MSI_REG_SIZE];
104 struct iproc_msi_grp *grps;
108 unsigned long *bitmap;
109 struct mutex bitmap_lock;
110 unsigned int nr_msi_vecs;
111 struct irq_domain *inner_domain;
112 struct irq_domain *msi_domain;
113 unsigned int nr_eq_region;
114 unsigned int nr_msi_region;
117 phys_addr_t msi_addr;
120 static const u16 iproc_msi_reg_paxb[NR_HW_IRQS][IPROC_MSI_REG_SIZE] = {
121 { 0x200, 0x2c0, 0x204, 0x2c4, 0x210, 0x250, 0x254, 0x208 },
122 { 0x200, 0x2c0, 0x204, 0x2c4, 0x214, 0x258, 0x25c, 0x208 },
123 { 0x200, 0x2c0, 0x204, 0x2c4, 0x218, 0x260, 0x264, 0x208 },
124 { 0x200, 0x2c0, 0x204, 0x2c4, 0x21c, 0x268, 0x26c, 0x208 },
125 { 0x200, 0x2c0, 0x204, 0x2c4, 0x220, 0x270, 0x274, 0x208 },
126 { 0x200, 0x2c0, 0x204, 0x2c4, 0x224, 0x278, 0x27c, 0x208 },
129 static const u16 iproc_msi_reg_paxc[NR_HW_IRQS][IPROC_MSI_REG_SIZE] = {
130 { 0xc00, 0xc04, 0xc08, 0xc0c, 0xc40, 0xc50, 0xc60 },
131 { 0xc10, 0xc14, 0xc18, 0xc1c, 0xc44, 0xc54, 0xc64 },
132 { 0xc20, 0xc24, 0xc28, 0xc2c, 0xc48, 0xc58, 0xc68 },
133 { 0xc30, 0xc34, 0xc38, 0xc3c, 0xc4c, 0xc5c, 0xc6c },
136 static inline u32 iproc_msi_read_reg(struct iproc_msi *msi,
137 enum iproc_msi_reg reg,
140 struct iproc_pcie *pcie = msi->pcie;
142 return readl_relaxed(pcie->base + msi->reg_offsets[eq][reg]);
145 static inline void iproc_msi_write_reg(struct iproc_msi *msi,
146 enum iproc_msi_reg reg,
149 struct iproc_pcie *pcie = msi->pcie;
151 writel_relaxed(val, pcie->base + msi->reg_offsets[eq][reg]);
154 static inline u32 hwirq_to_group(struct iproc_msi *msi, unsigned long hwirq)
156 return (hwirq % msi->nr_irqs);
159 static inline unsigned int iproc_msi_addr_offset(struct iproc_msi *msi,
162 if (msi->nr_msi_region > 1)
163 return hwirq_to_group(msi, hwirq) * MSI_MEM_REGION_SIZE;
165 return hwirq_to_group(msi, hwirq) * sizeof(u32);
168 static inline unsigned int iproc_msi_eq_offset(struct iproc_msi *msi, u32 eq)
170 if (msi->nr_eq_region > 1)
171 return eq * EQ_MEM_REGION_SIZE;
173 return eq * EQ_LEN * sizeof(u32);
176 static struct irq_chip iproc_msi_irq_chip = {
180 static struct msi_domain_info iproc_msi_domain_info = {
181 .flags = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
183 .chip = &iproc_msi_irq_chip,
187 * In iProc PCIe core, each MSI group is serviced by a GIC interrupt and a
188 * dedicated event queue. Each MSI group can support up to 64 MSI vectors.
190 * The number of MSI groups varies between different iProc SoCs. The total
191 * number of CPU cores also varies. To support MSI IRQ affinity, we
192 * distribute GIC interrupts across all available CPUs. MSI vector is moved
193 * from one GIC interrupt to another to steer to the target CPU.
196 * - the number of MSI groups is M
197 * - the number of CPU cores is N
198 * - M is always a multiple of N
200 * Total number of raw MSI vectors = M * 64
201 * Total number of supported MSI vectors = (M * 64) / N
203 static inline int hwirq_to_cpu(struct iproc_msi *msi, unsigned long hwirq)
205 return (hwirq % msi->nr_cpus);
208 static inline unsigned long hwirq_to_canonical_hwirq(struct iproc_msi *msi,
211 return (hwirq - hwirq_to_cpu(msi, hwirq));
214 static int iproc_msi_irq_set_affinity(struct irq_data *data,
215 const struct cpumask *mask, bool force)
217 struct iproc_msi *msi = irq_data_get_irq_chip_data(data);
218 int target_cpu = cpumask_first(mask);
221 curr_cpu = hwirq_to_cpu(msi, data->hwirq);
222 if (curr_cpu == target_cpu)
223 return IRQ_SET_MASK_OK_DONE;
225 /* steer MSI to the target CPU */
226 data->hwirq = hwirq_to_canonical_hwirq(msi, data->hwirq) + target_cpu;
228 return IRQ_SET_MASK_OK;
231 static void iproc_msi_irq_compose_msi_msg(struct irq_data *data,
234 struct iproc_msi *msi = irq_data_get_irq_chip_data(data);
237 addr = msi->msi_addr + iproc_msi_addr_offset(msi, data->hwirq);
238 msg->address_lo = lower_32_bits(addr);
239 msg->address_hi = upper_32_bits(addr);
240 msg->data = data->hwirq;
243 static struct irq_chip iproc_msi_bottom_irq_chip = {
245 .irq_set_affinity = iproc_msi_irq_set_affinity,
246 .irq_compose_msi_msg = iproc_msi_irq_compose_msi_msg,
249 static int iproc_msi_irq_domain_alloc(struct irq_domain *domain,
250 unsigned int virq, unsigned int nr_irqs,
253 struct iproc_msi *msi = domain->host_data;
256 mutex_lock(&msi->bitmap_lock);
258 /* Allocate 'nr_cpus' number of MSI vectors each time */
259 hwirq = bitmap_find_next_zero_area(msi->bitmap, msi->nr_msi_vecs, 0,
261 if (hwirq < msi->nr_msi_vecs) {
262 bitmap_set(msi->bitmap, hwirq, msi->nr_cpus);
264 mutex_unlock(&msi->bitmap_lock);
268 mutex_unlock(&msi->bitmap_lock);
270 irq_domain_set_info(domain, virq, hwirq, &iproc_msi_bottom_irq_chip,
271 domain->host_data, handle_simple_irq, NULL, NULL);
276 static void iproc_msi_irq_domain_free(struct irq_domain *domain,
277 unsigned int virq, unsigned int nr_irqs)
279 struct irq_data *data = irq_domain_get_irq_data(domain, virq);
280 struct iproc_msi *msi = irq_data_get_irq_chip_data(data);
283 mutex_lock(&msi->bitmap_lock);
285 hwirq = hwirq_to_canonical_hwirq(msi, data->hwirq);
286 bitmap_clear(msi->bitmap, hwirq, msi->nr_cpus);
288 mutex_unlock(&msi->bitmap_lock);
290 irq_domain_free_irqs_parent(domain, virq, nr_irqs);
293 static const struct irq_domain_ops msi_domain_ops = {
294 .alloc = iproc_msi_irq_domain_alloc,
295 .free = iproc_msi_irq_domain_free,
298 static inline u32 decode_msi_hwirq(struct iproc_msi *msi, u32 eq, u32 head)
303 offs = iproc_msi_eq_offset(msi, eq) + head * sizeof(u32);
304 msg = (u32 *)(msi->eq_cpu + offs);
305 hwirq = *msg & IPROC_MSI_EQ_MASK;
308 * Since we have multiple hwirq mapped to a single MSI vector,
309 * now we need to derive the hwirq at CPU0. It can then be used to
310 * mapped back to virq.
312 return hwirq_to_canonical_hwirq(msi, hwirq);
315 static void iproc_msi_handler(struct irq_desc *desc)
317 struct irq_chip *chip = irq_desc_get_chip(desc);
318 struct iproc_msi_grp *grp;
319 struct iproc_msi *msi;
320 struct iproc_pcie *pcie;
321 u32 eq, head, tail, nr_events;
325 chained_irq_enter(chip, desc);
327 grp = irq_desc_get_handler_data(desc);
333 * iProc MSI event queue is tracked by head and tail pointers. Head
334 * pointer indicates the next entry (MSI data) to be consumed by SW in
335 * the queue and needs to be updated by SW. iProc MSI core uses the
336 * tail pointer as the next data insertion point.
338 * Entries between head and tail pointers contain valid MSI data. MSI
339 * data is guaranteed to be in the event queue memory before the tail
340 * pointer is updated by the iProc MSI core.
342 head = iproc_msi_read_reg(msi, IPROC_MSI_EQ_HEAD,
343 eq) & IPROC_MSI_EQ_MASK;
345 tail = iproc_msi_read_reg(msi, IPROC_MSI_EQ_TAIL,
346 eq) & IPROC_MSI_EQ_MASK;
349 * Figure out total number of events (MSI data) to be
352 nr_events = (tail < head) ?
353 (EQ_LEN - (head - tail)) : (tail - head);
357 /* process all outstanding events */
358 while (nr_events--) {
359 hwirq = decode_msi_hwirq(msi, eq, head);
360 virq = irq_find_mapping(msi->inner_domain, hwirq);
361 generic_handle_irq(virq);
368 * Now all outstanding events have been processed. Update the
371 iproc_msi_write_reg(msi, IPROC_MSI_EQ_HEAD, eq, head);
374 * Now go read the tail pointer again to see if there are new
375 * oustanding events that came in during the above window.
379 chained_irq_exit(chip, desc);
382 static void iproc_msi_enable(struct iproc_msi *msi)
387 /* Program memory region for each event queue */
388 for (i = 0; i < msi->nr_eq_region; i++) {
389 dma_addr_t addr = msi->eq_dma + (i * EQ_MEM_REGION_SIZE);
391 iproc_msi_write_reg(msi, IPROC_MSI_EQ_PAGE, i,
392 lower_32_bits(addr));
393 iproc_msi_write_reg(msi, IPROC_MSI_EQ_PAGE_UPPER, i,
394 upper_32_bits(addr));
397 /* Program address region for MSI posted writes */
398 for (i = 0; i < msi->nr_msi_region; i++) {
399 phys_addr_t addr = msi->msi_addr + (i * MSI_MEM_REGION_SIZE);
401 iproc_msi_write_reg(msi, IPROC_MSI_PAGE, i,
402 lower_32_bits(addr));
403 iproc_msi_write_reg(msi, IPROC_MSI_PAGE_UPPER, i,
404 upper_32_bits(addr));
407 for (eq = 0; eq < msi->nr_irqs; eq++) {
408 /* Enable MSI event queue */
409 val = IPROC_MSI_INTR_EN | IPROC_MSI_INT_N_EVENT |
411 iproc_msi_write_reg(msi, IPROC_MSI_CTRL, eq, val);
414 * Some legacy platforms require the MSI interrupt enable
415 * register to be set explicitly.
417 if (msi->has_inten_reg) {
418 val = iproc_msi_read_reg(msi, IPROC_MSI_INTS_EN, eq);
420 iproc_msi_write_reg(msi, IPROC_MSI_INTS_EN, eq, val);
425 static void iproc_msi_disable(struct iproc_msi *msi)
429 for (eq = 0; eq < msi->nr_irqs; eq++) {
430 if (msi->has_inten_reg) {
431 val = iproc_msi_read_reg(msi, IPROC_MSI_INTS_EN, eq);
433 iproc_msi_write_reg(msi, IPROC_MSI_INTS_EN, eq, val);
436 val = iproc_msi_read_reg(msi, IPROC_MSI_CTRL, eq);
437 val &= ~(IPROC_MSI_INTR_EN | IPROC_MSI_INT_N_EVENT |
439 iproc_msi_write_reg(msi, IPROC_MSI_CTRL, eq, val);
443 static int iproc_msi_alloc_domains(struct device_node *node,
444 struct iproc_msi *msi)
446 msi->inner_domain = irq_domain_add_linear(NULL, msi->nr_msi_vecs,
447 &msi_domain_ops, msi);
448 if (!msi->inner_domain)
451 msi->msi_domain = pci_msi_create_irq_domain(of_node_to_fwnode(node),
452 &iproc_msi_domain_info,
454 if (!msi->msi_domain) {
455 irq_domain_remove(msi->inner_domain);
462 static void iproc_msi_free_domains(struct iproc_msi *msi)
465 irq_domain_remove(msi->msi_domain);
467 if (msi->inner_domain)
468 irq_domain_remove(msi->inner_domain);
471 static void iproc_msi_irq_free(struct iproc_msi *msi, unsigned int cpu)
475 for (i = cpu; i < msi->nr_irqs; i += msi->nr_cpus) {
476 irq_set_chained_handler_and_data(msi->grps[i].gic_irq,
481 static int iproc_msi_irq_setup(struct iproc_msi *msi, unsigned int cpu)
485 struct iproc_pcie *pcie = msi->pcie;
487 for (i = cpu; i < msi->nr_irqs; i += msi->nr_cpus) {
488 irq_set_chained_handler_and_data(msi->grps[i].gic_irq,
491 /* Dedicate GIC interrupt to each CPU core */
492 if (alloc_cpumask_var(&mask, GFP_KERNEL)) {
494 cpumask_set_cpu(cpu, mask);
495 ret = irq_set_affinity(msi->grps[i].gic_irq, mask);
498 "failed to set affinity for IRQ%d\n",
499 msi->grps[i].gic_irq);
500 free_cpumask_var(mask);
502 dev_err(pcie->dev, "failed to alloc CPU mask\n");
507 /* Free all configured/unconfigured IRQs */
508 iproc_msi_irq_free(msi, cpu);
516 int iproc_msi_init(struct iproc_pcie *pcie, struct device_node *node)
518 struct iproc_msi *msi;
522 if (!of_device_is_compatible(node, "brcm,iproc-msi"))
525 if (!of_find_property(node, "msi-controller", NULL))
531 msi = devm_kzalloc(pcie->dev, sizeof(*msi), GFP_KERNEL);
537 msi->msi_addr = pcie->base_addr;
538 mutex_init(&msi->bitmap_lock);
539 msi->nr_cpus = num_possible_cpus();
541 msi->nr_irqs = of_irq_count(node);
543 dev_err(pcie->dev, "found no MSI GIC interrupt\n");
547 if (msi->nr_irqs > NR_HW_IRQS) {
548 dev_warn(pcie->dev, "too many MSI GIC interrupts defined %d\n",
550 msi->nr_irqs = NR_HW_IRQS;
553 if (msi->nr_irqs < msi->nr_cpus) {
555 "not enough GIC interrupts for MSI affinity\n");
559 if (msi->nr_irqs % msi->nr_cpus != 0) {
560 msi->nr_irqs -= msi->nr_irqs % msi->nr_cpus;
561 dev_warn(pcie->dev, "Reducing number of interrupts to %d\n",
565 switch (pcie->type) {
566 case IPROC_PCIE_PAXB:
567 msi->reg_offsets = iproc_msi_reg_paxb;
568 msi->nr_eq_region = 1;
569 msi->nr_msi_region = 1;
571 case IPROC_PCIE_PAXC:
572 msi->reg_offsets = iproc_msi_reg_paxc;
573 msi->nr_eq_region = msi->nr_irqs;
574 msi->nr_msi_region = msi->nr_irqs;
577 dev_err(pcie->dev, "incompatible iProc PCIe interface\n");
581 if (of_find_property(node, "brcm,pcie-msi-inten", NULL))
582 msi->has_inten_reg = true;
584 msi->nr_msi_vecs = msi->nr_irqs * EQ_LEN;
585 msi->bitmap = devm_kcalloc(pcie->dev, BITS_TO_LONGS(msi->nr_msi_vecs),
586 sizeof(*msi->bitmap), GFP_KERNEL);
590 msi->grps = devm_kcalloc(pcie->dev, msi->nr_irqs, sizeof(*msi->grps),
595 for (i = 0; i < msi->nr_irqs; i++) {
596 unsigned int irq = irq_of_parse_and_map(node, i);
599 dev_err(pcie->dev, "unable to parse/map interrupt\n");
603 msi->grps[i].gic_irq = irq;
604 msi->grps[i].msi = msi;
608 /* Reserve memory for event queue and make sure memories are zeroed */
609 msi->eq_cpu = dma_zalloc_coherent(pcie->dev,
610 msi->nr_eq_region * EQ_MEM_REGION_SIZE,
611 &msi->eq_dma, GFP_KERNEL);
617 ret = iproc_msi_alloc_domains(node, msi);
619 dev_err(pcie->dev, "failed to create MSI domains\n");
623 for_each_online_cpu(cpu) {
624 ret = iproc_msi_irq_setup(msi, cpu);
629 iproc_msi_enable(msi);
634 for_each_online_cpu(cpu)
635 iproc_msi_irq_free(msi, cpu);
636 iproc_msi_free_domains(msi);
639 dma_free_coherent(pcie->dev, msi->nr_eq_region * EQ_MEM_REGION_SIZE,
640 msi->eq_cpu, msi->eq_dma);
643 for (i = 0; i < msi->nr_irqs; i++) {
644 if (msi->grps[i].gic_irq)
645 irq_dispose_mapping(msi->grps[i].gic_irq);
650 EXPORT_SYMBOL(iproc_msi_init);
652 void iproc_msi_exit(struct iproc_pcie *pcie)
654 struct iproc_msi *msi = pcie->msi;
660 iproc_msi_disable(msi);
662 for_each_online_cpu(cpu)
663 iproc_msi_irq_free(msi, cpu);
665 iproc_msi_free_domains(msi);
667 dma_free_coherent(pcie->dev, msi->nr_eq_region * EQ_MEM_REGION_SIZE,
668 msi->eq_cpu, msi->eq_dma);
670 for (i = 0; i < msi->nr_irqs; i++) {
671 if (msi->grps[i].gic_irq)
672 irq_dispose_mapping(msi->grps[i].gic_irq);
675 EXPORT_SYMBOL(iproc_msi_exit);