1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2015 Broadcom Corporation
6 #include <linux/interrupt.h>
7 #include <linux/irqchip/chained_irq.h>
8 #include <linux/irqdomain.h>
10 #include <linux/of_irq.h>
11 #include <linux/of_pci.h>
12 #include <linux/pci.h>
14 #include "pcie-iproc.h"
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)
23 #define IPROC_MSI_EQ_MASK 0x3f
25 /* Max number of GIC interrupts */
28 /* Number of entries in each event queue */
31 /* Size of each event queue memory region */
32 #define EQ_MEM_REGION_SIZE SZ_4K
34 /* Size of each MSI address region */
35 #define MSI_MEM_REGION_SIZE SZ_4K
38 IPROC_MSI_EQ_PAGE = 0,
39 IPROC_MSI_EQ_PAGE_UPPER,
54 * One MSI group is allocated per GIC interrupt, serviced by one iProc MSI
57 * @msi: pointer to iProc MSI data
58 * @gic_irq: GIC interrupt
59 * @eq: Event queue number
61 struct iproc_msi_grp {
62 struct iproc_msi *msi;
68 * iProc event queue based MSI
70 * Only meant to be used on platforms without MSI support integrated into the
73 * @pcie: pointer to iProc PCIe data
74 * @reg_offsets: MSI register offsets
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
87 * @nr_msi_region: required number of 4K aligned address region for MSI posted
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
94 struct iproc_pcie *pcie;
95 const u16 (*reg_offsets)[IPROC_MSI_REG_SIZE];
96 struct iproc_msi_grp *grps;
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;
109 phys_addr_t msi_addr;
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 },
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 },
128 static inline u32 iproc_msi_read_reg(struct iproc_msi *msi,
129 enum iproc_msi_reg reg,
132 struct iproc_pcie *pcie = msi->pcie;
134 return readl_relaxed(pcie->base + msi->reg_offsets[eq][reg]);
137 static inline void iproc_msi_write_reg(struct iproc_msi *msi,
138 enum iproc_msi_reg reg,
141 struct iproc_pcie *pcie = msi->pcie;
143 writel_relaxed(val, pcie->base + msi->reg_offsets[eq][reg]);
146 static inline u32 hwirq_to_group(struct iproc_msi *msi, unsigned long hwirq)
148 return (hwirq % msi->nr_irqs);
151 static inline unsigned int iproc_msi_addr_offset(struct iproc_msi *msi,
154 if (msi->nr_msi_region > 1)
155 return hwirq_to_group(msi, hwirq) * MSI_MEM_REGION_SIZE;
157 return hwirq_to_group(msi, hwirq) * sizeof(u32);
160 static inline unsigned int iproc_msi_eq_offset(struct iproc_msi *msi, u32 eq)
162 if (msi->nr_eq_region > 1)
163 return eq * EQ_MEM_REGION_SIZE;
165 return eq * EQ_LEN * sizeof(u32);
168 static struct irq_chip iproc_msi_irq_chip = {
172 static struct msi_domain_info iproc_msi_domain_info = {
173 .flags = MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
175 .chip = &iproc_msi_irq_chip,
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.
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.
188 * - the number of MSI groups is M
189 * - the number of CPU cores is N
190 * - M is always a multiple of N
192 * Total number of raw MSI vectors = M * 64
193 * Total number of supported MSI vectors = (M * 64) / N
195 static inline int hwirq_to_cpu(struct iproc_msi *msi, unsigned long hwirq)
197 return (hwirq % msi->nr_cpus);
200 static inline unsigned long hwirq_to_canonical_hwirq(struct iproc_msi *msi,
203 return (hwirq - hwirq_to_cpu(msi, hwirq));
206 static int iproc_msi_irq_set_affinity(struct irq_data *data,
207 const struct cpumask *mask, bool force)
209 struct iproc_msi *msi = irq_data_get_irq_chip_data(data);
210 int target_cpu = cpumask_first(mask);
214 curr_cpu = hwirq_to_cpu(msi, data->hwirq);
215 if (curr_cpu == target_cpu)
216 ret = IRQ_SET_MASK_OK_DONE;
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;
223 irq_data_update_effective_affinity(data, cpumask_of(target_cpu));
228 static void iproc_msi_irq_compose_msi_msg(struct irq_data *data,
231 struct iproc_msi *msi = irq_data_get_irq_chip_data(data);
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;
240 static struct irq_chip iproc_msi_bottom_irq_chip = {
242 .irq_set_affinity = iproc_msi_irq_set_affinity,
243 .irq_compose_msi_msg = iproc_msi_irq_compose_msi_msg,
246 static int iproc_msi_irq_domain_alloc(struct irq_domain *domain,
247 unsigned int virq, unsigned int nr_irqs,
250 struct iproc_msi *msi = domain->host_data;
253 if (msi->nr_cpus > 1 && nr_irqs > 1)
256 mutex_lock(&msi->bitmap_lock);
259 * Allocate 'nr_irqs' multiplied by 'nr_cpus' number of MSI vectors
262 hwirq = bitmap_find_free_region(msi->bitmap, msi->nr_msi_vecs,
263 order_base_2(msi->nr_cpus * nr_irqs));
265 mutex_unlock(&msi->bitmap_lock);
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,
280 static void iproc_msi_irq_domain_free(struct irq_domain *domain,
281 unsigned int virq, unsigned int nr_irqs)
283 struct irq_data *data = irq_domain_get_irq_data(domain, virq);
284 struct iproc_msi *msi = irq_data_get_irq_chip_data(data);
287 mutex_lock(&msi->bitmap_lock);
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));
293 mutex_unlock(&msi->bitmap_lock);
295 irq_domain_free_irqs_parent(domain, virq, nr_irqs);
298 static const struct irq_domain_ops msi_domain_ops = {
299 .alloc = iproc_msi_irq_domain_alloc,
300 .free = iproc_msi_irq_domain_free,
303 static inline u32 decode_msi_hwirq(struct iproc_msi *msi, u32 eq, u32 head)
309 offs = iproc_msi_eq_offset(msi, eq) + head * sizeof(u32);
310 msg = (u32 __iomem *)(msi->eq_cpu + offs);
312 hwirq = (hwirq >> 5) + (hwirq & 0x1f);
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.
319 return hwirq_to_canonical_hwirq(msi, hwirq);
322 static void iproc_msi_handler(struct irq_desc *desc)
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;
331 chained_irq_enter(chip, desc);
333 grp = irq_desc_get_handler_data(desc);
338 * iProc MSI event queue is tracked by head and tail pointers. Head
339 * pointer indicates the next entry (MSI data) to be consumed by SW in
340 * the queue and needs to be updated by SW. iProc MSI core uses the
341 * tail pointer as the next data insertion point.
343 * Entries between head and tail pointers contain valid MSI data. MSI
344 * data is guaranteed to be in the event queue memory before the tail
345 * pointer is updated by the iProc MSI core.
347 head = iproc_msi_read_reg(msi, IPROC_MSI_EQ_HEAD,
348 eq) & IPROC_MSI_EQ_MASK;
350 tail = iproc_msi_read_reg(msi, IPROC_MSI_EQ_TAIL,
351 eq) & IPROC_MSI_EQ_MASK;
354 * Figure out total number of events (MSI data) to be
357 nr_events = (tail < head) ?
358 (EQ_LEN - (head - tail)) : (tail - head);
362 /* process all outstanding events */
363 while (nr_events--) {
364 hwirq = decode_msi_hwirq(msi, eq, head);
365 virq = irq_find_mapping(msi->inner_domain, hwirq);
366 generic_handle_irq(virq);
373 * Now all outstanding events have been processed. Update the
376 iproc_msi_write_reg(msi, IPROC_MSI_EQ_HEAD, eq, head);
379 * Now go read the tail pointer again to see if there are new
380 * outstanding events that came in during the above window.
384 chained_irq_exit(chip, desc);
387 static void iproc_msi_enable(struct iproc_msi *msi)
392 /* Program memory region for each event queue */
393 for (i = 0; i < msi->nr_eq_region; i++) {
394 dma_addr_t addr = msi->eq_dma + (i * EQ_MEM_REGION_SIZE);
396 iproc_msi_write_reg(msi, IPROC_MSI_EQ_PAGE, i,
397 lower_32_bits(addr));
398 iproc_msi_write_reg(msi, IPROC_MSI_EQ_PAGE_UPPER, i,
399 upper_32_bits(addr));
402 /* Program address region for MSI posted writes */
403 for (i = 0; i < msi->nr_msi_region; i++) {
404 phys_addr_t addr = msi->msi_addr + (i * MSI_MEM_REGION_SIZE);
406 iproc_msi_write_reg(msi, IPROC_MSI_PAGE, i,
407 lower_32_bits(addr));
408 iproc_msi_write_reg(msi, IPROC_MSI_PAGE_UPPER, i,
409 upper_32_bits(addr));
412 for (eq = 0; eq < msi->nr_irqs; eq++) {
413 /* Enable MSI event queue */
414 val = IPROC_MSI_INTR_EN | IPROC_MSI_INT_N_EVENT |
416 iproc_msi_write_reg(msi, IPROC_MSI_CTRL, eq, val);
419 * Some legacy platforms require the MSI interrupt enable
420 * register to be set explicitly.
422 if (msi->has_inten_reg) {
423 val = iproc_msi_read_reg(msi, IPROC_MSI_INTS_EN, eq);
425 iproc_msi_write_reg(msi, IPROC_MSI_INTS_EN, eq, val);
430 static void iproc_msi_disable(struct iproc_msi *msi)
434 for (eq = 0; eq < msi->nr_irqs; eq++) {
435 if (msi->has_inten_reg) {
436 val = iproc_msi_read_reg(msi, IPROC_MSI_INTS_EN, eq);
438 iproc_msi_write_reg(msi, IPROC_MSI_INTS_EN, eq, val);
441 val = iproc_msi_read_reg(msi, IPROC_MSI_CTRL, eq);
442 val &= ~(IPROC_MSI_INTR_EN | IPROC_MSI_INT_N_EVENT |
444 iproc_msi_write_reg(msi, IPROC_MSI_CTRL, eq, val);
448 static int iproc_msi_alloc_domains(struct device_node *node,
449 struct iproc_msi *msi)
451 msi->inner_domain = irq_domain_add_linear(NULL, msi->nr_msi_vecs,
452 &msi_domain_ops, msi);
453 if (!msi->inner_domain)
456 msi->msi_domain = pci_msi_create_irq_domain(of_node_to_fwnode(node),
457 &iproc_msi_domain_info,
459 if (!msi->msi_domain) {
460 irq_domain_remove(msi->inner_domain);
467 static void iproc_msi_free_domains(struct iproc_msi *msi)
470 irq_domain_remove(msi->msi_domain);
472 if (msi->inner_domain)
473 irq_domain_remove(msi->inner_domain);
476 static void iproc_msi_irq_free(struct iproc_msi *msi, unsigned int cpu)
480 for (i = cpu; i < msi->nr_irqs; i += msi->nr_cpus) {
481 irq_set_chained_handler_and_data(msi->grps[i].gic_irq,
486 static int iproc_msi_irq_setup(struct iproc_msi *msi, unsigned int cpu)
490 struct iproc_pcie *pcie = msi->pcie;
492 for (i = cpu; i < msi->nr_irqs; i += msi->nr_cpus) {
493 irq_set_chained_handler_and_data(msi->grps[i].gic_irq,
496 /* Dedicate GIC interrupt to each CPU core */
497 if (alloc_cpumask_var(&mask, GFP_KERNEL)) {
499 cpumask_set_cpu(cpu, mask);
500 ret = irq_set_affinity(msi->grps[i].gic_irq, mask);
503 "failed to set affinity for IRQ%d\n",
504 msi->grps[i].gic_irq);
505 free_cpumask_var(mask);
507 dev_err(pcie->dev, "failed to alloc CPU mask\n");
512 /* Free all configured/unconfigured IRQs */
513 iproc_msi_irq_free(msi, cpu);
521 int iproc_msi_init(struct iproc_pcie *pcie, struct device_node *node)
523 struct iproc_msi *msi;
527 if (!of_device_is_compatible(node, "brcm,iproc-msi"))
530 if (!of_find_property(node, "msi-controller", NULL))
536 msi = devm_kzalloc(pcie->dev, sizeof(*msi), GFP_KERNEL);
542 msi->msi_addr = pcie->base_addr;
543 mutex_init(&msi->bitmap_lock);
544 msi->nr_cpus = num_possible_cpus();
546 if (msi->nr_cpus == 1)
547 iproc_msi_domain_info.flags |= MSI_FLAG_MULTI_PCI_MSI;
549 msi->nr_irqs = of_irq_count(node);
551 dev_err(pcie->dev, "found no MSI GIC interrupt\n");
555 if (msi->nr_irqs > NR_HW_IRQS) {
556 dev_warn(pcie->dev, "too many MSI GIC interrupts defined %d\n",
558 msi->nr_irqs = NR_HW_IRQS;
561 if (msi->nr_irqs < msi->nr_cpus) {
563 "not enough GIC interrupts for MSI affinity\n");
567 if (msi->nr_irqs % msi->nr_cpus != 0) {
568 msi->nr_irqs -= msi->nr_irqs % msi->nr_cpus;
569 dev_warn(pcie->dev, "Reducing number of interrupts to %d\n",
573 switch (pcie->type) {
574 case IPROC_PCIE_PAXB_BCMA:
575 case IPROC_PCIE_PAXB:
576 msi->reg_offsets = iproc_msi_reg_paxb;
577 msi->nr_eq_region = 1;
578 msi->nr_msi_region = 1;
580 case IPROC_PCIE_PAXC:
581 msi->reg_offsets = iproc_msi_reg_paxc;
582 msi->nr_eq_region = msi->nr_irqs;
583 msi->nr_msi_region = msi->nr_irqs;
586 dev_err(pcie->dev, "incompatible iProc PCIe interface\n");
590 if (of_find_property(node, "brcm,pcie-msi-inten", NULL))
591 msi->has_inten_reg = true;
593 msi->nr_msi_vecs = msi->nr_irqs * EQ_LEN;
594 msi->bitmap = devm_kcalloc(pcie->dev, BITS_TO_LONGS(msi->nr_msi_vecs),
595 sizeof(*msi->bitmap), GFP_KERNEL);
599 msi->grps = devm_kcalloc(pcie->dev, msi->nr_irqs, sizeof(*msi->grps),
604 for (i = 0; i < msi->nr_irqs; i++) {
605 unsigned int irq = irq_of_parse_and_map(node, i);
608 dev_err(pcie->dev, "unable to parse/map interrupt\n");
612 msi->grps[i].gic_irq = irq;
613 msi->grps[i].msi = msi;
617 /* Reserve memory for event queue and make sure memories are zeroed */
618 msi->eq_cpu = dma_alloc_coherent(pcie->dev,
619 msi->nr_eq_region * EQ_MEM_REGION_SIZE,
620 &msi->eq_dma, GFP_KERNEL);
626 ret = iproc_msi_alloc_domains(node, msi);
628 dev_err(pcie->dev, "failed to create MSI domains\n");
632 for_each_online_cpu(cpu) {
633 ret = iproc_msi_irq_setup(msi, cpu);
638 iproc_msi_enable(msi);
643 for_each_online_cpu(cpu)
644 iproc_msi_irq_free(msi, cpu);
645 iproc_msi_free_domains(msi);
648 dma_free_coherent(pcie->dev, msi->nr_eq_region * EQ_MEM_REGION_SIZE,
649 msi->eq_cpu, msi->eq_dma);
652 for (i = 0; i < msi->nr_irqs; i++) {
653 if (msi->grps[i].gic_irq)
654 irq_dispose_mapping(msi->grps[i].gic_irq);
659 EXPORT_SYMBOL(iproc_msi_init);
661 void iproc_msi_exit(struct iproc_pcie *pcie)
663 struct iproc_msi *msi = pcie->msi;
669 iproc_msi_disable(msi);
671 for_each_online_cpu(cpu)
672 iproc_msi_irq_free(msi, cpu);
674 iproc_msi_free_domains(msi);
676 dma_free_coherent(pcie->dev, msi->nr_eq_region * EQ_MEM_REGION_SIZE,
677 msi->eq_cpu, msi->eq_dma);
679 for (i = 0; i < msi->nr_irqs; i++) {
680 if (msi->grps[i].gic_irq)
681 irq_dispose_mapping(msi->grps[i].gic_irq);
684 EXPORT_SYMBOL(iproc_msi_exit);