x86/MSI: Use hierarchical irqdomains to manage MSI interrupts
[firefly-linux-kernel-4.4.55.git] / drivers / iommu / irq_remapping.c
1 #include <linux/seq_file.h>
2 #include <linux/cpumask.h>
3 #include <linux/kernel.h>
4 #include <linux/string.h>
5 #include <linux/errno.h>
6 #include <linux/msi.h>
7 #include <linux/irq.h>
8 #include <linux/pci.h>
9 #include <linux/irqdomain.h>
10
11 #include <asm/hw_irq.h>
12 #include <asm/irq_remapping.h>
13 #include <asm/processor.h>
14 #include <asm/x86_init.h>
15 #include <asm/apic.h>
16 #include <asm/hpet.h>
17
18 #include "irq_remapping.h"
19
20 int irq_remapping_enabled;
21 int irq_remap_broken;
22 int disable_sourceid_checking;
23 int no_x2apic_optout;
24
25 static int disable_irq_remap;
26 static struct irq_remap_ops *remap_ops;
27
28 static int msi_alloc_remapped_irq(struct pci_dev *pdev, int irq, int nvec);
29 static int msi_setup_remapped_irq(struct pci_dev *pdev, unsigned int irq,
30                                   int index, int sub_handle);
31 static int set_remapped_irq_affinity(struct irq_data *data,
32                                      const struct cpumask *mask,
33                                      bool force);
34
35 static bool irq_remapped(struct irq_cfg *cfg)
36 {
37         return (cfg->remapped == 1);
38 }
39
40 static void irq_remapping_disable_io_apic(void)
41 {
42         /*
43          * With interrupt-remapping, for now we will use virtual wire A
44          * mode, as virtual wire B is little complex (need to configure
45          * both IOAPIC RTE as well as interrupt-remapping table entry).
46          * As this gets called during crash dump, keep this simple for
47          * now.
48          */
49         if (cpu_has_apic || apic_from_smp_config())
50                 disconnect_bsp_APIC(0);
51 }
52
53 #ifndef CONFIG_GENERIC_IRQ_LEGACY_ALLOC_HWIRQ
54 static unsigned int irq_alloc_hwirqs(int cnt, int node)
55 {
56         return irq_domain_alloc_irqs(NULL, -1, cnt, node, NULL);
57 }
58
59 static void irq_free_hwirqs(unsigned int from, int cnt)
60 {
61         irq_domain_free_irqs(from, cnt);
62 }
63 #endif
64
65 static int do_setup_msi_irqs(struct pci_dev *dev, int nvec)
66 {
67         int ret, sub_handle, nvec_pow2, index = 0;
68         unsigned int irq;
69         struct msi_desc *msidesc;
70
71         msidesc = list_entry(dev->msi_list.next, struct msi_desc, list);
72
73         irq = irq_alloc_hwirqs(nvec, dev_to_node(&dev->dev));
74         if (irq == 0)
75                 return -ENOSPC;
76
77         nvec_pow2 = __roundup_pow_of_two(nvec);
78         for (sub_handle = 0; sub_handle < nvec; sub_handle++) {
79                 if (!sub_handle) {
80                         index = msi_alloc_remapped_irq(dev, irq, nvec_pow2);
81                         if (index < 0) {
82                                 ret = index;
83                                 goto error;
84                         }
85                 } else {
86                         ret = msi_setup_remapped_irq(dev, irq + sub_handle,
87                                                      index, sub_handle);
88                         if (ret < 0)
89                                 goto error;
90                 }
91                 ret = setup_msi_irq(dev, msidesc, irq, sub_handle);
92                 if (ret < 0)
93                         goto error;
94         }
95         return 0;
96
97 error:
98         irq_free_hwirqs(irq, nvec);
99
100         /*
101          * Restore altered MSI descriptor fields and prevent just destroyed
102          * IRQs from tearing down again in default_teardown_msi_irqs()
103          */
104         msidesc->irq = 0;
105
106         return ret;
107 }
108
109 static int do_setup_msix_irqs(struct pci_dev *dev, int nvec)
110 {
111         int node, ret, sub_handle, index = 0;
112         struct msi_desc *msidesc;
113         unsigned int irq;
114
115         node            = dev_to_node(&dev->dev);
116         sub_handle      = 0;
117
118         list_for_each_entry(msidesc, &dev->msi_list, list) {
119
120                 irq = irq_alloc_hwirqs(1, node);
121                 if (irq == 0)
122                         return -1;
123
124                 if (sub_handle == 0)
125                         ret = index = msi_alloc_remapped_irq(dev, irq, nvec);
126                 else
127                         ret = msi_setup_remapped_irq(dev, irq, index, sub_handle);
128
129                 if (ret < 0)
130                         goto error;
131
132                 ret = setup_msi_irq(dev, msidesc, irq, 0);
133                 if (ret < 0)
134                         goto error;
135
136                 sub_handle += 1;
137                 irq        += 1;
138         }
139
140         return 0;
141
142 error:
143         irq_free_hwirqs(irq, 1);
144         return ret;
145 }
146
147 static int irq_remapping_setup_msi_irqs(struct pci_dev *dev,
148                                         int nvec, int type)
149 {
150         if (type == PCI_CAP_ID_MSI)
151                 return do_setup_msi_irqs(dev, nvec);
152         else
153                 return do_setup_msix_irqs(dev, nvec);
154 }
155
156 static void eoi_ioapic_pin_remapped(int apic, int pin, int vector)
157 {
158         /*
159          * Intr-remapping uses pin number as the virtual vector
160          * in the RTE. Actual vector is programmed in
161          * intr-remapping table entry. Hence for the io-apic
162          * EOI we use the pin number.
163          */
164         io_apic_eoi(apic, pin);
165 }
166
167 static void __init irq_remapping_modify_x86_ops(void)
168 {
169         x86_io_apic_ops.disable         = irq_remapping_disable_io_apic;
170         x86_io_apic_ops.set_affinity    = set_remapped_irq_affinity;
171         x86_io_apic_ops.setup_entry     = setup_ioapic_remapped_entry;
172         x86_io_apic_ops.eoi_ioapic_pin  = eoi_ioapic_pin_remapped;
173         x86_msi.setup_hpet_msi          = setup_hpet_msi_remapped;
174         x86_msi.compose_msi_msg         = compose_remapped_msi_msg;
175 }
176
177 static __init int setup_nointremap(char *str)
178 {
179         disable_irq_remap = 1;
180         return 0;
181 }
182 early_param("nointremap", setup_nointremap);
183
184 static __init int setup_irqremap(char *str)
185 {
186         if (!str)
187                 return -EINVAL;
188
189         while (*str) {
190                 if (!strncmp(str, "on", 2))
191                         disable_irq_remap = 0;
192                 else if (!strncmp(str, "off", 3))
193                         disable_irq_remap = 1;
194                 else if (!strncmp(str, "nosid", 5))
195                         disable_sourceid_checking = 1;
196                 else if (!strncmp(str, "no_x2apic_optout", 16))
197                         no_x2apic_optout = 1;
198
199                 str += strcspn(str, ",");
200                 while (*str == ',')
201                         str++;
202         }
203
204         return 0;
205 }
206 early_param("intremap", setup_irqremap);
207
208 void set_irq_remapping_broken(void)
209 {
210         irq_remap_broken = 1;
211 }
212
213 int __init irq_remapping_prepare(void)
214 {
215         if (disable_irq_remap)
216                 return -ENOSYS;
217
218         if (intel_irq_remap_ops.prepare() == 0)
219                 remap_ops = &intel_irq_remap_ops;
220         else if (IS_ENABLED(CONFIG_AMD_IOMMU) &&
221                  amd_iommu_irq_ops.prepare() == 0)
222                 remap_ops = &amd_iommu_irq_ops;
223         else
224                 return -ENOSYS;
225
226         return 0;
227 }
228
229 int __init irq_remapping_enable(void)
230 {
231         int ret;
232
233         if (!remap_ops->enable)
234                 return -ENODEV;
235
236         ret = remap_ops->enable();
237
238         if (irq_remapping_enabled)
239                 irq_remapping_modify_x86_ops();
240
241         return ret;
242 }
243
244 void irq_remapping_disable(void)
245 {
246         if (irq_remapping_enabled && remap_ops->disable)
247                 remap_ops->disable();
248 }
249
250 int irq_remapping_reenable(int mode)
251 {
252         if (irq_remapping_enabled && remap_ops->reenable)
253                 return remap_ops->reenable(mode);
254
255         return 0;
256 }
257
258 int __init irq_remap_enable_fault_handling(void)
259 {
260         if (!irq_remapping_enabled)
261                 return 0;
262
263         if (!remap_ops->enable_faulting)
264                 return -ENODEV;
265
266         return remap_ops->enable_faulting();
267 }
268
269 int setup_ioapic_remapped_entry(int irq,
270                                 struct IO_APIC_route_entry *entry,
271                                 unsigned int destination, int vector,
272                                 struct io_apic_irq_attr *attr)
273 {
274         if (!remap_ops->setup_ioapic_entry)
275                 return -ENODEV;
276
277         return remap_ops->setup_ioapic_entry(irq, entry, destination,
278                                              vector, attr);
279 }
280
281 static int set_remapped_irq_affinity(struct irq_data *data,
282                                      const struct cpumask *mask, bool force)
283 {
284         if (!config_enabled(CONFIG_SMP) || !remap_ops->set_affinity)
285                 return 0;
286
287         return remap_ops->set_affinity(data, mask, force);
288 }
289
290 void free_remapped_irq(int irq)
291 {
292         struct irq_cfg *cfg = irq_cfg(irq);
293
294         if (irq_remapped(cfg) && remap_ops->free_irq)
295                 remap_ops->free_irq(irq);
296 }
297
298 void compose_remapped_msi_msg(struct pci_dev *pdev,
299                               unsigned int irq, unsigned int dest,
300                               struct msi_msg *msg, u8 hpet_id)
301 {
302         struct irq_cfg *cfg = irq_cfg(irq);
303
304         if (!irq_remapped(cfg))
305                 native_compose_msi_msg(pdev, irq, dest, msg, hpet_id);
306         else if (remap_ops->compose_msi_msg)
307                 remap_ops->compose_msi_msg(pdev, irq, dest, msg, hpet_id);
308 }
309
310 static int msi_alloc_remapped_irq(struct pci_dev *pdev, int irq, int nvec)
311 {
312         if (!remap_ops->msi_alloc_irq)
313                 return -ENODEV;
314
315         return remap_ops->msi_alloc_irq(pdev, irq, nvec);
316 }
317
318 static int msi_setup_remapped_irq(struct pci_dev *pdev, unsigned int irq,
319                                   int index, int sub_handle)
320 {
321         if (!remap_ops->msi_setup_irq)
322                 return -ENODEV;
323
324         return remap_ops->msi_setup_irq(pdev, irq, index, sub_handle);
325 }
326
327 int setup_hpet_msi_remapped(unsigned int irq, unsigned int id)
328 {
329         int ret;
330
331         if (!remap_ops->alloc_hpet_msi)
332                 return -ENODEV;
333
334         ret = remap_ops->alloc_hpet_msi(irq, id);
335         if (ret)
336                 return -EINVAL;
337
338         return default_setup_hpet_msi(irq, id);
339 }
340
341 void panic_if_irq_remap(const char *msg)
342 {
343         if (irq_remapping_enabled)
344                 panic(msg);
345 }
346
347 void ir_ack_apic_edge(struct irq_data *data)
348 {
349         ack_APIC_irq();
350 }
351
352 static void ir_ack_apic_level(struct irq_data *data)
353 {
354         ack_APIC_irq();
355         eoi_ioapic_irq(data->irq, irqd_cfg(data));
356 }
357
358 void irq_remapping_print_chip(struct irq_data *data, struct seq_file *p)
359 {
360         /*
361          * Assume interrupt is remapped if the parent irqdomain isn't the
362          * vector domain, which is true for MSI, HPET and IOAPIC on x86
363          * platforms.
364          */
365         if (data->domain && data->domain->parent != arch_get_ir_parent_domain())
366                 seq_printf(p, " IR-%s", data->chip->name);
367         else
368                 seq_printf(p, " %s", data->chip->name);
369 }
370
371 static void ir_print_prefix(struct irq_data *data, struct seq_file *p)
372 {
373         seq_printf(p, " IR-%s", data->chip->name);
374 }
375
376 void irq_remap_modify_chip_defaults(struct irq_chip *chip)
377 {
378         chip->irq_print_chip = ir_print_prefix;
379         chip->irq_ack = ir_ack_apic_edge;
380         chip->irq_eoi = ir_ack_apic_level;
381         chip->irq_set_affinity = x86_io_apic_ops.set_affinity;
382 }
383
384 bool setup_remapped_irq(int irq, struct irq_cfg *cfg, struct irq_chip *chip)
385 {
386         if (!irq_remapped(cfg))
387                 return false;
388         irq_set_status_flags(irq, IRQ_MOVE_PCNTXT);
389         irq_remap_modify_chip_defaults(chip);
390         return true;
391 }
392
393 /**
394  * irq_remapping_get_ir_irq_domain - Get the irqdomain associated with the IOMMU
395  *                                   device serving request @info
396  * @info: interrupt allocation information, used to identify the IOMMU device
397  *
398  * It's used to get parent irqdomain for HPET and IOAPIC irqdomains.
399  * Returns pointer to IRQ domain, or NULL on failure.
400  */
401 struct irq_domain *
402 irq_remapping_get_ir_irq_domain(struct irq_alloc_info *info)
403 {
404         if (!remap_ops || !remap_ops->get_ir_irq_domain)
405                 return NULL;
406
407         return remap_ops->get_ir_irq_domain(info);
408 }
409
410 /**
411  * irq_remapping_get_irq_domain - Get the irqdomain serving the request @info
412  * @info: interrupt allocation information, used to identify the IOMMU device
413  *
414  * There will be one PCI MSI/MSIX irqdomain associated with each interrupt
415  * remapping device, so this interface is used to retrieve the PCI MSI/MSIX
416  * irqdomain serving request @info.
417  * Returns pointer to IRQ domain, or NULL on failure.
418  */
419 struct irq_domain *
420 irq_remapping_get_irq_domain(struct irq_alloc_info *info)
421 {
422         if (!remap_ops || !remap_ops->get_irq_domain)
423                 return NULL;
424
425         return remap_ops->get_irq_domain(info);
426 }