xref: /openbmc/qemu/hw/pci/msix.c (revision b969526a)
1 /*
2  * MSI-X device support
3  *
4  * This module includes support for MSI-X in pci devices.
5  *
6  * Author: Michael S. Tsirkin <mst@redhat.com>
7  *
8  *  Copyright (c) 2009, Red Hat Inc, Michael S. Tsirkin (mst@redhat.com)
9  *
10  * This work is licensed under the terms of the GNU GPL, version 2.  See
11  * the COPYING file in the top-level directory.
12  *
13  * Contributions after 2012-01-13 are licensed under the terms of the
14  * GNU GPL, version 2 or (at your option) any later version.
15  */
16 
17 #include "hw/hw.h"
18 #include "hw/pci/msi.h"
19 #include "hw/pci/msix.h"
20 #include "hw/pci/pci.h"
21 #include "qemu/range.h"
22 
23 #define MSIX_CAP_LENGTH 12
24 
25 /* MSI enable bit and maskall bit are in byte 1 in FLAGS register */
26 #define MSIX_CONTROL_OFFSET (PCI_MSIX_FLAGS + 1)
27 #define MSIX_ENABLE_MASK (PCI_MSIX_FLAGS_ENABLE >> 8)
28 #define MSIX_MASKALL_MASK (PCI_MSIX_FLAGS_MASKALL >> 8)
29 
30 MSIMessage msix_get_message(PCIDevice *dev, unsigned vector)
31 {
32     uint8_t *table_entry = dev->msix_table + vector * PCI_MSIX_ENTRY_SIZE;
33     MSIMessage msg;
34 
35     msg.address = pci_get_quad(table_entry + PCI_MSIX_ENTRY_LOWER_ADDR);
36     msg.data = pci_get_long(table_entry + PCI_MSIX_ENTRY_DATA);
37     return msg;
38 }
39 
40 /*
41  * Special API for POWER to configure the vectors through
42  * a side channel. Should never be used by devices.
43  */
44 void msix_set_message(PCIDevice *dev, int vector, struct MSIMessage msg)
45 {
46     uint8_t *table_entry = dev->msix_table + vector * PCI_MSIX_ENTRY_SIZE;
47 
48     pci_set_quad(table_entry + PCI_MSIX_ENTRY_LOWER_ADDR, msg.address);
49     pci_set_long(table_entry + PCI_MSIX_ENTRY_DATA, msg.data);
50     table_entry[PCI_MSIX_ENTRY_VECTOR_CTRL] &= ~PCI_MSIX_ENTRY_CTRL_MASKBIT;
51 }
52 
53 static uint8_t msix_pending_mask(int vector)
54 {
55     return 1 << (vector % 8);
56 }
57 
58 static uint8_t *msix_pending_byte(PCIDevice *dev, int vector)
59 {
60     return dev->msix_pba + vector / 8;
61 }
62 
63 static int msix_is_pending(PCIDevice *dev, int vector)
64 {
65     return *msix_pending_byte(dev, vector) & msix_pending_mask(vector);
66 }
67 
68 void msix_set_pending(PCIDevice *dev, unsigned int vector)
69 {
70     *msix_pending_byte(dev, vector) |= msix_pending_mask(vector);
71 }
72 
73 static void msix_clr_pending(PCIDevice *dev, int vector)
74 {
75     *msix_pending_byte(dev, vector) &= ~msix_pending_mask(vector);
76 }
77 
78 static bool msix_vector_masked(PCIDevice *dev, unsigned int vector, bool fmask)
79 {
80     unsigned offset = vector * PCI_MSIX_ENTRY_SIZE + PCI_MSIX_ENTRY_VECTOR_CTRL;
81     return fmask || dev->msix_table[offset] & PCI_MSIX_ENTRY_CTRL_MASKBIT;
82 }
83 
84 bool msix_is_masked(PCIDevice *dev, unsigned int vector)
85 {
86     return msix_vector_masked(dev, vector, dev->msix_function_masked);
87 }
88 
89 static void msix_fire_vector_notifier(PCIDevice *dev,
90                                       unsigned int vector, bool is_masked)
91 {
92     MSIMessage msg;
93     int ret;
94 
95     if (!dev->msix_vector_use_notifier) {
96         return;
97     }
98     if (is_masked) {
99         dev->msix_vector_release_notifier(dev, vector);
100     } else {
101         msg = msix_get_message(dev, vector);
102         ret = dev->msix_vector_use_notifier(dev, vector, msg);
103         assert(ret >= 0);
104     }
105 }
106 
107 static void msix_handle_mask_update(PCIDevice *dev, int vector, bool was_masked)
108 {
109     bool is_masked = msix_is_masked(dev, vector);
110 
111     if (is_masked == was_masked) {
112         return;
113     }
114 
115     msix_fire_vector_notifier(dev, vector, is_masked);
116 
117     if (!is_masked && msix_is_pending(dev, vector)) {
118         msix_clr_pending(dev, vector);
119         msix_notify(dev, vector);
120     }
121 }
122 
123 static void msix_update_function_masked(PCIDevice *dev)
124 {
125     dev->msix_function_masked = !msix_enabled(dev) ||
126         (dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] & MSIX_MASKALL_MASK);
127 }
128 
129 /* Handle MSI-X capability config write. */
130 void msix_write_config(PCIDevice *dev, uint32_t addr,
131                        uint32_t val, int len)
132 {
133     unsigned enable_pos = dev->msix_cap + MSIX_CONTROL_OFFSET;
134     int vector;
135     bool was_masked;
136 
137     if (!msix_present(dev) || !range_covers_byte(addr, len, enable_pos)) {
138         return;
139     }
140 
141     was_masked = dev->msix_function_masked;
142     msix_update_function_masked(dev);
143 
144     if (!msix_enabled(dev)) {
145         return;
146     }
147 
148     pci_device_deassert_intx(dev);
149 
150     if (dev->msix_function_masked == was_masked) {
151         return;
152     }
153 
154     for (vector = 0; vector < dev->msix_entries_nr; ++vector) {
155         msix_handle_mask_update(dev, vector,
156                                 msix_vector_masked(dev, vector, was_masked));
157     }
158 }
159 
160 static uint64_t msix_table_mmio_read(void *opaque, hwaddr addr,
161                                      unsigned size)
162 {
163     PCIDevice *dev = opaque;
164 
165     return pci_get_long(dev->msix_table + addr);
166 }
167 
168 static void msix_table_mmio_write(void *opaque, hwaddr addr,
169                                   uint64_t val, unsigned size)
170 {
171     PCIDevice *dev = opaque;
172     int vector = addr / PCI_MSIX_ENTRY_SIZE;
173     bool was_masked;
174 
175     was_masked = msix_is_masked(dev, vector);
176     pci_set_long(dev->msix_table + addr, val);
177     msix_handle_mask_update(dev, vector, was_masked);
178 }
179 
180 static const MemoryRegionOps msix_table_mmio_ops = {
181     .read = msix_table_mmio_read,
182     .write = msix_table_mmio_write,
183     .endianness = DEVICE_LITTLE_ENDIAN,
184     .valid = {
185         .min_access_size = 4,
186         .max_access_size = 4,
187     },
188 };
189 
190 static uint64_t msix_pba_mmio_read(void *opaque, hwaddr addr,
191                                    unsigned size)
192 {
193     PCIDevice *dev = opaque;
194     if (dev->msix_vector_poll_notifier) {
195         unsigned vector_start = addr * 8;
196         unsigned vector_end = MIN(addr + size * 8, dev->msix_entries_nr);
197         dev->msix_vector_poll_notifier(dev, vector_start, vector_end);
198     }
199 
200     return pci_get_long(dev->msix_pba + addr);
201 }
202 
203 static void msix_pba_mmio_write(void *opaque, hwaddr addr,
204                                 uint64_t val, unsigned size)
205 {
206 }
207 
208 static const MemoryRegionOps msix_pba_mmio_ops = {
209     .read = msix_pba_mmio_read,
210     .write = msix_pba_mmio_write,
211     .endianness = DEVICE_LITTLE_ENDIAN,
212     .valid = {
213         .min_access_size = 4,
214         .max_access_size = 4,
215     },
216 };
217 
218 static void msix_mask_all(struct PCIDevice *dev, unsigned nentries)
219 {
220     int vector;
221 
222     for (vector = 0; vector < nentries; ++vector) {
223         unsigned offset =
224             vector * PCI_MSIX_ENTRY_SIZE + PCI_MSIX_ENTRY_VECTOR_CTRL;
225         bool was_masked = msix_is_masked(dev, vector);
226 
227         dev->msix_table[offset] |= PCI_MSIX_ENTRY_CTRL_MASKBIT;
228         msix_handle_mask_update(dev, vector, was_masked);
229     }
230 }
231 
232 /* Initialize the MSI-X structures */
233 int msix_init(struct PCIDevice *dev, unsigned short nentries,
234               MemoryRegion *table_bar, uint8_t table_bar_nr,
235               unsigned table_offset, MemoryRegion *pba_bar,
236               uint8_t pba_bar_nr, unsigned pba_offset, uint8_t cap_pos)
237 {
238     int cap;
239     unsigned table_size, pba_size;
240     uint8_t *config;
241 
242     /* Nothing to do if MSI is not supported by interrupt controller */
243     if (!msi_supported) {
244         return -ENOTSUP;
245     }
246 
247     if (nentries < 1 || nentries > PCI_MSIX_FLAGS_QSIZE + 1) {
248         return -EINVAL;
249     }
250 
251     table_size = nentries * PCI_MSIX_ENTRY_SIZE;
252     pba_size = QEMU_ALIGN_UP(nentries, 64) / 8;
253 
254     /* Sanity test: table & pba don't overlap, fit within BARs, min aligned */
255     if ((table_bar_nr == pba_bar_nr &&
256          ranges_overlap(table_offset, table_size, pba_offset, pba_size)) ||
257         table_offset + table_size > memory_region_size(table_bar) ||
258         pba_offset + pba_size > memory_region_size(pba_bar) ||
259         (table_offset | pba_offset) & PCI_MSIX_FLAGS_BIRMASK) {
260         return -EINVAL;
261     }
262 
263     cap = pci_add_capability(dev, PCI_CAP_ID_MSIX, cap_pos, MSIX_CAP_LENGTH);
264     if (cap < 0) {
265         return cap;
266     }
267 
268     dev->msix_cap = cap;
269     dev->cap_present |= QEMU_PCI_CAP_MSIX;
270     config = dev->config + cap;
271 
272     pci_set_word(config + PCI_MSIX_FLAGS, nentries - 1);
273     dev->msix_entries_nr = nentries;
274     dev->msix_function_masked = true;
275 
276     pci_set_long(config + PCI_MSIX_TABLE, table_offset | table_bar_nr);
277     pci_set_long(config + PCI_MSIX_PBA, pba_offset | pba_bar_nr);
278 
279     /* Make flags bit writable. */
280     dev->wmask[cap + MSIX_CONTROL_OFFSET] |= MSIX_ENABLE_MASK |
281                                              MSIX_MASKALL_MASK;
282 
283     dev->msix_table = g_malloc0(table_size);
284     dev->msix_pba = g_malloc0(pba_size);
285     dev->msix_entry_used = g_malloc0(nentries * sizeof *dev->msix_entry_used);
286 
287     msix_mask_all(dev, nentries);
288 
289     memory_region_init_io(&dev->msix_table_mmio, OBJECT(dev), &msix_table_mmio_ops, dev,
290                           "msix-table", table_size);
291     memory_region_add_subregion(table_bar, table_offset, &dev->msix_table_mmio);
292     memory_region_init_io(&dev->msix_pba_mmio, OBJECT(dev), &msix_pba_mmio_ops, dev,
293                           "msix-pba", pba_size);
294     memory_region_add_subregion(pba_bar, pba_offset, &dev->msix_pba_mmio);
295 
296     return 0;
297 }
298 
299 int msix_init_exclusive_bar(PCIDevice *dev, unsigned short nentries,
300                             uint8_t bar_nr)
301 {
302     int ret;
303     char *name;
304     uint32_t bar_size = 4096;
305     uint32_t bar_pba_offset = bar_size / 2;
306     uint32_t bar_pba_size = (nentries / 8 + 1) * 8;
307 
308     /*
309      * Migration compatibility dictates that this remains a 4k
310      * BAR with the vector table in the lower half and PBA in
311      * the upper half for nentries which is lower or equal to 128.
312      * No need to care about using more than 65 entries for legacy
313      * machine types who has at most 64 queues.
314      */
315     if (nentries * PCI_MSIX_ENTRY_SIZE > bar_pba_offset) {
316         bar_pba_offset = nentries * PCI_MSIX_ENTRY_SIZE;
317     }
318 
319     if (bar_pba_offset + bar_pba_size > 4096) {
320         bar_size = bar_pba_offset + bar_pba_size;
321     }
322 
323     bar_size = pow2ceil(bar_size);
324 
325     name = g_strdup_printf("%s-msix", dev->name);
326     memory_region_init(&dev->msix_exclusive_bar, OBJECT(dev), name, bar_size);
327     g_free(name);
328 
329     ret = msix_init(dev, nentries, &dev->msix_exclusive_bar, bar_nr,
330                     0, &dev->msix_exclusive_bar,
331                     bar_nr, bar_pba_offset,
332                     0);
333     if (ret) {
334         return ret;
335     }
336 
337     pci_register_bar(dev, bar_nr, PCI_BASE_ADDRESS_SPACE_MEMORY,
338                      &dev->msix_exclusive_bar);
339 
340     return 0;
341 }
342 
343 static void msix_free_irq_entries(PCIDevice *dev)
344 {
345     int vector;
346 
347     for (vector = 0; vector < dev->msix_entries_nr; ++vector) {
348         dev->msix_entry_used[vector] = 0;
349         msix_clr_pending(dev, vector);
350     }
351 }
352 
353 static void msix_clear_all_vectors(PCIDevice *dev)
354 {
355     int vector;
356 
357     for (vector = 0; vector < dev->msix_entries_nr; ++vector) {
358         msix_clr_pending(dev, vector);
359     }
360 }
361 
362 /* Clean up resources for the device. */
363 void msix_uninit(PCIDevice *dev, MemoryRegion *table_bar, MemoryRegion *pba_bar)
364 {
365     if (!msix_present(dev)) {
366         return;
367     }
368     pci_del_capability(dev, PCI_CAP_ID_MSIX, MSIX_CAP_LENGTH);
369     dev->msix_cap = 0;
370     msix_free_irq_entries(dev);
371     dev->msix_entries_nr = 0;
372     memory_region_del_subregion(pba_bar, &dev->msix_pba_mmio);
373     g_free(dev->msix_pba);
374     dev->msix_pba = NULL;
375     memory_region_del_subregion(table_bar, &dev->msix_table_mmio);
376     g_free(dev->msix_table);
377     dev->msix_table = NULL;
378     g_free(dev->msix_entry_used);
379     dev->msix_entry_used = NULL;
380     dev->cap_present &= ~QEMU_PCI_CAP_MSIX;
381 }
382 
383 void msix_uninit_exclusive_bar(PCIDevice *dev)
384 {
385     if (msix_present(dev)) {
386         msix_uninit(dev, &dev->msix_exclusive_bar, &dev->msix_exclusive_bar);
387     }
388 }
389 
390 void msix_save(PCIDevice *dev, QEMUFile *f)
391 {
392     unsigned n = dev->msix_entries_nr;
393 
394     if (!msix_present(dev)) {
395         return;
396     }
397 
398     qemu_put_buffer(f, dev->msix_table, n * PCI_MSIX_ENTRY_SIZE);
399     qemu_put_buffer(f, dev->msix_pba, (n + 7) / 8);
400 }
401 
402 /* Should be called after restoring the config space. */
403 void msix_load(PCIDevice *dev, QEMUFile *f)
404 {
405     unsigned n = dev->msix_entries_nr;
406     unsigned int vector;
407 
408     if (!msix_present(dev)) {
409         return;
410     }
411 
412     msix_clear_all_vectors(dev);
413     qemu_get_buffer(f, dev->msix_table, n * PCI_MSIX_ENTRY_SIZE);
414     qemu_get_buffer(f, dev->msix_pba, (n + 7) / 8);
415     msix_update_function_masked(dev);
416 
417     for (vector = 0; vector < n; vector++) {
418         msix_handle_mask_update(dev, vector, true);
419     }
420 }
421 
422 /* Does device support MSI-X? */
423 int msix_present(PCIDevice *dev)
424 {
425     return dev->cap_present & QEMU_PCI_CAP_MSIX;
426 }
427 
428 /* Is MSI-X enabled? */
429 int msix_enabled(PCIDevice *dev)
430 {
431     return (dev->cap_present & QEMU_PCI_CAP_MSIX) &&
432         (dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
433          MSIX_ENABLE_MASK);
434 }
435 
436 /* Send an MSI-X message */
437 void msix_notify(PCIDevice *dev, unsigned vector)
438 {
439     MSIMessage msg;
440 
441     if (vector >= dev->msix_entries_nr || !dev->msix_entry_used[vector])
442         return;
443     if (msix_is_masked(dev, vector)) {
444         msix_set_pending(dev, vector);
445         return;
446     }
447 
448     msg = msix_get_message(dev, vector);
449 
450     msi_send_message(dev, msg);
451 }
452 
453 void msix_reset(PCIDevice *dev)
454 {
455     if (!msix_present(dev)) {
456         return;
457     }
458     msix_clear_all_vectors(dev);
459     dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &=
460 	    ~dev->wmask[dev->msix_cap + MSIX_CONTROL_OFFSET];
461     memset(dev->msix_table, 0, dev->msix_entries_nr * PCI_MSIX_ENTRY_SIZE);
462     memset(dev->msix_pba, 0, QEMU_ALIGN_UP(dev->msix_entries_nr, 64) / 8);
463     msix_mask_all(dev, dev->msix_entries_nr);
464 }
465 
466 /* PCI spec suggests that devices make it possible for software to configure
467  * less vectors than supported by the device, but does not specify a standard
468  * mechanism for devices to do so.
469  *
470  * We support this by asking devices to declare vectors software is going to
471  * actually use, and checking this on the notification path. Devices that
472  * don't want to follow the spec suggestion can declare all vectors as used. */
473 
474 /* Mark vector as used. */
475 int msix_vector_use(PCIDevice *dev, unsigned vector)
476 {
477     if (vector >= dev->msix_entries_nr)
478         return -EINVAL;
479     dev->msix_entry_used[vector]++;
480     return 0;
481 }
482 
483 /* Mark vector as unused. */
484 void msix_vector_unuse(PCIDevice *dev, unsigned vector)
485 {
486     if (vector >= dev->msix_entries_nr || !dev->msix_entry_used[vector]) {
487         return;
488     }
489     if (--dev->msix_entry_used[vector]) {
490         return;
491     }
492     msix_clr_pending(dev, vector);
493 }
494 
495 void msix_unuse_all_vectors(PCIDevice *dev)
496 {
497     if (!msix_present(dev)) {
498         return;
499     }
500     msix_free_irq_entries(dev);
501 }
502 
503 unsigned int msix_nr_vectors_allocated(const PCIDevice *dev)
504 {
505     return dev->msix_entries_nr;
506 }
507 
508 static int msix_set_notifier_for_vector(PCIDevice *dev, unsigned int vector)
509 {
510     MSIMessage msg;
511 
512     if (msix_is_masked(dev, vector)) {
513         return 0;
514     }
515     msg = msix_get_message(dev, vector);
516     return dev->msix_vector_use_notifier(dev, vector, msg);
517 }
518 
519 static void msix_unset_notifier_for_vector(PCIDevice *dev, unsigned int vector)
520 {
521     if (msix_is_masked(dev, vector)) {
522         return;
523     }
524     dev->msix_vector_release_notifier(dev, vector);
525 }
526 
527 int msix_set_vector_notifiers(PCIDevice *dev,
528                               MSIVectorUseNotifier use_notifier,
529                               MSIVectorReleaseNotifier release_notifier,
530                               MSIVectorPollNotifier poll_notifier)
531 {
532     int vector, ret;
533 
534     assert(use_notifier && release_notifier);
535 
536     dev->msix_vector_use_notifier = use_notifier;
537     dev->msix_vector_release_notifier = release_notifier;
538     dev->msix_vector_poll_notifier = poll_notifier;
539 
540     if ((dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
541         (MSIX_ENABLE_MASK | MSIX_MASKALL_MASK)) == MSIX_ENABLE_MASK) {
542         for (vector = 0; vector < dev->msix_entries_nr; vector++) {
543             ret = msix_set_notifier_for_vector(dev, vector);
544             if (ret < 0) {
545                 goto undo;
546             }
547         }
548     }
549     if (dev->msix_vector_poll_notifier) {
550         dev->msix_vector_poll_notifier(dev, 0, dev->msix_entries_nr);
551     }
552     return 0;
553 
554 undo:
555     while (--vector >= 0) {
556         msix_unset_notifier_for_vector(dev, vector);
557     }
558     dev->msix_vector_use_notifier = NULL;
559     dev->msix_vector_release_notifier = NULL;
560     return ret;
561 }
562 
563 void msix_unset_vector_notifiers(PCIDevice *dev)
564 {
565     int vector;
566 
567     assert(dev->msix_vector_use_notifier &&
568            dev->msix_vector_release_notifier);
569 
570     if ((dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
571         (MSIX_ENABLE_MASK | MSIX_MASKALL_MASK)) == MSIX_ENABLE_MASK) {
572         for (vector = 0; vector < dev->msix_entries_nr; vector++) {
573             msix_unset_notifier_for_vector(dev, vector);
574         }
575     }
576     dev->msix_vector_use_notifier = NULL;
577     dev->msix_vector_release_notifier = NULL;
578     dev->msix_vector_poll_notifier = NULL;
579 }
580 
581 static void put_msix_state(QEMUFile *f, void *pv, size_t size)
582 {
583     msix_save(pv, f);
584 }
585 
586 static int get_msix_state(QEMUFile *f, void *pv, size_t size)
587 {
588     msix_load(pv, f);
589     return 0;
590 }
591 
592 static VMStateInfo vmstate_info_msix = {
593     .name = "msix state",
594     .get  = get_msix_state,
595     .put  = put_msix_state,
596 };
597 
598 const VMStateDescription vmstate_msix = {
599     .name = "msix",
600     .fields = (VMStateField[]) {
601         {
602             .name         = "msix",
603             .version_id   = 0,
604             .field_exists = NULL,
605             .size         = 0,   /* ouch */
606             .info         = &vmstate_info_msix,
607             .flags        = VMS_SINGLE,
608             .offset       = 0,
609         },
610         VMSTATE_END_OF_LIST()
611     }
612 };
613