xref: /openbmc/qemu/hw/pci/msix.c (revision 489653b5)
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 const MemoryRegionOps msix_pba_mmio_ops = {
204     .read = msix_pba_mmio_read,
205     .endianness = DEVICE_LITTLE_ENDIAN,
206     .valid = {
207         .min_access_size = 4,
208         .max_access_size = 4,
209     },
210 };
211 
212 static void msix_mask_all(struct PCIDevice *dev, unsigned nentries)
213 {
214     int vector;
215 
216     for (vector = 0; vector < nentries; ++vector) {
217         unsigned offset =
218             vector * PCI_MSIX_ENTRY_SIZE + PCI_MSIX_ENTRY_VECTOR_CTRL;
219         bool was_masked = msix_is_masked(dev, vector);
220 
221         dev->msix_table[offset] |= PCI_MSIX_ENTRY_CTRL_MASKBIT;
222         msix_handle_mask_update(dev, vector, was_masked);
223     }
224 }
225 
226 /* Initialize the MSI-X structures */
227 int msix_init(struct PCIDevice *dev, unsigned short nentries,
228               MemoryRegion *table_bar, uint8_t table_bar_nr,
229               unsigned table_offset, MemoryRegion *pba_bar,
230               uint8_t pba_bar_nr, unsigned pba_offset, uint8_t cap_pos)
231 {
232     int cap;
233     unsigned table_size, pba_size;
234     uint8_t *config;
235 
236     /* Nothing to do if MSI is not supported by interrupt controller */
237     if (!msi_supported) {
238         return -ENOTSUP;
239     }
240 
241     if (nentries < 1 || nentries > PCI_MSIX_FLAGS_QSIZE + 1) {
242         return -EINVAL;
243     }
244 
245     table_size = nentries * PCI_MSIX_ENTRY_SIZE;
246     pba_size = QEMU_ALIGN_UP(nentries, 64) / 8;
247 
248     /* Sanity test: table & pba don't overlap, fit within BARs, min aligned */
249     if ((table_bar_nr == pba_bar_nr &&
250          ranges_overlap(table_offset, table_size, pba_offset, pba_size)) ||
251         table_offset + table_size > memory_region_size(table_bar) ||
252         pba_offset + pba_size > memory_region_size(pba_bar) ||
253         (table_offset | pba_offset) & PCI_MSIX_FLAGS_BIRMASK) {
254         return -EINVAL;
255     }
256 
257     cap = pci_add_capability(dev, PCI_CAP_ID_MSIX, cap_pos, MSIX_CAP_LENGTH);
258     if (cap < 0) {
259         return cap;
260     }
261 
262     dev->msix_cap = cap;
263     dev->cap_present |= QEMU_PCI_CAP_MSIX;
264     config = dev->config + cap;
265 
266     pci_set_word(config + PCI_MSIX_FLAGS, nentries - 1);
267     dev->msix_entries_nr = nentries;
268     dev->msix_function_masked = true;
269 
270     pci_set_long(config + PCI_MSIX_TABLE, table_offset | table_bar_nr);
271     pci_set_long(config + PCI_MSIX_PBA, pba_offset | pba_bar_nr);
272 
273     /* Make flags bit writable. */
274     dev->wmask[cap + MSIX_CONTROL_OFFSET] |= MSIX_ENABLE_MASK |
275                                              MSIX_MASKALL_MASK;
276 
277     dev->msix_table = g_malloc0(table_size);
278     dev->msix_pba = g_malloc0(pba_size);
279     dev->msix_entry_used = g_malloc0(nentries * sizeof *dev->msix_entry_used);
280 
281     msix_mask_all(dev, nentries);
282 
283     memory_region_init_io(&dev->msix_table_mmio, OBJECT(dev), &msix_table_mmio_ops, dev,
284                           "msix-table", table_size);
285     memory_region_add_subregion(table_bar, table_offset, &dev->msix_table_mmio);
286     memory_region_init_io(&dev->msix_pba_mmio, OBJECT(dev), &msix_pba_mmio_ops, dev,
287                           "msix-pba", pba_size);
288     memory_region_add_subregion(pba_bar, pba_offset, &dev->msix_pba_mmio);
289 
290     return 0;
291 }
292 
293 int msix_init_exclusive_bar(PCIDevice *dev, unsigned short nentries,
294                             uint8_t bar_nr)
295 {
296     int ret;
297     char *name;
298     uint32_t bar_size = 4096;
299     uint32_t bar_pba_offset = bar_size / 2;
300     uint32_t bar_pba_size = (nentries / 8 + 1) * 8;
301 
302     /*
303      * Migration compatibility dictates that this remains a 4k
304      * BAR with the vector table in the lower half and PBA in
305      * the upper half for nentries which is lower or equal to 128.
306      * No need to care about using more than 65 entries for legacy
307      * machine types who has at most 64 queues.
308      */
309     if (nentries * PCI_MSIX_ENTRY_SIZE > bar_pba_offset) {
310         bar_pba_offset = nentries * PCI_MSIX_ENTRY_SIZE;
311     }
312 
313     if (bar_pba_offset + bar_pba_size > 4096) {
314         bar_size = bar_pba_offset + bar_pba_size;
315     }
316 
317     if (bar_size & (bar_size - 1)) {
318         bar_size = 1 << qemu_fls(bar_size);
319     }
320 
321     name = g_strdup_printf("%s-msix", dev->name);
322     memory_region_init(&dev->msix_exclusive_bar, OBJECT(dev), name, bar_size);
323     g_free(name);
324 
325     ret = msix_init(dev, nentries, &dev->msix_exclusive_bar, bar_nr,
326                     0, &dev->msix_exclusive_bar,
327                     bar_nr, bar_pba_offset,
328                     0);
329     if (ret) {
330         return ret;
331     }
332 
333     pci_register_bar(dev, bar_nr, PCI_BASE_ADDRESS_SPACE_MEMORY,
334                      &dev->msix_exclusive_bar);
335 
336     return 0;
337 }
338 
339 static void msix_free_irq_entries(PCIDevice *dev)
340 {
341     int vector;
342 
343     for (vector = 0; vector < dev->msix_entries_nr; ++vector) {
344         dev->msix_entry_used[vector] = 0;
345         msix_clr_pending(dev, vector);
346     }
347 }
348 
349 static void msix_clear_all_vectors(PCIDevice *dev)
350 {
351     int vector;
352 
353     for (vector = 0; vector < dev->msix_entries_nr; ++vector) {
354         msix_clr_pending(dev, vector);
355     }
356 }
357 
358 /* Clean up resources for the device. */
359 void msix_uninit(PCIDevice *dev, MemoryRegion *table_bar, MemoryRegion *pba_bar)
360 {
361     if (!msix_present(dev)) {
362         return;
363     }
364     pci_del_capability(dev, PCI_CAP_ID_MSIX, MSIX_CAP_LENGTH);
365     dev->msix_cap = 0;
366     msix_free_irq_entries(dev);
367     dev->msix_entries_nr = 0;
368     memory_region_del_subregion(pba_bar, &dev->msix_pba_mmio);
369     g_free(dev->msix_pba);
370     dev->msix_pba = NULL;
371     memory_region_del_subregion(table_bar, &dev->msix_table_mmio);
372     g_free(dev->msix_table);
373     dev->msix_table = NULL;
374     g_free(dev->msix_entry_used);
375     dev->msix_entry_used = NULL;
376     dev->cap_present &= ~QEMU_PCI_CAP_MSIX;
377 }
378 
379 void msix_uninit_exclusive_bar(PCIDevice *dev)
380 {
381     if (msix_present(dev)) {
382         msix_uninit(dev, &dev->msix_exclusive_bar, &dev->msix_exclusive_bar);
383     }
384 }
385 
386 void msix_save(PCIDevice *dev, QEMUFile *f)
387 {
388     unsigned n = dev->msix_entries_nr;
389 
390     if (!msix_present(dev)) {
391         return;
392     }
393 
394     qemu_put_buffer(f, dev->msix_table, n * PCI_MSIX_ENTRY_SIZE);
395     qemu_put_buffer(f, dev->msix_pba, (n + 7) / 8);
396 }
397 
398 /* Should be called after restoring the config space. */
399 void msix_load(PCIDevice *dev, QEMUFile *f)
400 {
401     unsigned n = dev->msix_entries_nr;
402     unsigned int vector;
403 
404     if (!msix_present(dev)) {
405         return;
406     }
407 
408     msix_clear_all_vectors(dev);
409     qemu_get_buffer(f, dev->msix_table, n * PCI_MSIX_ENTRY_SIZE);
410     qemu_get_buffer(f, dev->msix_pba, (n + 7) / 8);
411     msix_update_function_masked(dev);
412 
413     for (vector = 0; vector < n; vector++) {
414         msix_handle_mask_update(dev, vector, true);
415     }
416 }
417 
418 /* Does device support MSI-X? */
419 int msix_present(PCIDevice *dev)
420 {
421     return dev->cap_present & QEMU_PCI_CAP_MSIX;
422 }
423 
424 /* Is MSI-X enabled? */
425 int msix_enabled(PCIDevice *dev)
426 {
427     return (dev->cap_present & QEMU_PCI_CAP_MSIX) &&
428         (dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
429          MSIX_ENABLE_MASK);
430 }
431 
432 /* Send an MSI-X message */
433 void msix_notify(PCIDevice *dev, unsigned vector)
434 {
435     MSIMessage msg;
436 
437     if (vector >= dev->msix_entries_nr || !dev->msix_entry_used[vector])
438         return;
439     if (msix_is_masked(dev, vector)) {
440         msix_set_pending(dev, vector);
441         return;
442     }
443 
444     msg = msix_get_message(dev, vector);
445 
446     msi_send_message(dev, msg);
447 }
448 
449 void msix_reset(PCIDevice *dev)
450 {
451     if (!msix_present(dev)) {
452         return;
453     }
454     msix_clear_all_vectors(dev);
455     dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &=
456 	    ~dev->wmask[dev->msix_cap + MSIX_CONTROL_OFFSET];
457     memset(dev->msix_table, 0, dev->msix_entries_nr * PCI_MSIX_ENTRY_SIZE);
458     memset(dev->msix_pba, 0, QEMU_ALIGN_UP(dev->msix_entries_nr, 64) / 8);
459     msix_mask_all(dev, dev->msix_entries_nr);
460 }
461 
462 /* PCI spec suggests that devices make it possible for software to configure
463  * less vectors than supported by the device, but does not specify a standard
464  * mechanism for devices to do so.
465  *
466  * We support this by asking devices to declare vectors software is going to
467  * actually use, and checking this on the notification path. Devices that
468  * don't want to follow the spec suggestion can declare all vectors as used. */
469 
470 /* Mark vector as used. */
471 int msix_vector_use(PCIDevice *dev, unsigned vector)
472 {
473     if (vector >= dev->msix_entries_nr)
474         return -EINVAL;
475     dev->msix_entry_used[vector]++;
476     return 0;
477 }
478 
479 /* Mark vector as unused. */
480 void msix_vector_unuse(PCIDevice *dev, unsigned vector)
481 {
482     if (vector >= dev->msix_entries_nr || !dev->msix_entry_used[vector]) {
483         return;
484     }
485     if (--dev->msix_entry_used[vector]) {
486         return;
487     }
488     msix_clr_pending(dev, vector);
489 }
490 
491 void msix_unuse_all_vectors(PCIDevice *dev)
492 {
493     if (!msix_present(dev)) {
494         return;
495     }
496     msix_free_irq_entries(dev);
497 }
498 
499 unsigned int msix_nr_vectors_allocated(const PCIDevice *dev)
500 {
501     return dev->msix_entries_nr;
502 }
503 
504 static int msix_set_notifier_for_vector(PCIDevice *dev, unsigned int vector)
505 {
506     MSIMessage msg;
507 
508     if (msix_is_masked(dev, vector)) {
509         return 0;
510     }
511     msg = msix_get_message(dev, vector);
512     return dev->msix_vector_use_notifier(dev, vector, msg);
513 }
514 
515 static void msix_unset_notifier_for_vector(PCIDevice *dev, unsigned int vector)
516 {
517     if (msix_is_masked(dev, vector)) {
518         return;
519     }
520     dev->msix_vector_release_notifier(dev, vector);
521 }
522 
523 int msix_set_vector_notifiers(PCIDevice *dev,
524                               MSIVectorUseNotifier use_notifier,
525                               MSIVectorReleaseNotifier release_notifier,
526                               MSIVectorPollNotifier poll_notifier)
527 {
528     int vector, ret;
529 
530     assert(use_notifier && release_notifier);
531 
532     dev->msix_vector_use_notifier = use_notifier;
533     dev->msix_vector_release_notifier = release_notifier;
534     dev->msix_vector_poll_notifier = poll_notifier;
535 
536     if ((dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
537         (MSIX_ENABLE_MASK | MSIX_MASKALL_MASK)) == MSIX_ENABLE_MASK) {
538         for (vector = 0; vector < dev->msix_entries_nr; vector++) {
539             ret = msix_set_notifier_for_vector(dev, vector);
540             if (ret < 0) {
541                 goto undo;
542             }
543         }
544     }
545     if (dev->msix_vector_poll_notifier) {
546         dev->msix_vector_poll_notifier(dev, 0, dev->msix_entries_nr);
547     }
548     return 0;
549 
550 undo:
551     while (--vector >= 0) {
552         msix_unset_notifier_for_vector(dev, vector);
553     }
554     dev->msix_vector_use_notifier = NULL;
555     dev->msix_vector_release_notifier = NULL;
556     return ret;
557 }
558 
559 void msix_unset_vector_notifiers(PCIDevice *dev)
560 {
561     int vector;
562 
563     assert(dev->msix_vector_use_notifier &&
564            dev->msix_vector_release_notifier);
565 
566     if ((dev->config[dev->msix_cap + MSIX_CONTROL_OFFSET] &
567         (MSIX_ENABLE_MASK | MSIX_MASKALL_MASK)) == MSIX_ENABLE_MASK) {
568         for (vector = 0; vector < dev->msix_entries_nr; vector++) {
569             msix_unset_notifier_for_vector(dev, vector);
570         }
571     }
572     dev->msix_vector_use_notifier = NULL;
573     dev->msix_vector_release_notifier = NULL;
574     dev->msix_vector_poll_notifier = NULL;
575 }
576 
577 static void put_msix_state(QEMUFile *f, void *pv, size_t size)
578 {
579     msix_save(pv, f);
580 }
581 
582 static int get_msix_state(QEMUFile *f, void *pv, size_t size)
583 {
584     msix_load(pv, f);
585     return 0;
586 }
587 
588 static VMStateInfo vmstate_info_msix = {
589     .name = "msix state",
590     .get  = get_msix_state,
591     .put  = put_msix_state,
592 };
593 
594 const VMStateDescription vmstate_msix = {
595     .name = "msix",
596     .fields = (VMStateField[]) {
597         {
598             .name         = "msix",
599             .version_id   = 0,
600             .field_exists = NULL,
601             .size         = 0,   /* ouch */
602             .info         = &vmstate_info_msix,
603             .flags        = VMS_SINGLE,
604             .offset       = 0,
605         },
606         VMSTATE_END_OF_LIST()
607     }
608 };
609