4 * This module includes support for MSI-X in pci devices.
6 * Author: Michael S. Tsirkin <mst@redhat.com>
8 * Copyright (c) 2009, Red Hat Inc, Michael S. Tsirkin (mst@redhat.com)
10 * This work is licensed under the terms of the GNU GPL, version 2. See
11 * the COPYING file in the top-level directory.
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.
17 #include "qemu/osdep.h"
18 #include "hw/pci/msi.h"
19 #include "hw/pci/msix.h"
20 #include "hw/pci/pci.h"
21 #include "hw/xen/xen.h"
22 #include "sysemu/xen.h"
23 #include "migration/qemu-file-types.h"
24 #include "migration/vmstate.h"
25 #include "qemu/range.h"
26 #include "qapi/error.h"
29 /* MSI enable bit and maskall bit are in byte 1 in FLAGS register */
30 #define MSIX_CONTROL_OFFSET (PCI_MSIX_FLAGS + 1)
31 #define MSIX_ENABLE_MASK (PCI_MSIX_FLAGS_ENABLE >> 8)
32 #define MSIX_MASKALL_MASK (PCI_MSIX_FLAGS_MASKALL >> 8)
34 MSIMessage
msix_get_message(PCIDevice
*dev
, unsigned vector
)
36 uint8_t *table_entry
= dev
->msix_table
+ vector
* PCI_MSIX_ENTRY_SIZE
;
39 msg
.address
= pci_get_quad(table_entry
+ PCI_MSIX_ENTRY_LOWER_ADDR
);
40 msg
.data
= pci_get_long(table_entry
+ PCI_MSIX_ENTRY_DATA
);
45 * Special API for POWER to configure the vectors through
46 * a side channel. Should never be used by devices.
48 void msix_set_message(PCIDevice
*dev
, int vector
, struct MSIMessage msg
)
50 uint8_t *table_entry
= dev
->msix_table
+ vector
* PCI_MSIX_ENTRY_SIZE
;
52 pci_set_quad(table_entry
+ PCI_MSIX_ENTRY_LOWER_ADDR
, msg
.address
);
53 pci_set_long(table_entry
+ PCI_MSIX_ENTRY_DATA
, msg
.data
);
54 table_entry
[PCI_MSIX_ENTRY_VECTOR_CTRL
] &= ~PCI_MSIX_ENTRY_CTRL_MASKBIT
;
57 static uint8_t msix_pending_mask(int vector
)
59 return 1 << (vector
% 8);
62 static uint8_t *msix_pending_byte(PCIDevice
*dev
, int vector
)
64 return dev
->msix_pba
+ vector
/ 8;
67 static int msix_is_pending(PCIDevice
*dev
, int vector
)
69 return *msix_pending_byte(dev
, vector
) & msix_pending_mask(vector
);
72 void msix_set_pending(PCIDevice
*dev
, unsigned int vector
)
74 *msix_pending_byte(dev
, vector
) |= msix_pending_mask(vector
);
77 void msix_clr_pending(PCIDevice
*dev
, int vector
)
79 *msix_pending_byte(dev
, vector
) &= ~msix_pending_mask(vector
);
82 static bool msix_vector_masked(PCIDevice
*dev
, unsigned int vector
, bool fmask
)
84 unsigned offset
= vector
* PCI_MSIX_ENTRY_SIZE
;
85 uint8_t *data
= &dev
->msix_table
[offset
+ PCI_MSIX_ENTRY_DATA
];
86 /* MSIs on Xen can be remapped into pirqs. In those cases, masking
87 * and unmasking go through the PV evtchn path. */
88 if (xen_enabled() && xen_is_pirq_msi(pci_get_long(data
))) {
91 return fmask
|| dev
->msix_table
[offset
+ PCI_MSIX_ENTRY_VECTOR_CTRL
] &
92 PCI_MSIX_ENTRY_CTRL_MASKBIT
;
95 bool msix_is_masked(PCIDevice
*dev
, unsigned int vector
)
97 return msix_vector_masked(dev
, vector
, dev
->msix_function_masked
);
100 static void msix_fire_vector_notifier(PCIDevice
*dev
,
101 unsigned int vector
, bool is_masked
)
106 if (!dev
->msix_vector_use_notifier
) {
110 dev
->msix_vector_release_notifier(dev
, vector
);
112 msg
= msix_get_message(dev
, vector
);
113 ret
= dev
->msix_vector_use_notifier(dev
, vector
, msg
);
118 static void msix_handle_mask_update(PCIDevice
*dev
, int vector
, bool was_masked
)
120 bool is_masked
= msix_is_masked(dev
, vector
);
122 if (is_masked
== was_masked
) {
126 msix_fire_vector_notifier(dev
, vector
, is_masked
);
128 if (!is_masked
&& msix_is_pending(dev
, vector
)) {
129 msix_clr_pending(dev
, vector
);
130 msix_notify(dev
, vector
);
134 static bool msix_masked(PCIDevice
*dev
)
136 return dev
->config
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
] & MSIX_MASKALL_MASK
;
139 static void msix_update_function_masked(PCIDevice
*dev
)
141 dev
->msix_function_masked
= !msix_enabled(dev
) || msix_masked(dev
);
144 /* Handle MSI-X capability config write. */
145 void msix_write_config(PCIDevice
*dev
, uint32_t addr
,
146 uint32_t val
, int len
)
148 unsigned enable_pos
= dev
->msix_cap
+ MSIX_CONTROL_OFFSET
;
152 if (!msix_present(dev
) || !range_covers_byte(addr
, len
, enable_pos
)) {
156 trace_msix_write_config(dev
->name
, msix_enabled(dev
), msix_masked(dev
));
158 was_masked
= dev
->msix_function_masked
;
159 msix_update_function_masked(dev
);
161 if (!msix_enabled(dev
)) {
165 pci_device_deassert_intx(dev
);
167 if (dev
->msix_function_masked
== was_masked
) {
171 for (vector
= 0; vector
< dev
->msix_entries_nr
; ++vector
) {
172 msix_handle_mask_update(dev
, vector
,
173 msix_vector_masked(dev
, vector
, was_masked
));
177 static uint64_t msix_table_mmio_read(void *opaque
, hwaddr addr
,
180 PCIDevice
*dev
= opaque
;
182 assert(addr
+ size
<= dev
->msix_entries_nr
* PCI_MSIX_ENTRY_SIZE
);
183 return pci_get_long(dev
->msix_table
+ addr
);
186 static void msix_table_mmio_write(void *opaque
, hwaddr addr
,
187 uint64_t val
, unsigned size
)
189 PCIDevice
*dev
= opaque
;
190 int vector
= addr
/ PCI_MSIX_ENTRY_SIZE
;
193 assert(addr
+ size
<= dev
->msix_entries_nr
* PCI_MSIX_ENTRY_SIZE
);
195 was_masked
= msix_is_masked(dev
, vector
);
196 pci_set_long(dev
->msix_table
+ addr
, val
);
197 msix_handle_mask_update(dev
, vector
, was_masked
);
200 static const MemoryRegionOps msix_table_mmio_ops
= {
201 .read
= msix_table_mmio_read
,
202 .write
= msix_table_mmio_write
,
203 .endianness
= DEVICE_LITTLE_ENDIAN
,
205 .min_access_size
= 4,
206 .max_access_size
= 8,
209 .max_access_size
= 4,
213 static uint64_t msix_pba_mmio_read(void *opaque
, hwaddr addr
,
216 PCIDevice
*dev
= opaque
;
217 if (dev
->msix_vector_poll_notifier
) {
218 unsigned vector_start
= addr
* 8;
219 unsigned vector_end
= MIN(addr
+ size
* 8, dev
->msix_entries_nr
);
220 dev
->msix_vector_poll_notifier(dev
, vector_start
, vector_end
);
223 return pci_get_long(dev
->msix_pba
+ addr
);
226 static void msix_pba_mmio_write(void *opaque
, hwaddr addr
,
227 uint64_t val
, unsigned size
)
231 static const MemoryRegionOps msix_pba_mmio_ops
= {
232 .read
= msix_pba_mmio_read
,
233 .write
= msix_pba_mmio_write
,
234 .endianness
= DEVICE_LITTLE_ENDIAN
,
236 .min_access_size
= 4,
237 .max_access_size
= 8,
240 .max_access_size
= 4,
244 static void msix_mask_all(struct PCIDevice
*dev
, unsigned nentries
)
248 for (vector
= 0; vector
< nentries
; ++vector
) {
250 vector
* PCI_MSIX_ENTRY_SIZE
+ PCI_MSIX_ENTRY_VECTOR_CTRL
;
251 bool was_masked
= msix_is_masked(dev
, vector
);
253 dev
->msix_table
[offset
] |= PCI_MSIX_ENTRY_CTRL_MASKBIT
;
254 msix_handle_mask_update(dev
, vector
, was_masked
);
259 * Make PCI device @dev MSI-X capable
260 * @nentries is the max number of MSI-X vectors that the device support.
261 * @table_bar is the MemoryRegion that MSI-X table structure resides.
262 * @table_bar_nr is number of base address register corresponding to @table_bar.
263 * @table_offset indicates the offset that the MSI-X table structure starts with
265 * @pba_bar is the MemoryRegion that the Pending Bit Array structure resides.
266 * @pba_bar_nr is number of base address register corresponding to @pba_bar.
267 * @pba_offset indicates the offset that the Pending Bit Array structure
268 * starts with in @pba_bar.
269 * Non-zero @cap_pos puts capability MSI-X at that offset in PCI config space.
270 * @errp is for returning errors.
272 * Return 0 on success; set @errp and return -errno on error:
273 * -ENOTSUP means lacking msi support for a msi-capable platform.
274 * -EINVAL means capability overlap, happens when @cap_pos is non-zero,
275 * also means a programming error, except device assignment, which can check
276 * if a real HW is broken.
278 int msix_init(struct PCIDevice
*dev
, unsigned short nentries
,
279 MemoryRegion
*table_bar
, uint8_t table_bar_nr
,
280 unsigned table_offset
, MemoryRegion
*pba_bar
,
281 uint8_t pba_bar_nr
, unsigned pba_offset
, uint8_t cap_pos
,
285 unsigned table_size
, pba_size
;
288 /* Nothing to do if MSI is not supported by interrupt controller */
289 if (!msi_nonbroken
) {
290 error_setg(errp
, "MSI-X is not supported by interrupt controller");
294 if (nentries
< 1 || nentries
> PCI_MSIX_FLAGS_QSIZE
+ 1) {
295 error_setg(errp
, "The number of MSI-X vectors is invalid");
299 table_size
= nentries
* PCI_MSIX_ENTRY_SIZE
;
300 pba_size
= QEMU_ALIGN_UP(nentries
, 64) / 8;
302 /* Sanity test: table & pba don't overlap, fit within BARs, min aligned */
303 if ((table_bar_nr
== pba_bar_nr
&&
304 ranges_overlap(table_offset
, table_size
, pba_offset
, pba_size
)) ||
305 table_offset
+ table_size
> memory_region_size(table_bar
) ||
306 pba_offset
+ pba_size
> memory_region_size(pba_bar
) ||
307 (table_offset
| pba_offset
) & PCI_MSIX_FLAGS_BIRMASK
) {
308 error_setg(errp
, "table & pba overlap, or they don't fit in BARs,"
313 cap
= pci_add_capability(dev
, PCI_CAP_ID_MSIX
,
314 cap_pos
, MSIX_CAP_LENGTH
, errp
);
320 dev
->cap_present
|= QEMU_PCI_CAP_MSIX
;
321 config
= dev
->config
+ cap
;
323 pci_set_word(config
+ PCI_MSIX_FLAGS
, nentries
- 1);
324 dev
->msix_entries_nr
= nentries
;
325 dev
->msix_function_masked
= true;
327 pci_set_long(config
+ PCI_MSIX_TABLE
, table_offset
| table_bar_nr
);
328 pci_set_long(config
+ PCI_MSIX_PBA
, pba_offset
| pba_bar_nr
);
330 /* Make flags bit writable. */
331 dev
->wmask
[cap
+ MSIX_CONTROL_OFFSET
] |= MSIX_ENABLE_MASK
|
334 dev
->msix_table
= g_malloc0(table_size
);
335 dev
->msix_pba
= g_malloc0(pba_size
);
336 dev
->msix_entry_used
= g_malloc0(nentries
* sizeof *dev
->msix_entry_used
);
338 msix_mask_all(dev
, nentries
);
340 memory_region_init_io(&dev
->msix_table_mmio
, OBJECT(dev
), &msix_table_mmio_ops
, dev
,
341 "msix-table", table_size
);
342 memory_region_add_subregion(table_bar
, table_offset
, &dev
->msix_table_mmio
);
343 memory_region_init_io(&dev
->msix_pba_mmio
, OBJECT(dev
), &msix_pba_mmio_ops
, dev
,
344 "msix-pba", pba_size
);
345 memory_region_add_subregion(pba_bar
, pba_offset
, &dev
->msix_pba_mmio
);
350 int msix_init_exclusive_bar(PCIDevice
*dev
, unsigned short nentries
,
351 uint8_t bar_nr
, Error
**errp
)
355 uint32_t bar_size
= 4096;
356 uint32_t bar_pba_offset
= bar_size
/ 2;
357 uint32_t bar_pba_size
= QEMU_ALIGN_UP(nentries
, 64) / 8;
360 * Migration compatibility dictates that this remains a 4k
361 * BAR with the vector table in the lower half and PBA in
362 * the upper half for nentries which is lower or equal to 128.
363 * No need to care about using more than 65 entries for legacy
364 * machine types who has at most 64 queues.
366 if (nentries
* PCI_MSIX_ENTRY_SIZE
> bar_pba_offset
) {
367 bar_pba_offset
= nentries
* PCI_MSIX_ENTRY_SIZE
;
370 if (bar_pba_offset
+ bar_pba_size
> 4096) {
371 bar_size
= bar_pba_offset
+ bar_pba_size
;
374 bar_size
= pow2ceil(bar_size
);
376 name
= g_strdup_printf("%s-msix", dev
->name
);
377 memory_region_init(&dev
->msix_exclusive_bar
, OBJECT(dev
), name
, bar_size
);
380 ret
= msix_init(dev
, nentries
, &dev
->msix_exclusive_bar
, bar_nr
,
381 0, &dev
->msix_exclusive_bar
,
382 bar_nr
, bar_pba_offset
,
388 pci_register_bar(dev
, bar_nr
, PCI_BASE_ADDRESS_SPACE_MEMORY
,
389 &dev
->msix_exclusive_bar
);
394 static void msix_free_irq_entries(PCIDevice
*dev
)
398 for (vector
= 0; vector
< dev
->msix_entries_nr
; ++vector
) {
399 dev
->msix_entry_used
[vector
] = 0;
400 msix_clr_pending(dev
, vector
);
404 static void msix_clear_all_vectors(PCIDevice
*dev
)
408 for (vector
= 0; vector
< dev
->msix_entries_nr
; ++vector
) {
409 msix_clr_pending(dev
, vector
);
413 /* Clean up resources for the device. */
414 void msix_uninit(PCIDevice
*dev
, MemoryRegion
*table_bar
, MemoryRegion
*pba_bar
)
416 if (!msix_present(dev
)) {
419 pci_del_capability(dev
, PCI_CAP_ID_MSIX
, MSIX_CAP_LENGTH
);
421 msix_free_irq_entries(dev
);
422 dev
->msix_entries_nr
= 0;
423 memory_region_del_subregion(pba_bar
, &dev
->msix_pba_mmio
);
424 g_free(dev
->msix_pba
);
425 dev
->msix_pba
= NULL
;
426 memory_region_del_subregion(table_bar
, &dev
->msix_table_mmio
);
427 g_free(dev
->msix_table
);
428 dev
->msix_table
= NULL
;
429 g_free(dev
->msix_entry_used
);
430 dev
->msix_entry_used
= NULL
;
431 dev
->cap_present
&= ~QEMU_PCI_CAP_MSIX
;
434 void msix_uninit_exclusive_bar(PCIDevice
*dev
)
436 if (msix_present(dev
)) {
437 msix_uninit(dev
, &dev
->msix_exclusive_bar
, &dev
->msix_exclusive_bar
);
441 void msix_save(PCIDevice
*dev
, QEMUFile
*f
)
443 unsigned n
= dev
->msix_entries_nr
;
445 if (!msix_present(dev
)) {
449 qemu_put_buffer(f
, dev
->msix_table
, n
* PCI_MSIX_ENTRY_SIZE
);
450 qemu_put_buffer(f
, dev
->msix_pba
, DIV_ROUND_UP(n
, 8));
453 /* Should be called after restoring the config space. */
454 void msix_load(PCIDevice
*dev
, QEMUFile
*f
)
456 unsigned n
= dev
->msix_entries_nr
;
459 if (!msix_present(dev
)) {
463 msix_clear_all_vectors(dev
);
464 qemu_get_buffer(f
, dev
->msix_table
, n
* PCI_MSIX_ENTRY_SIZE
);
465 qemu_get_buffer(f
, dev
->msix_pba
, DIV_ROUND_UP(n
, 8));
466 msix_update_function_masked(dev
);
468 for (vector
= 0; vector
< n
; vector
++) {
469 msix_handle_mask_update(dev
, vector
, true);
473 /* Does device support MSI-X? */
474 int msix_present(PCIDevice
*dev
)
476 return dev
->cap_present
& QEMU_PCI_CAP_MSIX
;
479 /* Is MSI-X enabled? */
480 int msix_enabled(PCIDevice
*dev
)
482 return (dev
->cap_present
& QEMU_PCI_CAP_MSIX
) &&
483 (dev
->config
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
] &
487 /* Send an MSI-X message */
488 void msix_notify(PCIDevice
*dev
, unsigned vector
)
492 if (vector
>= dev
->msix_entries_nr
|| !dev
->msix_entry_used
[vector
]) {
496 if (msix_is_masked(dev
, vector
)) {
497 msix_set_pending(dev
, vector
);
501 msg
= msix_get_message(dev
, vector
);
503 msi_send_message(dev
, msg
);
506 void msix_reset(PCIDevice
*dev
)
508 if (!msix_present(dev
)) {
511 msix_clear_all_vectors(dev
);
512 dev
->config
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
] &=
513 ~dev
->wmask
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
];
514 memset(dev
->msix_table
, 0, dev
->msix_entries_nr
* PCI_MSIX_ENTRY_SIZE
);
515 memset(dev
->msix_pba
, 0, QEMU_ALIGN_UP(dev
->msix_entries_nr
, 64) / 8);
516 msix_mask_all(dev
, dev
->msix_entries_nr
);
519 /* PCI spec suggests that devices make it possible for software to configure
520 * less vectors than supported by the device, but does not specify a standard
521 * mechanism for devices to do so.
523 * We support this by asking devices to declare vectors software is going to
524 * actually use, and checking this on the notification path. Devices that
525 * don't want to follow the spec suggestion can declare all vectors as used. */
527 /* Mark vector as used. */
528 int msix_vector_use(PCIDevice
*dev
, unsigned vector
)
530 if (vector
>= dev
->msix_entries_nr
) {
534 dev
->msix_entry_used
[vector
]++;
538 /* Mark vector as unused. */
539 void msix_vector_unuse(PCIDevice
*dev
, unsigned vector
)
541 if (vector
>= dev
->msix_entries_nr
|| !dev
->msix_entry_used
[vector
]) {
544 if (--dev
->msix_entry_used
[vector
]) {
547 msix_clr_pending(dev
, vector
);
550 void msix_unuse_all_vectors(PCIDevice
*dev
)
552 if (!msix_present(dev
)) {
555 msix_free_irq_entries(dev
);
558 unsigned int msix_nr_vectors_allocated(const PCIDevice
*dev
)
560 return dev
->msix_entries_nr
;
563 static int msix_set_notifier_for_vector(PCIDevice
*dev
, unsigned int vector
)
567 if (msix_is_masked(dev
, vector
)) {
570 msg
= msix_get_message(dev
, vector
);
571 return dev
->msix_vector_use_notifier(dev
, vector
, msg
);
574 static void msix_unset_notifier_for_vector(PCIDevice
*dev
, unsigned int vector
)
576 if (msix_is_masked(dev
, vector
)) {
579 dev
->msix_vector_release_notifier(dev
, vector
);
582 int msix_set_vector_notifiers(PCIDevice
*dev
,
583 MSIVectorUseNotifier use_notifier
,
584 MSIVectorReleaseNotifier release_notifier
,
585 MSIVectorPollNotifier poll_notifier
)
589 assert(use_notifier
&& release_notifier
);
591 dev
->msix_vector_use_notifier
= use_notifier
;
592 dev
->msix_vector_release_notifier
= release_notifier
;
593 dev
->msix_vector_poll_notifier
= poll_notifier
;
595 if ((dev
->config
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
] &
596 (MSIX_ENABLE_MASK
| MSIX_MASKALL_MASK
)) == MSIX_ENABLE_MASK
) {
597 for (vector
= 0; vector
< dev
->msix_entries_nr
; vector
++) {
598 ret
= msix_set_notifier_for_vector(dev
, vector
);
604 if (dev
->msix_vector_poll_notifier
) {
605 dev
->msix_vector_poll_notifier(dev
, 0, dev
->msix_entries_nr
);
610 while (--vector
>= 0) {
611 msix_unset_notifier_for_vector(dev
, vector
);
613 dev
->msix_vector_use_notifier
= NULL
;
614 dev
->msix_vector_release_notifier
= NULL
;
618 void msix_unset_vector_notifiers(PCIDevice
*dev
)
622 assert(dev
->msix_vector_use_notifier
&&
623 dev
->msix_vector_release_notifier
);
625 if ((dev
->config
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
] &
626 (MSIX_ENABLE_MASK
| MSIX_MASKALL_MASK
)) == MSIX_ENABLE_MASK
) {
627 for (vector
= 0; vector
< dev
->msix_entries_nr
; vector
++) {
628 msix_unset_notifier_for_vector(dev
, vector
);
631 dev
->msix_vector_use_notifier
= NULL
;
632 dev
->msix_vector_release_notifier
= NULL
;
633 dev
->msix_vector_poll_notifier
= NULL
;
636 static int put_msix_state(QEMUFile
*f
, void *pv
, size_t size
,
637 const VMStateField
*field
, QJSON
*vmdesc
)
644 static int get_msix_state(QEMUFile
*f
, void *pv
, size_t size
,
645 const VMStateField
*field
)
651 static VMStateInfo vmstate_info_msix
= {
652 .name
= "msix state",
653 .get
= get_msix_state
,
654 .put
= put_msix_state
,
657 const VMStateDescription vmstate_msix
= {
659 .fields
= (VMStateField
[]) {
663 .field_exists
= NULL
,
664 .size
= 0, /* ouch */
665 .info
= &vmstate_info_msix
,
669 VMSTATE_END_OF_LIST()