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"
19 #include "hw/pci/msi.h"
20 #include "hw/pci/msix.h"
21 #include "hw/pci/pci.h"
22 #include "hw/xen/xen.h"
23 #include "qemu/range.h"
24 #include "qapi/error.h"
26 #define MSIX_CAP_LENGTH 12
28 /* MSI enable bit and maskall bit are in byte 1 in FLAGS register */
29 #define MSIX_CONTROL_OFFSET (PCI_MSIX_FLAGS + 1)
30 #define MSIX_ENABLE_MASK (PCI_MSIX_FLAGS_ENABLE >> 8)
31 #define MSIX_MASKALL_MASK (PCI_MSIX_FLAGS_MASKALL >> 8)
33 MSIMessage
msix_get_message(PCIDevice
*dev
, unsigned vector
)
35 uint8_t *table_entry
= dev
->msix_table
+ vector
* PCI_MSIX_ENTRY_SIZE
;
38 msg
.address
= pci_get_quad(table_entry
+ PCI_MSIX_ENTRY_LOWER_ADDR
);
39 msg
.data
= pci_get_long(table_entry
+ PCI_MSIX_ENTRY_DATA
);
44 * Special API for POWER to configure the vectors through
45 * a side channel. Should never be used by devices.
47 void msix_set_message(PCIDevice
*dev
, int vector
, struct MSIMessage msg
)
49 uint8_t *table_entry
= dev
->msix_table
+ vector
* PCI_MSIX_ENTRY_SIZE
;
51 pci_set_quad(table_entry
+ PCI_MSIX_ENTRY_LOWER_ADDR
, msg
.address
);
52 pci_set_long(table_entry
+ PCI_MSIX_ENTRY_DATA
, msg
.data
);
53 table_entry
[PCI_MSIX_ENTRY_VECTOR_CTRL
] &= ~PCI_MSIX_ENTRY_CTRL_MASKBIT
;
56 static uint8_t msix_pending_mask(int vector
)
58 return 1 << (vector
% 8);
61 static uint8_t *msix_pending_byte(PCIDevice
*dev
, int vector
)
63 return dev
->msix_pba
+ vector
/ 8;
66 static int msix_is_pending(PCIDevice
*dev
, int vector
)
68 return *msix_pending_byte(dev
, vector
) & msix_pending_mask(vector
);
71 void msix_set_pending(PCIDevice
*dev
, unsigned int vector
)
73 *msix_pending_byte(dev
, vector
) |= msix_pending_mask(vector
);
76 void msix_clr_pending(PCIDevice
*dev
, int vector
)
78 *msix_pending_byte(dev
, vector
) &= ~msix_pending_mask(vector
);
81 static bool msix_vector_masked(PCIDevice
*dev
, unsigned int vector
, bool fmask
)
83 unsigned offset
= vector
* PCI_MSIX_ENTRY_SIZE
;
84 uint8_t *data
= &dev
->msix_table
[offset
+ PCI_MSIX_ENTRY_DATA
];
85 /* MSIs on Xen can be remapped into pirqs. In those cases, masking
86 * and unmasking go through the PV evtchn path. */
87 if (xen_enabled() && xen_is_pirq_msi(pci_get_long(data
))) {
90 return fmask
|| dev
->msix_table
[offset
+ PCI_MSIX_ENTRY_VECTOR_CTRL
] &
91 PCI_MSIX_ENTRY_CTRL_MASKBIT
;
94 bool msix_is_masked(PCIDevice
*dev
, unsigned int vector
)
96 return msix_vector_masked(dev
, vector
, dev
->msix_function_masked
);
99 static void msix_fire_vector_notifier(PCIDevice
*dev
,
100 unsigned int vector
, bool is_masked
)
105 if (!dev
->msix_vector_use_notifier
) {
109 dev
->msix_vector_release_notifier(dev
, vector
);
111 msg
= msix_get_message(dev
, vector
);
112 ret
= dev
->msix_vector_use_notifier(dev
, vector
, msg
);
117 static void msix_handle_mask_update(PCIDevice
*dev
, int vector
, bool was_masked
)
119 bool is_masked
= msix_is_masked(dev
, vector
);
121 if (is_masked
== was_masked
) {
125 msix_fire_vector_notifier(dev
, vector
, is_masked
);
127 if (!is_masked
&& msix_is_pending(dev
, vector
)) {
128 msix_clr_pending(dev
, vector
);
129 msix_notify(dev
, vector
);
133 static void msix_update_function_masked(PCIDevice
*dev
)
135 dev
->msix_function_masked
= !msix_enabled(dev
) ||
136 (dev
->config
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
] & MSIX_MASKALL_MASK
);
139 /* Handle MSI-X capability config write. */
140 void msix_write_config(PCIDevice
*dev
, uint32_t addr
,
141 uint32_t val
, int len
)
143 unsigned enable_pos
= dev
->msix_cap
+ MSIX_CONTROL_OFFSET
;
147 if (!msix_present(dev
) || !range_covers_byte(addr
, len
, enable_pos
)) {
151 was_masked
= dev
->msix_function_masked
;
152 msix_update_function_masked(dev
);
154 if (!msix_enabled(dev
)) {
158 pci_device_deassert_intx(dev
);
160 if (dev
->msix_function_masked
== was_masked
) {
164 for (vector
= 0; vector
< dev
->msix_entries_nr
; ++vector
) {
165 msix_handle_mask_update(dev
, vector
,
166 msix_vector_masked(dev
, vector
, was_masked
));
170 static uint64_t msix_table_mmio_read(void *opaque
, hwaddr addr
,
173 PCIDevice
*dev
= opaque
;
175 return pci_get_long(dev
->msix_table
+ addr
);
178 static void msix_table_mmio_write(void *opaque
, hwaddr addr
,
179 uint64_t val
, unsigned size
)
181 PCIDevice
*dev
= opaque
;
182 int vector
= addr
/ PCI_MSIX_ENTRY_SIZE
;
185 was_masked
= msix_is_masked(dev
, vector
);
186 pci_set_long(dev
->msix_table
+ addr
, val
);
187 msix_handle_mask_update(dev
, vector
, was_masked
);
190 static const MemoryRegionOps msix_table_mmio_ops
= {
191 .read
= msix_table_mmio_read
,
192 .write
= msix_table_mmio_write
,
193 .endianness
= DEVICE_LITTLE_ENDIAN
,
195 .min_access_size
= 4,
196 .max_access_size
= 4,
200 static uint64_t msix_pba_mmio_read(void *opaque
, hwaddr addr
,
203 PCIDevice
*dev
= opaque
;
204 if (dev
->msix_vector_poll_notifier
) {
205 unsigned vector_start
= addr
* 8;
206 unsigned vector_end
= MIN(addr
+ size
* 8, dev
->msix_entries_nr
);
207 dev
->msix_vector_poll_notifier(dev
, vector_start
, vector_end
);
210 return pci_get_long(dev
->msix_pba
+ addr
);
213 static void msix_pba_mmio_write(void *opaque
, hwaddr addr
,
214 uint64_t val
, unsigned size
)
218 static const MemoryRegionOps msix_pba_mmio_ops
= {
219 .read
= msix_pba_mmio_read
,
220 .write
= msix_pba_mmio_write
,
221 .endianness
= DEVICE_LITTLE_ENDIAN
,
223 .min_access_size
= 4,
224 .max_access_size
= 4,
228 static void msix_mask_all(struct PCIDevice
*dev
, unsigned nentries
)
232 for (vector
= 0; vector
< nentries
; ++vector
) {
234 vector
* PCI_MSIX_ENTRY_SIZE
+ PCI_MSIX_ENTRY_VECTOR_CTRL
;
235 bool was_masked
= msix_is_masked(dev
, vector
);
237 dev
->msix_table
[offset
] |= PCI_MSIX_ENTRY_CTRL_MASKBIT
;
238 msix_handle_mask_update(dev
, vector
, was_masked
);
243 * Make PCI device @dev MSI-X capable
244 * @nentries is the max number of MSI-X vectors that the device support.
245 * @table_bar is the MemoryRegion that MSI-X table structure resides.
246 * @table_bar_nr is number of base address register corresponding to @table_bar.
247 * @table_offset indicates the offset that the MSI-X table structure starts with
249 * @pba_bar is the MemoryRegion that the Pending Bit Array structure resides.
250 * @pba_bar_nr is number of base address register corresponding to @pba_bar.
251 * @pba_offset indicates the offset that the Pending Bit Array structure
252 * starts with in @pba_bar.
253 * Non-zero @cap_pos puts capability MSI-X at that offset in PCI config space.
254 * @errp is for returning errors.
256 * Return 0 on success; set @errp and return -errno on error:
257 * -ENOTSUP means lacking msi support for a msi-capable platform.
258 * -EINVAL means capability overlap, happens when @cap_pos is non-zero,
259 * also means a programming error, except device assignment, which can check
260 * if a real HW is broken.
262 int msix_init(struct PCIDevice
*dev
, unsigned short nentries
,
263 MemoryRegion
*table_bar
, uint8_t table_bar_nr
,
264 unsigned table_offset
, MemoryRegion
*pba_bar
,
265 uint8_t pba_bar_nr
, unsigned pba_offset
, uint8_t cap_pos
,
269 unsigned table_size
, pba_size
;
272 /* Nothing to do if MSI is not supported by interrupt controller */
273 if (!msi_nonbroken
) {
274 error_setg(errp
, "MSI-X is not supported by interrupt controller");
278 if (nentries
< 1 || nentries
> PCI_MSIX_FLAGS_QSIZE
+ 1) {
279 error_setg(errp
, "The number of MSI-X vectors is invalid");
283 table_size
= nentries
* PCI_MSIX_ENTRY_SIZE
;
284 pba_size
= QEMU_ALIGN_UP(nentries
, 64) / 8;
286 /* Sanity test: table & pba don't overlap, fit within BARs, min aligned */
287 if ((table_bar_nr
== pba_bar_nr
&&
288 ranges_overlap(table_offset
, table_size
, pba_offset
, pba_size
)) ||
289 table_offset
+ table_size
> memory_region_size(table_bar
) ||
290 pba_offset
+ pba_size
> memory_region_size(pba_bar
) ||
291 (table_offset
| pba_offset
) & PCI_MSIX_FLAGS_BIRMASK
) {
292 error_setg(errp
, "table & pba overlap, or they don't fit in BARs,"
297 cap
= pci_add_capability2(dev
, PCI_CAP_ID_MSIX
,
298 cap_pos
, MSIX_CAP_LENGTH
, errp
);
304 dev
->cap_present
|= QEMU_PCI_CAP_MSIX
;
305 config
= dev
->config
+ cap
;
307 pci_set_word(config
+ PCI_MSIX_FLAGS
, nentries
- 1);
308 dev
->msix_entries_nr
= nentries
;
309 dev
->msix_function_masked
= true;
311 pci_set_long(config
+ PCI_MSIX_TABLE
, table_offset
| table_bar_nr
);
312 pci_set_long(config
+ PCI_MSIX_PBA
, pba_offset
| pba_bar_nr
);
314 /* Make flags bit writable. */
315 dev
->wmask
[cap
+ MSIX_CONTROL_OFFSET
] |= MSIX_ENABLE_MASK
|
318 dev
->msix_table
= g_malloc0(table_size
);
319 dev
->msix_pba
= g_malloc0(pba_size
);
320 dev
->msix_entry_used
= g_malloc0(nentries
* sizeof *dev
->msix_entry_used
);
322 msix_mask_all(dev
, nentries
);
324 memory_region_init_io(&dev
->msix_table_mmio
, OBJECT(dev
), &msix_table_mmio_ops
, dev
,
325 "msix-table", table_size
);
326 memory_region_add_subregion(table_bar
, table_offset
, &dev
->msix_table_mmio
);
327 memory_region_init_io(&dev
->msix_pba_mmio
, OBJECT(dev
), &msix_pba_mmio_ops
, dev
,
328 "msix-pba", pba_size
);
329 memory_region_add_subregion(pba_bar
, pba_offset
, &dev
->msix_pba_mmio
);
334 int msix_init_exclusive_bar(PCIDevice
*dev
, unsigned short nentries
,
335 uint8_t bar_nr
, Error
**errp
)
339 uint32_t bar_size
= 4096;
340 uint32_t bar_pba_offset
= bar_size
/ 2;
341 uint32_t bar_pba_size
= (nentries
/ 8 + 1) * 8;
344 * Migration compatibility dictates that this remains a 4k
345 * BAR with the vector table in the lower half and PBA in
346 * the upper half for nentries which is lower or equal to 128.
347 * No need to care about using more than 65 entries for legacy
348 * machine types who has at most 64 queues.
350 if (nentries
* PCI_MSIX_ENTRY_SIZE
> bar_pba_offset
) {
351 bar_pba_offset
= nentries
* PCI_MSIX_ENTRY_SIZE
;
354 if (bar_pba_offset
+ bar_pba_size
> 4096) {
355 bar_size
= bar_pba_offset
+ bar_pba_size
;
358 bar_size
= pow2ceil(bar_size
);
360 name
= g_strdup_printf("%s-msix", dev
->name
);
361 memory_region_init(&dev
->msix_exclusive_bar
, OBJECT(dev
), name
, bar_size
);
364 ret
= msix_init(dev
, nentries
, &dev
->msix_exclusive_bar
, bar_nr
,
365 0, &dev
->msix_exclusive_bar
,
366 bar_nr
, bar_pba_offset
,
372 pci_register_bar(dev
, bar_nr
, PCI_BASE_ADDRESS_SPACE_MEMORY
,
373 &dev
->msix_exclusive_bar
);
378 static void msix_free_irq_entries(PCIDevice
*dev
)
382 for (vector
= 0; vector
< dev
->msix_entries_nr
; ++vector
) {
383 dev
->msix_entry_used
[vector
] = 0;
384 msix_clr_pending(dev
, vector
);
388 static void msix_clear_all_vectors(PCIDevice
*dev
)
392 for (vector
= 0; vector
< dev
->msix_entries_nr
; ++vector
) {
393 msix_clr_pending(dev
, vector
);
397 /* Clean up resources for the device. */
398 void msix_uninit(PCIDevice
*dev
, MemoryRegion
*table_bar
, MemoryRegion
*pba_bar
)
400 if (!msix_present(dev
)) {
403 pci_del_capability(dev
, PCI_CAP_ID_MSIX
, MSIX_CAP_LENGTH
);
405 msix_free_irq_entries(dev
);
406 dev
->msix_entries_nr
= 0;
407 memory_region_del_subregion(pba_bar
, &dev
->msix_pba_mmio
);
408 g_free(dev
->msix_pba
);
409 dev
->msix_pba
= NULL
;
410 memory_region_del_subregion(table_bar
, &dev
->msix_table_mmio
);
411 g_free(dev
->msix_table
);
412 dev
->msix_table
= NULL
;
413 g_free(dev
->msix_entry_used
);
414 dev
->msix_entry_used
= NULL
;
415 dev
->cap_present
&= ~QEMU_PCI_CAP_MSIX
;
418 void msix_uninit_exclusive_bar(PCIDevice
*dev
)
420 if (msix_present(dev
)) {
421 msix_uninit(dev
, &dev
->msix_exclusive_bar
, &dev
->msix_exclusive_bar
);
425 void msix_save(PCIDevice
*dev
, QEMUFile
*f
)
427 unsigned n
= dev
->msix_entries_nr
;
429 if (!msix_present(dev
)) {
433 qemu_put_buffer(f
, dev
->msix_table
, n
* PCI_MSIX_ENTRY_SIZE
);
434 qemu_put_buffer(f
, dev
->msix_pba
, (n
+ 7) / 8);
437 /* Should be called after restoring the config space. */
438 void msix_load(PCIDevice
*dev
, QEMUFile
*f
)
440 unsigned n
= dev
->msix_entries_nr
;
443 if (!msix_present(dev
)) {
447 msix_clear_all_vectors(dev
);
448 qemu_get_buffer(f
, dev
->msix_table
, n
* PCI_MSIX_ENTRY_SIZE
);
449 qemu_get_buffer(f
, dev
->msix_pba
, (n
+ 7) / 8);
450 msix_update_function_masked(dev
);
452 for (vector
= 0; vector
< n
; vector
++) {
453 msix_handle_mask_update(dev
, vector
, true);
457 /* Does device support MSI-X? */
458 int msix_present(PCIDevice
*dev
)
460 return dev
->cap_present
& QEMU_PCI_CAP_MSIX
;
463 /* Is MSI-X enabled? */
464 int msix_enabled(PCIDevice
*dev
)
466 return (dev
->cap_present
& QEMU_PCI_CAP_MSIX
) &&
467 (dev
->config
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
] &
471 /* Send an MSI-X message */
472 void msix_notify(PCIDevice
*dev
, unsigned vector
)
476 if (vector
>= dev
->msix_entries_nr
|| !dev
->msix_entry_used
[vector
]) {
480 if (msix_is_masked(dev
, vector
)) {
481 msix_set_pending(dev
, vector
);
485 msg
= msix_get_message(dev
, vector
);
487 msi_send_message(dev
, msg
);
490 void msix_reset(PCIDevice
*dev
)
492 if (!msix_present(dev
)) {
495 msix_clear_all_vectors(dev
);
496 dev
->config
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
] &=
497 ~dev
->wmask
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
];
498 memset(dev
->msix_table
, 0, dev
->msix_entries_nr
* PCI_MSIX_ENTRY_SIZE
);
499 memset(dev
->msix_pba
, 0, QEMU_ALIGN_UP(dev
->msix_entries_nr
, 64) / 8);
500 msix_mask_all(dev
, dev
->msix_entries_nr
);
503 /* PCI spec suggests that devices make it possible for software to configure
504 * less vectors than supported by the device, but does not specify a standard
505 * mechanism for devices to do so.
507 * We support this by asking devices to declare vectors software is going to
508 * actually use, and checking this on the notification path. Devices that
509 * don't want to follow the spec suggestion can declare all vectors as used. */
511 /* Mark vector as used. */
512 int msix_vector_use(PCIDevice
*dev
, unsigned vector
)
514 if (vector
>= dev
->msix_entries_nr
) {
518 dev
->msix_entry_used
[vector
]++;
522 /* Mark vector as unused. */
523 void msix_vector_unuse(PCIDevice
*dev
, unsigned vector
)
525 if (vector
>= dev
->msix_entries_nr
|| !dev
->msix_entry_used
[vector
]) {
528 if (--dev
->msix_entry_used
[vector
]) {
531 msix_clr_pending(dev
, vector
);
534 void msix_unuse_all_vectors(PCIDevice
*dev
)
536 if (!msix_present(dev
)) {
539 msix_free_irq_entries(dev
);
542 unsigned int msix_nr_vectors_allocated(const PCIDevice
*dev
)
544 return dev
->msix_entries_nr
;
547 static int msix_set_notifier_for_vector(PCIDevice
*dev
, unsigned int vector
)
551 if (msix_is_masked(dev
, vector
)) {
554 msg
= msix_get_message(dev
, vector
);
555 return dev
->msix_vector_use_notifier(dev
, vector
, msg
);
558 static void msix_unset_notifier_for_vector(PCIDevice
*dev
, unsigned int vector
)
560 if (msix_is_masked(dev
, vector
)) {
563 dev
->msix_vector_release_notifier(dev
, vector
);
566 int msix_set_vector_notifiers(PCIDevice
*dev
,
567 MSIVectorUseNotifier use_notifier
,
568 MSIVectorReleaseNotifier release_notifier
,
569 MSIVectorPollNotifier poll_notifier
)
573 assert(use_notifier
&& release_notifier
);
575 dev
->msix_vector_use_notifier
= use_notifier
;
576 dev
->msix_vector_release_notifier
= release_notifier
;
577 dev
->msix_vector_poll_notifier
= poll_notifier
;
579 if ((dev
->config
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
] &
580 (MSIX_ENABLE_MASK
| MSIX_MASKALL_MASK
)) == MSIX_ENABLE_MASK
) {
581 for (vector
= 0; vector
< dev
->msix_entries_nr
; vector
++) {
582 ret
= msix_set_notifier_for_vector(dev
, vector
);
588 if (dev
->msix_vector_poll_notifier
) {
589 dev
->msix_vector_poll_notifier(dev
, 0, dev
->msix_entries_nr
);
594 while (--vector
>= 0) {
595 msix_unset_notifier_for_vector(dev
, vector
);
597 dev
->msix_vector_use_notifier
= NULL
;
598 dev
->msix_vector_release_notifier
= NULL
;
602 void msix_unset_vector_notifiers(PCIDevice
*dev
)
606 assert(dev
->msix_vector_use_notifier
&&
607 dev
->msix_vector_release_notifier
);
609 if ((dev
->config
[dev
->msix_cap
+ MSIX_CONTROL_OFFSET
] &
610 (MSIX_ENABLE_MASK
| MSIX_MASKALL_MASK
)) == MSIX_ENABLE_MASK
) {
611 for (vector
= 0; vector
< dev
->msix_entries_nr
; vector
++) {
612 msix_unset_notifier_for_vector(dev
, vector
);
615 dev
->msix_vector_use_notifier
= NULL
;
616 dev
->msix_vector_release_notifier
= NULL
;
617 dev
->msix_vector_poll_notifier
= NULL
;
620 static int put_msix_state(QEMUFile
*f
, void *pv
, size_t size
,
621 VMStateField
*field
, QJSON
*vmdesc
)
628 static int get_msix_state(QEMUFile
*f
, void *pv
, size_t size
,
635 static VMStateInfo vmstate_info_msix
= {
636 .name
= "msix state",
637 .get
= get_msix_state
,
638 .put
= put_msix_state
,
641 const VMStateDescription vmstate_msix
= {
643 .fields
= (VMStateField
[]) {
647 .field_exists
= NULL
,
648 .size
= 0, /* ouch */
649 .info
= &vmstate_info_msix
,
653 VMSTATE_END_OF_LIST()