xref: /openbmc/qemu/hw/vfio/pci-quirks.c (revision 2f44bea9)
1 /*
2  * device quirks for PCI devices
3  *
4  * Copyright Red Hat, Inc. 2012-2015
5  *
6  * Authors:
7  *  Alex Williamson <alex.williamson@redhat.com>
8  *
9  * This work is licensed under the terms of the GNU GPL, version 2.  See
10  * the COPYING file in the top-level directory.
11  */
12 
13 #include "qemu/osdep.h"
14 #include CONFIG_DEVICES
15 #include "exec/memop.h"
16 #include "qemu/units.h"
17 #include "qemu/log.h"
18 #include "qemu/error-report.h"
19 #include "qemu/main-loop.h"
20 #include "qemu/module.h"
21 #include "qemu/range.h"
22 #include "qapi/error.h"
23 #include "qapi/visitor.h"
24 #include <sys/ioctl.h>
25 #include "hw/nvram/fw_cfg.h"
26 #include "hw/qdev-properties.h"
27 #include "pci.h"
28 #include "trace.h"
29 
30 /*
31  * List of device ids/vendor ids for which to disable
32  * option rom loading. This avoids the guest hangs during rom
33  * execution as noticed with the BCM 57810 card for lack of a
34  * more better way to handle such issues.
35  * The  user can still override by specifying a romfile or
36  * rombar=1.
37  * Please see https://bugs.launchpad.net/qemu/+bug/1284874
38  * for an analysis of the 57810 card hang. When adding
39  * a new vendor id/device id combination below, please also add
40  * your card/environment details and information that could
41  * help in debugging to the bug tracking this issue
42  */
43 static const struct {
44     uint32_t vendor;
45     uint32_t device;
46 } rom_denylist[] = {
47     { 0x14e4, 0x168e }, /* Broadcom BCM 57810 */
48 };
49 
50 bool vfio_opt_rom_in_denylist(VFIOPCIDevice *vdev)
51 {
52     int i;
53 
54     for (i = 0 ; i < ARRAY_SIZE(rom_denylist); i++) {
55         if (vfio_pci_is(vdev, rom_denylist[i].vendor, rom_denylist[i].device)) {
56             trace_vfio_quirk_rom_in_denylist(vdev->vbasedev.name,
57                                              rom_denylist[i].vendor,
58                                              rom_denylist[i].device);
59             return true;
60         }
61     }
62     return false;
63 }
64 
65 /*
66  * Device specific region quirks (mostly backdoors to PCI config space)
67  */
68 
69 /*
70  * The generic window quirks operate on an address and data register,
71  * vfio_generic_window_address_quirk handles the address register and
72  * vfio_generic_window_data_quirk handles the data register.  These ops
73  * pass reads and writes through to hardware until a value matching the
74  * stored address match/mask is written.  When this occurs, the data
75  * register access emulated PCI config space for the device rather than
76  * passing through accesses.  This enables devices where PCI config space
77  * is accessible behind a window register to maintain the virtualization
78  * provided through vfio.
79  */
80 typedef struct VFIOConfigWindowMatch {
81     uint32_t match;
82     uint32_t mask;
83 } VFIOConfigWindowMatch;
84 
85 typedef struct VFIOConfigWindowQuirk {
86     struct VFIOPCIDevice *vdev;
87 
88     uint32_t address_val;
89 
90     uint32_t address_offset;
91     uint32_t data_offset;
92 
93     bool window_enabled;
94     uint8_t bar;
95 
96     MemoryRegion *addr_mem;
97     MemoryRegion *data_mem;
98 
99     uint32_t nr_matches;
100     VFIOConfigWindowMatch matches[];
101 } VFIOConfigWindowQuirk;
102 
103 static uint64_t vfio_generic_window_quirk_address_read(void *opaque,
104                                                        hwaddr addr,
105                                                        unsigned size)
106 {
107     VFIOConfigWindowQuirk *window = opaque;
108     VFIOPCIDevice *vdev = window->vdev;
109 
110     return vfio_region_read(&vdev->bars[window->bar].region,
111                             addr + window->address_offset, size);
112 }
113 
114 static void vfio_generic_window_quirk_address_write(void *opaque, hwaddr addr,
115                                                     uint64_t data,
116                                                     unsigned size)
117 {
118     VFIOConfigWindowQuirk *window = opaque;
119     VFIOPCIDevice *vdev = window->vdev;
120     int i;
121 
122     window->window_enabled = false;
123 
124     vfio_region_write(&vdev->bars[window->bar].region,
125                       addr + window->address_offset, data, size);
126 
127     for (i = 0; i < window->nr_matches; i++) {
128         if ((data & ~window->matches[i].mask) == window->matches[i].match) {
129             window->window_enabled = true;
130             window->address_val = data & window->matches[i].mask;
131             trace_vfio_quirk_generic_window_address_write(vdev->vbasedev.name,
132                                     memory_region_name(window->addr_mem), data);
133             break;
134         }
135     }
136 }
137 
138 static const MemoryRegionOps vfio_generic_window_address_quirk = {
139     .read = vfio_generic_window_quirk_address_read,
140     .write = vfio_generic_window_quirk_address_write,
141     .endianness = DEVICE_LITTLE_ENDIAN,
142 };
143 
144 static uint64_t vfio_generic_window_quirk_data_read(void *opaque,
145                                                     hwaddr addr, unsigned size)
146 {
147     VFIOConfigWindowQuirk *window = opaque;
148     VFIOPCIDevice *vdev = window->vdev;
149     uint64_t data;
150 
151     /* Always read data reg, discard if window enabled */
152     data = vfio_region_read(&vdev->bars[window->bar].region,
153                             addr + window->data_offset, size);
154 
155     if (window->window_enabled) {
156         data = vfio_pci_read_config(&vdev->pdev, window->address_val, size);
157         trace_vfio_quirk_generic_window_data_read(vdev->vbasedev.name,
158                                     memory_region_name(window->data_mem), data);
159     }
160 
161     return data;
162 }
163 
164 static void vfio_generic_window_quirk_data_write(void *opaque, hwaddr addr,
165                                                  uint64_t data, unsigned size)
166 {
167     VFIOConfigWindowQuirk *window = opaque;
168     VFIOPCIDevice *vdev = window->vdev;
169 
170     if (window->window_enabled) {
171         vfio_pci_write_config(&vdev->pdev, window->address_val, data, size);
172         trace_vfio_quirk_generic_window_data_write(vdev->vbasedev.name,
173                                     memory_region_name(window->data_mem), data);
174         return;
175     }
176 
177     vfio_region_write(&vdev->bars[window->bar].region,
178                       addr + window->data_offset, data, size);
179 }
180 
181 static const MemoryRegionOps vfio_generic_window_data_quirk = {
182     .read = vfio_generic_window_quirk_data_read,
183     .write = vfio_generic_window_quirk_data_write,
184     .endianness = DEVICE_LITTLE_ENDIAN,
185 };
186 
187 /*
188  * The generic mirror quirk handles devices which expose PCI config space
189  * through a region within a BAR.  When enabled, reads and writes are
190  * redirected through to emulated PCI config space.  XXX if PCI config space
191  * used memory regions, this could just be an alias.
192  */
193 typedef struct VFIOConfigMirrorQuirk {
194     struct VFIOPCIDevice *vdev;
195     uint32_t offset;
196     uint8_t bar;
197     MemoryRegion *mem;
198     uint8_t data[];
199 } VFIOConfigMirrorQuirk;
200 
201 static uint64_t vfio_generic_quirk_mirror_read(void *opaque,
202                                                hwaddr addr, unsigned size)
203 {
204     VFIOConfigMirrorQuirk *mirror = opaque;
205     VFIOPCIDevice *vdev = mirror->vdev;
206     uint64_t data;
207 
208     /* Read and discard in case the hardware cares */
209     (void)vfio_region_read(&vdev->bars[mirror->bar].region,
210                            addr + mirror->offset, size);
211 
212     data = vfio_pci_read_config(&vdev->pdev, addr, size);
213     trace_vfio_quirk_generic_mirror_read(vdev->vbasedev.name,
214                                          memory_region_name(mirror->mem),
215                                          addr, data);
216     return data;
217 }
218 
219 static void vfio_generic_quirk_mirror_write(void *opaque, hwaddr addr,
220                                             uint64_t data, unsigned size)
221 {
222     VFIOConfigMirrorQuirk *mirror = opaque;
223     VFIOPCIDevice *vdev = mirror->vdev;
224 
225     vfio_pci_write_config(&vdev->pdev, addr, data, size);
226     trace_vfio_quirk_generic_mirror_write(vdev->vbasedev.name,
227                                           memory_region_name(mirror->mem),
228                                           addr, data);
229 }
230 
231 static const MemoryRegionOps vfio_generic_mirror_quirk = {
232     .read = vfio_generic_quirk_mirror_read,
233     .write = vfio_generic_quirk_mirror_write,
234     .endianness = DEVICE_LITTLE_ENDIAN,
235 };
236 
237 /* Is range1 fully contained within range2?  */
238 static bool vfio_range_contained(uint64_t first1, uint64_t len1,
239                                  uint64_t first2, uint64_t len2) {
240     return (first1 >= first2 && first1 + len1 <= first2 + len2);
241 }
242 
243 #define PCI_VENDOR_ID_ATI               0x1002
244 
245 /*
246  * Radeon HD cards (HD5450 & HD7850) report the upper byte of the I/O port BAR
247  * through VGA register 0x3c3.  On newer cards, the I/O port BAR is always
248  * BAR4 (older cards like the X550 used BAR1, but we don't care to support
249  * those).  Note that on bare metal, a read of 0x3c3 doesn't always return the
250  * I/O port BAR address.  Originally this was coded to return the virtual BAR
251  * address only if the physical register read returns the actual BAR address,
252  * but users have reported greater success if we return the virtual address
253  * unconditionally.
254  */
255 static uint64_t vfio_ati_3c3_quirk_read(void *opaque,
256                                         hwaddr addr, unsigned size)
257 {
258     VFIOPCIDevice *vdev = opaque;
259     uint64_t data = vfio_pci_read_config(&vdev->pdev,
260                                          PCI_BASE_ADDRESS_4 + 1, size);
261 
262     trace_vfio_quirk_ati_3c3_read(vdev->vbasedev.name, data);
263 
264     return data;
265 }
266 
267 static void vfio_ati_3c3_quirk_write(void *opaque, hwaddr addr,
268                                         uint64_t data, unsigned size)
269 {
270     qemu_log_mask(LOG_GUEST_ERROR, "%s: invalid access\n", __func__);
271 }
272 
273 static const MemoryRegionOps vfio_ati_3c3_quirk = {
274     .read = vfio_ati_3c3_quirk_read,
275     .write = vfio_ati_3c3_quirk_write,
276     .endianness = DEVICE_LITTLE_ENDIAN,
277 };
278 
279 VFIOQuirk *vfio_quirk_alloc(int nr_mem)
280 {
281     VFIOQuirk *quirk = g_new0(VFIOQuirk, 1);
282     QLIST_INIT(&quirk->ioeventfds);
283     quirk->mem = g_new0(MemoryRegion, nr_mem);
284     quirk->nr_mem = nr_mem;
285 
286     return quirk;
287 }
288 
289 static void vfio_ioeventfd_exit(VFIOPCIDevice *vdev, VFIOIOEventFD *ioeventfd)
290 {
291     QLIST_REMOVE(ioeventfd, next);
292     memory_region_del_eventfd(ioeventfd->mr, ioeventfd->addr, ioeventfd->size,
293                               true, ioeventfd->data, &ioeventfd->e);
294 
295     if (ioeventfd->vfio) {
296         struct vfio_device_ioeventfd vfio_ioeventfd;
297 
298         vfio_ioeventfd.argsz = sizeof(vfio_ioeventfd);
299         vfio_ioeventfd.flags = ioeventfd->size;
300         vfio_ioeventfd.data = ioeventfd->data;
301         vfio_ioeventfd.offset = ioeventfd->region->fd_offset +
302                                 ioeventfd->region_addr;
303         vfio_ioeventfd.fd = -1;
304 
305         if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_IOEVENTFD, &vfio_ioeventfd)) {
306             error_report("Failed to remove vfio ioeventfd for %s+0x%"
307                          HWADDR_PRIx"[%d]:0x%"PRIx64" (%m)",
308                          memory_region_name(ioeventfd->mr), ioeventfd->addr,
309                          ioeventfd->size, ioeventfd->data);
310         }
311     } else {
312         qemu_set_fd_handler(event_notifier_get_fd(&ioeventfd->e),
313                             NULL, NULL, NULL);
314     }
315 
316     event_notifier_cleanup(&ioeventfd->e);
317     trace_vfio_ioeventfd_exit(memory_region_name(ioeventfd->mr),
318                               (uint64_t)ioeventfd->addr, ioeventfd->size,
319                               ioeventfd->data);
320     g_free(ioeventfd);
321 }
322 
323 static void vfio_drop_dynamic_eventfds(VFIOPCIDevice *vdev, VFIOQuirk *quirk)
324 {
325     VFIOIOEventFD *ioeventfd, *tmp;
326 
327     QLIST_FOREACH_SAFE(ioeventfd, &quirk->ioeventfds, next, tmp) {
328         if (ioeventfd->dynamic) {
329             vfio_ioeventfd_exit(vdev, ioeventfd);
330         }
331     }
332 }
333 
334 static void vfio_ioeventfd_handler(void *opaque)
335 {
336     VFIOIOEventFD *ioeventfd = opaque;
337 
338     if (event_notifier_test_and_clear(&ioeventfd->e)) {
339         vfio_region_write(ioeventfd->region, ioeventfd->region_addr,
340                           ioeventfd->data, ioeventfd->size);
341         trace_vfio_ioeventfd_handler(memory_region_name(ioeventfd->mr),
342                                      (uint64_t)ioeventfd->addr, ioeventfd->size,
343                                      ioeventfd->data);
344     }
345 }
346 
347 static VFIOIOEventFD *vfio_ioeventfd_init(VFIOPCIDevice *vdev,
348                                           MemoryRegion *mr, hwaddr addr,
349                                           unsigned size, uint64_t data,
350                                           VFIORegion *region,
351                                           hwaddr region_addr, bool dynamic)
352 {
353     VFIOIOEventFD *ioeventfd;
354 
355     if (vdev->no_kvm_ioeventfd) {
356         return NULL;
357     }
358 
359     ioeventfd = g_malloc0(sizeof(*ioeventfd));
360 
361     if (event_notifier_init(&ioeventfd->e, 0)) {
362         g_free(ioeventfd);
363         return NULL;
364     }
365 
366     /*
367      * MemoryRegion and relative offset, plus additional ioeventfd setup
368      * parameters for configuring and later tearing down KVM ioeventfd.
369      */
370     ioeventfd->mr = mr;
371     ioeventfd->addr = addr;
372     ioeventfd->size = size;
373     ioeventfd->data = data;
374     ioeventfd->dynamic = dynamic;
375     /*
376      * VFIORegion and relative offset for implementing the userspace
377      * handler.  data & size fields shared for both uses.
378      */
379     ioeventfd->region = region;
380     ioeventfd->region_addr = region_addr;
381 
382     if (!vdev->no_vfio_ioeventfd) {
383         struct vfio_device_ioeventfd vfio_ioeventfd;
384 
385         vfio_ioeventfd.argsz = sizeof(vfio_ioeventfd);
386         vfio_ioeventfd.flags = ioeventfd->size;
387         vfio_ioeventfd.data = ioeventfd->data;
388         vfio_ioeventfd.offset = ioeventfd->region->fd_offset +
389                                 ioeventfd->region_addr;
390         vfio_ioeventfd.fd = event_notifier_get_fd(&ioeventfd->e);
391 
392         ioeventfd->vfio = !ioctl(vdev->vbasedev.fd,
393                                  VFIO_DEVICE_IOEVENTFD, &vfio_ioeventfd);
394     }
395 
396     if (!ioeventfd->vfio) {
397         qemu_set_fd_handler(event_notifier_get_fd(&ioeventfd->e),
398                             vfio_ioeventfd_handler, NULL, ioeventfd);
399     }
400 
401     memory_region_add_eventfd(ioeventfd->mr, ioeventfd->addr, ioeventfd->size,
402                               true, ioeventfd->data, &ioeventfd->e);
403     trace_vfio_ioeventfd_init(memory_region_name(mr), (uint64_t)addr,
404                               size, data, ioeventfd->vfio);
405 
406     return ioeventfd;
407 }
408 
409 static void vfio_vga_probe_ati_3c3_quirk(VFIOPCIDevice *vdev)
410 {
411     VFIOQuirk *quirk;
412 
413     /*
414      * As long as the BAR is >= 256 bytes it will be aligned such that the
415      * lower byte is always zero.  Filter out anything else, if it exists.
416      */
417     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_ATI, PCI_ANY_ID) ||
418         !vdev->bars[4].ioport || vdev->bars[4].region.size < 256) {
419         return;
420     }
421 
422     quirk = vfio_quirk_alloc(1);
423 
424     memory_region_init_io(quirk->mem, OBJECT(vdev), &vfio_ati_3c3_quirk, vdev,
425                           "vfio-ati-3c3-quirk", 1);
426     memory_region_add_subregion(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
427                                 3 /* offset 3 bytes from 0x3c0 */, quirk->mem);
428 
429     QLIST_INSERT_HEAD(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].quirks,
430                       quirk, next);
431 
432     trace_vfio_quirk_ati_3c3_probe(vdev->vbasedev.name);
433 }
434 
435 /*
436  * Newer ATI/AMD devices, including HD5450 and HD7850, have a mirror to PCI
437  * config space through MMIO BAR2 at offset 0x4000.  Nothing seems to access
438  * the MMIO space directly, but a window to this space is provided through
439  * I/O port BAR4.  Offset 0x0 is the address register and offset 0x4 is the
440  * data register.  When the address is programmed to a range of 0x4000-0x4fff
441  * PCI configuration space is available.  Experimentation seems to indicate
442  * that read-only may be provided by hardware.
443  */
444 static void vfio_probe_ati_bar4_quirk(VFIOPCIDevice *vdev, int nr)
445 {
446     VFIOQuirk *quirk;
447     VFIOConfigWindowQuirk *window;
448 
449     /* This windows doesn't seem to be used except by legacy VGA code */
450     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_ATI, PCI_ANY_ID) ||
451         !vdev->vga || nr != 4) {
452         return;
453     }
454 
455     quirk = vfio_quirk_alloc(2);
456     window = quirk->data = g_malloc0(sizeof(*window) +
457                                      sizeof(VFIOConfigWindowMatch));
458     window->vdev = vdev;
459     window->address_offset = 0;
460     window->data_offset = 4;
461     window->nr_matches = 1;
462     window->matches[0].match = 0x4000;
463     window->matches[0].mask = vdev->config_size - 1;
464     window->bar = nr;
465     window->addr_mem = &quirk->mem[0];
466     window->data_mem = &quirk->mem[1];
467 
468     memory_region_init_io(window->addr_mem, OBJECT(vdev),
469                           &vfio_generic_window_address_quirk, window,
470                           "vfio-ati-bar4-window-address-quirk", 4);
471     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
472                                         window->address_offset,
473                                         window->addr_mem, 1);
474 
475     memory_region_init_io(window->data_mem, OBJECT(vdev),
476                           &vfio_generic_window_data_quirk, window,
477                           "vfio-ati-bar4-window-data-quirk", 4);
478     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
479                                         window->data_offset,
480                                         window->data_mem, 1);
481 
482     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
483 
484     trace_vfio_quirk_ati_bar4_probe(vdev->vbasedev.name);
485 }
486 
487 /*
488  * Trap the BAR2 MMIO mirror to config space as well.
489  */
490 static void vfio_probe_ati_bar2_quirk(VFIOPCIDevice *vdev, int nr)
491 {
492     VFIOQuirk *quirk;
493     VFIOConfigMirrorQuirk *mirror;
494 
495     /* Only enable on newer devices where BAR2 is 64bit */
496     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_ATI, PCI_ANY_ID) ||
497         !vdev->vga || nr != 2 || !vdev->bars[2].mem64) {
498         return;
499     }
500 
501     quirk = vfio_quirk_alloc(1);
502     mirror = quirk->data = g_malloc0(sizeof(*mirror));
503     mirror->mem = quirk->mem;
504     mirror->vdev = vdev;
505     mirror->offset = 0x4000;
506     mirror->bar = nr;
507 
508     memory_region_init_io(mirror->mem, OBJECT(vdev),
509                           &vfio_generic_mirror_quirk, mirror,
510                           "vfio-ati-bar2-4000-quirk", PCI_CONFIG_SPACE_SIZE);
511     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
512                                         mirror->offset, mirror->mem, 1);
513 
514     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
515 
516     trace_vfio_quirk_ati_bar2_probe(vdev->vbasedev.name);
517 }
518 
519 /*
520  * Older ATI/AMD cards like the X550 have a similar window to that above.
521  * I/O port BAR1 provides a window to a mirror of PCI config space located
522  * in BAR2 at offset 0xf00.  We don't care to support such older cards, but
523  * note it for future reference.
524  */
525 
526 /*
527  * Nvidia has several different methods to get to config space, the
528  * nouveu project has several of these documented here:
529  * https://github.com/pathscale/envytools/tree/master/hwdocs
530  *
531  * The first quirk is actually not documented in envytools and is found
532  * on 10de:01d1 (NVIDIA Corporation G72 [GeForce 7300 LE]).  This is an
533  * NV46 chipset.  The backdoor uses the legacy VGA I/O ports to access
534  * the mirror of PCI config space found at BAR0 offset 0x1800.  The access
535  * sequence first writes 0x338 to I/O port 0x3d4.  The target offset is
536  * then written to 0x3d0.  Finally 0x538 is written for a read and 0x738
537  * is written for a write to 0x3d4.  The BAR0 offset is then accessible
538  * through 0x3d0.  This quirk doesn't seem to be necessary on newer cards
539  * that use the I/O port BAR5 window but it doesn't hurt to leave it.
540  */
541 typedef enum {NONE = 0, SELECT, WINDOW, READ, WRITE} VFIONvidia3d0State;
542 static const char *nv3d0_states[] = { "NONE", "SELECT",
543                                       "WINDOW", "READ", "WRITE" };
544 
545 typedef struct VFIONvidia3d0Quirk {
546     VFIOPCIDevice *vdev;
547     VFIONvidia3d0State state;
548     uint32_t offset;
549 } VFIONvidia3d0Quirk;
550 
551 static uint64_t vfio_nvidia_3d4_quirk_read(void *opaque,
552                                            hwaddr addr, unsigned size)
553 {
554     VFIONvidia3d0Quirk *quirk = opaque;
555     VFIOPCIDevice *vdev = quirk->vdev;
556 
557     quirk->state = NONE;
558 
559     return vfio_vga_read(&vdev->vga->region[QEMU_PCI_VGA_IO_HI],
560                          addr + 0x14, size);
561 }
562 
563 static void vfio_nvidia_3d4_quirk_write(void *opaque, hwaddr addr,
564                                         uint64_t data, unsigned size)
565 {
566     VFIONvidia3d0Quirk *quirk = opaque;
567     VFIOPCIDevice *vdev = quirk->vdev;
568     VFIONvidia3d0State old_state = quirk->state;
569 
570     quirk->state = NONE;
571 
572     switch (data) {
573     case 0x338:
574         if (old_state == NONE) {
575             quirk->state = SELECT;
576             trace_vfio_quirk_nvidia_3d0_state(vdev->vbasedev.name,
577                                               nv3d0_states[quirk->state]);
578         }
579         break;
580     case 0x538:
581         if (old_state == WINDOW) {
582             quirk->state = READ;
583             trace_vfio_quirk_nvidia_3d0_state(vdev->vbasedev.name,
584                                               nv3d0_states[quirk->state]);
585         }
586         break;
587     case 0x738:
588         if (old_state == WINDOW) {
589             quirk->state = WRITE;
590             trace_vfio_quirk_nvidia_3d0_state(vdev->vbasedev.name,
591                                               nv3d0_states[quirk->state]);
592         }
593         break;
594     }
595 
596     vfio_vga_write(&vdev->vga->region[QEMU_PCI_VGA_IO_HI],
597                    addr + 0x14, data, size);
598 }
599 
600 static const MemoryRegionOps vfio_nvidia_3d4_quirk = {
601     .read = vfio_nvidia_3d4_quirk_read,
602     .write = vfio_nvidia_3d4_quirk_write,
603     .endianness = DEVICE_LITTLE_ENDIAN,
604 };
605 
606 static uint64_t vfio_nvidia_3d0_quirk_read(void *opaque,
607                                            hwaddr addr, unsigned size)
608 {
609     VFIONvidia3d0Quirk *quirk = opaque;
610     VFIOPCIDevice *vdev = quirk->vdev;
611     VFIONvidia3d0State old_state = quirk->state;
612     uint64_t data = vfio_vga_read(&vdev->vga->region[QEMU_PCI_VGA_IO_HI],
613                                   addr + 0x10, size);
614 
615     quirk->state = NONE;
616 
617     if (old_state == READ &&
618         (quirk->offset & ~(PCI_CONFIG_SPACE_SIZE - 1)) == 0x1800) {
619         uint8_t offset = quirk->offset & (PCI_CONFIG_SPACE_SIZE - 1);
620 
621         data = vfio_pci_read_config(&vdev->pdev, offset, size);
622         trace_vfio_quirk_nvidia_3d0_read(vdev->vbasedev.name,
623                                          offset, size, data);
624     }
625 
626     return data;
627 }
628 
629 static void vfio_nvidia_3d0_quirk_write(void *opaque, hwaddr addr,
630                                         uint64_t data, unsigned size)
631 {
632     VFIONvidia3d0Quirk *quirk = opaque;
633     VFIOPCIDevice *vdev = quirk->vdev;
634     VFIONvidia3d0State old_state = quirk->state;
635 
636     quirk->state = NONE;
637 
638     if (old_state == SELECT) {
639         quirk->offset = (uint32_t)data;
640         quirk->state = WINDOW;
641         trace_vfio_quirk_nvidia_3d0_state(vdev->vbasedev.name,
642                                           nv3d0_states[quirk->state]);
643     } else if (old_state == WRITE) {
644         if ((quirk->offset & ~(PCI_CONFIG_SPACE_SIZE - 1)) == 0x1800) {
645             uint8_t offset = quirk->offset & (PCI_CONFIG_SPACE_SIZE - 1);
646 
647             vfio_pci_write_config(&vdev->pdev, offset, data, size);
648             trace_vfio_quirk_nvidia_3d0_write(vdev->vbasedev.name,
649                                               offset, data, size);
650             return;
651         }
652     }
653 
654     vfio_vga_write(&vdev->vga->region[QEMU_PCI_VGA_IO_HI],
655                    addr + 0x10, data, size);
656 }
657 
658 static const MemoryRegionOps vfio_nvidia_3d0_quirk = {
659     .read = vfio_nvidia_3d0_quirk_read,
660     .write = vfio_nvidia_3d0_quirk_write,
661     .endianness = DEVICE_LITTLE_ENDIAN,
662 };
663 
664 static void vfio_vga_probe_nvidia_3d0_quirk(VFIOPCIDevice *vdev)
665 {
666     VFIOQuirk *quirk;
667     VFIONvidia3d0Quirk *data;
668 
669     if (vdev->no_geforce_quirks ||
670         !vfio_pci_is(vdev, PCI_VENDOR_ID_NVIDIA, PCI_ANY_ID) ||
671         !vdev->bars[1].region.size) {
672         return;
673     }
674 
675     quirk = vfio_quirk_alloc(2);
676     quirk->data = data = g_malloc0(sizeof(*data));
677     data->vdev = vdev;
678 
679     memory_region_init_io(&quirk->mem[0], OBJECT(vdev), &vfio_nvidia_3d4_quirk,
680                           data, "vfio-nvidia-3d4-quirk", 2);
681     memory_region_add_subregion(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
682                                 0x14 /* 0x3c0 + 0x14 */, &quirk->mem[0]);
683 
684     memory_region_init_io(&quirk->mem[1], OBJECT(vdev), &vfio_nvidia_3d0_quirk,
685                           data, "vfio-nvidia-3d0-quirk", 2);
686     memory_region_add_subregion(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
687                                 0x10 /* 0x3c0 + 0x10 */, &quirk->mem[1]);
688 
689     QLIST_INSERT_HEAD(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].quirks,
690                       quirk, next);
691 
692     trace_vfio_quirk_nvidia_3d0_probe(vdev->vbasedev.name);
693 }
694 
695 /*
696  * The second quirk is documented in envytools.  The I/O port BAR5 is just
697  * a set of address/data ports to the MMIO BARs.  The BAR we care about is
698  * again BAR0.  This backdoor is apparently a bit newer than the one above
699  * so we need to not only trap 256 bytes @0x1800, but all of PCI config
700  * space, including extended space is available at the 4k @0x88000.
701  */
702 typedef struct VFIONvidiaBAR5Quirk {
703     uint32_t master;
704     uint32_t enable;
705     MemoryRegion *addr_mem;
706     MemoryRegion *data_mem;
707     bool enabled;
708     VFIOConfigWindowQuirk window; /* last for match data */
709 } VFIONvidiaBAR5Quirk;
710 
711 static void vfio_nvidia_bar5_enable(VFIONvidiaBAR5Quirk *bar5)
712 {
713     VFIOPCIDevice *vdev = bar5->window.vdev;
714 
715     if (((bar5->master & bar5->enable) & 0x1) == bar5->enabled) {
716         return;
717     }
718 
719     bar5->enabled = !bar5->enabled;
720     trace_vfio_quirk_nvidia_bar5_state(vdev->vbasedev.name,
721                                        bar5->enabled ?  "Enable" : "Disable");
722     memory_region_set_enabled(bar5->addr_mem, bar5->enabled);
723     memory_region_set_enabled(bar5->data_mem, bar5->enabled);
724 }
725 
726 static uint64_t vfio_nvidia_bar5_quirk_master_read(void *opaque,
727                                                    hwaddr addr, unsigned size)
728 {
729     VFIONvidiaBAR5Quirk *bar5 = opaque;
730     VFIOPCIDevice *vdev = bar5->window.vdev;
731 
732     return vfio_region_read(&vdev->bars[5].region, addr, size);
733 }
734 
735 static void vfio_nvidia_bar5_quirk_master_write(void *opaque, hwaddr addr,
736                                                 uint64_t data, unsigned size)
737 {
738     VFIONvidiaBAR5Quirk *bar5 = opaque;
739     VFIOPCIDevice *vdev = bar5->window.vdev;
740 
741     vfio_region_write(&vdev->bars[5].region, addr, data, size);
742 
743     bar5->master = data;
744     vfio_nvidia_bar5_enable(bar5);
745 }
746 
747 static const MemoryRegionOps vfio_nvidia_bar5_quirk_master = {
748     .read = vfio_nvidia_bar5_quirk_master_read,
749     .write = vfio_nvidia_bar5_quirk_master_write,
750     .endianness = DEVICE_LITTLE_ENDIAN,
751 };
752 
753 static uint64_t vfio_nvidia_bar5_quirk_enable_read(void *opaque,
754                                                    hwaddr addr, unsigned size)
755 {
756     VFIONvidiaBAR5Quirk *bar5 = opaque;
757     VFIOPCIDevice *vdev = bar5->window.vdev;
758 
759     return vfio_region_read(&vdev->bars[5].region, addr + 4, size);
760 }
761 
762 static void vfio_nvidia_bar5_quirk_enable_write(void *opaque, hwaddr addr,
763                                                 uint64_t data, unsigned size)
764 {
765     VFIONvidiaBAR5Quirk *bar5 = opaque;
766     VFIOPCIDevice *vdev = bar5->window.vdev;
767 
768     vfio_region_write(&vdev->bars[5].region, addr + 4, data, size);
769 
770     bar5->enable = data;
771     vfio_nvidia_bar5_enable(bar5);
772 }
773 
774 static const MemoryRegionOps vfio_nvidia_bar5_quirk_enable = {
775     .read = vfio_nvidia_bar5_quirk_enable_read,
776     .write = vfio_nvidia_bar5_quirk_enable_write,
777     .endianness = DEVICE_LITTLE_ENDIAN,
778 };
779 
780 static void vfio_probe_nvidia_bar5_quirk(VFIOPCIDevice *vdev, int nr)
781 {
782     VFIOQuirk *quirk;
783     VFIONvidiaBAR5Quirk *bar5;
784     VFIOConfigWindowQuirk *window;
785 
786     if (vdev->no_geforce_quirks ||
787         !vfio_pci_is(vdev, PCI_VENDOR_ID_NVIDIA, PCI_ANY_ID) ||
788         !vdev->vga || nr != 5 || !vdev->bars[5].ioport) {
789         return;
790     }
791 
792     quirk = vfio_quirk_alloc(4);
793     bar5 = quirk->data = g_malloc0(sizeof(*bar5) +
794                                    (sizeof(VFIOConfigWindowMatch) * 2));
795     window = &bar5->window;
796 
797     window->vdev = vdev;
798     window->address_offset = 0x8;
799     window->data_offset = 0xc;
800     window->nr_matches = 2;
801     window->matches[0].match = 0x1800;
802     window->matches[0].mask = PCI_CONFIG_SPACE_SIZE - 1;
803     window->matches[1].match = 0x88000;
804     window->matches[1].mask = vdev->config_size - 1;
805     window->bar = nr;
806     window->addr_mem = bar5->addr_mem = &quirk->mem[0];
807     window->data_mem = bar5->data_mem = &quirk->mem[1];
808 
809     memory_region_init_io(window->addr_mem, OBJECT(vdev),
810                           &vfio_generic_window_address_quirk, window,
811                           "vfio-nvidia-bar5-window-address-quirk", 4);
812     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
813                                         window->address_offset,
814                                         window->addr_mem, 1);
815     memory_region_set_enabled(window->addr_mem, false);
816 
817     memory_region_init_io(window->data_mem, OBJECT(vdev),
818                           &vfio_generic_window_data_quirk, window,
819                           "vfio-nvidia-bar5-window-data-quirk", 4);
820     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
821                                         window->data_offset,
822                                         window->data_mem, 1);
823     memory_region_set_enabled(window->data_mem, false);
824 
825     memory_region_init_io(&quirk->mem[2], OBJECT(vdev),
826                           &vfio_nvidia_bar5_quirk_master, bar5,
827                           "vfio-nvidia-bar5-master-quirk", 4);
828     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
829                                         0, &quirk->mem[2], 1);
830 
831     memory_region_init_io(&quirk->mem[3], OBJECT(vdev),
832                           &vfio_nvidia_bar5_quirk_enable, bar5,
833                           "vfio-nvidia-bar5-enable-quirk", 4);
834     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
835                                         4, &quirk->mem[3], 1);
836 
837     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
838 
839     trace_vfio_quirk_nvidia_bar5_probe(vdev->vbasedev.name);
840 }
841 
842 typedef struct LastDataSet {
843     VFIOQuirk *quirk;
844     hwaddr addr;
845     uint64_t data;
846     unsigned size;
847     int hits;
848     int added;
849 } LastDataSet;
850 
851 #define MAX_DYN_IOEVENTFD 10
852 #define HITS_FOR_IOEVENTFD 10
853 
854 /*
855  * Finally, BAR0 itself.  We want to redirect any accesses to either
856  * 0x1800 or 0x88000 through the PCI config space access functions.
857  */
858 static void vfio_nvidia_quirk_mirror_write(void *opaque, hwaddr addr,
859                                            uint64_t data, unsigned size)
860 {
861     VFIOConfigMirrorQuirk *mirror = opaque;
862     VFIOPCIDevice *vdev = mirror->vdev;
863     PCIDevice *pdev = &vdev->pdev;
864     LastDataSet *last = (LastDataSet *)&mirror->data;
865 
866     vfio_generic_quirk_mirror_write(opaque, addr, data, size);
867 
868     /*
869      * Nvidia seems to acknowledge MSI interrupts by writing 0xff to the
870      * MSI capability ID register.  Both the ID and next register are
871      * read-only, so we allow writes covering either of those to real hw.
872      */
873     if ((pdev->cap_present & QEMU_PCI_CAP_MSI) &&
874         vfio_range_contained(addr, size, pdev->msi_cap, PCI_MSI_FLAGS)) {
875         vfio_region_write(&vdev->bars[mirror->bar].region,
876                           addr + mirror->offset, data, size);
877         trace_vfio_quirk_nvidia_bar0_msi_ack(vdev->vbasedev.name);
878     }
879 
880     /*
881      * Automatically add an ioeventfd to handle any repeated write with the
882      * same data and size above the standard PCI config space header.  This is
883      * primarily expected to accelerate the MSI-ACK behavior, such as noted
884      * above.  Current hardware/drivers should trigger an ioeventfd at config
885      * offset 0x704 (region offset 0x88704), with data 0x0, size 4.
886      *
887      * The criteria of 10 successive hits is arbitrary but reliably adds the
888      * MSI-ACK region.  Note that as some writes are bypassed via the ioeventfd,
889      * the remaining ones have a greater chance of being seen successively.
890      * To avoid the pathological case of burning up all of QEMU's open file
891      * handles, arbitrarily limit this algorithm from adding no more than 10
892      * ioeventfds, print an error if we would have added an 11th, and then
893      * stop counting.
894      */
895     if (!vdev->no_kvm_ioeventfd &&
896         addr >= PCI_STD_HEADER_SIZEOF && last->added <= MAX_DYN_IOEVENTFD) {
897         if (addr != last->addr || data != last->data || size != last->size) {
898             last->addr = addr;
899             last->data = data;
900             last->size = size;
901             last->hits = 1;
902         } else if (++last->hits >= HITS_FOR_IOEVENTFD) {
903             if (last->added < MAX_DYN_IOEVENTFD) {
904                 VFIOIOEventFD *ioeventfd;
905                 ioeventfd = vfio_ioeventfd_init(vdev, mirror->mem, addr, size,
906                                         data, &vdev->bars[mirror->bar].region,
907                                         mirror->offset + addr, true);
908                 if (ioeventfd) {
909                     VFIOQuirk *quirk = last->quirk;
910 
911                     QLIST_INSERT_HEAD(&quirk->ioeventfds, ioeventfd, next);
912                     last->added++;
913                 }
914             } else {
915                 last->added++;
916                 warn_report("NVIDIA ioeventfd queue full for %s, unable to "
917                             "accelerate 0x%"HWADDR_PRIx", data 0x%"PRIx64", "
918                             "size %u", vdev->vbasedev.name, addr, data, size);
919             }
920         }
921     }
922 }
923 
924 static const MemoryRegionOps vfio_nvidia_mirror_quirk = {
925     .read = vfio_generic_quirk_mirror_read,
926     .write = vfio_nvidia_quirk_mirror_write,
927     .endianness = DEVICE_LITTLE_ENDIAN,
928 };
929 
930 static void vfio_nvidia_bar0_quirk_reset(VFIOPCIDevice *vdev, VFIOQuirk *quirk)
931 {
932     VFIOConfigMirrorQuirk *mirror = quirk->data;
933     LastDataSet *last = (LastDataSet *)&mirror->data;
934 
935     last->addr = last->data = last->size = last->hits = last->added = 0;
936 
937     vfio_drop_dynamic_eventfds(vdev, quirk);
938 }
939 
940 static void vfio_probe_nvidia_bar0_quirk(VFIOPCIDevice *vdev, int nr)
941 {
942     VFIOQuirk *quirk;
943     VFIOConfigMirrorQuirk *mirror;
944     LastDataSet *last;
945 
946     if (vdev->no_geforce_quirks ||
947         !vfio_pci_is(vdev, PCI_VENDOR_ID_NVIDIA, PCI_ANY_ID) ||
948         !vfio_is_vga(vdev) || nr != 0) {
949         return;
950     }
951 
952     quirk = vfio_quirk_alloc(1);
953     quirk->reset = vfio_nvidia_bar0_quirk_reset;
954     mirror = quirk->data = g_malloc0(sizeof(*mirror) + sizeof(LastDataSet));
955     mirror->mem = quirk->mem;
956     mirror->vdev = vdev;
957     mirror->offset = 0x88000;
958     mirror->bar = nr;
959     last = (LastDataSet *)&mirror->data;
960     last->quirk = quirk;
961 
962     memory_region_init_io(mirror->mem, OBJECT(vdev),
963                           &vfio_nvidia_mirror_quirk, mirror,
964                           "vfio-nvidia-bar0-88000-mirror-quirk",
965                           vdev->config_size);
966     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
967                                         mirror->offset, mirror->mem, 1);
968 
969     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
970 
971     /* The 0x1800 offset mirror only seems to get used by legacy VGA */
972     if (vdev->vga) {
973         quirk = vfio_quirk_alloc(1);
974         quirk->reset = vfio_nvidia_bar0_quirk_reset;
975         mirror = quirk->data = g_malloc0(sizeof(*mirror) + sizeof(LastDataSet));
976         mirror->mem = quirk->mem;
977         mirror->vdev = vdev;
978         mirror->offset = 0x1800;
979         mirror->bar = nr;
980         last = (LastDataSet *)&mirror->data;
981         last->quirk = quirk;
982 
983         memory_region_init_io(mirror->mem, OBJECT(vdev),
984                               &vfio_nvidia_mirror_quirk, mirror,
985                               "vfio-nvidia-bar0-1800-mirror-quirk",
986                               PCI_CONFIG_SPACE_SIZE);
987         memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
988                                             mirror->offset, mirror->mem, 1);
989 
990         QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
991     }
992 
993     trace_vfio_quirk_nvidia_bar0_probe(vdev->vbasedev.name);
994 }
995 
996 /*
997  * TODO - Some Nvidia devices provide config access to their companion HDA
998  * device and even to their parent bridge via these config space mirrors.
999  * Add quirks for those regions.
1000  */
1001 
1002 #define PCI_VENDOR_ID_REALTEK 0x10ec
1003 
1004 /*
1005  * RTL8168 devices have a backdoor that can access the MSI-X table.  At BAR2
1006  * offset 0x70 there is a dword data register, offset 0x74 is a dword address
1007  * register.  According to the Linux r8169 driver, the MSI-X table is addressed
1008  * when the "type" portion of the address register is set to 0x1.  This appears
1009  * to be bits 16:30.  Bit 31 is both a write indicator and some sort of
1010  * "address latched" indicator.  Bits 12:15 are a mask field, which we can
1011  * ignore because the MSI-X table should always be accessed as a dword (full
1012  * mask).  Bits 0:11 is offset within the type.
1013  *
1014  * Example trace:
1015  *
1016  * Read from MSI-X table offset 0
1017  * vfio: vfio_bar_write(0000:05:00.0:BAR2+0x74, 0x1f000, 4) // store read addr
1018  * vfio: vfio_bar_read(0000:05:00.0:BAR2+0x74, 4) = 0x8001f000 // latch
1019  * vfio: vfio_bar_read(0000:05:00.0:BAR2+0x70, 4) = 0xfee00398 // read data
1020  *
1021  * Write 0xfee00000 to MSI-X table offset 0
1022  * vfio: vfio_bar_write(0000:05:00.0:BAR2+0x70, 0xfee00000, 4) // write data
1023  * vfio: vfio_bar_write(0000:05:00.0:BAR2+0x74, 0x8001f000, 4) // do write
1024  * vfio: vfio_bar_read(0000:05:00.0:BAR2+0x74, 4) = 0x1f000 // complete
1025  */
1026 typedef struct VFIOrtl8168Quirk {
1027     VFIOPCIDevice *vdev;
1028     uint32_t addr;
1029     uint32_t data;
1030     bool enabled;
1031 } VFIOrtl8168Quirk;
1032 
1033 static uint64_t vfio_rtl8168_quirk_address_read(void *opaque,
1034                                                 hwaddr addr, unsigned size)
1035 {
1036     VFIOrtl8168Quirk *rtl = opaque;
1037     VFIOPCIDevice *vdev = rtl->vdev;
1038     uint64_t data = vfio_region_read(&vdev->bars[2].region, addr + 0x74, size);
1039 
1040     if (rtl->enabled) {
1041         data = rtl->addr ^ 0x80000000U; /* latch/complete */
1042         trace_vfio_quirk_rtl8168_fake_latch(vdev->vbasedev.name, data);
1043     }
1044 
1045     return data;
1046 }
1047 
1048 static void vfio_rtl8168_quirk_address_write(void *opaque, hwaddr addr,
1049                                              uint64_t data, unsigned size)
1050 {
1051     VFIOrtl8168Quirk *rtl = opaque;
1052     VFIOPCIDevice *vdev = rtl->vdev;
1053 
1054     rtl->enabled = false;
1055 
1056     if ((data & 0x7fff0000) == 0x10000) { /* MSI-X table */
1057         rtl->enabled = true;
1058         rtl->addr = (uint32_t)data;
1059 
1060         if (data & 0x80000000U) { /* Do write */
1061             if (vdev->pdev.cap_present & QEMU_PCI_CAP_MSIX) {
1062                 hwaddr offset = data & 0xfff;
1063                 uint64_t val = rtl->data;
1064 
1065                 trace_vfio_quirk_rtl8168_msix_write(vdev->vbasedev.name,
1066                                                     (uint16_t)offset, val);
1067 
1068                 /* Write to the proper guest MSI-X table instead */
1069                 memory_region_dispatch_write(&vdev->pdev.msix_table_mmio,
1070                                              offset, val,
1071                                              size_memop(size) | MO_LE,
1072                                              MEMTXATTRS_UNSPECIFIED);
1073             }
1074             return; /* Do not write guest MSI-X data to hardware */
1075         }
1076     }
1077 
1078     vfio_region_write(&vdev->bars[2].region, addr + 0x74, data, size);
1079 }
1080 
1081 static const MemoryRegionOps vfio_rtl_address_quirk = {
1082     .read = vfio_rtl8168_quirk_address_read,
1083     .write = vfio_rtl8168_quirk_address_write,
1084     .valid = {
1085         .min_access_size = 4,
1086         .max_access_size = 4,
1087         .unaligned = false,
1088     },
1089     .endianness = DEVICE_LITTLE_ENDIAN,
1090 };
1091 
1092 static uint64_t vfio_rtl8168_quirk_data_read(void *opaque,
1093                                              hwaddr addr, unsigned size)
1094 {
1095     VFIOrtl8168Quirk *rtl = opaque;
1096     VFIOPCIDevice *vdev = rtl->vdev;
1097     uint64_t data = vfio_region_read(&vdev->bars[2].region, addr + 0x70, size);
1098 
1099     if (rtl->enabled && (vdev->pdev.cap_present & QEMU_PCI_CAP_MSIX)) {
1100         hwaddr offset = rtl->addr & 0xfff;
1101         memory_region_dispatch_read(&vdev->pdev.msix_table_mmio, offset,
1102                                     &data, size_memop(size) | MO_LE,
1103                                     MEMTXATTRS_UNSPECIFIED);
1104         trace_vfio_quirk_rtl8168_msix_read(vdev->vbasedev.name, offset, data);
1105     }
1106 
1107     return data;
1108 }
1109 
1110 static void vfio_rtl8168_quirk_data_write(void *opaque, hwaddr addr,
1111                                           uint64_t data, unsigned size)
1112 {
1113     VFIOrtl8168Quirk *rtl = opaque;
1114     VFIOPCIDevice *vdev = rtl->vdev;
1115 
1116     rtl->data = (uint32_t)data;
1117 
1118     vfio_region_write(&vdev->bars[2].region, addr + 0x70, data, size);
1119 }
1120 
1121 static const MemoryRegionOps vfio_rtl_data_quirk = {
1122     .read = vfio_rtl8168_quirk_data_read,
1123     .write = vfio_rtl8168_quirk_data_write,
1124     .valid = {
1125         .min_access_size = 4,
1126         .max_access_size = 4,
1127         .unaligned = false,
1128     },
1129     .endianness = DEVICE_LITTLE_ENDIAN,
1130 };
1131 
1132 static void vfio_probe_rtl8168_bar2_quirk(VFIOPCIDevice *vdev, int nr)
1133 {
1134     VFIOQuirk *quirk;
1135     VFIOrtl8168Quirk *rtl;
1136 
1137     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_REALTEK, 0x8168) || nr != 2) {
1138         return;
1139     }
1140 
1141     quirk = vfio_quirk_alloc(2);
1142     quirk->data = rtl = g_malloc0(sizeof(*rtl));
1143     rtl->vdev = vdev;
1144 
1145     memory_region_init_io(&quirk->mem[0], OBJECT(vdev),
1146                           &vfio_rtl_address_quirk, rtl,
1147                           "vfio-rtl8168-window-address-quirk", 4);
1148     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
1149                                         0x74, &quirk->mem[0], 1);
1150 
1151     memory_region_init_io(&quirk->mem[1], OBJECT(vdev),
1152                           &vfio_rtl_data_quirk, rtl,
1153                           "vfio-rtl8168-window-data-quirk", 4);
1154     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
1155                                         0x70, &quirk->mem[1], 1);
1156 
1157     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1158 
1159     trace_vfio_quirk_rtl8168_probe(vdev->vbasedev.name);
1160 }
1161 
1162 #define IGD_ASLS 0xfc /* ASL Storage Register */
1163 
1164 /*
1165  * The OpRegion includes the Video BIOS Table, which seems important for
1166  * telling the driver what sort of outputs it has.  Without this, the device
1167  * may work in the guest, but we may not get output.  This also requires BIOS
1168  * support to reserve and populate a section of guest memory sufficient for
1169  * the table and to write the base address of that memory to the ASLS register
1170  * of the IGD device.
1171  */
1172 int vfio_pci_igd_opregion_init(VFIOPCIDevice *vdev,
1173                                struct vfio_region_info *info, Error **errp)
1174 {
1175     int ret;
1176 
1177     vdev->igd_opregion = g_malloc0(info->size);
1178     ret = pread(vdev->vbasedev.fd, vdev->igd_opregion,
1179                 info->size, info->offset);
1180     if (ret != info->size) {
1181         error_setg(errp, "failed to read IGD OpRegion");
1182         g_free(vdev->igd_opregion);
1183         vdev->igd_opregion = NULL;
1184         return -EINVAL;
1185     }
1186 
1187     /*
1188      * Provide fw_cfg with a copy of the OpRegion which the VM firmware is to
1189      * allocate 32bit reserved memory for, copy these contents into, and write
1190      * the reserved memory base address to the device ASLS register at 0xFC.
1191      * Alignment of this reserved region seems flexible, but using a 4k page
1192      * alignment seems to work well.  This interface assumes a single IGD
1193      * device, which may be at VM address 00:02.0 in legacy mode or another
1194      * address in UPT mode.
1195      *
1196      * NB, there may be future use cases discovered where the VM should have
1197      * direct interaction with the host OpRegion, in which case the write to
1198      * the ASLS register would trigger MemoryRegion setup to enable that.
1199      */
1200     fw_cfg_add_file(fw_cfg_find(), "etc/igd-opregion",
1201                     vdev->igd_opregion, info->size);
1202 
1203     trace_vfio_pci_igd_opregion_enabled(vdev->vbasedev.name);
1204 
1205     pci_set_long(vdev->pdev.config + IGD_ASLS, 0);
1206     pci_set_long(vdev->pdev.wmask + IGD_ASLS, ~0);
1207     pci_set_long(vdev->emulated_config_bits + IGD_ASLS, ~0);
1208 
1209     return 0;
1210 }
1211 
1212 /*
1213  * Common quirk probe entry points.
1214  */
1215 void vfio_vga_quirk_setup(VFIOPCIDevice *vdev)
1216 {
1217     vfio_vga_probe_ati_3c3_quirk(vdev);
1218     vfio_vga_probe_nvidia_3d0_quirk(vdev);
1219 }
1220 
1221 void vfio_vga_quirk_exit(VFIOPCIDevice *vdev)
1222 {
1223     VFIOQuirk *quirk;
1224     int i, j;
1225 
1226     for (i = 0; i < ARRAY_SIZE(vdev->vga->region); i++) {
1227         QLIST_FOREACH(quirk, &vdev->vga->region[i].quirks, next) {
1228             for (j = 0; j < quirk->nr_mem; j++) {
1229                 memory_region_del_subregion(&vdev->vga->region[i].mem,
1230                                             &quirk->mem[j]);
1231             }
1232         }
1233     }
1234 }
1235 
1236 void vfio_vga_quirk_finalize(VFIOPCIDevice *vdev)
1237 {
1238     int i, j;
1239 
1240     for (i = 0; i < ARRAY_SIZE(vdev->vga->region); i++) {
1241         while (!QLIST_EMPTY(&vdev->vga->region[i].quirks)) {
1242             VFIOQuirk *quirk = QLIST_FIRST(&vdev->vga->region[i].quirks);
1243             QLIST_REMOVE(quirk, next);
1244             for (j = 0; j < quirk->nr_mem; j++) {
1245                 object_unparent(OBJECT(&quirk->mem[j]));
1246             }
1247             g_free(quirk->mem);
1248             g_free(quirk->data);
1249             g_free(quirk);
1250         }
1251     }
1252 }
1253 
1254 void vfio_bar_quirk_setup(VFIOPCIDevice *vdev, int nr)
1255 {
1256     vfio_probe_ati_bar4_quirk(vdev, nr);
1257     vfio_probe_ati_bar2_quirk(vdev, nr);
1258     vfio_probe_nvidia_bar5_quirk(vdev, nr);
1259     vfio_probe_nvidia_bar0_quirk(vdev, nr);
1260     vfio_probe_rtl8168_bar2_quirk(vdev, nr);
1261 #ifdef CONFIG_VFIO_IGD
1262     vfio_probe_igd_bar4_quirk(vdev, nr);
1263 #endif
1264 }
1265 
1266 void vfio_bar_quirk_exit(VFIOPCIDevice *vdev, int nr)
1267 {
1268     VFIOBAR *bar = &vdev->bars[nr];
1269     VFIOQuirk *quirk;
1270     int i;
1271 
1272     QLIST_FOREACH(quirk, &bar->quirks, next) {
1273         while (!QLIST_EMPTY(&quirk->ioeventfds)) {
1274             vfio_ioeventfd_exit(vdev, QLIST_FIRST(&quirk->ioeventfds));
1275         }
1276 
1277         for (i = 0; i < quirk->nr_mem; i++) {
1278             memory_region_del_subregion(bar->region.mem, &quirk->mem[i]);
1279         }
1280     }
1281 }
1282 
1283 void vfio_bar_quirk_finalize(VFIOPCIDevice *vdev, int nr)
1284 {
1285     VFIOBAR *bar = &vdev->bars[nr];
1286     int i;
1287 
1288     while (!QLIST_EMPTY(&bar->quirks)) {
1289         VFIOQuirk *quirk = QLIST_FIRST(&bar->quirks);
1290         QLIST_REMOVE(quirk, next);
1291         for (i = 0; i < quirk->nr_mem; i++) {
1292             object_unparent(OBJECT(&quirk->mem[i]));
1293         }
1294         g_free(quirk->mem);
1295         g_free(quirk->data);
1296         g_free(quirk);
1297     }
1298 }
1299 
1300 /*
1301  * Reset quirks
1302  */
1303 void vfio_quirk_reset(VFIOPCIDevice *vdev)
1304 {
1305     int i;
1306 
1307     for (i = 0; i < PCI_ROM_SLOT; i++) {
1308         VFIOQuirk *quirk;
1309         VFIOBAR *bar = &vdev->bars[i];
1310 
1311         QLIST_FOREACH(quirk, &bar->quirks, next) {
1312             if (quirk->reset) {
1313                 quirk->reset(vdev, quirk);
1314             }
1315         }
1316     }
1317 }
1318 
1319 /*
1320  * AMD Radeon PCI config reset, based on Linux:
1321  *   drivers/gpu/drm/radeon/ci_smc.c:ci_is_smc_running()
1322  *   drivers/gpu/drm/radeon/radeon_device.c:radeon_pci_config_reset
1323  *   drivers/gpu/drm/radeon/ci_smc.c:ci_reset_smc()
1324  *   drivers/gpu/drm/radeon/ci_smc.c:ci_stop_smc_clock()
1325  * IDs: include/drm/drm_pciids.h
1326  * Registers: http://cgit.freedesktop.org/~agd5f/linux/commit/?id=4e2aa447f6f0
1327  *
1328  * Bonaire and Hawaii GPUs do not respond to a bus reset.  This is a bug in the
1329  * hardware that should be fixed on future ASICs.  The symptom of this is that
1330  * once the accerlated driver loads, Windows guests will bsod on subsequent
1331  * attmpts to load the driver, such as after VM reset or shutdown/restart.  To
1332  * work around this, we do an AMD specific PCI config reset, followed by an SMC
1333  * reset.  The PCI config reset only works if SMC firmware is running, so we
1334  * have a dependency on the state of the device as to whether this reset will
1335  * be effective.  There are still cases where we won't be able to kick the
1336  * device into working, but this greatly improves the usability overall.  The
1337  * config reset magic is relatively common on AMD GPUs, but the setup and SMC
1338  * poking is largely ASIC specific.
1339  */
1340 static bool vfio_radeon_smc_is_running(VFIOPCIDevice *vdev)
1341 {
1342     uint32_t clk, pc_c;
1343 
1344     /*
1345      * Registers 200h and 204h are index and data registers for accessing
1346      * indirect configuration registers within the device.
1347      */
1348     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000004, 4);
1349     clk = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
1350     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000370, 4);
1351     pc_c = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
1352 
1353     return (!(clk & 1) && (0x20100 <= pc_c));
1354 }
1355 
1356 /*
1357  * The scope of a config reset is controlled by a mode bit in the misc register
1358  * and a fuse, exposed as a bit in another register.  The fuse is the default
1359  * (0 = GFX, 1 = whole GPU), the misc bit is a toggle, with the formula
1360  * scope = !(misc ^ fuse), where the resulting scope is defined the same as
1361  * the fuse.  A truth table therefore tells us that if misc == fuse, we need
1362  * to flip the value of the bit in the misc register.
1363  */
1364 static void vfio_radeon_set_gfx_only_reset(VFIOPCIDevice *vdev)
1365 {
1366     uint32_t misc, fuse;
1367     bool a, b;
1368 
1369     vfio_region_write(&vdev->bars[5].region, 0x200, 0xc00c0000, 4);
1370     fuse = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
1371     b = fuse & 64;
1372 
1373     vfio_region_write(&vdev->bars[5].region, 0x200, 0xc0000010, 4);
1374     misc = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
1375     a = misc & 2;
1376 
1377     if (a == b) {
1378         vfio_region_write(&vdev->bars[5].region, 0x204, misc ^ 2, 4);
1379         vfio_region_read(&vdev->bars[5].region, 0x204, 4); /* flush */
1380     }
1381 }
1382 
1383 static int vfio_radeon_reset(VFIOPCIDevice *vdev)
1384 {
1385     PCIDevice *pdev = &vdev->pdev;
1386     int i, ret = 0;
1387     uint32_t data;
1388 
1389     /* Defer to a kernel implemented reset */
1390     if (vdev->vbasedev.reset_works) {
1391         trace_vfio_quirk_ati_bonaire_reset_skipped(vdev->vbasedev.name);
1392         return -ENODEV;
1393     }
1394 
1395     /* Enable only memory BAR access */
1396     vfio_pci_write_config(pdev, PCI_COMMAND, PCI_COMMAND_MEMORY, 2);
1397 
1398     /* Reset only works if SMC firmware is loaded and running */
1399     if (!vfio_radeon_smc_is_running(vdev)) {
1400         ret = -EINVAL;
1401         trace_vfio_quirk_ati_bonaire_reset_no_smc(vdev->vbasedev.name);
1402         goto out;
1403     }
1404 
1405     /* Make sure only the GFX function is reset */
1406     vfio_radeon_set_gfx_only_reset(vdev);
1407 
1408     /* AMD PCI config reset */
1409     vfio_pci_write_config(pdev, 0x7c, 0x39d5e86b, 4);
1410     usleep(100);
1411 
1412     /* Read back the memory size to make sure we're out of reset */
1413     for (i = 0; i < 100000; i++) {
1414         if (vfio_region_read(&vdev->bars[5].region, 0x5428, 4) != 0xffffffff) {
1415             goto reset_smc;
1416         }
1417         usleep(1);
1418     }
1419 
1420     trace_vfio_quirk_ati_bonaire_reset_timeout(vdev->vbasedev.name);
1421 
1422 reset_smc:
1423     /* Reset SMC */
1424     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000000, 4);
1425     data = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
1426     data |= 1;
1427     vfio_region_write(&vdev->bars[5].region, 0x204, data, 4);
1428 
1429     /* Disable SMC clock */
1430     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000004, 4);
1431     data = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
1432     data |= 1;
1433     vfio_region_write(&vdev->bars[5].region, 0x204, data, 4);
1434 
1435     trace_vfio_quirk_ati_bonaire_reset_done(vdev->vbasedev.name);
1436 
1437 out:
1438     /* Restore PCI command register */
1439     vfio_pci_write_config(pdev, PCI_COMMAND, 0, 2);
1440 
1441     return ret;
1442 }
1443 
1444 void vfio_setup_resetfn_quirk(VFIOPCIDevice *vdev)
1445 {
1446     switch (vdev->vendor_id) {
1447     case 0x1002:
1448         switch (vdev->device_id) {
1449         /* Bonaire */
1450         case 0x6649: /* Bonaire [FirePro W5100] */
1451         case 0x6650:
1452         case 0x6651:
1453         case 0x6658: /* Bonaire XTX [Radeon R7 260X] */
1454         case 0x665c: /* Bonaire XT [Radeon HD 7790/8770 / R9 260 OEM] */
1455         case 0x665d: /* Bonaire [Radeon R7 200 Series] */
1456         /* Hawaii */
1457         case 0x67A0: /* Hawaii XT GL [FirePro W9100] */
1458         case 0x67A1: /* Hawaii PRO GL [FirePro W8100] */
1459         case 0x67A2:
1460         case 0x67A8:
1461         case 0x67A9:
1462         case 0x67AA:
1463         case 0x67B0: /* Hawaii XT [Radeon R9 290X] */
1464         case 0x67B1: /* Hawaii PRO [Radeon R9 290] */
1465         case 0x67B8:
1466         case 0x67B9:
1467         case 0x67BA:
1468         case 0x67BE:
1469             vdev->resetfn = vfio_radeon_reset;
1470             trace_vfio_quirk_ati_bonaire_reset(vdev->vbasedev.name);
1471             break;
1472         }
1473         break;
1474     }
1475 }
1476 
1477 /*
1478  * The NVIDIA GPUDirect P2P Vendor capability allows the user to specify
1479  * devices as a member of a clique.  Devices within the same clique ID
1480  * are capable of direct P2P.  It's the user's responsibility that this
1481  * is correct.  The spec says that this may reside at any unused config
1482  * offset, but reserves and recommends hypervisors place this at C8h.
1483  * The spec also states that the hypervisor should place this capability
1484  * at the end of the capability list, thus next is defined as 0h.
1485  *
1486  * +----------------+----------------+----------------+----------------+
1487  * | sig 7:0 ('P')  |  vndr len (8h) |    next (0h)   |   cap id (9h)  |
1488  * +----------------+----------------+----------------+----------------+
1489  * | rsvd 15:7(0h),id 6:3,ver 2:0(0h)|          sig 23:8 ('P2')        |
1490  * +---------------------------------+---------------------------------+
1491  *
1492  * https://lists.gnu.org/archive/html/qemu-devel/2017-08/pdfUda5iEpgOS.pdf
1493  */
1494 static void get_nv_gpudirect_clique_id(Object *obj, Visitor *v,
1495                                        const char *name, void *opaque,
1496                                        Error **errp)
1497 {
1498     Property *prop = opaque;
1499     uint8_t *ptr = object_field_prop_ptr(obj, prop);
1500 
1501     visit_type_uint8(v, name, ptr, errp);
1502 }
1503 
1504 static void set_nv_gpudirect_clique_id(Object *obj, Visitor *v,
1505                                        const char *name, void *opaque,
1506                                        Error **errp)
1507 {
1508     Property *prop = opaque;
1509     uint8_t value, *ptr = object_field_prop_ptr(obj, prop);
1510 
1511     if (!visit_type_uint8(v, name, &value, errp)) {
1512         return;
1513     }
1514 
1515     if (value & ~0xF) {
1516         error_setg(errp, "Property %s: valid range 0-15", name);
1517         return;
1518     }
1519 
1520     *ptr = value;
1521 }
1522 
1523 const PropertyInfo qdev_prop_nv_gpudirect_clique = {
1524     .name = "uint4",
1525     .description = "NVIDIA GPUDirect Clique ID (0 - 15)",
1526     .get = get_nv_gpudirect_clique_id,
1527     .set = set_nv_gpudirect_clique_id,
1528 };
1529 
1530 static int vfio_add_nv_gpudirect_cap(VFIOPCIDevice *vdev, Error **errp)
1531 {
1532     PCIDevice *pdev = &vdev->pdev;
1533     int ret, pos = 0xC8;
1534 
1535     if (vdev->nv_gpudirect_clique == 0xFF) {
1536         return 0;
1537     }
1538 
1539     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_NVIDIA, PCI_ANY_ID)) {
1540         error_setg(errp, "NVIDIA GPUDirect Clique ID: invalid device vendor");
1541         return -EINVAL;
1542     }
1543 
1544     if (pci_get_byte(pdev->config + PCI_CLASS_DEVICE + 1) !=
1545         PCI_BASE_CLASS_DISPLAY) {
1546         error_setg(errp, "NVIDIA GPUDirect Clique ID: unsupported PCI class");
1547         return -EINVAL;
1548     }
1549 
1550     ret = pci_add_capability(pdev, PCI_CAP_ID_VNDR, pos, 8, errp);
1551     if (ret < 0) {
1552         error_prepend(errp, "Failed to add NVIDIA GPUDirect cap: ");
1553         return ret;
1554     }
1555 
1556     memset(vdev->emulated_config_bits + pos, 0xFF, 8);
1557     pos += PCI_CAP_FLAGS;
1558     pci_set_byte(pdev->config + pos++, 8);
1559     pci_set_byte(pdev->config + pos++, 'P');
1560     pci_set_byte(pdev->config + pos++, '2');
1561     pci_set_byte(pdev->config + pos++, 'P');
1562     pci_set_byte(pdev->config + pos++, vdev->nv_gpudirect_clique << 3);
1563     pci_set_byte(pdev->config + pos, 0);
1564 
1565     return 0;
1566 }
1567 
1568 static void vfio_pci_nvlink2_get_tgt(Object *obj, Visitor *v,
1569                                      const char *name,
1570                                      void *opaque, Error **errp)
1571 {
1572     uint64_t tgt = (uintptr_t) opaque;
1573     visit_type_uint64(v, name, &tgt, errp);
1574 }
1575 
1576 static void vfio_pci_nvlink2_get_link_speed(Object *obj, Visitor *v,
1577                                                  const char *name,
1578                                                  void *opaque, Error **errp)
1579 {
1580     uint32_t link_speed = (uint32_t)(uintptr_t) opaque;
1581     visit_type_uint32(v, name, &link_speed, errp);
1582 }
1583 
1584 int vfio_pci_nvidia_v100_ram_init(VFIOPCIDevice *vdev, Error **errp)
1585 {
1586     int ret;
1587     void *p;
1588     struct vfio_region_info *nv2reg = NULL;
1589     struct vfio_info_cap_header *hdr;
1590     struct vfio_region_info_cap_nvlink2_ssatgt *cap;
1591     VFIOQuirk *quirk;
1592 
1593     ret = vfio_get_dev_region_info(&vdev->vbasedev,
1594                                    VFIO_REGION_TYPE_PCI_VENDOR_TYPE |
1595                                    PCI_VENDOR_ID_NVIDIA,
1596                                    VFIO_REGION_SUBTYPE_NVIDIA_NVLINK2_RAM,
1597                                    &nv2reg);
1598     if (ret) {
1599         return ret;
1600     }
1601 
1602     hdr = vfio_get_region_info_cap(nv2reg, VFIO_REGION_INFO_CAP_NVLINK2_SSATGT);
1603     if (!hdr) {
1604         ret = -ENODEV;
1605         goto free_exit;
1606     }
1607     cap = (void *) hdr;
1608 
1609     p = mmap(NULL, nv2reg->size, PROT_READ | PROT_WRITE,
1610              MAP_SHARED, vdev->vbasedev.fd, nv2reg->offset);
1611     if (p == MAP_FAILED) {
1612         ret = -errno;
1613         goto free_exit;
1614     }
1615 
1616     quirk = vfio_quirk_alloc(1);
1617     memory_region_init_ram_ptr(&quirk->mem[0], OBJECT(vdev), "nvlink2-mr",
1618                                nv2reg->size, p);
1619     QLIST_INSERT_HEAD(&vdev->bars[0].quirks, quirk, next);
1620 
1621     object_property_add(OBJECT(vdev), "nvlink2-tgt", "uint64",
1622                         vfio_pci_nvlink2_get_tgt, NULL, NULL,
1623                         (void *) (uintptr_t) cap->tgt);
1624     trace_vfio_pci_nvidia_gpu_setup_quirk(vdev->vbasedev.name, cap->tgt,
1625                                           nv2reg->size);
1626 free_exit:
1627     g_free(nv2reg);
1628 
1629     return ret;
1630 }
1631 
1632 int vfio_pci_nvlink2_init(VFIOPCIDevice *vdev, Error **errp)
1633 {
1634     int ret;
1635     void *p;
1636     struct vfio_region_info *atsdreg = NULL;
1637     struct vfio_info_cap_header *hdr;
1638     struct vfio_region_info_cap_nvlink2_ssatgt *captgt;
1639     struct vfio_region_info_cap_nvlink2_lnkspd *capspeed;
1640     VFIOQuirk *quirk;
1641 
1642     ret = vfio_get_dev_region_info(&vdev->vbasedev,
1643                                    VFIO_REGION_TYPE_PCI_VENDOR_TYPE |
1644                                    PCI_VENDOR_ID_IBM,
1645                                    VFIO_REGION_SUBTYPE_IBM_NVLINK2_ATSD,
1646                                    &atsdreg);
1647     if (ret) {
1648         return ret;
1649     }
1650 
1651     hdr = vfio_get_region_info_cap(atsdreg,
1652                                    VFIO_REGION_INFO_CAP_NVLINK2_SSATGT);
1653     if (!hdr) {
1654         ret = -ENODEV;
1655         goto free_exit;
1656     }
1657     captgt = (void *) hdr;
1658 
1659     hdr = vfio_get_region_info_cap(atsdreg,
1660                                    VFIO_REGION_INFO_CAP_NVLINK2_LNKSPD);
1661     if (!hdr) {
1662         ret = -ENODEV;
1663         goto free_exit;
1664     }
1665     capspeed = (void *) hdr;
1666 
1667     /* Some NVLink bridges may not have assigned ATSD */
1668     if (atsdreg->size) {
1669         p = mmap(NULL, atsdreg->size, PROT_READ | PROT_WRITE,
1670                  MAP_SHARED, vdev->vbasedev.fd, atsdreg->offset);
1671         if (p == MAP_FAILED) {
1672             ret = -errno;
1673             goto free_exit;
1674         }
1675 
1676         quirk = vfio_quirk_alloc(1);
1677         memory_region_init_ram_device_ptr(&quirk->mem[0], OBJECT(vdev),
1678                                           "nvlink2-atsd-mr", atsdreg->size, p);
1679         QLIST_INSERT_HEAD(&vdev->bars[0].quirks, quirk, next);
1680     }
1681 
1682     object_property_add(OBJECT(vdev), "nvlink2-tgt", "uint64",
1683                         vfio_pci_nvlink2_get_tgt, NULL, NULL,
1684                         (void *) (uintptr_t) captgt->tgt);
1685     trace_vfio_pci_nvlink2_setup_quirk_ssatgt(vdev->vbasedev.name, captgt->tgt,
1686                                               atsdreg->size);
1687 
1688     object_property_add(OBJECT(vdev), "nvlink2-link-speed", "uint32",
1689                         vfio_pci_nvlink2_get_link_speed, NULL, NULL,
1690                         (void *) (uintptr_t) capspeed->link_speed);
1691     trace_vfio_pci_nvlink2_setup_quirk_lnkspd(vdev->vbasedev.name,
1692                                               capspeed->link_speed);
1693 free_exit:
1694     g_free(atsdreg);
1695 
1696     return ret;
1697 }
1698 
1699 /*
1700  * The VMD endpoint provides a real PCIe domain to the guest and the guest
1701  * kernel performs enumeration of the VMD sub-device domain. Guest transactions
1702  * to VMD sub-devices go through MMU translation from guest addresses to
1703  * physical addresses. When MMIO goes to an endpoint after being translated to
1704  * physical addresses, the bridge rejects the transaction because the window
1705  * has been programmed with guest addresses.
1706  *
1707  * VMD can use the Host Physical Address in order to correctly program the
1708  * bridge windows in its PCIe domain. VMD device 28C0 has HPA shadow registers
1709  * located at offset 0x2000 in MEMBAR2 (BAR 4). This quirk provides the HPA
1710  * shadow registers in a vendor-specific capability register for devices
1711  * without native support. The position of 0xE8-0xFF is in the reserved range
1712  * of the VMD device capability space following the Power Management
1713  * Capability.
1714  */
1715 #define VMD_SHADOW_CAP_VER 1
1716 #define VMD_SHADOW_CAP_LEN 24
1717 static int vfio_add_vmd_shadow_cap(VFIOPCIDevice *vdev, Error **errp)
1718 {
1719     uint8_t membar_phys[16];
1720     int ret, pos = 0xE8;
1721 
1722     if (!(vfio_pci_is(vdev, PCI_VENDOR_ID_INTEL, 0x201D) ||
1723           vfio_pci_is(vdev, PCI_VENDOR_ID_INTEL, 0x467F) ||
1724           vfio_pci_is(vdev, PCI_VENDOR_ID_INTEL, 0x4C3D) ||
1725           vfio_pci_is(vdev, PCI_VENDOR_ID_INTEL, 0x9A0B))) {
1726         return 0;
1727     }
1728 
1729     ret = pread(vdev->vbasedev.fd, membar_phys, 16,
1730                 vdev->config_offset + PCI_BASE_ADDRESS_2);
1731     if (ret != 16) {
1732         error_report("VMD %s cannot read MEMBARs (%d)",
1733                      vdev->vbasedev.name, ret);
1734         return -EFAULT;
1735     }
1736 
1737     ret = pci_add_capability(&vdev->pdev, PCI_CAP_ID_VNDR, pos,
1738                              VMD_SHADOW_CAP_LEN, errp);
1739     if (ret < 0) {
1740         error_prepend(errp, "Failed to add VMD MEMBAR Shadow cap: ");
1741         return ret;
1742     }
1743 
1744     memset(vdev->emulated_config_bits + pos, 0xFF, VMD_SHADOW_CAP_LEN);
1745     pos += PCI_CAP_FLAGS;
1746     pci_set_byte(vdev->pdev.config + pos++, VMD_SHADOW_CAP_LEN);
1747     pci_set_byte(vdev->pdev.config + pos++, VMD_SHADOW_CAP_VER);
1748     pci_set_long(vdev->pdev.config + pos, 0x53484457); /* SHDW */
1749     memcpy(vdev->pdev.config + pos + 4, membar_phys, 16);
1750 
1751     return 0;
1752 }
1753 
1754 int vfio_add_virt_caps(VFIOPCIDevice *vdev, Error **errp)
1755 {
1756     int ret;
1757 
1758     ret = vfio_add_nv_gpudirect_cap(vdev, errp);
1759     if (ret) {
1760         return ret;
1761     }
1762 
1763     ret = vfio_add_vmd_shadow_cap(vdev, errp);
1764     if (ret) {
1765         return ret;
1766     }
1767 
1768     return 0;
1769 }
1770