xref: /openbmc/qemu/hw/vfio/pci-quirks.c (revision 0b8fa32f)
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 "qemu/units.h"
15 #include "qemu/error-report.h"
16 #include "qemu/main-loop.h"
17 #include "qemu/module.h"
18 #include "qemu/range.h"
19 #include "qapi/error.h"
20 #include "qapi/visitor.h"
21 #include <sys/ioctl.h>
22 #include "hw/nvram/fw_cfg.h"
23 #include "pci.h"
24 #include "trace.h"
25 
26 /* Use uin32_t for vendor & device so PCI_ANY_ID expands and cannot match hw */
27 static bool vfio_pci_is(VFIOPCIDevice *vdev, uint32_t vendor, uint32_t device)
28 {
29     return (vendor == PCI_ANY_ID || vendor == vdev->vendor_id) &&
30            (device == PCI_ANY_ID || device == vdev->device_id);
31 }
32 
33 static bool vfio_is_vga(VFIOPCIDevice *vdev)
34 {
35     PCIDevice *pdev = &vdev->pdev;
36     uint16_t class = pci_get_word(pdev->config + PCI_CLASS_DEVICE);
37 
38     return class == PCI_CLASS_DISPLAY_VGA;
39 }
40 
41 /*
42  * List of device ids/vendor ids for which to disable
43  * option rom loading. This avoids the guest hangs during rom
44  * execution as noticed with the BCM 57810 card for lack of a
45  * more better way to handle such issues.
46  * The  user can still override by specifying a romfile or
47  * rombar=1.
48  * Please see https://bugs.launchpad.net/qemu/+bug/1284874
49  * for an analysis of the 57810 card hang. When adding
50  * a new vendor id/device id combination below, please also add
51  * your card/environment details and information that could
52  * help in debugging to the bug tracking this issue
53  */
54 static const struct {
55     uint32_t vendor;
56     uint32_t device;
57 } romblacklist[] = {
58     { 0x14e4, 0x168e }, /* Broadcom BCM 57810 */
59 };
60 
61 bool vfio_blacklist_opt_rom(VFIOPCIDevice *vdev)
62 {
63     int i;
64 
65     for (i = 0 ; i < ARRAY_SIZE(romblacklist); i++) {
66         if (vfio_pci_is(vdev, romblacklist[i].vendor, romblacklist[i].device)) {
67             trace_vfio_quirk_rom_blacklisted(vdev->vbasedev.name,
68                                              romblacklist[i].vendor,
69                                              romblacklist[i].device);
70             return true;
71         }
72     }
73     return false;
74 }
75 
76 /*
77  * Device specific region quirks (mostly backdoors to PCI config space)
78  */
79 
80 /*
81  * The generic window quirks operate on an address and data register,
82  * vfio_generic_window_address_quirk handles the address register and
83  * vfio_generic_window_data_quirk handles the data register.  These ops
84  * pass reads and writes through to hardware until a value matching the
85  * stored address match/mask is written.  When this occurs, the data
86  * register access emulated PCI config space for the device rather than
87  * passing through accesses.  This enables devices where PCI config space
88  * is accessible behind a window register to maintain the virtualization
89  * provided through vfio.
90  */
91 typedef struct VFIOConfigWindowMatch {
92     uint32_t match;
93     uint32_t mask;
94 } VFIOConfigWindowMatch;
95 
96 typedef struct VFIOConfigWindowQuirk {
97     struct VFIOPCIDevice *vdev;
98 
99     uint32_t address_val;
100 
101     uint32_t address_offset;
102     uint32_t data_offset;
103 
104     bool window_enabled;
105     uint8_t bar;
106 
107     MemoryRegion *addr_mem;
108     MemoryRegion *data_mem;
109 
110     uint32_t nr_matches;
111     VFIOConfigWindowMatch matches[];
112 } VFIOConfigWindowQuirk;
113 
114 static uint64_t vfio_generic_window_quirk_address_read(void *opaque,
115                                                        hwaddr addr,
116                                                        unsigned size)
117 {
118     VFIOConfigWindowQuirk *window = opaque;
119     VFIOPCIDevice *vdev = window->vdev;
120 
121     return vfio_region_read(&vdev->bars[window->bar].region,
122                             addr + window->address_offset, size);
123 }
124 
125 static void vfio_generic_window_quirk_address_write(void *opaque, hwaddr addr,
126                                                     uint64_t data,
127                                                     unsigned size)
128 {
129     VFIOConfigWindowQuirk *window = opaque;
130     VFIOPCIDevice *vdev = window->vdev;
131     int i;
132 
133     window->window_enabled = false;
134 
135     vfio_region_write(&vdev->bars[window->bar].region,
136                       addr + window->address_offset, data, size);
137 
138     for (i = 0; i < window->nr_matches; i++) {
139         if ((data & ~window->matches[i].mask) == window->matches[i].match) {
140             window->window_enabled = true;
141             window->address_val = data & window->matches[i].mask;
142             trace_vfio_quirk_generic_window_address_write(vdev->vbasedev.name,
143                                     memory_region_name(window->addr_mem), data);
144             break;
145         }
146     }
147 }
148 
149 static const MemoryRegionOps vfio_generic_window_address_quirk = {
150     .read = vfio_generic_window_quirk_address_read,
151     .write = vfio_generic_window_quirk_address_write,
152     .endianness = DEVICE_LITTLE_ENDIAN,
153 };
154 
155 static uint64_t vfio_generic_window_quirk_data_read(void *opaque,
156                                                     hwaddr addr, unsigned size)
157 {
158     VFIOConfigWindowQuirk *window = opaque;
159     VFIOPCIDevice *vdev = window->vdev;
160     uint64_t data;
161 
162     /* Always read data reg, discard if window enabled */
163     data = vfio_region_read(&vdev->bars[window->bar].region,
164                             addr + window->data_offset, size);
165 
166     if (window->window_enabled) {
167         data = vfio_pci_read_config(&vdev->pdev, window->address_val, size);
168         trace_vfio_quirk_generic_window_data_read(vdev->vbasedev.name,
169                                     memory_region_name(window->data_mem), data);
170     }
171 
172     return data;
173 }
174 
175 static void vfio_generic_window_quirk_data_write(void *opaque, hwaddr addr,
176                                                  uint64_t data, unsigned size)
177 {
178     VFIOConfigWindowQuirk *window = opaque;
179     VFIOPCIDevice *vdev = window->vdev;
180 
181     if (window->window_enabled) {
182         vfio_pci_write_config(&vdev->pdev, window->address_val, data, size);
183         trace_vfio_quirk_generic_window_data_write(vdev->vbasedev.name,
184                                     memory_region_name(window->data_mem), data);
185         return;
186     }
187 
188     vfio_region_write(&vdev->bars[window->bar].region,
189                       addr + window->data_offset, data, size);
190 }
191 
192 static const MemoryRegionOps vfio_generic_window_data_quirk = {
193     .read = vfio_generic_window_quirk_data_read,
194     .write = vfio_generic_window_quirk_data_write,
195     .endianness = DEVICE_LITTLE_ENDIAN,
196 };
197 
198 /*
199  * The generic mirror quirk handles devices which expose PCI config space
200  * through a region within a BAR.  When enabled, reads and writes are
201  * redirected through to emulated PCI config space.  XXX if PCI config space
202  * used memory regions, this could just be an alias.
203  */
204 typedef struct VFIOConfigMirrorQuirk {
205     struct VFIOPCIDevice *vdev;
206     uint32_t offset;
207     uint8_t bar;
208     MemoryRegion *mem;
209     uint8_t data[];
210 } VFIOConfigMirrorQuirk;
211 
212 static uint64_t vfio_generic_quirk_mirror_read(void *opaque,
213                                                hwaddr addr, unsigned size)
214 {
215     VFIOConfigMirrorQuirk *mirror = opaque;
216     VFIOPCIDevice *vdev = mirror->vdev;
217     uint64_t data;
218 
219     /* Read and discard in case the hardware cares */
220     (void)vfio_region_read(&vdev->bars[mirror->bar].region,
221                            addr + mirror->offset, size);
222 
223     data = vfio_pci_read_config(&vdev->pdev, addr, size);
224     trace_vfio_quirk_generic_mirror_read(vdev->vbasedev.name,
225                                          memory_region_name(mirror->mem),
226                                          addr, data);
227     return data;
228 }
229 
230 static void vfio_generic_quirk_mirror_write(void *opaque, hwaddr addr,
231                                             uint64_t data, unsigned size)
232 {
233     VFIOConfigMirrorQuirk *mirror = opaque;
234     VFIOPCIDevice *vdev = mirror->vdev;
235 
236     vfio_pci_write_config(&vdev->pdev, addr, data, size);
237     trace_vfio_quirk_generic_mirror_write(vdev->vbasedev.name,
238                                           memory_region_name(mirror->mem),
239                                           addr, data);
240 }
241 
242 static const MemoryRegionOps vfio_generic_mirror_quirk = {
243     .read = vfio_generic_quirk_mirror_read,
244     .write = vfio_generic_quirk_mirror_write,
245     .endianness = DEVICE_LITTLE_ENDIAN,
246 };
247 
248 /* Is range1 fully contained within range2?  */
249 static bool vfio_range_contained(uint64_t first1, uint64_t len1,
250                                  uint64_t first2, uint64_t len2) {
251     return (first1 >= first2 && first1 + len1 <= first2 + len2);
252 }
253 
254 #define PCI_VENDOR_ID_ATI               0x1002
255 
256 /*
257  * Radeon HD cards (HD5450 & HD7850) report the upper byte of the I/O port BAR
258  * through VGA register 0x3c3.  On newer cards, the I/O port BAR is always
259  * BAR4 (older cards like the X550 used BAR1, but we don't care to support
260  * those).  Note that on bare metal, a read of 0x3c3 doesn't always return the
261  * I/O port BAR address.  Originally this was coded to return the virtual BAR
262  * address only if the physical register read returns the actual BAR address,
263  * but users have reported greater success if we return the virtual address
264  * unconditionally.
265  */
266 static uint64_t vfio_ati_3c3_quirk_read(void *opaque,
267                                         hwaddr addr, unsigned size)
268 {
269     VFIOPCIDevice *vdev = opaque;
270     uint64_t data = vfio_pci_read_config(&vdev->pdev,
271                                          PCI_BASE_ADDRESS_4 + 1, size);
272 
273     trace_vfio_quirk_ati_3c3_read(vdev->vbasedev.name, data);
274 
275     return data;
276 }
277 
278 static const MemoryRegionOps vfio_ati_3c3_quirk = {
279     .read = vfio_ati_3c3_quirk_read,
280     .endianness = DEVICE_LITTLE_ENDIAN,
281 };
282 
283 static VFIOQuirk *vfio_quirk_alloc(int nr_mem)
284 {
285     VFIOQuirk *quirk = g_new0(VFIOQuirk, 1);
286     QLIST_INIT(&quirk->ioeventfds);
287     quirk->mem = g_new0(MemoryRegion, nr_mem);
288     quirk->nr_mem = nr_mem;
289 
290     return quirk;
291 }
292 
293 static void vfio_ioeventfd_exit(VFIOPCIDevice *vdev, VFIOIOEventFD *ioeventfd)
294 {
295     QLIST_REMOVE(ioeventfd, next);
296     memory_region_del_eventfd(ioeventfd->mr, ioeventfd->addr, ioeventfd->size,
297                               true, ioeventfd->data, &ioeventfd->e);
298 
299     if (ioeventfd->vfio) {
300         struct vfio_device_ioeventfd vfio_ioeventfd;
301 
302         vfio_ioeventfd.argsz = sizeof(vfio_ioeventfd);
303         vfio_ioeventfd.flags = ioeventfd->size;
304         vfio_ioeventfd.data = ioeventfd->data;
305         vfio_ioeventfd.offset = ioeventfd->region->fd_offset +
306                                 ioeventfd->region_addr;
307         vfio_ioeventfd.fd = -1;
308 
309         if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_IOEVENTFD, &vfio_ioeventfd)) {
310             error_report("Failed to remove vfio ioeventfd for %s+0x%"
311                          HWADDR_PRIx"[%d]:0x%"PRIx64" (%m)",
312                          memory_region_name(ioeventfd->mr), ioeventfd->addr,
313                          ioeventfd->size, ioeventfd->data);
314         }
315     } else {
316         qemu_set_fd_handler(event_notifier_get_fd(&ioeventfd->e),
317                             NULL, NULL, NULL);
318     }
319 
320     event_notifier_cleanup(&ioeventfd->e);
321     trace_vfio_ioeventfd_exit(memory_region_name(ioeventfd->mr),
322                               (uint64_t)ioeventfd->addr, ioeventfd->size,
323                               ioeventfd->data);
324     g_free(ioeventfd);
325 }
326 
327 static void vfio_drop_dynamic_eventfds(VFIOPCIDevice *vdev, VFIOQuirk *quirk)
328 {
329     VFIOIOEventFD *ioeventfd, *tmp;
330 
331     QLIST_FOREACH_SAFE(ioeventfd, &quirk->ioeventfds, next, tmp) {
332         if (ioeventfd->dynamic) {
333             vfio_ioeventfd_exit(vdev, ioeventfd);
334         }
335     }
336 }
337 
338 static void vfio_ioeventfd_handler(void *opaque)
339 {
340     VFIOIOEventFD *ioeventfd = opaque;
341 
342     if (event_notifier_test_and_clear(&ioeventfd->e)) {
343         vfio_region_write(ioeventfd->region, ioeventfd->region_addr,
344                           ioeventfd->data, ioeventfd->size);
345         trace_vfio_ioeventfd_handler(memory_region_name(ioeventfd->mr),
346                                      (uint64_t)ioeventfd->addr, ioeventfd->size,
347                                      ioeventfd->data);
348     }
349 }
350 
351 static VFIOIOEventFD *vfio_ioeventfd_init(VFIOPCIDevice *vdev,
352                                           MemoryRegion *mr, hwaddr addr,
353                                           unsigned size, uint64_t data,
354                                           VFIORegion *region,
355                                           hwaddr region_addr, bool dynamic)
356 {
357     VFIOIOEventFD *ioeventfd;
358 
359     if (vdev->no_kvm_ioeventfd) {
360         return NULL;
361     }
362 
363     ioeventfd = g_malloc0(sizeof(*ioeventfd));
364 
365     if (event_notifier_init(&ioeventfd->e, 0)) {
366         g_free(ioeventfd);
367         return NULL;
368     }
369 
370     /*
371      * MemoryRegion and relative offset, plus additional ioeventfd setup
372      * parameters for configuring and later tearing down KVM ioeventfd.
373      */
374     ioeventfd->mr = mr;
375     ioeventfd->addr = addr;
376     ioeventfd->size = size;
377     ioeventfd->data = data;
378     ioeventfd->dynamic = dynamic;
379     /*
380      * VFIORegion and relative offset for implementing the userspace
381      * handler.  data & size fields shared for both uses.
382      */
383     ioeventfd->region = region;
384     ioeventfd->region_addr = region_addr;
385 
386     if (!vdev->no_vfio_ioeventfd) {
387         struct vfio_device_ioeventfd vfio_ioeventfd;
388 
389         vfio_ioeventfd.argsz = sizeof(vfio_ioeventfd);
390         vfio_ioeventfd.flags = ioeventfd->size;
391         vfio_ioeventfd.data = ioeventfd->data;
392         vfio_ioeventfd.offset = ioeventfd->region->fd_offset +
393                                 ioeventfd->region_addr;
394         vfio_ioeventfd.fd = event_notifier_get_fd(&ioeventfd->e);
395 
396         ioeventfd->vfio = !ioctl(vdev->vbasedev.fd,
397                                  VFIO_DEVICE_IOEVENTFD, &vfio_ioeventfd);
398     }
399 
400     if (!ioeventfd->vfio) {
401         qemu_set_fd_handler(event_notifier_get_fd(&ioeventfd->e),
402                             vfio_ioeventfd_handler, NULL, ioeventfd);
403     }
404 
405     memory_region_add_eventfd(ioeventfd->mr, ioeventfd->addr, ioeventfd->size,
406                               true, ioeventfd->data, &ioeventfd->e);
407     trace_vfio_ioeventfd_init(memory_region_name(mr), (uint64_t)addr,
408                               size, data, ioeventfd->vfio);
409 
410     return ioeventfd;
411 }
412 
413 static void vfio_vga_probe_ati_3c3_quirk(VFIOPCIDevice *vdev)
414 {
415     VFIOQuirk *quirk;
416 
417     /*
418      * As long as the BAR is >= 256 bytes it will be aligned such that the
419      * lower byte is always zero.  Filter out anything else, if it exists.
420      */
421     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_ATI, PCI_ANY_ID) ||
422         !vdev->bars[4].ioport || vdev->bars[4].region.size < 256) {
423         return;
424     }
425 
426     quirk = vfio_quirk_alloc(1);
427 
428     memory_region_init_io(quirk->mem, OBJECT(vdev), &vfio_ati_3c3_quirk, vdev,
429                           "vfio-ati-3c3-quirk", 1);
430     memory_region_add_subregion(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
431                                 3 /* offset 3 bytes from 0x3c0 */, quirk->mem);
432 
433     QLIST_INSERT_HEAD(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].quirks,
434                       quirk, next);
435 
436     trace_vfio_quirk_ati_3c3_probe(vdev->vbasedev.name);
437 }
438 
439 /*
440  * Newer ATI/AMD devices, including HD5450 and HD7850, have a mirror to PCI
441  * config space through MMIO BAR2 at offset 0x4000.  Nothing seems to access
442  * the MMIO space directly, but a window to this space is provided through
443  * I/O port BAR4.  Offset 0x0 is the address register and offset 0x4 is the
444  * data register.  When the address is programmed to a range of 0x4000-0x4fff
445  * PCI configuration space is available.  Experimentation seems to indicate
446  * that read-only may be provided by hardware.
447  */
448 static void vfio_probe_ati_bar4_quirk(VFIOPCIDevice *vdev, int nr)
449 {
450     VFIOQuirk *quirk;
451     VFIOConfigWindowQuirk *window;
452 
453     /* This windows doesn't seem to be used except by legacy VGA code */
454     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_ATI, PCI_ANY_ID) ||
455         !vdev->vga || nr != 4) {
456         return;
457     }
458 
459     quirk = vfio_quirk_alloc(2);
460     window = quirk->data = g_malloc0(sizeof(*window) +
461                                      sizeof(VFIOConfigWindowMatch));
462     window->vdev = vdev;
463     window->address_offset = 0;
464     window->data_offset = 4;
465     window->nr_matches = 1;
466     window->matches[0].match = 0x4000;
467     window->matches[0].mask = vdev->config_size - 1;
468     window->bar = nr;
469     window->addr_mem = &quirk->mem[0];
470     window->data_mem = &quirk->mem[1];
471 
472     memory_region_init_io(window->addr_mem, OBJECT(vdev),
473                           &vfio_generic_window_address_quirk, window,
474                           "vfio-ati-bar4-window-address-quirk", 4);
475     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
476                                         window->address_offset,
477                                         window->addr_mem, 1);
478 
479     memory_region_init_io(window->data_mem, OBJECT(vdev),
480                           &vfio_generic_window_data_quirk, window,
481                           "vfio-ati-bar4-window-data-quirk", 4);
482     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
483                                         window->data_offset,
484                                         window->data_mem, 1);
485 
486     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
487 
488     trace_vfio_quirk_ati_bar4_probe(vdev->vbasedev.name);
489 }
490 
491 /*
492  * Trap the BAR2 MMIO mirror to config space as well.
493  */
494 static void vfio_probe_ati_bar2_quirk(VFIOPCIDevice *vdev, int nr)
495 {
496     VFIOQuirk *quirk;
497     VFIOConfigMirrorQuirk *mirror;
498 
499     /* Only enable on newer devices where BAR2 is 64bit */
500     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_ATI, PCI_ANY_ID) ||
501         !vdev->vga || nr != 2 || !vdev->bars[2].mem64) {
502         return;
503     }
504 
505     quirk = vfio_quirk_alloc(1);
506     mirror = quirk->data = g_malloc0(sizeof(*mirror));
507     mirror->mem = quirk->mem;
508     mirror->vdev = vdev;
509     mirror->offset = 0x4000;
510     mirror->bar = nr;
511 
512     memory_region_init_io(mirror->mem, OBJECT(vdev),
513                           &vfio_generic_mirror_quirk, mirror,
514                           "vfio-ati-bar2-4000-quirk", PCI_CONFIG_SPACE_SIZE);
515     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
516                                         mirror->offset, mirror->mem, 1);
517 
518     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
519 
520     trace_vfio_quirk_ati_bar2_probe(vdev->vbasedev.name);
521 }
522 
523 /*
524  * Older ATI/AMD cards like the X550 have a similar window to that above.
525  * I/O port BAR1 provides a window to a mirror of PCI config space located
526  * in BAR2 at offset 0xf00.  We don't care to support such older cards, but
527  * note it for future reference.
528  */
529 
530 /*
531  * Nvidia has several different methods to get to config space, the
532  * nouveu project has several of these documented here:
533  * https://github.com/pathscale/envytools/tree/master/hwdocs
534  *
535  * The first quirk is actually not documented in envytools and is found
536  * on 10de:01d1 (NVIDIA Corporation G72 [GeForce 7300 LE]).  This is an
537  * NV46 chipset.  The backdoor uses the legacy VGA I/O ports to access
538  * the mirror of PCI config space found at BAR0 offset 0x1800.  The access
539  * sequence first writes 0x338 to I/O port 0x3d4.  The target offset is
540  * then written to 0x3d0.  Finally 0x538 is written for a read and 0x738
541  * is written for a write to 0x3d4.  The BAR0 offset is then accessible
542  * through 0x3d0.  This quirk doesn't seem to be necessary on newer cards
543  * that use the I/O port BAR5 window but it doesn't hurt to leave it.
544  */
545 typedef enum {NONE = 0, SELECT, WINDOW, READ, WRITE} VFIONvidia3d0State;
546 static const char *nv3d0_states[] = { "NONE", "SELECT",
547                                       "WINDOW", "READ", "WRITE" };
548 
549 typedef struct VFIONvidia3d0Quirk {
550     VFIOPCIDevice *vdev;
551     VFIONvidia3d0State state;
552     uint32_t offset;
553 } VFIONvidia3d0Quirk;
554 
555 static uint64_t vfio_nvidia_3d4_quirk_read(void *opaque,
556                                            hwaddr addr, unsigned size)
557 {
558     VFIONvidia3d0Quirk *quirk = opaque;
559     VFIOPCIDevice *vdev = quirk->vdev;
560 
561     quirk->state = NONE;
562 
563     return vfio_vga_read(&vdev->vga->region[QEMU_PCI_VGA_IO_HI],
564                          addr + 0x14, size);
565 }
566 
567 static void vfio_nvidia_3d4_quirk_write(void *opaque, hwaddr addr,
568                                         uint64_t data, unsigned size)
569 {
570     VFIONvidia3d0Quirk *quirk = opaque;
571     VFIOPCIDevice *vdev = quirk->vdev;
572     VFIONvidia3d0State old_state = quirk->state;
573 
574     quirk->state = NONE;
575 
576     switch (data) {
577     case 0x338:
578         if (old_state == NONE) {
579             quirk->state = SELECT;
580             trace_vfio_quirk_nvidia_3d0_state(vdev->vbasedev.name,
581                                               nv3d0_states[quirk->state]);
582         }
583         break;
584     case 0x538:
585         if (old_state == WINDOW) {
586             quirk->state = READ;
587             trace_vfio_quirk_nvidia_3d0_state(vdev->vbasedev.name,
588                                               nv3d0_states[quirk->state]);
589         }
590         break;
591     case 0x738:
592         if (old_state == WINDOW) {
593             quirk->state = WRITE;
594             trace_vfio_quirk_nvidia_3d0_state(vdev->vbasedev.name,
595                                               nv3d0_states[quirk->state]);
596         }
597         break;
598     }
599 
600     vfio_vga_write(&vdev->vga->region[QEMU_PCI_VGA_IO_HI],
601                    addr + 0x14, data, size);
602 }
603 
604 static const MemoryRegionOps vfio_nvidia_3d4_quirk = {
605     .read = vfio_nvidia_3d4_quirk_read,
606     .write = vfio_nvidia_3d4_quirk_write,
607     .endianness = DEVICE_LITTLE_ENDIAN,
608 };
609 
610 static uint64_t vfio_nvidia_3d0_quirk_read(void *opaque,
611                                            hwaddr addr, unsigned size)
612 {
613     VFIONvidia3d0Quirk *quirk = opaque;
614     VFIOPCIDevice *vdev = quirk->vdev;
615     VFIONvidia3d0State old_state = quirk->state;
616     uint64_t data = vfio_vga_read(&vdev->vga->region[QEMU_PCI_VGA_IO_HI],
617                                   addr + 0x10, size);
618 
619     quirk->state = NONE;
620 
621     if (old_state == READ &&
622         (quirk->offset & ~(PCI_CONFIG_SPACE_SIZE - 1)) == 0x1800) {
623         uint8_t offset = quirk->offset & (PCI_CONFIG_SPACE_SIZE - 1);
624 
625         data = vfio_pci_read_config(&vdev->pdev, offset, size);
626         trace_vfio_quirk_nvidia_3d0_read(vdev->vbasedev.name,
627                                          offset, size, data);
628     }
629 
630     return data;
631 }
632 
633 static void vfio_nvidia_3d0_quirk_write(void *opaque, hwaddr addr,
634                                         uint64_t data, unsigned size)
635 {
636     VFIONvidia3d0Quirk *quirk = opaque;
637     VFIOPCIDevice *vdev = quirk->vdev;
638     VFIONvidia3d0State old_state = quirk->state;
639 
640     quirk->state = NONE;
641 
642     if (old_state == SELECT) {
643         quirk->offset = (uint32_t)data;
644         quirk->state = WINDOW;
645         trace_vfio_quirk_nvidia_3d0_state(vdev->vbasedev.name,
646                                           nv3d0_states[quirk->state]);
647     } else if (old_state == WRITE) {
648         if ((quirk->offset & ~(PCI_CONFIG_SPACE_SIZE - 1)) == 0x1800) {
649             uint8_t offset = quirk->offset & (PCI_CONFIG_SPACE_SIZE - 1);
650 
651             vfio_pci_write_config(&vdev->pdev, offset, data, size);
652             trace_vfio_quirk_nvidia_3d0_write(vdev->vbasedev.name,
653                                               offset, data, size);
654             return;
655         }
656     }
657 
658     vfio_vga_write(&vdev->vga->region[QEMU_PCI_VGA_IO_HI],
659                    addr + 0x10, data, size);
660 }
661 
662 static const MemoryRegionOps vfio_nvidia_3d0_quirk = {
663     .read = vfio_nvidia_3d0_quirk_read,
664     .write = vfio_nvidia_3d0_quirk_write,
665     .endianness = DEVICE_LITTLE_ENDIAN,
666 };
667 
668 static void vfio_vga_probe_nvidia_3d0_quirk(VFIOPCIDevice *vdev)
669 {
670     VFIOQuirk *quirk;
671     VFIONvidia3d0Quirk *data;
672 
673     if (vdev->no_geforce_quirks ||
674         !vfio_pci_is(vdev, PCI_VENDOR_ID_NVIDIA, PCI_ANY_ID) ||
675         !vdev->bars[1].region.size) {
676         return;
677     }
678 
679     quirk = vfio_quirk_alloc(2);
680     quirk->data = data = g_malloc0(sizeof(*data));
681     data->vdev = vdev;
682 
683     memory_region_init_io(&quirk->mem[0], OBJECT(vdev), &vfio_nvidia_3d4_quirk,
684                           data, "vfio-nvidia-3d4-quirk", 2);
685     memory_region_add_subregion(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
686                                 0x14 /* 0x3c0 + 0x14 */, &quirk->mem[0]);
687 
688     memory_region_init_io(&quirk->mem[1], OBJECT(vdev), &vfio_nvidia_3d0_quirk,
689                           data, "vfio-nvidia-3d0-quirk", 2);
690     memory_region_add_subregion(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem,
691                                 0x10 /* 0x3c0 + 0x10 */, &quirk->mem[1]);
692 
693     QLIST_INSERT_HEAD(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].quirks,
694                       quirk, next);
695 
696     trace_vfio_quirk_nvidia_3d0_probe(vdev->vbasedev.name);
697 }
698 
699 /*
700  * The second quirk is documented in envytools.  The I/O port BAR5 is just
701  * a set of address/data ports to the MMIO BARs.  The BAR we care about is
702  * again BAR0.  This backdoor is apparently a bit newer than the one above
703  * so we need to not only trap 256 bytes @0x1800, but all of PCI config
704  * space, including extended space is available at the 4k @0x88000.
705  */
706 typedef struct VFIONvidiaBAR5Quirk {
707     uint32_t master;
708     uint32_t enable;
709     MemoryRegion *addr_mem;
710     MemoryRegion *data_mem;
711     bool enabled;
712     VFIOConfigWindowQuirk window; /* last for match data */
713 } VFIONvidiaBAR5Quirk;
714 
715 static void vfio_nvidia_bar5_enable(VFIONvidiaBAR5Quirk *bar5)
716 {
717     VFIOPCIDevice *vdev = bar5->window.vdev;
718 
719     if (((bar5->master & bar5->enable) & 0x1) == bar5->enabled) {
720         return;
721     }
722 
723     bar5->enabled = !bar5->enabled;
724     trace_vfio_quirk_nvidia_bar5_state(vdev->vbasedev.name,
725                                        bar5->enabled ?  "Enable" : "Disable");
726     memory_region_set_enabled(bar5->addr_mem, bar5->enabled);
727     memory_region_set_enabled(bar5->data_mem, bar5->enabled);
728 }
729 
730 static uint64_t vfio_nvidia_bar5_quirk_master_read(void *opaque,
731                                                    hwaddr addr, unsigned size)
732 {
733     VFIONvidiaBAR5Quirk *bar5 = opaque;
734     VFIOPCIDevice *vdev = bar5->window.vdev;
735 
736     return vfio_region_read(&vdev->bars[5].region, addr, size);
737 }
738 
739 static void vfio_nvidia_bar5_quirk_master_write(void *opaque, hwaddr addr,
740                                                 uint64_t data, unsigned size)
741 {
742     VFIONvidiaBAR5Quirk *bar5 = opaque;
743     VFIOPCIDevice *vdev = bar5->window.vdev;
744 
745     vfio_region_write(&vdev->bars[5].region, addr, data, size);
746 
747     bar5->master = data;
748     vfio_nvidia_bar5_enable(bar5);
749 }
750 
751 static const MemoryRegionOps vfio_nvidia_bar5_quirk_master = {
752     .read = vfio_nvidia_bar5_quirk_master_read,
753     .write = vfio_nvidia_bar5_quirk_master_write,
754     .endianness = DEVICE_LITTLE_ENDIAN,
755 };
756 
757 static uint64_t vfio_nvidia_bar5_quirk_enable_read(void *opaque,
758                                                    hwaddr addr, unsigned size)
759 {
760     VFIONvidiaBAR5Quirk *bar5 = opaque;
761     VFIOPCIDevice *vdev = bar5->window.vdev;
762 
763     return vfio_region_read(&vdev->bars[5].region, addr + 4, size);
764 }
765 
766 static void vfio_nvidia_bar5_quirk_enable_write(void *opaque, hwaddr addr,
767                                                 uint64_t data, unsigned size)
768 {
769     VFIONvidiaBAR5Quirk *bar5 = opaque;
770     VFIOPCIDevice *vdev = bar5->window.vdev;
771 
772     vfio_region_write(&vdev->bars[5].region, addr + 4, data, size);
773 
774     bar5->enable = data;
775     vfio_nvidia_bar5_enable(bar5);
776 }
777 
778 static const MemoryRegionOps vfio_nvidia_bar5_quirk_enable = {
779     .read = vfio_nvidia_bar5_quirk_enable_read,
780     .write = vfio_nvidia_bar5_quirk_enable_write,
781     .endianness = DEVICE_LITTLE_ENDIAN,
782 };
783 
784 static void vfio_probe_nvidia_bar5_quirk(VFIOPCIDevice *vdev, int nr)
785 {
786     VFIOQuirk *quirk;
787     VFIONvidiaBAR5Quirk *bar5;
788     VFIOConfigWindowQuirk *window;
789 
790     if (vdev->no_geforce_quirks ||
791         !vfio_pci_is(vdev, PCI_VENDOR_ID_NVIDIA, PCI_ANY_ID) ||
792         !vdev->vga || nr != 5 || !vdev->bars[5].ioport) {
793         return;
794     }
795 
796     quirk = vfio_quirk_alloc(4);
797     bar5 = quirk->data = g_malloc0(sizeof(*bar5) +
798                                    (sizeof(VFIOConfigWindowMatch) * 2));
799     window = &bar5->window;
800 
801     window->vdev = vdev;
802     window->address_offset = 0x8;
803     window->data_offset = 0xc;
804     window->nr_matches = 2;
805     window->matches[0].match = 0x1800;
806     window->matches[0].mask = PCI_CONFIG_SPACE_SIZE - 1;
807     window->matches[1].match = 0x88000;
808     window->matches[1].mask = vdev->config_size - 1;
809     window->bar = nr;
810     window->addr_mem = bar5->addr_mem = &quirk->mem[0];
811     window->data_mem = bar5->data_mem = &quirk->mem[1];
812 
813     memory_region_init_io(window->addr_mem, OBJECT(vdev),
814                           &vfio_generic_window_address_quirk, window,
815                           "vfio-nvidia-bar5-window-address-quirk", 4);
816     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
817                                         window->address_offset,
818                                         window->addr_mem, 1);
819     memory_region_set_enabled(window->addr_mem, false);
820 
821     memory_region_init_io(window->data_mem, OBJECT(vdev),
822                           &vfio_generic_window_data_quirk, window,
823                           "vfio-nvidia-bar5-window-data-quirk", 4);
824     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
825                                         window->data_offset,
826                                         window->data_mem, 1);
827     memory_region_set_enabled(window->data_mem, false);
828 
829     memory_region_init_io(&quirk->mem[2], OBJECT(vdev),
830                           &vfio_nvidia_bar5_quirk_master, bar5,
831                           "vfio-nvidia-bar5-master-quirk", 4);
832     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
833                                         0, &quirk->mem[2], 1);
834 
835     memory_region_init_io(&quirk->mem[3], OBJECT(vdev),
836                           &vfio_nvidia_bar5_quirk_enable, bar5,
837                           "vfio-nvidia-bar5-enable-quirk", 4);
838     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
839                                         4, &quirk->mem[3], 1);
840 
841     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
842 
843     trace_vfio_quirk_nvidia_bar5_probe(vdev->vbasedev.name);
844 }
845 
846 typedef struct LastDataSet {
847     VFIOQuirk *quirk;
848     hwaddr addr;
849     uint64_t data;
850     unsigned size;
851     int hits;
852     int added;
853 } LastDataSet;
854 
855 #define MAX_DYN_IOEVENTFD 10
856 #define HITS_FOR_IOEVENTFD 10
857 
858 /*
859  * Finally, BAR0 itself.  We want to redirect any accesses to either
860  * 0x1800 or 0x88000 through the PCI config space access functions.
861  */
862 static void vfio_nvidia_quirk_mirror_write(void *opaque, hwaddr addr,
863                                            uint64_t data, unsigned size)
864 {
865     VFIOConfigMirrorQuirk *mirror = opaque;
866     VFIOPCIDevice *vdev = mirror->vdev;
867     PCIDevice *pdev = &vdev->pdev;
868     LastDataSet *last = (LastDataSet *)&mirror->data;
869 
870     vfio_generic_quirk_mirror_write(opaque, addr, data, size);
871 
872     /*
873      * Nvidia seems to acknowledge MSI interrupts by writing 0xff to the
874      * MSI capability ID register.  Both the ID and next register are
875      * read-only, so we allow writes covering either of those to real hw.
876      */
877     if ((pdev->cap_present & QEMU_PCI_CAP_MSI) &&
878         vfio_range_contained(addr, size, pdev->msi_cap, PCI_MSI_FLAGS)) {
879         vfio_region_write(&vdev->bars[mirror->bar].region,
880                           addr + mirror->offset, data, size);
881         trace_vfio_quirk_nvidia_bar0_msi_ack(vdev->vbasedev.name);
882     }
883 
884     /*
885      * Automatically add an ioeventfd to handle any repeated write with the
886      * same data and size above the standard PCI config space header.  This is
887      * primarily expected to accelerate the MSI-ACK behavior, such as noted
888      * above.  Current hardware/drivers should trigger an ioeventfd at config
889      * offset 0x704 (region offset 0x88704), with data 0x0, size 4.
890      *
891      * The criteria of 10 successive hits is arbitrary but reliably adds the
892      * MSI-ACK region.  Note that as some writes are bypassed via the ioeventfd,
893      * the remaining ones have a greater chance of being seen successively.
894      * To avoid the pathological case of burning up all of QEMU's open file
895      * handles, arbitrarily limit this algorithm from adding no more than 10
896      * ioeventfds, print an error if we would have added an 11th, and then
897      * stop counting.
898      */
899     if (!vdev->no_kvm_ioeventfd &&
900         addr >= PCI_STD_HEADER_SIZEOF && last->added <= MAX_DYN_IOEVENTFD) {
901         if (addr != last->addr || data != last->data || size != last->size) {
902             last->addr = addr;
903             last->data = data;
904             last->size = size;
905             last->hits = 1;
906         } else if (++last->hits >= HITS_FOR_IOEVENTFD) {
907             if (last->added < MAX_DYN_IOEVENTFD) {
908                 VFIOIOEventFD *ioeventfd;
909                 ioeventfd = vfio_ioeventfd_init(vdev, mirror->mem, addr, size,
910                                         data, &vdev->bars[mirror->bar].region,
911                                         mirror->offset + addr, true);
912                 if (ioeventfd) {
913                     VFIOQuirk *quirk = last->quirk;
914 
915                     QLIST_INSERT_HEAD(&quirk->ioeventfds, ioeventfd, next);
916                     last->added++;
917                 }
918             } else {
919                 last->added++;
920                 warn_report("NVIDIA ioeventfd queue full for %s, unable to "
921                             "accelerate 0x%"HWADDR_PRIx", data 0x%"PRIx64", "
922                             "size %u", vdev->vbasedev.name, addr, data, size);
923             }
924         }
925     }
926 }
927 
928 static const MemoryRegionOps vfio_nvidia_mirror_quirk = {
929     .read = vfio_generic_quirk_mirror_read,
930     .write = vfio_nvidia_quirk_mirror_write,
931     .endianness = DEVICE_LITTLE_ENDIAN,
932 };
933 
934 static void vfio_nvidia_bar0_quirk_reset(VFIOPCIDevice *vdev, VFIOQuirk *quirk)
935 {
936     VFIOConfigMirrorQuirk *mirror = quirk->data;
937     LastDataSet *last = (LastDataSet *)&mirror->data;
938 
939     last->addr = last->data = last->size = last->hits = last->added = 0;
940 
941     vfio_drop_dynamic_eventfds(vdev, quirk);
942 }
943 
944 static void vfio_probe_nvidia_bar0_quirk(VFIOPCIDevice *vdev, int nr)
945 {
946     VFIOQuirk *quirk;
947     VFIOConfigMirrorQuirk *mirror;
948     LastDataSet *last;
949 
950     if (vdev->no_geforce_quirks ||
951         !vfio_pci_is(vdev, PCI_VENDOR_ID_NVIDIA, PCI_ANY_ID) ||
952         !vfio_is_vga(vdev) || nr != 0) {
953         return;
954     }
955 
956     quirk = vfio_quirk_alloc(1);
957     quirk->reset = vfio_nvidia_bar0_quirk_reset;
958     mirror = quirk->data = g_malloc0(sizeof(*mirror) + sizeof(LastDataSet));
959     mirror->mem = quirk->mem;
960     mirror->vdev = vdev;
961     mirror->offset = 0x88000;
962     mirror->bar = nr;
963     last = (LastDataSet *)&mirror->data;
964     last->quirk = quirk;
965 
966     memory_region_init_io(mirror->mem, OBJECT(vdev),
967                           &vfio_nvidia_mirror_quirk, mirror,
968                           "vfio-nvidia-bar0-88000-mirror-quirk",
969                           vdev->config_size);
970     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
971                                         mirror->offset, mirror->mem, 1);
972 
973     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
974 
975     /* The 0x1800 offset mirror only seems to get used by legacy VGA */
976     if (vdev->vga) {
977         quirk = vfio_quirk_alloc(1);
978         quirk->reset = vfio_nvidia_bar0_quirk_reset;
979         mirror = quirk->data = g_malloc0(sizeof(*mirror) + sizeof(LastDataSet));
980         mirror->mem = quirk->mem;
981         mirror->vdev = vdev;
982         mirror->offset = 0x1800;
983         mirror->bar = nr;
984         last = (LastDataSet *)&mirror->data;
985         last->quirk = quirk;
986 
987         memory_region_init_io(mirror->mem, OBJECT(vdev),
988                               &vfio_nvidia_mirror_quirk, mirror,
989                               "vfio-nvidia-bar0-1800-mirror-quirk",
990                               PCI_CONFIG_SPACE_SIZE);
991         memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
992                                             mirror->offset, mirror->mem, 1);
993 
994         QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
995     }
996 
997     trace_vfio_quirk_nvidia_bar0_probe(vdev->vbasedev.name);
998 }
999 
1000 /*
1001  * TODO - Some Nvidia devices provide config access to their companion HDA
1002  * device and even to their parent bridge via these config space mirrors.
1003  * Add quirks for those regions.
1004  */
1005 
1006 #define PCI_VENDOR_ID_REALTEK 0x10ec
1007 
1008 /*
1009  * RTL8168 devices have a backdoor that can access the MSI-X table.  At BAR2
1010  * offset 0x70 there is a dword data register, offset 0x74 is a dword address
1011  * register.  According to the Linux r8169 driver, the MSI-X table is addressed
1012  * when the "type" portion of the address register is set to 0x1.  This appears
1013  * to be bits 16:30.  Bit 31 is both a write indicator and some sort of
1014  * "address latched" indicator.  Bits 12:15 are a mask field, which we can
1015  * ignore because the MSI-X table should always be accessed as a dword (full
1016  * mask).  Bits 0:11 is offset within the type.
1017  *
1018  * Example trace:
1019  *
1020  * Read from MSI-X table offset 0
1021  * vfio: vfio_bar_write(0000:05:00.0:BAR2+0x74, 0x1f000, 4) // store read addr
1022  * vfio: vfio_bar_read(0000:05:00.0:BAR2+0x74, 4) = 0x8001f000 // latch
1023  * vfio: vfio_bar_read(0000:05:00.0:BAR2+0x70, 4) = 0xfee00398 // read data
1024  *
1025  * Write 0xfee00000 to MSI-X table offset 0
1026  * vfio: vfio_bar_write(0000:05:00.0:BAR2+0x70, 0xfee00000, 4) // write data
1027  * vfio: vfio_bar_write(0000:05:00.0:BAR2+0x74, 0x8001f000, 4) // do write
1028  * vfio: vfio_bar_read(0000:05:00.0:BAR2+0x74, 4) = 0x1f000 // complete
1029  */
1030 typedef struct VFIOrtl8168Quirk {
1031     VFIOPCIDevice *vdev;
1032     uint32_t addr;
1033     uint32_t data;
1034     bool enabled;
1035 } VFIOrtl8168Quirk;
1036 
1037 static uint64_t vfio_rtl8168_quirk_address_read(void *opaque,
1038                                                 hwaddr addr, unsigned size)
1039 {
1040     VFIOrtl8168Quirk *rtl = opaque;
1041     VFIOPCIDevice *vdev = rtl->vdev;
1042     uint64_t data = vfio_region_read(&vdev->bars[2].region, addr + 0x74, size);
1043 
1044     if (rtl->enabled) {
1045         data = rtl->addr ^ 0x80000000U; /* latch/complete */
1046         trace_vfio_quirk_rtl8168_fake_latch(vdev->vbasedev.name, data);
1047     }
1048 
1049     return data;
1050 }
1051 
1052 static void vfio_rtl8168_quirk_address_write(void *opaque, hwaddr addr,
1053                                              uint64_t data, unsigned size)
1054 {
1055     VFIOrtl8168Quirk *rtl = opaque;
1056     VFIOPCIDevice *vdev = rtl->vdev;
1057 
1058     rtl->enabled = false;
1059 
1060     if ((data & 0x7fff0000) == 0x10000) { /* MSI-X table */
1061         rtl->enabled = true;
1062         rtl->addr = (uint32_t)data;
1063 
1064         if (data & 0x80000000U) { /* Do write */
1065             if (vdev->pdev.cap_present & QEMU_PCI_CAP_MSIX) {
1066                 hwaddr offset = data & 0xfff;
1067                 uint64_t val = rtl->data;
1068 
1069                 trace_vfio_quirk_rtl8168_msix_write(vdev->vbasedev.name,
1070                                                     (uint16_t)offset, val);
1071 
1072                 /* Write to the proper guest MSI-X table instead */
1073                 memory_region_dispatch_write(&vdev->pdev.msix_table_mmio,
1074                                              offset, val, size,
1075                                              MEMTXATTRS_UNSPECIFIED);
1076             }
1077             return; /* Do not write guest MSI-X data to hardware */
1078         }
1079     }
1080 
1081     vfio_region_write(&vdev->bars[2].region, addr + 0x74, data, size);
1082 }
1083 
1084 static const MemoryRegionOps vfio_rtl_address_quirk = {
1085     .read = vfio_rtl8168_quirk_address_read,
1086     .write = vfio_rtl8168_quirk_address_write,
1087     .valid = {
1088         .min_access_size = 4,
1089         .max_access_size = 4,
1090         .unaligned = false,
1091     },
1092     .endianness = DEVICE_LITTLE_ENDIAN,
1093 };
1094 
1095 static uint64_t vfio_rtl8168_quirk_data_read(void *opaque,
1096                                              hwaddr addr, unsigned size)
1097 {
1098     VFIOrtl8168Quirk *rtl = opaque;
1099     VFIOPCIDevice *vdev = rtl->vdev;
1100     uint64_t data = vfio_region_read(&vdev->bars[2].region, addr + 0x70, size);
1101 
1102     if (rtl->enabled && (vdev->pdev.cap_present & QEMU_PCI_CAP_MSIX)) {
1103         hwaddr offset = rtl->addr & 0xfff;
1104         memory_region_dispatch_read(&vdev->pdev.msix_table_mmio, offset,
1105                                     &data, size, MEMTXATTRS_UNSPECIFIED);
1106         trace_vfio_quirk_rtl8168_msix_read(vdev->vbasedev.name, offset, data);
1107     }
1108 
1109     return data;
1110 }
1111 
1112 static void vfio_rtl8168_quirk_data_write(void *opaque, hwaddr addr,
1113                                           uint64_t data, unsigned size)
1114 {
1115     VFIOrtl8168Quirk *rtl = opaque;
1116     VFIOPCIDevice *vdev = rtl->vdev;
1117 
1118     rtl->data = (uint32_t)data;
1119 
1120     vfio_region_write(&vdev->bars[2].region, addr + 0x70, data, size);
1121 }
1122 
1123 static const MemoryRegionOps vfio_rtl_data_quirk = {
1124     .read = vfio_rtl8168_quirk_data_read,
1125     .write = vfio_rtl8168_quirk_data_write,
1126     .valid = {
1127         .min_access_size = 4,
1128         .max_access_size = 4,
1129         .unaligned = false,
1130     },
1131     .endianness = DEVICE_LITTLE_ENDIAN,
1132 };
1133 
1134 static void vfio_probe_rtl8168_bar2_quirk(VFIOPCIDevice *vdev, int nr)
1135 {
1136     VFIOQuirk *quirk;
1137     VFIOrtl8168Quirk *rtl;
1138 
1139     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_REALTEK, 0x8168) || nr != 2) {
1140         return;
1141     }
1142 
1143     quirk = vfio_quirk_alloc(2);
1144     quirk->data = rtl = g_malloc0(sizeof(*rtl));
1145     rtl->vdev = vdev;
1146 
1147     memory_region_init_io(&quirk->mem[0], OBJECT(vdev),
1148                           &vfio_rtl_address_quirk, rtl,
1149                           "vfio-rtl8168-window-address-quirk", 4);
1150     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
1151                                         0x74, &quirk->mem[0], 1);
1152 
1153     memory_region_init_io(&quirk->mem[1], OBJECT(vdev),
1154                           &vfio_rtl_data_quirk, rtl,
1155                           "vfio-rtl8168-window-data-quirk", 4);
1156     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
1157                                         0x70, &quirk->mem[1], 1);
1158 
1159     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1160 
1161     trace_vfio_quirk_rtl8168_probe(vdev->vbasedev.name);
1162 }
1163 
1164 /*
1165  * Intel IGD support
1166  *
1167  * Obviously IGD is not a discrete device, this is evidenced not only by it
1168  * being integrated into the CPU, but by the various chipset and BIOS
1169  * dependencies that it brings along with it.  Intel is trying to move away
1170  * from this and Broadwell and newer devices can run in what Intel calls
1171  * "Universal Pass-Through" mode, or UPT.  Theoretically in UPT mode, nothing
1172  * more is required beyond assigning the IGD device to a VM.  There are
1173  * however support limitations to this mode.  It only supports IGD as a
1174  * secondary graphics device in the VM and it doesn't officially support any
1175  * physical outputs.
1176  *
1177  * The code here attempts to enable what we'll call legacy mode assignment,
1178  * IGD retains most of the capabilities we expect for it to have on bare
1179  * metal.  To enable this mode, the IGD device must be assigned to the VM
1180  * at PCI address 00:02.0, it must have a ROM, it very likely needs VGA
1181  * support, we must have VM BIOS support for reserving and populating some
1182  * of the required tables, and we need to tweak the chipset with revisions
1183  * and IDs and an LPC/ISA bridge device.  The intention is to make all of
1184  * this happen automatically by installing the device at the correct VM PCI
1185  * bus address.  If any of the conditions are not met, we cross our fingers
1186  * and hope the user knows better.
1187  *
1188  * NB - It is possible to enable physical outputs in UPT mode by supplying
1189  * an OpRegion table.  We don't do this by default because the guest driver
1190  * behaves differently if an OpRegion is provided and no monitor is attached
1191  * vs no OpRegion and a monitor being attached or not.  Effectively, if a
1192  * headless setup is desired, the OpRegion gets in the way of that.
1193  */
1194 
1195 /*
1196  * This presumes the device is already known to be an Intel VGA device, so we
1197  * take liberties in which device ID bits match which generation.  This should
1198  * not be taken as an indication that all the devices are supported, or even
1199  * supportable, some of them don't even support VT-d.
1200  * See linux:include/drm/i915_pciids.h for IDs.
1201  */
1202 static int igd_gen(VFIOPCIDevice *vdev)
1203 {
1204     if ((vdev->device_id & 0xfff) == 0xa84) {
1205         return 8; /* Broxton */
1206     }
1207 
1208     switch (vdev->device_id & 0xff00) {
1209     /* Old, untested, unavailable, unknown */
1210     case 0x0000:
1211     case 0x2500:
1212     case 0x2700:
1213     case 0x2900:
1214     case 0x2a00:
1215     case 0x2e00:
1216     case 0x3500:
1217     case 0xa000:
1218         return -1;
1219     /* SandyBridge, IvyBridge, ValleyView, Haswell */
1220     case 0x0100:
1221     case 0x0400:
1222     case 0x0a00:
1223     case 0x0c00:
1224     case 0x0d00:
1225     case 0x0f00:
1226         return 6;
1227     /* BroadWell, CherryView, SkyLake, KabyLake */
1228     case 0x1600:
1229     case 0x1900:
1230     case 0x2200:
1231     case 0x5900:
1232         return 8;
1233     }
1234 
1235     return 8; /* Assume newer is compatible */
1236 }
1237 
1238 typedef struct VFIOIGDQuirk {
1239     struct VFIOPCIDevice *vdev;
1240     uint32_t index;
1241     uint32_t bdsm;
1242 } VFIOIGDQuirk;
1243 
1244 #define IGD_GMCH 0x50 /* Graphics Control Register */
1245 #define IGD_BDSM 0x5c /* Base Data of Stolen Memory */
1246 #define IGD_ASLS 0xfc /* ASL Storage Register */
1247 
1248 /*
1249  * The OpRegion includes the Video BIOS Table, which seems important for
1250  * telling the driver what sort of outputs it has.  Without this, the device
1251  * may work in the guest, but we may not get output.  This also requires BIOS
1252  * support to reserve and populate a section of guest memory sufficient for
1253  * the table and to write the base address of that memory to the ASLS register
1254  * of the IGD device.
1255  */
1256 int vfio_pci_igd_opregion_init(VFIOPCIDevice *vdev,
1257                                struct vfio_region_info *info, Error **errp)
1258 {
1259     int ret;
1260 
1261     vdev->igd_opregion = g_malloc0(info->size);
1262     ret = pread(vdev->vbasedev.fd, vdev->igd_opregion,
1263                 info->size, info->offset);
1264     if (ret != info->size) {
1265         error_setg(errp, "failed to read IGD OpRegion");
1266         g_free(vdev->igd_opregion);
1267         vdev->igd_opregion = NULL;
1268         return -EINVAL;
1269     }
1270 
1271     /*
1272      * Provide fw_cfg with a copy of the OpRegion which the VM firmware is to
1273      * allocate 32bit reserved memory for, copy these contents into, and write
1274      * the reserved memory base address to the device ASLS register at 0xFC.
1275      * Alignment of this reserved region seems flexible, but using a 4k page
1276      * alignment seems to work well.  This interface assumes a single IGD
1277      * device, which may be at VM address 00:02.0 in legacy mode or another
1278      * address in UPT mode.
1279      *
1280      * NB, there may be future use cases discovered where the VM should have
1281      * direct interaction with the host OpRegion, in which case the write to
1282      * the ASLS register would trigger MemoryRegion setup to enable that.
1283      */
1284     fw_cfg_add_file(fw_cfg_find(), "etc/igd-opregion",
1285                     vdev->igd_opregion, info->size);
1286 
1287     trace_vfio_pci_igd_opregion_enabled(vdev->vbasedev.name);
1288 
1289     pci_set_long(vdev->pdev.config + IGD_ASLS, 0);
1290     pci_set_long(vdev->pdev.wmask + IGD_ASLS, ~0);
1291     pci_set_long(vdev->emulated_config_bits + IGD_ASLS, ~0);
1292 
1293     return 0;
1294 }
1295 
1296 /*
1297  * The rather short list of registers that we copy from the host devices.
1298  * The LPC/ISA bridge values are definitely needed to support the vBIOS, the
1299  * host bridge values may or may not be needed depending on the guest OS.
1300  * Since we're only munging revision and subsystem values on the host bridge,
1301  * we don't require our own device.  The LPC/ISA bridge needs to be our very
1302  * own though.
1303  */
1304 typedef struct {
1305     uint8_t offset;
1306     uint8_t len;
1307 } IGDHostInfo;
1308 
1309 static const IGDHostInfo igd_host_bridge_infos[] = {
1310     {PCI_REVISION_ID,         2},
1311     {PCI_SUBSYSTEM_VENDOR_ID, 2},
1312     {PCI_SUBSYSTEM_ID,        2},
1313 };
1314 
1315 static const IGDHostInfo igd_lpc_bridge_infos[] = {
1316     {PCI_VENDOR_ID,           2},
1317     {PCI_DEVICE_ID,           2},
1318     {PCI_REVISION_ID,         2},
1319     {PCI_SUBSYSTEM_VENDOR_ID, 2},
1320     {PCI_SUBSYSTEM_ID,        2},
1321 };
1322 
1323 static int vfio_pci_igd_copy(VFIOPCIDevice *vdev, PCIDevice *pdev,
1324                              struct vfio_region_info *info,
1325                              const IGDHostInfo *list, int len)
1326 {
1327     int i, ret;
1328 
1329     for (i = 0; i < len; i++) {
1330         ret = pread(vdev->vbasedev.fd, pdev->config + list[i].offset,
1331                     list[i].len, info->offset + list[i].offset);
1332         if (ret != list[i].len) {
1333             error_report("IGD copy failed: %m");
1334             return -errno;
1335         }
1336     }
1337 
1338     return 0;
1339 }
1340 
1341 /*
1342  * Stuff a few values into the host bridge.
1343  */
1344 static int vfio_pci_igd_host_init(VFIOPCIDevice *vdev,
1345                                   struct vfio_region_info *info)
1346 {
1347     PCIBus *bus;
1348     PCIDevice *host_bridge;
1349     int ret;
1350 
1351     bus = pci_device_root_bus(&vdev->pdev);
1352     host_bridge = pci_find_device(bus, 0, PCI_DEVFN(0, 0));
1353 
1354     if (!host_bridge) {
1355         error_report("Can't find host bridge");
1356         return -ENODEV;
1357     }
1358 
1359     ret = vfio_pci_igd_copy(vdev, host_bridge, info, igd_host_bridge_infos,
1360                             ARRAY_SIZE(igd_host_bridge_infos));
1361     if (!ret) {
1362         trace_vfio_pci_igd_host_bridge_enabled(vdev->vbasedev.name);
1363     }
1364 
1365     return ret;
1366 }
1367 
1368 /*
1369  * IGD LPC/ISA bridge support code.  The vBIOS needs this, but we can't write
1370  * arbitrary values into just any bridge, so we must create our own.  We try
1371  * to handle if the user has created it for us, which they might want to do
1372  * to enable multifunction so we don't occupy the whole PCI slot.
1373  */
1374 static void vfio_pci_igd_lpc_bridge_realize(PCIDevice *pdev, Error **errp)
1375 {
1376     if (pdev->devfn != PCI_DEVFN(0x1f, 0)) {
1377         error_setg(errp, "VFIO dummy ISA/LPC bridge must have address 1f.0");
1378     }
1379 }
1380 
1381 static void vfio_pci_igd_lpc_bridge_class_init(ObjectClass *klass, void *data)
1382 {
1383     DeviceClass *dc = DEVICE_CLASS(klass);
1384     PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
1385 
1386     set_bit(DEVICE_CATEGORY_BRIDGE, dc->categories);
1387     dc->desc = "VFIO dummy ISA/LPC bridge for IGD assignment";
1388     dc->hotpluggable = false;
1389     k->realize = vfio_pci_igd_lpc_bridge_realize;
1390     k->class_id = PCI_CLASS_BRIDGE_ISA;
1391 }
1392 
1393 static TypeInfo vfio_pci_igd_lpc_bridge_info = {
1394     .name = "vfio-pci-igd-lpc-bridge",
1395     .parent = TYPE_PCI_DEVICE,
1396     .class_init = vfio_pci_igd_lpc_bridge_class_init,
1397     .interfaces = (InterfaceInfo[]) {
1398         { INTERFACE_CONVENTIONAL_PCI_DEVICE },
1399         { },
1400     },
1401 };
1402 
1403 static void vfio_pci_igd_register_types(void)
1404 {
1405     type_register_static(&vfio_pci_igd_lpc_bridge_info);
1406 }
1407 
1408 type_init(vfio_pci_igd_register_types)
1409 
1410 static int vfio_pci_igd_lpc_init(VFIOPCIDevice *vdev,
1411                                  struct vfio_region_info *info)
1412 {
1413     PCIDevice *lpc_bridge;
1414     int ret;
1415 
1416     lpc_bridge = pci_find_device(pci_device_root_bus(&vdev->pdev),
1417                                  0, PCI_DEVFN(0x1f, 0));
1418     if (!lpc_bridge) {
1419         lpc_bridge = pci_create_simple(pci_device_root_bus(&vdev->pdev),
1420                                  PCI_DEVFN(0x1f, 0), "vfio-pci-igd-lpc-bridge");
1421     }
1422 
1423     ret = vfio_pci_igd_copy(vdev, lpc_bridge, info, igd_lpc_bridge_infos,
1424                             ARRAY_SIZE(igd_lpc_bridge_infos));
1425     if (!ret) {
1426         trace_vfio_pci_igd_lpc_bridge_enabled(vdev->vbasedev.name);
1427     }
1428 
1429     return ret;
1430 }
1431 
1432 /*
1433  * IGD Gen8 and newer support up to 8MB for the GTT and use a 64bit PTE
1434  * entry, older IGDs use 2MB and 32bit.  Each PTE maps a 4k page.  Therefore
1435  * we either have 2M/4k * 4 = 2k or 8M/4k * 8 = 16k as the maximum iobar index
1436  * for programming the GTT.
1437  *
1438  * See linux:include/drm/i915_drm.h for shift and mask values.
1439  */
1440 static int vfio_igd_gtt_max(VFIOPCIDevice *vdev)
1441 {
1442     uint32_t gmch = vfio_pci_read_config(&vdev->pdev, IGD_GMCH, sizeof(gmch));
1443     int ggms, gen = igd_gen(vdev);
1444 
1445     gmch = vfio_pci_read_config(&vdev->pdev, IGD_GMCH, sizeof(gmch));
1446     ggms = (gmch >> (gen < 8 ? 8 : 6)) & 0x3;
1447     if (gen > 6) {
1448         ggms = 1 << ggms;
1449     }
1450 
1451     ggms *= MiB;
1452 
1453     return (ggms / (4 * KiB)) * (gen < 8 ? 4 : 8);
1454 }
1455 
1456 /*
1457  * The IGD ROM will make use of stolen memory (GGMS) for support of VESA modes.
1458  * Somehow the host stolen memory range is used for this, but how the ROM gets
1459  * it is a mystery, perhaps it's hardcoded into the ROM.  Thankfully though, it
1460  * reprograms the GTT through the IOBAR where we can trap it and transpose the
1461  * programming to the VM allocated buffer.  That buffer gets reserved by the VM
1462  * firmware via the fw_cfg entry added below.  Here we're just monitoring the
1463  * IOBAR address and data registers to detect a write sequence targeting the
1464  * GTTADR.  This code is developed by observed behavior and doesn't have a
1465  * direct spec reference, unfortunately.
1466  */
1467 static uint64_t vfio_igd_quirk_data_read(void *opaque,
1468                                          hwaddr addr, unsigned size)
1469 {
1470     VFIOIGDQuirk *igd = opaque;
1471     VFIOPCIDevice *vdev = igd->vdev;
1472 
1473     igd->index = ~0;
1474 
1475     return vfio_region_read(&vdev->bars[4].region, addr + 4, size);
1476 }
1477 
1478 static void vfio_igd_quirk_data_write(void *opaque, hwaddr addr,
1479                                       uint64_t data, unsigned size)
1480 {
1481     VFIOIGDQuirk *igd = opaque;
1482     VFIOPCIDevice *vdev = igd->vdev;
1483     uint64_t val = data;
1484     int gen = igd_gen(vdev);
1485 
1486     /*
1487      * Programming the GGMS starts at index 0x1 and uses every 4th index (ie.
1488      * 0x1, 0x5, 0x9, 0xd,...).  For pre-Gen8 each 4-byte write is a whole PTE
1489      * entry, with 0th bit enable set.  For Gen8 and up, PTEs are 64bit, so
1490      * entries 0x5 & 0xd are the high dword, in our case zero.  Each PTE points
1491      * to a 4k page, which we translate to a page from the VM allocated region,
1492      * pointed to by the BDSM register.  If this is not set, we fail.
1493      *
1494      * We trap writes to the full configured GTT size, but we typically only
1495      * see the vBIOS writing up to (nearly) the 1MB barrier.  In fact it often
1496      * seems to miss the last entry for an even 1MB GTT.  Doing a gratuitous
1497      * write of that last entry does work, but is hopefully unnecessary since
1498      * we clear the previous GTT on initialization.
1499      */
1500     if ((igd->index % 4 == 1) && igd->index < vfio_igd_gtt_max(vdev)) {
1501         if (gen < 8 || (igd->index % 8 == 1)) {
1502             uint32_t base;
1503 
1504             base = pci_get_long(vdev->pdev.config + IGD_BDSM);
1505             if (!base) {
1506                 hw_error("vfio-igd: Guest attempted to program IGD GTT before "
1507                          "BIOS reserved stolen memory.  Unsupported BIOS?");
1508             }
1509 
1510             val = data - igd->bdsm + base;
1511         } else {
1512             val = 0; /* upper 32bits of pte, we only enable below 4G PTEs */
1513         }
1514 
1515         trace_vfio_pci_igd_bar4_write(vdev->vbasedev.name,
1516                                       igd->index, data, val);
1517     }
1518 
1519     vfio_region_write(&vdev->bars[4].region, addr + 4, val, size);
1520 
1521     igd->index = ~0;
1522 }
1523 
1524 static const MemoryRegionOps vfio_igd_data_quirk = {
1525     .read = vfio_igd_quirk_data_read,
1526     .write = vfio_igd_quirk_data_write,
1527     .endianness = DEVICE_LITTLE_ENDIAN,
1528 };
1529 
1530 static uint64_t vfio_igd_quirk_index_read(void *opaque,
1531                                           hwaddr addr, unsigned size)
1532 {
1533     VFIOIGDQuirk *igd = opaque;
1534     VFIOPCIDevice *vdev = igd->vdev;
1535 
1536     igd->index = ~0;
1537 
1538     return vfio_region_read(&vdev->bars[4].region, addr, size);
1539 }
1540 
1541 static void vfio_igd_quirk_index_write(void *opaque, hwaddr addr,
1542                                        uint64_t data, unsigned size)
1543 {
1544     VFIOIGDQuirk *igd = opaque;
1545     VFIOPCIDevice *vdev = igd->vdev;
1546 
1547     igd->index = data;
1548 
1549     vfio_region_write(&vdev->bars[4].region, addr, data, size);
1550 }
1551 
1552 static const MemoryRegionOps vfio_igd_index_quirk = {
1553     .read = vfio_igd_quirk_index_read,
1554     .write = vfio_igd_quirk_index_write,
1555     .endianness = DEVICE_LITTLE_ENDIAN,
1556 };
1557 
1558 static void vfio_probe_igd_bar4_quirk(VFIOPCIDevice *vdev, int nr)
1559 {
1560     struct vfio_region_info *rom = NULL, *opregion = NULL,
1561                             *host = NULL, *lpc = NULL;
1562     VFIOQuirk *quirk;
1563     VFIOIGDQuirk *igd;
1564     PCIDevice *lpc_bridge;
1565     int i, ret, ggms_mb, gms_mb = 0, gen;
1566     uint64_t *bdsm_size;
1567     uint32_t gmch;
1568     uint16_t cmd_orig, cmd;
1569     Error *err = NULL;
1570 
1571     /*
1572      * This must be an Intel VGA device at address 00:02.0 for us to even
1573      * consider enabling legacy mode.  The vBIOS has dependencies on the
1574      * PCI bus address.
1575      */
1576     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_INTEL, PCI_ANY_ID) ||
1577         !vfio_is_vga(vdev) || nr != 4 ||
1578         &vdev->pdev != pci_find_device(pci_device_root_bus(&vdev->pdev),
1579                                        0, PCI_DEVFN(0x2, 0))) {
1580         return;
1581     }
1582 
1583     /*
1584      * We need to create an LPC/ISA bridge at PCI bus address 00:1f.0 that we
1585      * can stuff host values into, so if there's already one there and it's not
1586      * one we can hack on, legacy mode is no-go.  Sorry Q35.
1587      */
1588     lpc_bridge = pci_find_device(pci_device_root_bus(&vdev->pdev),
1589                                  0, PCI_DEVFN(0x1f, 0));
1590     if (lpc_bridge && !object_dynamic_cast(OBJECT(lpc_bridge),
1591                                            "vfio-pci-igd-lpc-bridge")) {
1592         error_report("IGD device %s cannot support legacy mode due to existing "
1593                      "devices at address 1f.0", vdev->vbasedev.name);
1594         return;
1595     }
1596 
1597     /*
1598      * IGD is not a standard, they like to change their specs often.  We
1599      * only attempt to support back to SandBridge and we hope that newer
1600      * devices maintain compatibility with generation 8.
1601      */
1602     gen = igd_gen(vdev);
1603     if (gen != 6 && gen != 8) {
1604         error_report("IGD device %s is unsupported in legacy mode, "
1605                      "try SandyBridge or newer", vdev->vbasedev.name);
1606         return;
1607     }
1608 
1609     /*
1610      * Most of what we're doing here is to enable the ROM to run, so if
1611      * there's no ROM, there's no point in setting up this quirk.
1612      * NB. We only seem to get BIOS ROMs, so a UEFI VM would need CSM support.
1613      */
1614     ret = vfio_get_region_info(&vdev->vbasedev,
1615                                VFIO_PCI_ROM_REGION_INDEX, &rom);
1616     if ((ret || !rom->size) && !vdev->pdev.romfile) {
1617         error_report("IGD device %s has no ROM, legacy mode disabled",
1618                      vdev->vbasedev.name);
1619         goto out;
1620     }
1621 
1622     /*
1623      * Ignore the hotplug corner case, mark the ROM failed, we can't
1624      * create the devices we need for legacy mode in the hotplug scenario.
1625      */
1626     if (vdev->pdev.qdev.hotplugged) {
1627         error_report("IGD device %s hotplugged, ROM disabled, "
1628                      "legacy mode disabled", vdev->vbasedev.name);
1629         vdev->rom_read_failed = true;
1630         goto out;
1631     }
1632 
1633     /*
1634      * Check whether we have all the vfio device specific regions to
1635      * support legacy mode (added in Linux v4.6).  If not, bail.
1636      */
1637     ret = vfio_get_dev_region_info(&vdev->vbasedev,
1638                         VFIO_REGION_TYPE_PCI_VENDOR_TYPE | PCI_VENDOR_ID_INTEL,
1639                         VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION, &opregion);
1640     if (ret) {
1641         error_report("IGD device %s does not support OpRegion access,"
1642                      "legacy mode disabled", vdev->vbasedev.name);
1643         goto out;
1644     }
1645 
1646     ret = vfio_get_dev_region_info(&vdev->vbasedev,
1647                         VFIO_REGION_TYPE_PCI_VENDOR_TYPE | PCI_VENDOR_ID_INTEL,
1648                         VFIO_REGION_SUBTYPE_INTEL_IGD_HOST_CFG, &host);
1649     if (ret) {
1650         error_report("IGD device %s does not support host bridge access,"
1651                      "legacy mode disabled", vdev->vbasedev.name);
1652         goto out;
1653     }
1654 
1655     ret = vfio_get_dev_region_info(&vdev->vbasedev,
1656                         VFIO_REGION_TYPE_PCI_VENDOR_TYPE | PCI_VENDOR_ID_INTEL,
1657                         VFIO_REGION_SUBTYPE_INTEL_IGD_LPC_CFG, &lpc);
1658     if (ret) {
1659         error_report("IGD device %s does not support LPC bridge access,"
1660                      "legacy mode disabled", vdev->vbasedev.name);
1661         goto out;
1662     }
1663 
1664     gmch = vfio_pci_read_config(&vdev->pdev, IGD_GMCH, 4);
1665 
1666     /*
1667      * If IGD VGA Disable is clear (expected) and VGA is not already enabled,
1668      * try to enable it.  Probably shouldn't be using legacy mode without VGA,
1669      * but also no point in us enabling VGA if disabled in hardware.
1670      */
1671     if (!(gmch & 0x2) && !vdev->vga && vfio_populate_vga(vdev, &err)) {
1672         error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
1673         error_report("IGD device %s failed to enable VGA access, "
1674                      "legacy mode disabled", vdev->vbasedev.name);
1675         goto out;
1676     }
1677 
1678     /* Create our LPC/ISA bridge */
1679     ret = vfio_pci_igd_lpc_init(vdev, lpc);
1680     if (ret) {
1681         error_report("IGD device %s failed to create LPC bridge, "
1682                      "legacy mode disabled", vdev->vbasedev.name);
1683         goto out;
1684     }
1685 
1686     /* Stuff some host values into the VM PCI host bridge */
1687     ret = vfio_pci_igd_host_init(vdev, host);
1688     if (ret) {
1689         error_report("IGD device %s failed to modify host bridge, "
1690                      "legacy mode disabled", vdev->vbasedev.name);
1691         goto out;
1692     }
1693 
1694     /* Setup OpRegion access */
1695     ret = vfio_pci_igd_opregion_init(vdev, opregion, &err);
1696     if (ret) {
1697         error_append_hint(&err, "IGD legacy mode disabled\n");
1698         error_reportf_err(err, VFIO_MSG_PREFIX, vdev->vbasedev.name);
1699         goto out;
1700     }
1701 
1702     /* Setup our quirk to munge GTT addresses to the VM allocated buffer */
1703     quirk = vfio_quirk_alloc(2);
1704     igd = quirk->data = g_malloc0(sizeof(*igd));
1705     igd->vdev = vdev;
1706     igd->index = ~0;
1707     igd->bdsm = vfio_pci_read_config(&vdev->pdev, IGD_BDSM, 4);
1708     igd->bdsm &= ~((1 * MiB) - 1); /* 1MB aligned */
1709 
1710     memory_region_init_io(&quirk->mem[0], OBJECT(vdev), &vfio_igd_index_quirk,
1711                           igd, "vfio-igd-index-quirk", 4);
1712     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
1713                                         0, &quirk->mem[0], 1);
1714 
1715     memory_region_init_io(&quirk->mem[1], OBJECT(vdev), &vfio_igd_data_quirk,
1716                           igd, "vfio-igd-data-quirk", 4);
1717     memory_region_add_subregion_overlap(vdev->bars[nr].region.mem,
1718                                         4, &quirk->mem[1], 1);
1719 
1720     QLIST_INSERT_HEAD(&vdev->bars[nr].quirks, quirk, next);
1721 
1722     /* Determine the size of stolen memory needed for GTT */
1723     ggms_mb = (gmch >> (gen < 8 ? 8 : 6)) & 0x3;
1724     if (gen > 6) {
1725         ggms_mb = 1 << ggms_mb;
1726     }
1727 
1728     /*
1729      * Assume we have no GMS memory, but allow it to be overrided by device
1730      * option (experimental).  The spec doesn't actually allow zero GMS when
1731      * when IVD (IGD VGA Disable) is clear, but the claim is that it's unused,
1732      * so let's not waste VM memory for it.
1733      */
1734     gmch &= ~((gen < 8 ? 0x1f : 0xff) << (gen < 8 ? 3 : 8));
1735 
1736     if (vdev->igd_gms) {
1737         if (vdev->igd_gms <= 0x10) {
1738             gms_mb = vdev->igd_gms * 32;
1739             gmch |= vdev->igd_gms << (gen < 8 ? 3 : 8);
1740         } else {
1741             error_report("Unsupported IGD GMS value 0x%x", vdev->igd_gms);
1742             vdev->igd_gms = 0;
1743         }
1744     }
1745 
1746     /*
1747      * Request reserved memory for stolen memory via fw_cfg.  VM firmware
1748      * must allocate a 1MB aligned reserved memory region below 4GB with
1749      * the requested size (in bytes) for use by the Intel PCI class VGA
1750      * device at VM address 00:02.0.  The base address of this reserved
1751      * memory region must be written to the device BDSM regsiter at PCI
1752      * config offset 0x5C.
1753      */
1754     bdsm_size = g_malloc(sizeof(*bdsm_size));
1755     *bdsm_size = cpu_to_le64((ggms_mb + gms_mb) * MiB);
1756     fw_cfg_add_file(fw_cfg_find(), "etc/igd-bdsm-size",
1757                     bdsm_size, sizeof(*bdsm_size));
1758 
1759     /* GMCH is read-only, emulated */
1760     pci_set_long(vdev->pdev.config + IGD_GMCH, gmch);
1761     pci_set_long(vdev->pdev.wmask + IGD_GMCH, 0);
1762     pci_set_long(vdev->emulated_config_bits + IGD_GMCH, ~0);
1763 
1764     /* BDSM is read-write, emulated.  The BIOS needs to be able to write it */
1765     pci_set_long(vdev->pdev.config + IGD_BDSM, 0);
1766     pci_set_long(vdev->pdev.wmask + IGD_BDSM, ~0);
1767     pci_set_long(vdev->emulated_config_bits + IGD_BDSM, ~0);
1768 
1769     /*
1770      * This IOBAR gives us access to GTTADR, which allows us to write to
1771      * the GTT itself.  So let's go ahead and write zero to all the GTT
1772      * entries to avoid spurious DMA faults.  Be sure I/O access is enabled
1773      * before talking to the device.
1774      */
1775     if (pread(vdev->vbasedev.fd, &cmd_orig, sizeof(cmd_orig),
1776               vdev->config_offset + PCI_COMMAND) != sizeof(cmd_orig)) {
1777         error_report("IGD device %s - failed to read PCI command register",
1778                      vdev->vbasedev.name);
1779     }
1780 
1781     cmd = cmd_orig | PCI_COMMAND_IO;
1782 
1783     if (pwrite(vdev->vbasedev.fd, &cmd, sizeof(cmd),
1784                vdev->config_offset + PCI_COMMAND) != sizeof(cmd)) {
1785         error_report("IGD device %s - failed to write PCI command register",
1786                      vdev->vbasedev.name);
1787     }
1788 
1789     for (i = 1; i < vfio_igd_gtt_max(vdev); i += 4) {
1790         vfio_region_write(&vdev->bars[4].region, 0, i, 4);
1791         vfio_region_write(&vdev->bars[4].region, 4, 0, 4);
1792     }
1793 
1794     if (pwrite(vdev->vbasedev.fd, &cmd_orig, sizeof(cmd_orig),
1795                vdev->config_offset + PCI_COMMAND) != sizeof(cmd_orig)) {
1796         error_report("IGD device %s - failed to restore PCI command register",
1797                      vdev->vbasedev.name);
1798     }
1799 
1800     trace_vfio_pci_igd_bdsm_enabled(vdev->vbasedev.name, ggms_mb + gms_mb);
1801 
1802 out:
1803     g_free(rom);
1804     g_free(opregion);
1805     g_free(host);
1806     g_free(lpc);
1807 }
1808 
1809 /*
1810  * Common quirk probe entry points.
1811  */
1812 void vfio_vga_quirk_setup(VFIOPCIDevice *vdev)
1813 {
1814     vfio_vga_probe_ati_3c3_quirk(vdev);
1815     vfio_vga_probe_nvidia_3d0_quirk(vdev);
1816 }
1817 
1818 void vfio_vga_quirk_exit(VFIOPCIDevice *vdev)
1819 {
1820     VFIOQuirk *quirk;
1821     int i, j;
1822 
1823     for (i = 0; i < ARRAY_SIZE(vdev->vga->region); i++) {
1824         QLIST_FOREACH(quirk, &vdev->vga->region[i].quirks, next) {
1825             for (j = 0; j < quirk->nr_mem; j++) {
1826                 memory_region_del_subregion(&vdev->vga->region[i].mem,
1827                                             &quirk->mem[j]);
1828             }
1829         }
1830     }
1831 }
1832 
1833 void vfio_vga_quirk_finalize(VFIOPCIDevice *vdev)
1834 {
1835     int i, j;
1836 
1837     for (i = 0; i < ARRAY_SIZE(vdev->vga->region); i++) {
1838         while (!QLIST_EMPTY(&vdev->vga->region[i].quirks)) {
1839             VFIOQuirk *quirk = QLIST_FIRST(&vdev->vga->region[i].quirks);
1840             QLIST_REMOVE(quirk, next);
1841             for (j = 0; j < quirk->nr_mem; j++) {
1842                 object_unparent(OBJECT(&quirk->mem[j]));
1843             }
1844             g_free(quirk->mem);
1845             g_free(quirk->data);
1846             g_free(quirk);
1847         }
1848     }
1849 }
1850 
1851 void vfio_bar_quirk_setup(VFIOPCIDevice *vdev, int nr)
1852 {
1853     vfio_probe_ati_bar4_quirk(vdev, nr);
1854     vfio_probe_ati_bar2_quirk(vdev, nr);
1855     vfio_probe_nvidia_bar5_quirk(vdev, nr);
1856     vfio_probe_nvidia_bar0_quirk(vdev, nr);
1857     vfio_probe_rtl8168_bar2_quirk(vdev, nr);
1858     vfio_probe_igd_bar4_quirk(vdev, nr);
1859 }
1860 
1861 void vfio_bar_quirk_exit(VFIOPCIDevice *vdev, int nr)
1862 {
1863     VFIOBAR *bar = &vdev->bars[nr];
1864     VFIOQuirk *quirk;
1865     int i;
1866 
1867     QLIST_FOREACH(quirk, &bar->quirks, next) {
1868         while (!QLIST_EMPTY(&quirk->ioeventfds)) {
1869             vfio_ioeventfd_exit(vdev, QLIST_FIRST(&quirk->ioeventfds));
1870         }
1871 
1872         for (i = 0; i < quirk->nr_mem; i++) {
1873             memory_region_del_subregion(bar->region.mem, &quirk->mem[i]);
1874         }
1875     }
1876 }
1877 
1878 void vfio_bar_quirk_finalize(VFIOPCIDevice *vdev, int nr)
1879 {
1880     VFIOBAR *bar = &vdev->bars[nr];
1881     int i;
1882 
1883     while (!QLIST_EMPTY(&bar->quirks)) {
1884         VFIOQuirk *quirk = QLIST_FIRST(&bar->quirks);
1885         QLIST_REMOVE(quirk, next);
1886         for (i = 0; i < quirk->nr_mem; i++) {
1887             object_unparent(OBJECT(&quirk->mem[i]));
1888         }
1889         g_free(quirk->mem);
1890         g_free(quirk->data);
1891         g_free(quirk);
1892     }
1893 }
1894 
1895 /*
1896  * Reset quirks
1897  */
1898 void vfio_quirk_reset(VFIOPCIDevice *vdev)
1899 {
1900     int i;
1901 
1902     for (i = 0; i < PCI_ROM_SLOT; i++) {
1903         VFIOQuirk *quirk;
1904         VFIOBAR *bar = &vdev->bars[i];
1905 
1906         QLIST_FOREACH(quirk, &bar->quirks, next) {
1907             if (quirk->reset) {
1908                 quirk->reset(vdev, quirk);
1909             }
1910         }
1911     }
1912 }
1913 
1914 /*
1915  * AMD Radeon PCI config reset, based on Linux:
1916  *   drivers/gpu/drm/radeon/ci_smc.c:ci_is_smc_running()
1917  *   drivers/gpu/drm/radeon/radeon_device.c:radeon_pci_config_reset
1918  *   drivers/gpu/drm/radeon/ci_smc.c:ci_reset_smc()
1919  *   drivers/gpu/drm/radeon/ci_smc.c:ci_stop_smc_clock()
1920  * IDs: include/drm/drm_pciids.h
1921  * Registers: http://cgit.freedesktop.org/~agd5f/linux/commit/?id=4e2aa447f6f0
1922  *
1923  * Bonaire and Hawaii GPUs do not respond to a bus reset.  This is a bug in the
1924  * hardware that should be fixed on future ASICs.  The symptom of this is that
1925  * once the accerlated driver loads, Windows guests will bsod on subsequent
1926  * attmpts to load the driver, such as after VM reset or shutdown/restart.  To
1927  * work around this, we do an AMD specific PCI config reset, followed by an SMC
1928  * reset.  The PCI config reset only works if SMC firmware is running, so we
1929  * have a dependency on the state of the device as to whether this reset will
1930  * be effective.  There are still cases where we won't be able to kick the
1931  * device into working, but this greatly improves the usability overall.  The
1932  * config reset magic is relatively common on AMD GPUs, but the setup and SMC
1933  * poking is largely ASIC specific.
1934  */
1935 static bool vfio_radeon_smc_is_running(VFIOPCIDevice *vdev)
1936 {
1937     uint32_t clk, pc_c;
1938 
1939     /*
1940      * Registers 200h and 204h are index and data registers for accessing
1941      * indirect configuration registers within the device.
1942      */
1943     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000004, 4);
1944     clk = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
1945     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000370, 4);
1946     pc_c = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
1947 
1948     return (!(clk & 1) && (0x20100 <= pc_c));
1949 }
1950 
1951 /*
1952  * The scope of a config reset is controlled by a mode bit in the misc register
1953  * and a fuse, exposed as a bit in another register.  The fuse is the default
1954  * (0 = GFX, 1 = whole GPU), the misc bit is a toggle, with the forumula
1955  * scope = !(misc ^ fuse), where the resulting scope is defined the same as
1956  * the fuse.  A truth table therefore tells us that if misc == fuse, we need
1957  * to flip the value of the bit in the misc register.
1958  */
1959 static void vfio_radeon_set_gfx_only_reset(VFIOPCIDevice *vdev)
1960 {
1961     uint32_t misc, fuse;
1962     bool a, b;
1963 
1964     vfio_region_write(&vdev->bars[5].region, 0x200, 0xc00c0000, 4);
1965     fuse = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
1966     b = fuse & 64;
1967 
1968     vfio_region_write(&vdev->bars[5].region, 0x200, 0xc0000010, 4);
1969     misc = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
1970     a = misc & 2;
1971 
1972     if (a == b) {
1973         vfio_region_write(&vdev->bars[5].region, 0x204, misc ^ 2, 4);
1974         vfio_region_read(&vdev->bars[5].region, 0x204, 4); /* flush */
1975     }
1976 }
1977 
1978 static int vfio_radeon_reset(VFIOPCIDevice *vdev)
1979 {
1980     PCIDevice *pdev = &vdev->pdev;
1981     int i, ret = 0;
1982     uint32_t data;
1983 
1984     /* Defer to a kernel implemented reset */
1985     if (vdev->vbasedev.reset_works) {
1986         trace_vfio_quirk_ati_bonaire_reset_skipped(vdev->vbasedev.name);
1987         return -ENODEV;
1988     }
1989 
1990     /* Enable only memory BAR access */
1991     vfio_pci_write_config(pdev, PCI_COMMAND, PCI_COMMAND_MEMORY, 2);
1992 
1993     /* Reset only works if SMC firmware is loaded and running */
1994     if (!vfio_radeon_smc_is_running(vdev)) {
1995         ret = -EINVAL;
1996         trace_vfio_quirk_ati_bonaire_reset_no_smc(vdev->vbasedev.name);
1997         goto out;
1998     }
1999 
2000     /* Make sure only the GFX function is reset */
2001     vfio_radeon_set_gfx_only_reset(vdev);
2002 
2003     /* AMD PCI config reset */
2004     vfio_pci_write_config(pdev, 0x7c, 0x39d5e86b, 4);
2005     usleep(100);
2006 
2007     /* Read back the memory size to make sure we're out of reset */
2008     for (i = 0; i < 100000; i++) {
2009         if (vfio_region_read(&vdev->bars[5].region, 0x5428, 4) != 0xffffffff) {
2010             goto reset_smc;
2011         }
2012         usleep(1);
2013     }
2014 
2015     trace_vfio_quirk_ati_bonaire_reset_timeout(vdev->vbasedev.name);
2016 
2017 reset_smc:
2018     /* Reset SMC */
2019     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000000, 4);
2020     data = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
2021     data |= 1;
2022     vfio_region_write(&vdev->bars[5].region, 0x204, data, 4);
2023 
2024     /* Disable SMC clock */
2025     vfio_region_write(&vdev->bars[5].region, 0x200, 0x80000004, 4);
2026     data = vfio_region_read(&vdev->bars[5].region, 0x204, 4);
2027     data |= 1;
2028     vfio_region_write(&vdev->bars[5].region, 0x204, data, 4);
2029 
2030     trace_vfio_quirk_ati_bonaire_reset_done(vdev->vbasedev.name);
2031 
2032 out:
2033     /* Restore PCI command register */
2034     vfio_pci_write_config(pdev, PCI_COMMAND, 0, 2);
2035 
2036     return ret;
2037 }
2038 
2039 void vfio_setup_resetfn_quirk(VFIOPCIDevice *vdev)
2040 {
2041     switch (vdev->vendor_id) {
2042     case 0x1002:
2043         switch (vdev->device_id) {
2044         /* Bonaire */
2045         case 0x6649: /* Bonaire [FirePro W5100] */
2046         case 0x6650:
2047         case 0x6651:
2048         case 0x6658: /* Bonaire XTX [Radeon R7 260X] */
2049         case 0x665c: /* Bonaire XT [Radeon HD 7790/8770 / R9 260 OEM] */
2050         case 0x665d: /* Bonaire [Radeon R7 200 Series] */
2051         /* Hawaii */
2052         case 0x67A0: /* Hawaii XT GL [FirePro W9100] */
2053         case 0x67A1: /* Hawaii PRO GL [FirePro W8100] */
2054         case 0x67A2:
2055         case 0x67A8:
2056         case 0x67A9:
2057         case 0x67AA:
2058         case 0x67B0: /* Hawaii XT [Radeon R9 290X] */
2059         case 0x67B1: /* Hawaii PRO [Radeon R9 290] */
2060         case 0x67B8:
2061         case 0x67B9:
2062         case 0x67BA:
2063         case 0x67BE:
2064             vdev->resetfn = vfio_radeon_reset;
2065             trace_vfio_quirk_ati_bonaire_reset(vdev->vbasedev.name);
2066             break;
2067         }
2068         break;
2069     }
2070 }
2071 
2072 /*
2073  * The NVIDIA GPUDirect P2P Vendor capability allows the user to specify
2074  * devices as a member of a clique.  Devices within the same clique ID
2075  * are capable of direct P2P.  It's the user's responsibility that this
2076  * is correct.  The spec says that this may reside at any unused config
2077  * offset, but reserves and recommends hypervisors place this at C8h.
2078  * The spec also states that the hypervisor should place this capability
2079  * at the end of the capability list, thus next is defined as 0h.
2080  *
2081  * +----------------+----------------+----------------+----------------+
2082  * | sig 7:0 ('P')  |  vndr len (8h) |    next (0h)   |   cap id (9h)  |
2083  * +----------------+----------------+----------------+----------------+
2084  * | rsvd 15:7(0h),id 6:3,ver 2:0(0h)|          sig 23:8 ('P2')        |
2085  * +---------------------------------+---------------------------------+
2086  *
2087  * https://lists.gnu.org/archive/html/qemu-devel/2017-08/pdfUda5iEpgOS.pdf
2088  */
2089 static void get_nv_gpudirect_clique_id(Object *obj, Visitor *v,
2090                                        const char *name, void *opaque,
2091                                        Error **errp)
2092 {
2093     DeviceState *dev = DEVICE(obj);
2094     Property *prop = opaque;
2095     uint8_t *ptr = qdev_get_prop_ptr(dev, prop);
2096 
2097     visit_type_uint8(v, name, ptr, errp);
2098 }
2099 
2100 static void set_nv_gpudirect_clique_id(Object *obj, Visitor *v,
2101                                        const char *name, void *opaque,
2102                                        Error **errp)
2103 {
2104     DeviceState *dev = DEVICE(obj);
2105     Property *prop = opaque;
2106     uint8_t value, *ptr = qdev_get_prop_ptr(dev, prop);
2107     Error *local_err = NULL;
2108 
2109     if (dev->realized) {
2110         qdev_prop_set_after_realize(dev, name, errp);
2111         return;
2112     }
2113 
2114     visit_type_uint8(v, name, &value, &local_err);
2115     if (local_err) {
2116         error_propagate(errp, local_err);
2117         return;
2118     }
2119 
2120     if (value & ~0xF) {
2121         error_setg(errp, "Property %s: valid range 0-15", name);
2122         return;
2123     }
2124 
2125     *ptr = value;
2126 }
2127 
2128 const PropertyInfo qdev_prop_nv_gpudirect_clique = {
2129     .name = "uint4",
2130     .description = "NVIDIA GPUDirect Clique ID (0 - 15)",
2131     .get = get_nv_gpudirect_clique_id,
2132     .set = set_nv_gpudirect_clique_id,
2133 };
2134 
2135 static int vfio_add_nv_gpudirect_cap(VFIOPCIDevice *vdev, Error **errp)
2136 {
2137     PCIDevice *pdev = &vdev->pdev;
2138     int ret, pos = 0xC8;
2139 
2140     if (vdev->nv_gpudirect_clique == 0xFF) {
2141         return 0;
2142     }
2143 
2144     if (!vfio_pci_is(vdev, PCI_VENDOR_ID_NVIDIA, PCI_ANY_ID)) {
2145         error_setg(errp, "NVIDIA GPUDirect Clique ID: invalid device vendor");
2146         return -EINVAL;
2147     }
2148 
2149     if (pci_get_byte(pdev->config + PCI_CLASS_DEVICE + 1) !=
2150         PCI_BASE_CLASS_DISPLAY) {
2151         error_setg(errp, "NVIDIA GPUDirect Clique ID: unsupported PCI class");
2152         return -EINVAL;
2153     }
2154 
2155     ret = pci_add_capability(pdev, PCI_CAP_ID_VNDR, pos, 8, errp);
2156     if (ret < 0) {
2157         error_prepend(errp, "Failed to add NVIDIA GPUDirect cap: ");
2158         return ret;
2159     }
2160 
2161     memset(vdev->emulated_config_bits + pos, 0xFF, 8);
2162     pos += PCI_CAP_FLAGS;
2163     pci_set_byte(pdev->config + pos++, 8);
2164     pci_set_byte(pdev->config + pos++, 'P');
2165     pci_set_byte(pdev->config + pos++, '2');
2166     pci_set_byte(pdev->config + pos++, 'P');
2167     pci_set_byte(pdev->config + pos++, vdev->nv_gpudirect_clique << 3);
2168     pci_set_byte(pdev->config + pos, 0);
2169 
2170     return 0;
2171 }
2172 
2173 int vfio_add_virt_caps(VFIOPCIDevice *vdev, Error **errp)
2174 {
2175     int ret;
2176 
2177     ret = vfio_add_nv_gpudirect_cap(vdev, errp);
2178     if (ret) {
2179         return ret;
2180     }
2181 
2182     return 0;
2183 }
2184 
2185 static void vfio_pci_nvlink2_get_tgt(Object *obj, Visitor *v,
2186                                      const char *name,
2187                                      void *opaque, Error **errp)
2188 {
2189     uint64_t tgt = (uintptr_t) opaque;
2190     visit_type_uint64(v, name, &tgt, errp);
2191 }
2192 
2193 static void vfio_pci_nvlink2_get_link_speed(Object *obj, Visitor *v,
2194                                                  const char *name,
2195                                                  void *opaque, Error **errp)
2196 {
2197     uint32_t link_speed = (uint32_t)(uintptr_t) opaque;
2198     visit_type_uint32(v, name, &link_speed, errp);
2199 }
2200 
2201 int vfio_pci_nvidia_v100_ram_init(VFIOPCIDevice *vdev, Error **errp)
2202 {
2203     int ret;
2204     void *p;
2205     struct vfio_region_info *nv2reg = NULL;
2206     struct vfio_info_cap_header *hdr;
2207     struct vfio_region_info_cap_nvlink2_ssatgt *cap;
2208     VFIOQuirk *quirk;
2209 
2210     ret = vfio_get_dev_region_info(&vdev->vbasedev,
2211                                    VFIO_REGION_TYPE_PCI_VENDOR_TYPE |
2212                                    PCI_VENDOR_ID_NVIDIA,
2213                                    VFIO_REGION_SUBTYPE_NVIDIA_NVLINK2_RAM,
2214                                    &nv2reg);
2215     if (ret) {
2216         return ret;
2217     }
2218 
2219     hdr = vfio_get_region_info_cap(nv2reg, VFIO_REGION_INFO_CAP_NVLINK2_SSATGT);
2220     if (!hdr) {
2221         ret = -ENODEV;
2222         goto free_exit;
2223     }
2224     cap = (void *) hdr;
2225 
2226     p = mmap(NULL, nv2reg->size, PROT_READ | PROT_WRITE | PROT_EXEC,
2227              MAP_SHARED, vdev->vbasedev.fd, nv2reg->offset);
2228     if (p == MAP_FAILED) {
2229         ret = -errno;
2230         goto free_exit;
2231     }
2232 
2233     quirk = vfio_quirk_alloc(1);
2234     memory_region_init_ram_ptr(&quirk->mem[0], OBJECT(vdev), "nvlink2-mr",
2235                                nv2reg->size, p);
2236     QLIST_INSERT_HEAD(&vdev->bars[0].quirks, quirk, next);
2237 
2238     object_property_add(OBJECT(vdev), "nvlink2-tgt", "uint64",
2239                         vfio_pci_nvlink2_get_tgt, NULL, NULL,
2240                         (void *) (uintptr_t) cap->tgt, NULL);
2241     trace_vfio_pci_nvidia_gpu_setup_quirk(vdev->vbasedev.name, cap->tgt,
2242                                           nv2reg->size);
2243 free_exit:
2244     g_free(nv2reg);
2245 
2246     return ret;
2247 }
2248 
2249 int vfio_pci_nvlink2_init(VFIOPCIDevice *vdev, Error **errp)
2250 {
2251     int ret;
2252     void *p;
2253     struct vfio_region_info *atsdreg = NULL;
2254     struct vfio_info_cap_header *hdr;
2255     struct vfio_region_info_cap_nvlink2_ssatgt *captgt;
2256     struct vfio_region_info_cap_nvlink2_lnkspd *capspeed;
2257     VFIOQuirk *quirk;
2258 
2259     ret = vfio_get_dev_region_info(&vdev->vbasedev,
2260                                    VFIO_REGION_TYPE_PCI_VENDOR_TYPE |
2261                                    PCI_VENDOR_ID_IBM,
2262                                    VFIO_REGION_SUBTYPE_IBM_NVLINK2_ATSD,
2263                                    &atsdreg);
2264     if (ret) {
2265         return ret;
2266     }
2267 
2268     hdr = vfio_get_region_info_cap(atsdreg,
2269                                    VFIO_REGION_INFO_CAP_NVLINK2_SSATGT);
2270     if (!hdr) {
2271         ret = -ENODEV;
2272         goto free_exit;
2273     }
2274     captgt = (void *) hdr;
2275 
2276     hdr = vfio_get_region_info_cap(atsdreg,
2277                                    VFIO_REGION_INFO_CAP_NVLINK2_LNKSPD);
2278     if (!hdr) {
2279         ret = -ENODEV;
2280         goto free_exit;
2281     }
2282     capspeed = (void *) hdr;
2283 
2284     /* Some NVLink bridges may not have assigned ATSD */
2285     if (atsdreg->size) {
2286         p = mmap(NULL, atsdreg->size, PROT_READ | PROT_WRITE | PROT_EXEC,
2287                  MAP_SHARED, vdev->vbasedev.fd, atsdreg->offset);
2288         if (p == MAP_FAILED) {
2289             ret = -errno;
2290             goto free_exit;
2291         }
2292 
2293         quirk = vfio_quirk_alloc(1);
2294         memory_region_init_ram_device_ptr(&quirk->mem[0], OBJECT(vdev),
2295                                           "nvlink2-atsd-mr", atsdreg->size, p);
2296         QLIST_INSERT_HEAD(&vdev->bars[0].quirks, quirk, next);
2297     }
2298 
2299     object_property_add(OBJECT(vdev), "nvlink2-tgt", "uint64",
2300                         vfio_pci_nvlink2_get_tgt, NULL, NULL,
2301                         (void *) (uintptr_t) captgt->tgt, NULL);
2302     trace_vfio_pci_nvlink2_setup_quirk_ssatgt(vdev->vbasedev.name, captgt->tgt,
2303                                               atsdreg->size);
2304 
2305     object_property_add(OBJECT(vdev), "nvlink2-link-speed", "uint32",
2306                         vfio_pci_nvlink2_get_link_speed, NULL, NULL,
2307                         (void *) (uintptr_t) capspeed->link_speed, NULL);
2308     trace_vfio_pci_nvlink2_setup_quirk_lnkspd(vdev->vbasedev.name,
2309                                               capspeed->link_speed);
2310 free_exit:
2311     g_free(atsdreg);
2312 
2313     return ret;
2314 }
2315