1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * VFIO PCI config space virtualization
4  *
5  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
6  *     Author: Alex Williamson <alex.williamson@redhat.com>
7  *
8  * Derived from original vfio:
9  * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
10  * Author: Tom Lyon, pugs@cisco.com
11  */
12 
13 /*
14  * This code handles reading and writing of PCI configuration registers.
15  * This is hairy because we want to allow a lot of flexibility to the
16  * user driver, but cannot trust it with all of the config fields.
17  * Tables determine which fields can be read and written, as well as
18  * which fields are 'virtualized' - special actions and translations to
19  * make it appear to the user that he has control, when in fact things
20  * must be negotiated with the underlying OS.
21  */
22 
23 #include <linux/fs.h>
24 #include <linux/pci.h>
25 #include <linux/uaccess.h>
26 #include <linux/vfio.h>
27 #include <linux/slab.h>
28 
29 #include <linux/vfio_pci_core.h>
30 
31 /* Fake capability ID for standard config space */
32 #define PCI_CAP_ID_BASIC	0
33 
34 #define is_bar(offset)	\
35 	((offset >= PCI_BASE_ADDRESS_0 && offset < PCI_BASE_ADDRESS_5 + 4) || \
36 	 (offset >= PCI_ROM_ADDRESS && offset < PCI_ROM_ADDRESS + 4))
37 
38 /*
39  * Lengths of PCI Config Capabilities
40  *   0: Removed from the user visible capability list
41  *   FF: Variable length
42  */
43 static const u8 pci_cap_length[PCI_CAP_ID_MAX + 1] = {
44 	[PCI_CAP_ID_BASIC]	= PCI_STD_HEADER_SIZEOF, /* pci config header */
45 	[PCI_CAP_ID_PM]		= PCI_PM_SIZEOF,
46 	[PCI_CAP_ID_AGP]	= PCI_AGP_SIZEOF,
47 	[PCI_CAP_ID_VPD]	= PCI_CAP_VPD_SIZEOF,
48 	[PCI_CAP_ID_SLOTID]	= 0,		/* bridge - don't care */
49 	[PCI_CAP_ID_MSI]	= 0xFF,		/* 10, 14, 20, or 24 */
50 	[PCI_CAP_ID_CHSWP]	= 0,		/* cpci - not yet */
51 	[PCI_CAP_ID_PCIX]	= 0xFF,		/* 8 or 24 */
52 	[PCI_CAP_ID_HT]		= 0xFF,		/* hypertransport */
53 	[PCI_CAP_ID_VNDR]	= 0xFF,		/* variable */
54 	[PCI_CAP_ID_DBG]	= 0,		/* debug - don't care */
55 	[PCI_CAP_ID_CCRC]	= 0,		/* cpci - not yet */
56 	[PCI_CAP_ID_SHPC]	= 0,		/* hotswap - not yet */
57 	[PCI_CAP_ID_SSVID]	= 0,		/* bridge - don't care */
58 	[PCI_CAP_ID_AGP3]	= 0,		/* AGP8x - not yet */
59 	[PCI_CAP_ID_SECDEV]	= 0,		/* secure device not yet */
60 	[PCI_CAP_ID_EXP]	= 0xFF,		/* 20 or 44 */
61 	[PCI_CAP_ID_MSIX]	= PCI_CAP_MSIX_SIZEOF,
62 	[PCI_CAP_ID_SATA]	= 0xFF,
63 	[PCI_CAP_ID_AF]		= PCI_CAP_AF_SIZEOF,
64 };
65 
66 /*
67  * Lengths of PCIe/PCI-X Extended Config Capabilities
68  *   0: Removed or masked from the user visible capability list
69  *   FF: Variable length
70  */
71 static const u16 pci_ext_cap_length[PCI_EXT_CAP_ID_MAX + 1] = {
72 	[PCI_EXT_CAP_ID_ERR]	=	PCI_ERR_ROOT_COMMAND,
73 	[PCI_EXT_CAP_ID_VC]	=	0xFF,
74 	[PCI_EXT_CAP_ID_DSN]	=	PCI_EXT_CAP_DSN_SIZEOF,
75 	[PCI_EXT_CAP_ID_PWR]	=	PCI_EXT_CAP_PWR_SIZEOF,
76 	[PCI_EXT_CAP_ID_RCLD]	=	0,	/* root only - don't care */
77 	[PCI_EXT_CAP_ID_RCILC]	=	0,	/* root only - don't care */
78 	[PCI_EXT_CAP_ID_RCEC]	=	0,	/* root only - don't care */
79 	[PCI_EXT_CAP_ID_MFVC]	=	0xFF,
80 	[PCI_EXT_CAP_ID_VC9]	=	0xFF,	/* same as CAP_ID_VC */
81 	[PCI_EXT_CAP_ID_RCRB]	=	0,	/* root only - don't care */
82 	[PCI_EXT_CAP_ID_VNDR]	=	0xFF,
83 	[PCI_EXT_CAP_ID_CAC]	=	0,	/* obsolete */
84 	[PCI_EXT_CAP_ID_ACS]	=	0xFF,
85 	[PCI_EXT_CAP_ID_ARI]	=	PCI_EXT_CAP_ARI_SIZEOF,
86 	[PCI_EXT_CAP_ID_ATS]	=	PCI_EXT_CAP_ATS_SIZEOF,
87 	[PCI_EXT_CAP_ID_SRIOV]	=	PCI_EXT_CAP_SRIOV_SIZEOF,
88 	[PCI_EXT_CAP_ID_MRIOV]	=	0,	/* not yet */
89 	[PCI_EXT_CAP_ID_MCAST]	=	PCI_EXT_CAP_MCAST_ENDPOINT_SIZEOF,
90 	[PCI_EXT_CAP_ID_PRI]	=	PCI_EXT_CAP_PRI_SIZEOF,
91 	[PCI_EXT_CAP_ID_AMD_XXX] =	0,	/* not yet */
92 	[PCI_EXT_CAP_ID_REBAR]	=	0xFF,
93 	[PCI_EXT_CAP_ID_DPA]	=	0xFF,
94 	[PCI_EXT_CAP_ID_TPH]	=	0xFF,
95 	[PCI_EXT_CAP_ID_LTR]	=	PCI_EXT_CAP_LTR_SIZEOF,
96 	[PCI_EXT_CAP_ID_SECPCI]	=	0,	/* not yet */
97 	[PCI_EXT_CAP_ID_PMUX]	=	0,	/* not yet */
98 	[PCI_EXT_CAP_ID_PASID]	=	0,	/* not yet */
99 };
100 
101 /*
102  * Read/Write Permission Bits - one bit for each bit in capability
103  * Any field can be read if it exists, but what is read depends on
104  * whether the field is 'virtualized', or just pass through to the
105  * hardware.  Any virtualized field is also virtualized for writes.
106  * Writes are only permitted if they have a 1 bit here.
107  */
108 struct perm_bits {
109 	u8	*virt;		/* read/write virtual data, not hw */
110 	u8	*write;		/* writeable bits */
111 	int	(*readfn)(struct vfio_pci_core_device *vdev, int pos, int count,
112 			  struct perm_bits *perm, int offset, __le32 *val);
113 	int	(*writefn)(struct vfio_pci_core_device *vdev, int pos, int count,
114 			   struct perm_bits *perm, int offset, __le32 val);
115 };
116 
117 #define	NO_VIRT		0
118 #define	ALL_VIRT	0xFFFFFFFFU
119 #define	NO_WRITE	0
120 #define	ALL_WRITE	0xFFFFFFFFU
121 
122 static int vfio_user_config_read(struct pci_dev *pdev, int offset,
123 				 __le32 *val, int count)
124 {
125 	int ret = -EINVAL;
126 	u32 tmp_val = 0;
127 
128 	switch (count) {
129 	case 1:
130 	{
131 		u8 tmp;
132 		ret = pci_user_read_config_byte(pdev, offset, &tmp);
133 		tmp_val = tmp;
134 		break;
135 	}
136 	case 2:
137 	{
138 		u16 tmp;
139 		ret = pci_user_read_config_word(pdev, offset, &tmp);
140 		tmp_val = tmp;
141 		break;
142 	}
143 	case 4:
144 		ret = pci_user_read_config_dword(pdev, offset, &tmp_val);
145 		break;
146 	}
147 
148 	*val = cpu_to_le32(tmp_val);
149 
150 	return ret;
151 }
152 
153 static int vfio_user_config_write(struct pci_dev *pdev, int offset,
154 				  __le32 val, int count)
155 {
156 	int ret = -EINVAL;
157 	u32 tmp_val = le32_to_cpu(val);
158 
159 	switch (count) {
160 	case 1:
161 		ret = pci_user_write_config_byte(pdev, offset, tmp_val);
162 		break;
163 	case 2:
164 		ret = pci_user_write_config_word(pdev, offset, tmp_val);
165 		break;
166 	case 4:
167 		ret = pci_user_write_config_dword(pdev, offset, tmp_val);
168 		break;
169 	}
170 
171 	return ret;
172 }
173 
174 static int vfio_default_config_read(struct vfio_pci_core_device *vdev, int pos,
175 				    int count, struct perm_bits *perm,
176 				    int offset, __le32 *val)
177 {
178 	__le32 virt = 0;
179 
180 	memcpy(val, vdev->vconfig + pos, count);
181 
182 	memcpy(&virt, perm->virt + offset, count);
183 
184 	/* Any non-virtualized bits? */
185 	if (cpu_to_le32(~0U >> (32 - (count * 8))) != virt) {
186 		struct pci_dev *pdev = vdev->pdev;
187 		__le32 phys_val = 0;
188 		int ret;
189 
190 		ret = vfio_user_config_read(pdev, pos, &phys_val, count);
191 		if (ret)
192 			return ret;
193 
194 		*val = (phys_val & ~virt) | (*val & virt);
195 	}
196 
197 	return count;
198 }
199 
200 static int vfio_default_config_write(struct vfio_pci_core_device *vdev, int pos,
201 				     int count, struct perm_bits *perm,
202 				     int offset, __le32 val)
203 {
204 	__le32 virt = 0, write = 0;
205 
206 	memcpy(&write, perm->write + offset, count);
207 
208 	if (!write)
209 		return count; /* drop, no writable bits */
210 
211 	memcpy(&virt, perm->virt + offset, count);
212 
213 	/* Virtualized and writable bits go to vconfig */
214 	if (write & virt) {
215 		__le32 virt_val = 0;
216 
217 		memcpy(&virt_val, vdev->vconfig + pos, count);
218 
219 		virt_val &= ~(write & virt);
220 		virt_val |= (val & (write & virt));
221 
222 		memcpy(vdev->vconfig + pos, &virt_val, count);
223 	}
224 
225 	/* Non-virtualzed and writable bits go to hardware */
226 	if (write & ~virt) {
227 		struct pci_dev *pdev = vdev->pdev;
228 		__le32 phys_val = 0;
229 		int ret;
230 
231 		ret = vfio_user_config_read(pdev, pos, &phys_val, count);
232 		if (ret)
233 			return ret;
234 
235 		phys_val &= ~(write & ~virt);
236 		phys_val |= (val & (write & ~virt));
237 
238 		ret = vfio_user_config_write(pdev, pos, phys_val, count);
239 		if (ret)
240 			return ret;
241 	}
242 
243 	return count;
244 }
245 
246 /* Allow direct read from hardware, except for capability next pointer */
247 static int vfio_direct_config_read(struct vfio_pci_core_device *vdev, int pos,
248 				   int count, struct perm_bits *perm,
249 				   int offset, __le32 *val)
250 {
251 	int ret;
252 
253 	ret = vfio_user_config_read(vdev->pdev, pos, val, count);
254 	if (ret)
255 		return ret;
256 
257 	if (pos >= PCI_CFG_SPACE_SIZE) { /* Extended cap header mangling */
258 		if (offset < 4)
259 			memcpy(val, vdev->vconfig + pos, count);
260 	} else if (pos >= PCI_STD_HEADER_SIZEOF) { /* Std cap mangling */
261 		if (offset == PCI_CAP_LIST_ID && count > 1)
262 			memcpy(val, vdev->vconfig + pos,
263 			       min(PCI_CAP_FLAGS, count));
264 		else if (offset == PCI_CAP_LIST_NEXT)
265 			memcpy(val, vdev->vconfig + pos, 1);
266 	}
267 
268 	return count;
269 }
270 
271 /* Raw access skips any kind of virtualization */
272 static int vfio_raw_config_write(struct vfio_pci_core_device *vdev, int pos,
273 				 int count, struct perm_bits *perm,
274 				 int offset, __le32 val)
275 {
276 	int ret;
277 
278 	ret = vfio_user_config_write(vdev->pdev, pos, val, count);
279 	if (ret)
280 		return ret;
281 
282 	return count;
283 }
284 
285 static int vfio_raw_config_read(struct vfio_pci_core_device *vdev, int pos,
286 				int count, struct perm_bits *perm,
287 				int offset, __le32 *val)
288 {
289 	int ret;
290 
291 	ret = vfio_user_config_read(vdev->pdev, pos, val, count);
292 	if (ret)
293 		return ret;
294 
295 	return count;
296 }
297 
298 /* Virt access uses only virtualization */
299 static int vfio_virt_config_write(struct vfio_pci_core_device *vdev, int pos,
300 				  int count, struct perm_bits *perm,
301 				  int offset, __le32 val)
302 {
303 	memcpy(vdev->vconfig + pos, &val, count);
304 	return count;
305 }
306 
307 static int vfio_virt_config_read(struct vfio_pci_core_device *vdev, int pos,
308 				 int count, struct perm_bits *perm,
309 				 int offset, __le32 *val)
310 {
311 	memcpy(val, vdev->vconfig + pos, count);
312 	return count;
313 }
314 
315 /* Default capability regions to read-only, no-virtualization */
316 static struct perm_bits cap_perms[PCI_CAP_ID_MAX + 1] = {
317 	[0 ... PCI_CAP_ID_MAX] = { .readfn = vfio_direct_config_read }
318 };
319 static struct perm_bits ecap_perms[PCI_EXT_CAP_ID_MAX + 1] = {
320 	[0 ... PCI_EXT_CAP_ID_MAX] = { .readfn = vfio_direct_config_read }
321 };
322 /*
323  * Default unassigned regions to raw read-write access.  Some devices
324  * require this to function as they hide registers between the gaps in
325  * config space (be2net).  Like MMIO and I/O port registers, we have
326  * to trust the hardware isolation.
327  */
328 static struct perm_bits unassigned_perms = {
329 	.readfn = vfio_raw_config_read,
330 	.writefn = vfio_raw_config_write
331 };
332 
333 static struct perm_bits virt_perms = {
334 	.readfn = vfio_virt_config_read,
335 	.writefn = vfio_virt_config_write
336 };
337 
338 static void free_perm_bits(struct perm_bits *perm)
339 {
340 	kfree(perm->virt);
341 	kfree(perm->write);
342 	perm->virt = NULL;
343 	perm->write = NULL;
344 }
345 
346 static int alloc_perm_bits(struct perm_bits *perm, int size)
347 {
348 	/*
349 	 * Round up all permission bits to the next dword, this lets us
350 	 * ignore whether a read/write exceeds the defined capability
351 	 * structure.  We can do this because:
352 	 *  - Standard config space is already dword aligned
353 	 *  - Capabilities are all dword aligned (bits 0:1 of next reserved)
354 	 *  - Express capabilities defined as dword aligned
355 	 */
356 	size = round_up(size, 4);
357 
358 	/*
359 	 * Zero state is
360 	 * - All Readable, None Writeable, None Virtualized
361 	 */
362 	perm->virt = kzalloc(size, GFP_KERNEL);
363 	perm->write = kzalloc(size, GFP_KERNEL);
364 	if (!perm->virt || !perm->write) {
365 		free_perm_bits(perm);
366 		return -ENOMEM;
367 	}
368 
369 	perm->readfn = vfio_default_config_read;
370 	perm->writefn = vfio_default_config_write;
371 
372 	return 0;
373 }
374 
375 /*
376  * Helper functions for filling in permission tables
377  */
378 static inline void p_setb(struct perm_bits *p, int off, u8 virt, u8 write)
379 {
380 	p->virt[off] = virt;
381 	p->write[off] = write;
382 }
383 
384 /* Handle endian-ness - pci and tables are little-endian */
385 static inline void p_setw(struct perm_bits *p, int off, u16 virt, u16 write)
386 {
387 	*(__le16 *)(&p->virt[off]) = cpu_to_le16(virt);
388 	*(__le16 *)(&p->write[off]) = cpu_to_le16(write);
389 }
390 
391 /* Handle endian-ness - pci and tables are little-endian */
392 static inline void p_setd(struct perm_bits *p, int off, u32 virt, u32 write)
393 {
394 	*(__le32 *)(&p->virt[off]) = cpu_to_le32(virt);
395 	*(__le32 *)(&p->write[off]) = cpu_to_le32(write);
396 }
397 
398 /* Caller should hold memory_lock semaphore */
399 bool __vfio_pci_memory_enabled(struct vfio_pci_core_device *vdev)
400 {
401 	struct pci_dev *pdev = vdev->pdev;
402 	u16 cmd = le16_to_cpu(*(__le16 *)&vdev->vconfig[PCI_COMMAND]);
403 
404 	/*
405 	 * Memory region cannot be accessed if device power state is D3.
406 	 *
407 	 * SR-IOV VF memory enable is handled by the MSE bit in the
408 	 * PF SR-IOV capability, there's therefore no need to trigger
409 	 * faults based on the virtual value.
410 	 */
411 	return pdev->current_state < PCI_D3hot &&
412 	       (pdev->no_command_memory || (cmd & PCI_COMMAND_MEMORY));
413 }
414 
415 /*
416  * Restore the *real* BARs after we detect a FLR or backdoor reset.
417  * (backdoor = some device specific technique that we didn't catch)
418  */
419 static void vfio_bar_restore(struct vfio_pci_core_device *vdev)
420 {
421 	struct pci_dev *pdev = vdev->pdev;
422 	u32 *rbar = vdev->rbar;
423 	u16 cmd;
424 	int i;
425 
426 	if (pdev->is_virtfn)
427 		return;
428 
429 	pci_info(pdev, "%s: reset recovery - restoring BARs\n", __func__);
430 
431 	for (i = PCI_BASE_ADDRESS_0; i <= PCI_BASE_ADDRESS_5; i += 4, rbar++)
432 		pci_user_write_config_dword(pdev, i, *rbar);
433 
434 	pci_user_write_config_dword(pdev, PCI_ROM_ADDRESS, *rbar);
435 
436 	if (vdev->nointx) {
437 		pci_user_read_config_word(pdev, PCI_COMMAND, &cmd);
438 		cmd |= PCI_COMMAND_INTX_DISABLE;
439 		pci_user_write_config_word(pdev, PCI_COMMAND, cmd);
440 	}
441 }
442 
443 static __le32 vfio_generate_bar_flags(struct pci_dev *pdev, int bar)
444 {
445 	unsigned long flags = pci_resource_flags(pdev, bar);
446 	u32 val;
447 
448 	if (flags & IORESOURCE_IO)
449 		return cpu_to_le32(PCI_BASE_ADDRESS_SPACE_IO);
450 
451 	val = PCI_BASE_ADDRESS_SPACE_MEMORY;
452 
453 	if (flags & IORESOURCE_PREFETCH)
454 		val |= PCI_BASE_ADDRESS_MEM_PREFETCH;
455 
456 	if (flags & IORESOURCE_MEM_64)
457 		val |= PCI_BASE_ADDRESS_MEM_TYPE_64;
458 
459 	return cpu_to_le32(val);
460 }
461 
462 /*
463  * Pretend we're hardware and tweak the values of the *virtual* PCI BARs
464  * to reflect the hardware capabilities.  This implements BAR sizing.
465  */
466 static void vfio_bar_fixup(struct vfio_pci_core_device *vdev)
467 {
468 	struct pci_dev *pdev = vdev->pdev;
469 	int i;
470 	__le32 *vbar;
471 	u64 mask;
472 
473 	if (!vdev->bardirty)
474 		return;
475 
476 	vbar = (__le32 *)&vdev->vconfig[PCI_BASE_ADDRESS_0];
477 
478 	for (i = 0; i < PCI_STD_NUM_BARS; i++, vbar++) {
479 		int bar = i + PCI_STD_RESOURCES;
480 
481 		if (!pci_resource_start(pdev, bar)) {
482 			*vbar = 0; /* Unmapped by host = unimplemented to user */
483 			continue;
484 		}
485 
486 		mask = ~(pci_resource_len(pdev, bar) - 1);
487 
488 		*vbar &= cpu_to_le32((u32)mask);
489 		*vbar |= vfio_generate_bar_flags(pdev, bar);
490 
491 		if (*vbar & cpu_to_le32(PCI_BASE_ADDRESS_MEM_TYPE_64)) {
492 			vbar++;
493 			*vbar &= cpu_to_le32((u32)(mask >> 32));
494 			i++;
495 		}
496 	}
497 
498 	vbar = (__le32 *)&vdev->vconfig[PCI_ROM_ADDRESS];
499 
500 	/*
501 	 * NB. REGION_INFO will have reported zero size if we weren't able
502 	 * to read the ROM, but we still return the actual BAR size here if
503 	 * it exists (or the shadow ROM space).
504 	 */
505 	if (pci_resource_start(pdev, PCI_ROM_RESOURCE)) {
506 		mask = ~(pci_resource_len(pdev, PCI_ROM_RESOURCE) - 1);
507 		mask |= PCI_ROM_ADDRESS_ENABLE;
508 		*vbar &= cpu_to_le32((u32)mask);
509 	} else if (pdev->resource[PCI_ROM_RESOURCE].flags &
510 					IORESOURCE_ROM_SHADOW) {
511 		mask = ~(0x20000 - 1);
512 		mask |= PCI_ROM_ADDRESS_ENABLE;
513 		*vbar &= cpu_to_le32((u32)mask);
514 	} else
515 		*vbar = 0;
516 
517 	vdev->bardirty = false;
518 }
519 
520 static int vfio_basic_config_read(struct vfio_pci_core_device *vdev, int pos,
521 				  int count, struct perm_bits *perm,
522 				  int offset, __le32 *val)
523 {
524 	if (is_bar(offset)) /* pos == offset for basic config */
525 		vfio_bar_fixup(vdev);
526 
527 	count = vfio_default_config_read(vdev, pos, count, perm, offset, val);
528 
529 	/* Mask in virtual memory enable */
530 	if (offset == PCI_COMMAND && vdev->pdev->no_command_memory) {
531 		u16 cmd = le16_to_cpu(*(__le16 *)&vdev->vconfig[PCI_COMMAND]);
532 		u32 tmp_val = le32_to_cpu(*val);
533 
534 		tmp_val |= cmd & PCI_COMMAND_MEMORY;
535 		*val = cpu_to_le32(tmp_val);
536 	}
537 
538 	return count;
539 }
540 
541 /* Test whether BARs match the value we think they should contain */
542 static bool vfio_need_bar_restore(struct vfio_pci_core_device *vdev)
543 {
544 	int i = 0, pos = PCI_BASE_ADDRESS_0, ret;
545 	u32 bar;
546 
547 	for (; pos <= PCI_BASE_ADDRESS_5; i++, pos += 4) {
548 		if (vdev->rbar[i]) {
549 			ret = pci_user_read_config_dword(vdev->pdev, pos, &bar);
550 			if (ret || vdev->rbar[i] != bar)
551 				return true;
552 		}
553 	}
554 
555 	return false;
556 }
557 
558 static int vfio_basic_config_write(struct vfio_pci_core_device *vdev, int pos,
559 				   int count, struct perm_bits *perm,
560 				   int offset, __le32 val)
561 {
562 	struct pci_dev *pdev = vdev->pdev;
563 	__le16 *virt_cmd;
564 	u16 new_cmd = 0;
565 	int ret;
566 
567 	virt_cmd = (__le16 *)&vdev->vconfig[PCI_COMMAND];
568 
569 	if (offset == PCI_COMMAND) {
570 		bool phys_mem, virt_mem, new_mem, phys_io, virt_io, new_io;
571 		u16 phys_cmd;
572 
573 		ret = pci_user_read_config_word(pdev, PCI_COMMAND, &phys_cmd);
574 		if (ret)
575 			return ret;
576 
577 		new_cmd = le32_to_cpu(val);
578 
579 		phys_io = !!(phys_cmd & PCI_COMMAND_IO);
580 		virt_io = !!(le16_to_cpu(*virt_cmd) & PCI_COMMAND_IO);
581 		new_io = !!(new_cmd & PCI_COMMAND_IO);
582 
583 		phys_mem = !!(phys_cmd & PCI_COMMAND_MEMORY);
584 		virt_mem = !!(le16_to_cpu(*virt_cmd) & PCI_COMMAND_MEMORY);
585 		new_mem = !!(new_cmd & PCI_COMMAND_MEMORY);
586 
587 		if (!new_mem)
588 			vfio_pci_zap_and_down_write_memory_lock(vdev);
589 		else
590 			down_write(&vdev->memory_lock);
591 
592 		/*
593 		 * If the user is writing mem/io enable (new_mem/io) and we
594 		 * think it's already enabled (virt_mem/io), but the hardware
595 		 * shows it disabled (phys_mem/io, then the device has
596 		 * undergone some kind of backdoor reset and needs to be
597 		 * restored before we allow it to enable the bars.
598 		 * SR-IOV devices will trigger this - for mem enable let's
599 		 * catch this now and for io enable it will be caught later
600 		 */
601 		if ((new_mem && virt_mem && !phys_mem &&
602 		     !pdev->no_command_memory) ||
603 		    (new_io && virt_io && !phys_io) ||
604 		    vfio_need_bar_restore(vdev))
605 			vfio_bar_restore(vdev);
606 	}
607 
608 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
609 	if (count < 0) {
610 		if (offset == PCI_COMMAND)
611 			up_write(&vdev->memory_lock);
612 		return count;
613 	}
614 
615 	/*
616 	 * Save current memory/io enable bits in vconfig to allow for
617 	 * the test above next time.
618 	 */
619 	if (offset == PCI_COMMAND) {
620 		u16 mask = PCI_COMMAND_MEMORY | PCI_COMMAND_IO;
621 
622 		*virt_cmd &= cpu_to_le16(~mask);
623 		*virt_cmd |= cpu_to_le16(new_cmd & mask);
624 
625 		up_write(&vdev->memory_lock);
626 	}
627 
628 	/* Emulate INTx disable */
629 	if (offset >= PCI_COMMAND && offset <= PCI_COMMAND + 1) {
630 		bool virt_intx_disable;
631 
632 		virt_intx_disable = !!(le16_to_cpu(*virt_cmd) &
633 				       PCI_COMMAND_INTX_DISABLE);
634 
635 		if (virt_intx_disable && !vdev->virq_disabled) {
636 			vdev->virq_disabled = true;
637 			vfio_pci_intx_mask(vdev);
638 		} else if (!virt_intx_disable && vdev->virq_disabled) {
639 			vdev->virq_disabled = false;
640 			vfio_pci_intx_unmask(vdev);
641 		}
642 	}
643 
644 	if (is_bar(offset))
645 		vdev->bardirty = true;
646 
647 	return count;
648 }
649 
650 /* Permissions for the Basic PCI Header */
651 static int __init init_pci_cap_basic_perm(struct perm_bits *perm)
652 {
653 	if (alloc_perm_bits(perm, PCI_STD_HEADER_SIZEOF))
654 		return -ENOMEM;
655 
656 	perm->readfn = vfio_basic_config_read;
657 	perm->writefn = vfio_basic_config_write;
658 
659 	/* Virtualized for SR-IOV functions, which just have FFFF */
660 	p_setw(perm, PCI_VENDOR_ID, (u16)ALL_VIRT, NO_WRITE);
661 	p_setw(perm, PCI_DEVICE_ID, (u16)ALL_VIRT, NO_WRITE);
662 
663 	/*
664 	 * Virtualize INTx disable, we use it internally for interrupt
665 	 * control and can emulate it for non-PCI 2.3 devices.
666 	 */
667 	p_setw(perm, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE, (u16)ALL_WRITE);
668 
669 	/* Virtualize capability list, we might want to skip/disable */
670 	p_setw(perm, PCI_STATUS, PCI_STATUS_CAP_LIST, NO_WRITE);
671 
672 	/* No harm to write */
673 	p_setb(perm, PCI_CACHE_LINE_SIZE, NO_VIRT, (u8)ALL_WRITE);
674 	p_setb(perm, PCI_LATENCY_TIMER, NO_VIRT, (u8)ALL_WRITE);
675 	p_setb(perm, PCI_BIST, NO_VIRT, (u8)ALL_WRITE);
676 
677 	/* Virtualize all bars, can't touch the real ones */
678 	p_setd(perm, PCI_BASE_ADDRESS_0, ALL_VIRT, ALL_WRITE);
679 	p_setd(perm, PCI_BASE_ADDRESS_1, ALL_VIRT, ALL_WRITE);
680 	p_setd(perm, PCI_BASE_ADDRESS_2, ALL_VIRT, ALL_WRITE);
681 	p_setd(perm, PCI_BASE_ADDRESS_3, ALL_VIRT, ALL_WRITE);
682 	p_setd(perm, PCI_BASE_ADDRESS_4, ALL_VIRT, ALL_WRITE);
683 	p_setd(perm, PCI_BASE_ADDRESS_5, ALL_VIRT, ALL_WRITE);
684 	p_setd(perm, PCI_ROM_ADDRESS, ALL_VIRT, ALL_WRITE);
685 
686 	/* Allow us to adjust capability chain */
687 	p_setb(perm, PCI_CAPABILITY_LIST, (u8)ALL_VIRT, NO_WRITE);
688 
689 	/* Sometimes used by sw, just virtualize */
690 	p_setb(perm, PCI_INTERRUPT_LINE, (u8)ALL_VIRT, (u8)ALL_WRITE);
691 
692 	/* Virtualize interrupt pin to allow hiding INTx */
693 	p_setb(perm, PCI_INTERRUPT_PIN, (u8)ALL_VIRT, (u8)NO_WRITE);
694 
695 	return 0;
696 }
697 
698 /*
699  * It takes all the required locks to protect the access of power related
700  * variables and then invokes vfio_pci_set_power_state().
701  */
702 static void vfio_lock_and_set_power_state(struct vfio_pci_core_device *vdev,
703 					  pci_power_t state)
704 {
705 	if (state >= PCI_D3hot)
706 		vfio_pci_zap_and_down_write_memory_lock(vdev);
707 	else
708 		down_write(&vdev->memory_lock);
709 
710 	vfio_pci_set_power_state(vdev, state);
711 	up_write(&vdev->memory_lock);
712 }
713 
714 static int vfio_pm_config_write(struct vfio_pci_core_device *vdev, int pos,
715 				int count, struct perm_bits *perm,
716 				int offset, __le32 val)
717 {
718 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
719 	if (count < 0)
720 		return count;
721 
722 	if (offset == PCI_PM_CTRL) {
723 		pci_power_t state;
724 
725 		switch (le32_to_cpu(val) & PCI_PM_CTRL_STATE_MASK) {
726 		case 0:
727 			state = PCI_D0;
728 			break;
729 		case 1:
730 			state = PCI_D1;
731 			break;
732 		case 2:
733 			state = PCI_D2;
734 			break;
735 		case 3:
736 			state = PCI_D3hot;
737 			break;
738 		}
739 
740 		vfio_lock_and_set_power_state(vdev, state);
741 	}
742 
743 	return count;
744 }
745 
746 /* Permissions for the Power Management capability */
747 static int __init init_pci_cap_pm_perm(struct perm_bits *perm)
748 {
749 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_PM]))
750 		return -ENOMEM;
751 
752 	perm->writefn = vfio_pm_config_write;
753 
754 	/*
755 	 * We always virtualize the next field so we can remove
756 	 * capabilities from the chain if we want to.
757 	 */
758 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
759 
760 	/*
761 	 * The guests can't process PME events. If any PME event will be
762 	 * generated, then it will be mostly handled in the host and the
763 	 * host will clear the PME_STATUS. So virtualize PME_Support bits.
764 	 * The vconfig bits will be cleared during device capability
765 	 * initialization.
766 	 */
767 	p_setw(perm, PCI_PM_PMC, PCI_PM_CAP_PME_MASK, NO_WRITE);
768 
769 	/*
770 	 * Power management is defined *per function*, so we can let
771 	 * the user change power state, but we trap and initiate the
772 	 * change ourselves, so the state bits are read-only.
773 	 *
774 	 * The guest can't process PME from D3cold so virtualize PME_Status
775 	 * and PME_En bits. The vconfig bits will be cleared during device
776 	 * capability initialization.
777 	 */
778 	p_setd(perm, PCI_PM_CTRL,
779 	       PCI_PM_CTRL_PME_ENABLE | PCI_PM_CTRL_PME_STATUS,
780 	       ~(PCI_PM_CTRL_PME_ENABLE | PCI_PM_CTRL_PME_STATUS |
781 		 PCI_PM_CTRL_STATE_MASK));
782 
783 	return 0;
784 }
785 
786 static int vfio_vpd_config_write(struct vfio_pci_core_device *vdev, int pos,
787 				 int count, struct perm_bits *perm,
788 				 int offset, __le32 val)
789 {
790 	struct pci_dev *pdev = vdev->pdev;
791 	__le16 *paddr = (__le16 *)(vdev->vconfig + pos - offset + PCI_VPD_ADDR);
792 	__le32 *pdata = (__le32 *)(vdev->vconfig + pos - offset + PCI_VPD_DATA);
793 	u16 addr;
794 	u32 data;
795 
796 	/*
797 	 * Write through to emulation.  If the write includes the upper byte
798 	 * of PCI_VPD_ADDR, then the PCI_VPD_ADDR_F bit is written and we
799 	 * have work to do.
800 	 */
801 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
802 	if (count < 0 || offset > PCI_VPD_ADDR + 1 ||
803 	    offset + count <= PCI_VPD_ADDR + 1)
804 		return count;
805 
806 	addr = le16_to_cpu(*paddr);
807 
808 	if (addr & PCI_VPD_ADDR_F) {
809 		data = le32_to_cpu(*pdata);
810 		if (pci_write_vpd(pdev, addr & ~PCI_VPD_ADDR_F, 4, &data) != 4)
811 			return count;
812 	} else {
813 		data = 0;
814 		if (pci_read_vpd(pdev, addr, 4, &data) < 0)
815 			return count;
816 		*pdata = cpu_to_le32(data);
817 	}
818 
819 	/*
820 	 * Toggle PCI_VPD_ADDR_F in the emulated PCI_VPD_ADDR register to
821 	 * signal completion.  If an error occurs above, we assume that not
822 	 * toggling this bit will induce a driver timeout.
823 	 */
824 	addr ^= PCI_VPD_ADDR_F;
825 	*paddr = cpu_to_le16(addr);
826 
827 	return count;
828 }
829 
830 /* Permissions for Vital Product Data capability */
831 static int __init init_pci_cap_vpd_perm(struct perm_bits *perm)
832 {
833 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_VPD]))
834 		return -ENOMEM;
835 
836 	perm->writefn = vfio_vpd_config_write;
837 
838 	/*
839 	 * We always virtualize the next field so we can remove
840 	 * capabilities from the chain if we want to.
841 	 */
842 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
843 
844 	/*
845 	 * Both the address and data registers are virtualized to
846 	 * enable access through the pci_vpd_read/write functions
847 	 */
848 	p_setw(perm, PCI_VPD_ADDR, (u16)ALL_VIRT, (u16)ALL_WRITE);
849 	p_setd(perm, PCI_VPD_DATA, ALL_VIRT, ALL_WRITE);
850 
851 	return 0;
852 }
853 
854 /* Permissions for PCI-X capability */
855 static int __init init_pci_cap_pcix_perm(struct perm_bits *perm)
856 {
857 	/* Alloc 24, but only 8 are used in v0 */
858 	if (alloc_perm_bits(perm, PCI_CAP_PCIX_SIZEOF_V2))
859 		return -ENOMEM;
860 
861 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
862 
863 	p_setw(perm, PCI_X_CMD, NO_VIRT, (u16)ALL_WRITE);
864 	p_setd(perm, PCI_X_ECC_CSR, NO_VIRT, ALL_WRITE);
865 	return 0;
866 }
867 
868 static int vfio_exp_config_write(struct vfio_pci_core_device *vdev, int pos,
869 				 int count, struct perm_bits *perm,
870 				 int offset, __le32 val)
871 {
872 	__le16 *ctrl = (__le16 *)(vdev->vconfig + pos -
873 				  offset + PCI_EXP_DEVCTL);
874 	int readrq = le16_to_cpu(*ctrl) & PCI_EXP_DEVCTL_READRQ;
875 
876 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
877 	if (count < 0)
878 		return count;
879 
880 	/*
881 	 * The FLR bit is virtualized, if set and the device supports PCIe
882 	 * FLR, issue a reset_function.  Regardless, clear the bit, the spec
883 	 * requires it to be always read as zero.  NB, reset_function might
884 	 * not use a PCIe FLR, we don't have that level of granularity.
885 	 */
886 	if (*ctrl & cpu_to_le16(PCI_EXP_DEVCTL_BCR_FLR)) {
887 		u32 cap;
888 		int ret;
889 
890 		*ctrl &= ~cpu_to_le16(PCI_EXP_DEVCTL_BCR_FLR);
891 
892 		ret = pci_user_read_config_dword(vdev->pdev,
893 						 pos - offset + PCI_EXP_DEVCAP,
894 						 &cap);
895 
896 		if (!ret && (cap & PCI_EXP_DEVCAP_FLR)) {
897 			vfio_pci_zap_and_down_write_memory_lock(vdev);
898 			pci_try_reset_function(vdev->pdev);
899 			up_write(&vdev->memory_lock);
900 		}
901 	}
902 
903 	/*
904 	 * MPS is virtualized to the user, writes do not change the physical
905 	 * register since determining a proper MPS value requires a system wide
906 	 * device view.  The MRRS is largely independent of MPS, but since the
907 	 * user does not have that system-wide view, they might set a safe, but
908 	 * inefficiently low value.  Here we allow writes through to hardware,
909 	 * but we set the floor to the physical device MPS setting, so that
910 	 * we can at least use full TLPs, as defined by the MPS value.
911 	 *
912 	 * NB, if any devices actually depend on an artificially low MRRS
913 	 * setting, this will need to be revisited, perhaps with a quirk
914 	 * though pcie_set_readrq().
915 	 */
916 	if (readrq != (le16_to_cpu(*ctrl) & PCI_EXP_DEVCTL_READRQ)) {
917 		readrq = 128 <<
918 			((le16_to_cpu(*ctrl) & PCI_EXP_DEVCTL_READRQ) >> 12);
919 		readrq = max(readrq, pcie_get_mps(vdev->pdev));
920 
921 		pcie_set_readrq(vdev->pdev, readrq);
922 	}
923 
924 	return count;
925 }
926 
927 /* Permissions for PCI Express capability */
928 static int __init init_pci_cap_exp_perm(struct perm_bits *perm)
929 {
930 	/* Alloc largest of possible sizes */
931 	if (alloc_perm_bits(perm, PCI_CAP_EXP_ENDPOINT_SIZEOF_V2))
932 		return -ENOMEM;
933 
934 	perm->writefn = vfio_exp_config_write;
935 
936 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
937 
938 	/*
939 	 * Allow writes to device control fields, except devctl_phantom,
940 	 * which could confuse IOMMU, MPS, which can break communication
941 	 * with other physical devices, and the ARI bit in devctl2, which
942 	 * is set at probe time.  FLR and MRRS get virtualized via our
943 	 * writefn.
944 	 */
945 	p_setw(perm, PCI_EXP_DEVCTL,
946 	       PCI_EXP_DEVCTL_BCR_FLR | PCI_EXP_DEVCTL_PAYLOAD |
947 	       PCI_EXP_DEVCTL_READRQ, ~PCI_EXP_DEVCTL_PHANTOM);
948 	p_setw(perm, PCI_EXP_DEVCTL2, NO_VIRT, ~PCI_EXP_DEVCTL2_ARI);
949 	return 0;
950 }
951 
952 static int vfio_af_config_write(struct vfio_pci_core_device *vdev, int pos,
953 				int count, struct perm_bits *perm,
954 				int offset, __le32 val)
955 {
956 	u8 *ctrl = vdev->vconfig + pos - offset + PCI_AF_CTRL;
957 
958 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
959 	if (count < 0)
960 		return count;
961 
962 	/*
963 	 * The FLR bit is virtualized, if set and the device supports AF
964 	 * FLR, issue a reset_function.  Regardless, clear the bit, the spec
965 	 * requires it to be always read as zero.  NB, reset_function might
966 	 * not use an AF FLR, we don't have that level of granularity.
967 	 */
968 	if (*ctrl & PCI_AF_CTRL_FLR) {
969 		u8 cap;
970 		int ret;
971 
972 		*ctrl &= ~PCI_AF_CTRL_FLR;
973 
974 		ret = pci_user_read_config_byte(vdev->pdev,
975 						pos - offset + PCI_AF_CAP,
976 						&cap);
977 
978 		if (!ret && (cap & PCI_AF_CAP_FLR) && (cap & PCI_AF_CAP_TP)) {
979 			vfio_pci_zap_and_down_write_memory_lock(vdev);
980 			pci_try_reset_function(vdev->pdev);
981 			up_write(&vdev->memory_lock);
982 		}
983 	}
984 
985 	return count;
986 }
987 
988 /* Permissions for Advanced Function capability */
989 static int __init init_pci_cap_af_perm(struct perm_bits *perm)
990 {
991 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_AF]))
992 		return -ENOMEM;
993 
994 	perm->writefn = vfio_af_config_write;
995 
996 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
997 	p_setb(perm, PCI_AF_CTRL, PCI_AF_CTRL_FLR, PCI_AF_CTRL_FLR);
998 	return 0;
999 }
1000 
1001 /* Permissions for Advanced Error Reporting extended capability */
1002 static int __init init_pci_ext_cap_err_perm(struct perm_bits *perm)
1003 {
1004 	u32 mask;
1005 
1006 	if (alloc_perm_bits(perm, pci_ext_cap_length[PCI_EXT_CAP_ID_ERR]))
1007 		return -ENOMEM;
1008 
1009 	/*
1010 	 * Virtualize the first dword of all express capabilities
1011 	 * because it includes the next pointer.  This lets us later
1012 	 * remove capabilities from the chain if we need to.
1013 	 */
1014 	p_setd(perm, 0, ALL_VIRT, NO_WRITE);
1015 
1016 	/* Writable bits mask */
1017 	mask =	PCI_ERR_UNC_UND |		/* Undefined */
1018 		PCI_ERR_UNC_DLP |		/* Data Link Protocol */
1019 		PCI_ERR_UNC_SURPDN |		/* Surprise Down */
1020 		PCI_ERR_UNC_POISON_TLP |	/* Poisoned TLP */
1021 		PCI_ERR_UNC_FCP |		/* Flow Control Protocol */
1022 		PCI_ERR_UNC_COMP_TIME |		/* Completion Timeout */
1023 		PCI_ERR_UNC_COMP_ABORT |	/* Completer Abort */
1024 		PCI_ERR_UNC_UNX_COMP |		/* Unexpected Completion */
1025 		PCI_ERR_UNC_RX_OVER |		/* Receiver Overflow */
1026 		PCI_ERR_UNC_MALF_TLP |		/* Malformed TLP */
1027 		PCI_ERR_UNC_ECRC |		/* ECRC Error Status */
1028 		PCI_ERR_UNC_UNSUP |		/* Unsupported Request */
1029 		PCI_ERR_UNC_ACSV |		/* ACS Violation */
1030 		PCI_ERR_UNC_INTN |		/* internal error */
1031 		PCI_ERR_UNC_MCBTLP |		/* MC blocked TLP */
1032 		PCI_ERR_UNC_ATOMEG |		/* Atomic egress blocked */
1033 		PCI_ERR_UNC_TLPPRE;		/* TLP prefix blocked */
1034 	p_setd(perm, PCI_ERR_UNCOR_STATUS, NO_VIRT, mask);
1035 	p_setd(perm, PCI_ERR_UNCOR_MASK, NO_VIRT, mask);
1036 	p_setd(perm, PCI_ERR_UNCOR_SEVER, NO_VIRT, mask);
1037 
1038 	mask =	PCI_ERR_COR_RCVR |		/* Receiver Error Status */
1039 		PCI_ERR_COR_BAD_TLP |		/* Bad TLP Status */
1040 		PCI_ERR_COR_BAD_DLLP |		/* Bad DLLP Status */
1041 		PCI_ERR_COR_REP_ROLL |		/* REPLAY_NUM Rollover */
1042 		PCI_ERR_COR_REP_TIMER |		/* Replay Timer Timeout */
1043 		PCI_ERR_COR_ADV_NFAT |		/* Advisory Non-Fatal */
1044 		PCI_ERR_COR_INTERNAL |		/* Corrected Internal */
1045 		PCI_ERR_COR_LOG_OVER;		/* Header Log Overflow */
1046 	p_setd(perm, PCI_ERR_COR_STATUS, NO_VIRT, mask);
1047 	p_setd(perm, PCI_ERR_COR_MASK, NO_VIRT, mask);
1048 
1049 	mask =	PCI_ERR_CAP_ECRC_GENE |		/* ECRC Generation Enable */
1050 		PCI_ERR_CAP_ECRC_CHKE;		/* ECRC Check Enable */
1051 	p_setd(perm, PCI_ERR_CAP, NO_VIRT, mask);
1052 	return 0;
1053 }
1054 
1055 /* Permissions for Power Budgeting extended capability */
1056 static int __init init_pci_ext_cap_pwr_perm(struct perm_bits *perm)
1057 {
1058 	if (alloc_perm_bits(perm, pci_ext_cap_length[PCI_EXT_CAP_ID_PWR]))
1059 		return -ENOMEM;
1060 
1061 	p_setd(perm, 0, ALL_VIRT, NO_WRITE);
1062 
1063 	/* Writing the data selector is OK, the info is still read-only */
1064 	p_setb(perm, PCI_PWR_DATA, NO_VIRT, (u8)ALL_WRITE);
1065 	return 0;
1066 }
1067 
1068 /*
1069  * Initialize the shared permission tables
1070  */
1071 void vfio_pci_uninit_perm_bits(void)
1072 {
1073 	free_perm_bits(&cap_perms[PCI_CAP_ID_BASIC]);
1074 
1075 	free_perm_bits(&cap_perms[PCI_CAP_ID_PM]);
1076 	free_perm_bits(&cap_perms[PCI_CAP_ID_VPD]);
1077 	free_perm_bits(&cap_perms[PCI_CAP_ID_PCIX]);
1078 	free_perm_bits(&cap_perms[PCI_CAP_ID_EXP]);
1079 	free_perm_bits(&cap_perms[PCI_CAP_ID_AF]);
1080 
1081 	free_perm_bits(&ecap_perms[PCI_EXT_CAP_ID_ERR]);
1082 	free_perm_bits(&ecap_perms[PCI_EXT_CAP_ID_PWR]);
1083 }
1084 
1085 int __init vfio_pci_init_perm_bits(void)
1086 {
1087 	int ret;
1088 
1089 	/* Basic config space */
1090 	ret = init_pci_cap_basic_perm(&cap_perms[PCI_CAP_ID_BASIC]);
1091 
1092 	/* Capabilities */
1093 	ret |= init_pci_cap_pm_perm(&cap_perms[PCI_CAP_ID_PM]);
1094 	ret |= init_pci_cap_vpd_perm(&cap_perms[PCI_CAP_ID_VPD]);
1095 	ret |= init_pci_cap_pcix_perm(&cap_perms[PCI_CAP_ID_PCIX]);
1096 	cap_perms[PCI_CAP_ID_VNDR].writefn = vfio_raw_config_write;
1097 	ret |= init_pci_cap_exp_perm(&cap_perms[PCI_CAP_ID_EXP]);
1098 	ret |= init_pci_cap_af_perm(&cap_perms[PCI_CAP_ID_AF]);
1099 
1100 	/* Extended capabilities */
1101 	ret |= init_pci_ext_cap_err_perm(&ecap_perms[PCI_EXT_CAP_ID_ERR]);
1102 	ret |= init_pci_ext_cap_pwr_perm(&ecap_perms[PCI_EXT_CAP_ID_PWR]);
1103 	ecap_perms[PCI_EXT_CAP_ID_VNDR].writefn = vfio_raw_config_write;
1104 
1105 	if (ret)
1106 		vfio_pci_uninit_perm_bits();
1107 
1108 	return ret;
1109 }
1110 
1111 static int vfio_find_cap_start(struct vfio_pci_core_device *vdev, int pos)
1112 {
1113 	u8 cap;
1114 	int base = (pos >= PCI_CFG_SPACE_SIZE) ? PCI_CFG_SPACE_SIZE :
1115 						 PCI_STD_HEADER_SIZEOF;
1116 	cap = vdev->pci_config_map[pos];
1117 
1118 	if (cap == PCI_CAP_ID_BASIC)
1119 		return 0;
1120 
1121 	/* XXX Can we have to abutting capabilities of the same type? */
1122 	while (pos - 1 >= base && vdev->pci_config_map[pos - 1] == cap)
1123 		pos--;
1124 
1125 	return pos;
1126 }
1127 
1128 static int vfio_msi_config_read(struct vfio_pci_core_device *vdev, int pos,
1129 				int count, struct perm_bits *perm,
1130 				int offset, __le32 *val)
1131 {
1132 	/* Update max available queue size from msi_qmax */
1133 	if (offset <= PCI_MSI_FLAGS && offset + count >= PCI_MSI_FLAGS) {
1134 		__le16 *flags;
1135 		int start;
1136 
1137 		start = vfio_find_cap_start(vdev, pos);
1138 
1139 		flags = (__le16 *)&vdev->vconfig[start];
1140 
1141 		*flags &= cpu_to_le16(~PCI_MSI_FLAGS_QMASK);
1142 		*flags |= cpu_to_le16(vdev->msi_qmax << 1);
1143 	}
1144 
1145 	return vfio_default_config_read(vdev, pos, count, perm, offset, val);
1146 }
1147 
1148 static int vfio_msi_config_write(struct vfio_pci_core_device *vdev, int pos,
1149 				 int count, struct perm_bits *perm,
1150 				 int offset, __le32 val)
1151 {
1152 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
1153 	if (count < 0)
1154 		return count;
1155 
1156 	/* Fixup and write configured queue size and enable to hardware */
1157 	if (offset <= PCI_MSI_FLAGS && offset + count >= PCI_MSI_FLAGS) {
1158 		__le16 *pflags;
1159 		u16 flags;
1160 		int start, ret;
1161 
1162 		start = vfio_find_cap_start(vdev, pos);
1163 
1164 		pflags = (__le16 *)&vdev->vconfig[start + PCI_MSI_FLAGS];
1165 
1166 		flags = le16_to_cpu(*pflags);
1167 
1168 		/* MSI is enabled via ioctl */
1169 		if  (!is_msi(vdev))
1170 			flags &= ~PCI_MSI_FLAGS_ENABLE;
1171 
1172 		/* Check queue size */
1173 		if ((flags & PCI_MSI_FLAGS_QSIZE) >> 4 > vdev->msi_qmax) {
1174 			flags &= ~PCI_MSI_FLAGS_QSIZE;
1175 			flags |= vdev->msi_qmax << 4;
1176 		}
1177 
1178 		/* Write back to virt and to hardware */
1179 		*pflags = cpu_to_le16(flags);
1180 		ret = pci_user_write_config_word(vdev->pdev,
1181 						 start + PCI_MSI_FLAGS,
1182 						 flags);
1183 		if (ret)
1184 			return ret;
1185 	}
1186 
1187 	return count;
1188 }
1189 
1190 /*
1191  * MSI determination is per-device, so this routine gets used beyond
1192  * initialization time. Don't add __init
1193  */
1194 static int init_pci_cap_msi_perm(struct perm_bits *perm, int len, u16 flags)
1195 {
1196 	if (alloc_perm_bits(perm, len))
1197 		return -ENOMEM;
1198 
1199 	perm->readfn = vfio_msi_config_read;
1200 	perm->writefn = vfio_msi_config_write;
1201 
1202 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
1203 
1204 	/*
1205 	 * The upper byte of the control register is reserved,
1206 	 * just setup the lower byte.
1207 	 */
1208 	p_setb(perm, PCI_MSI_FLAGS, (u8)ALL_VIRT, (u8)ALL_WRITE);
1209 	p_setd(perm, PCI_MSI_ADDRESS_LO, ALL_VIRT, ALL_WRITE);
1210 	if (flags & PCI_MSI_FLAGS_64BIT) {
1211 		p_setd(perm, PCI_MSI_ADDRESS_HI, ALL_VIRT, ALL_WRITE);
1212 		p_setw(perm, PCI_MSI_DATA_64, (u16)ALL_VIRT, (u16)ALL_WRITE);
1213 		if (flags & PCI_MSI_FLAGS_MASKBIT) {
1214 			p_setd(perm, PCI_MSI_MASK_64, NO_VIRT, ALL_WRITE);
1215 			p_setd(perm, PCI_MSI_PENDING_64, NO_VIRT, ALL_WRITE);
1216 		}
1217 	} else {
1218 		p_setw(perm, PCI_MSI_DATA_32, (u16)ALL_VIRT, (u16)ALL_WRITE);
1219 		if (flags & PCI_MSI_FLAGS_MASKBIT) {
1220 			p_setd(perm, PCI_MSI_MASK_32, NO_VIRT, ALL_WRITE);
1221 			p_setd(perm, PCI_MSI_PENDING_32, NO_VIRT, ALL_WRITE);
1222 		}
1223 	}
1224 	return 0;
1225 }
1226 
1227 /* Determine MSI CAP field length; initialize msi_perms on 1st call per vdev */
1228 static int vfio_msi_cap_len(struct vfio_pci_core_device *vdev, u8 pos)
1229 {
1230 	struct pci_dev *pdev = vdev->pdev;
1231 	int len, ret;
1232 	u16 flags;
1233 
1234 	ret = pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &flags);
1235 	if (ret)
1236 		return pcibios_err_to_errno(ret);
1237 
1238 	len = 10; /* Minimum size */
1239 	if (flags & PCI_MSI_FLAGS_64BIT)
1240 		len += 4;
1241 	if (flags & PCI_MSI_FLAGS_MASKBIT)
1242 		len += 10;
1243 
1244 	if (vdev->msi_perm)
1245 		return len;
1246 
1247 	vdev->msi_perm = kmalloc(sizeof(struct perm_bits), GFP_KERNEL);
1248 	if (!vdev->msi_perm)
1249 		return -ENOMEM;
1250 
1251 	ret = init_pci_cap_msi_perm(vdev->msi_perm, len, flags);
1252 	if (ret) {
1253 		kfree(vdev->msi_perm);
1254 		return ret;
1255 	}
1256 
1257 	return len;
1258 }
1259 
1260 /* Determine extended capability length for VC (2 & 9) and MFVC */
1261 static int vfio_vc_cap_len(struct vfio_pci_core_device *vdev, u16 pos)
1262 {
1263 	struct pci_dev *pdev = vdev->pdev;
1264 	u32 tmp;
1265 	int ret, evcc, phases, vc_arb;
1266 	int len = PCI_CAP_VC_BASE_SIZEOF;
1267 
1268 	ret = pci_read_config_dword(pdev, pos + PCI_VC_PORT_CAP1, &tmp);
1269 	if (ret)
1270 		return pcibios_err_to_errno(ret);
1271 
1272 	evcc = tmp & PCI_VC_CAP1_EVCC; /* extended vc count */
1273 	ret = pci_read_config_dword(pdev, pos + PCI_VC_PORT_CAP2, &tmp);
1274 	if (ret)
1275 		return pcibios_err_to_errno(ret);
1276 
1277 	if (tmp & PCI_VC_CAP2_128_PHASE)
1278 		phases = 128;
1279 	else if (tmp & PCI_VC_CAP2_64_PHASE)
1280 		phases = 64;
1281 	else if (tmp & PCI_VC_CAP2_32_PHASE)
1282 		phases = 32;
1283 	else
1284 		phases = 0;
1285 
1286 	vc_arb = phases * 4;
1287 
1288 	/*
1289 	 * Port arbitration tables are root & switch only;
1290 	 * function arbitration tables are function 0 only.
1291 	 * In either case, we'll never let user write them so
1292 	 * we don't care how big they are
1293 	 */
1294 	len += (1 + evcc) * PCI_CAP_VC_PER_VC_SIZEOF;
1295 	if (vc_arb) {
1296 		len = round_up(len, 16);
1297 		len += vc_arb / 8;
1298 	}
1299 	return len;
1300 }
1301 
1302 static int vfio_cap_len(struct vfio_pci_core_device *vdev, u8 cap, u8 pos)
1303 {
1304 	struct pci_dev *pdev = vdev->pdev;
1305 	u32 dword;
1306 	u16 word;
1307 	u8 byte;
1308 	int ret;
1309 
1310 	switch (cap) {
1311 	case PCI_CAP_ID_MSI:
1312 		return vfio_msi_cap_len(vdev, pos);
1313 	case PCI_CAP_ID_PCIX:
1314 		ret = pci_read_config_word(pdev, pos + PCI_X_CMD, &word);
1315 		if (ret)
1316 			return pcibios_err_to_errno(ret);
1317 
1318 		if (PCI_X_CMD_VERSION(word)) {
1319 			if (pdev->cfg_size > PCI_CFG_SPACE_SIZE) {
1320 				/* Test for extended capabilities */
1321 				pci_read_config_dword(pdev, PCI_CFG_SPACE_SIZE,
1322 						      &dword);
1323 				vdev->extended_caps = (dword != 0);
1324 			}
1325 			return PCI_CAP_PCIX_SIZEOF_V2;
1326 		} else
1327 			return PCI_CAP_PCIX_SIZEOF_V0;
1328 	case PCI_CAP_ID_VNDR:
1329 		/* length follows next field */
1330 		ret = pci_read_config_byte(pdev, pos + PCI_CAP_FLAGS, &byte);
1331 		if (ret)
1332 			return pcibios_err_to_errno(ret);
1333 
1334 		return byte;
1335 	case PCI_CAP_ID_EXP:
1336 		if (pdev->cfg_size > PCI_CFG_SPACE_SIZE) {
1337 			/* Test for extended capabilities */
1338 			pci_read_config_dword(pdev, PCI_CFG_SPACE_SIZE, &dword);
1339 			vdev->extended_caps = (dword != 0);
1340 		}
1341 
1342 		/* length based on version and type */
1343 		if ((pcie_caps_reg(pdev) & PCI_EXP_FLAGS_VERS) == 1) {
1344 			if (pci_pcie_type(pdev) == PCI_EXP_TYPE_RC_END)
1345 				return 0xc; /* "All Devices" only, no link */
1346 			return PCI_CAP_EXP_ENDPOINT_SIZEOF_V1;
1347 		} else {
1348 			if (pci_pcie_type(pdev) == PCI_EXP_TYPE_RC_END)
1349 				return 0x2c; /* No link */
1350 			return PCI_CAP_EXP_ENDPOINT_SIZEOF_V2;
1351 		}
1352 	case PCI_CAP_ID_HT:
1353 		ret = pci_read_config_byte(pdev, pos + 3, &byte);
1354 		if (ret)
1355 			return pcibios_err_to_errno(ret);
1356 
1357 		return (byte & HT_3BIT_CAP_MASK) ?
1358 			HT_CAP_SIZEOF_SHORT : HT_CAP_SIZEOF_LONG;
1359 	case PCI_CAP_ID_SATA:
1360 		ret = pci_read_config_byte(pdev, pos + PCI_SATA_REGS, &byte);
1361 		if (ret)
1362 			return pcibios_err_to_errno(ret);
1363 
1364 		byte &= PCI_SATA_REGS_MASK;
1365 		if (byte == PCI_SATA_REGS_INLINE)
1366 			return PCI_SATA_SIZEOF_LONG;
1367 		else
1368 			return PCI_SATA_SIZEOF_SHORT;
1369 	default:
1370 		pci_warn(pdev, "%s: unknown length for PCI cap %#x@%#x\n",
1371 			 __func__, cap, pos);
1372 	}
1373 
1374 	return 0;
1375 }
1376 
1377 static int vfio_ext_cap_len(struct vfio_pci_core_device *vdev, u16 ecap, u16 epos)
1378 {
1379 	struct pci_dev *pdev = vdev->pdev;
1380 	u8 byte;
1381 	u32 dword;
1382 	int ret;
1383 
1384 	switch (ecap) {
1385 	case PCI_EXT_CAP_ID_VNDR:
1386 		ret = pci_read_config_dword(pdev, epos + PCI_VSEC_HDR, &dword);
1387 		if (ret)
1388 			return pcibios_err_to_errno(ret);
1389 
1390 		return dword >> PCI_VSEC_HDR_LEN_SHIFT;
1391 	case PCI_EXT_CAP_ID_VC:
1392 	case PCI_EXT_CAP_ID_VC9:
1393 	case PCI_EXT_CAP_ID_MFVC:
1394 		return vfio_vc_cap_len(vdev, epos);
1395 	case PCI_EXT_CAP_ID_ACS:
1396 		ret = pci_read_config_byte(pdev, epos + PCI_ACS_CAP, &byte);
1397 		if (ret)
1398 			return pcibios_err_to_errno(ret);
1399 
1400 		if (byte & PCI_ACS_EC) {
1401 			int bits;
1402 
1403 			ret = pci_read_config_byte(pdev,
1404 						   epos + PCI_ACS_EGRESS_BITS,
1405 						   &byte);
1406 			if (ret)
1407 				return pcibios_err_to_errno(ret);
1408 
1409 			bits = byte ? round_up(byte, 32) : 256;
1410 			return 8 + (bits / 8);
1411 		}
1412 		return 8;
1413 
1414 	case PCI_EXT_CAP_ID_REBAR:
1415 		ret = pci_read_config_byte(pdev, epos + PCI_REBAR_CTRL, &byte);
1416 		if (ret)
1417 			return pcibios_err_to_errno(ret);
1418 
1419 		byte &= PCI_REBAR_CTRL_NBAR_MASK;
1420 		byte >>= PCI_REBAR_CTRL_NBAR_SHIFT;
1421 
1422 		return 4 + (byte * 8);
1423 	case PCI_EXT_CAP_ID_DPA:
1424 		ret = pci_read_config_byte(pdev, epos + PCI_DPA_CAP, &byte);
1425 		if (ret)
1426 			return pcibios_err_to_errno(ret);
1427 
1428 		byte &= PCI_DPA_CAP_SUBSTATE_MASK;
1429 		return PCI_DPA_BASE_SIZEOF + byte + 1;
1430 	case PCI_EXT_CAP_ID_TPH:
1431 		ret = pci_read_config_dword(pdev, epos + PCI_TPH_CAP, &dword);
1432 		if (ret)
1433 			return pcibios_err_to_errno(ret);
1434 
1435 		if ((dword & PCI_TPH_CAP_LOC_MASK) == PCI_TPH_LOC_CAP) {
1436 			int sts;
1437 
1438 			sts = dword & PCI_TPH_CAP_ST_MASK;
1439 			sts >>= PCI_TPH_CAP_ST_SHIFT;
1440 			return PCI_TPH_BASE_SIZEOF + (sts * 2) + 2;
1441 		}
1442 		return PCI_TPH_BASE_SIZEOF;
1443 	default:
1444 		pci_warn(pdev, "%s: unknown length for PCI ecap %#x@%#x\n",
1445 			 __func__, ecap, epos);
1446 	}
1447 
1448 	return 0;
1449 }
1450 
1451 static void vfio_update_pm_vconfig_bytes(struct vfio_pci_core_device *vdev,
1452 					 int offset)
1453 {
1454 	__le16 *pmc = (__le16 *)&vdev->vconfig[offset + PCI_PM_PMC];
1455 	__le16 *ctrl = (__le16 *)&vdev->vconfig[offset + PCI_PM_CTRL];
1456 
1457 	/* Clear vconfig PME_Support, PME_Status, and PME_En bits */
1458 	*pmc &= ~cpu_to_le16(PCI_PM_CAP_PME_MASK);
1459 	*ctrl &= ~cpu_to_le16(PCI_PM_CTRL_PME_ENABLE | PCI_PM_CTRL_PME_STATUS);
1460 }
1461 
1462 static int vfio_fill_vconfig_bytes(struct vfio_pci_core_device *vdev,
1463 				   int offset, int size)
1464 {
1465 	struct pci_dev *pdev = vdev->pdev;
1466 	int ret = 0;
1467 
1468 	/*
1469 	 * We try to read physical config space in the largest chunks
1470 	 * we can, assuming that all of the fields support dword access.
1471 	 * pci_save_state() makes this same assumption and seems to do ok.
1472 	 */
1473 	while (size) {
1474 		int filled;
1475 
1476 		if (size >= 4 && !(offset % 4)) {
1477 			__le32 *dwordp = (__le32 *)&vdev->vconfig[offset];
1478 			u32 dword;
1479 
1480 			ret = pci_read_config_dword(pdev, offset, &dword);
1481 			if (ret)
1482 				return ret;
1483 			*dwordp = cpu_to_le32(dword);
1484 			filled = 4;
1485 		} else if (size >= 2 && !(offset % 2)) {
1486 			__le16 *wordp = (__le16 *)&vdev->vconfig[offset];
1487 			u16 word;
1488 
1489 			ret = pci_read_config_word(pdev, offset, &word);
1490 			if (ret)
1491 				return ret;
1492 			*wordp = cpu_to_le16(word);
1493 			filled = 2;
1494 		} else {
1495 			u8 *byte = &vdev->vconfig[offset];
1496 			ret = pci_read_config_byte(pdev, offset, byte);
1497 			if (ret)
1498 				return ret;
1499 			filled = 1;
1500 		}
1501 
1502 		offset += filled;
1503 		size -= filled;
1504 	}
1505 
1506 	return ret;
1507 }
1508 
1509 static int vfio_cap_init(struct vfio_pci_core_device *vdev)
1510 {
1511 	struct pci_dev *pdev = vdev->pdev;
1512 	u8 *map = vdev->pci_config_map;
1513 	u16 status;
1514 	u8 pos, *prev, cap;
1515 	int loops, ret, caps = 0;
1516 
1517 	/* Any capabilities? */
1518 	ret = pci_read_config_word(pdev, PCI_STATUS, &status);
1519 	if (ret)
1520 		return ret;
1521 
1522 	if (!(status & PCI_STATUS_CAP_LIST))
1523 		return 0; /* Done */
1524 
1525 	ret = pci_read_config_byte(pdev, PCI_CAPABILITY_LIST, &pos);
1526 	if (ret)
1527 		return ret;
1528 
1529 	/* Mark the previous position in case we want to skip a capability */
1530 	prev = &vdev->vconfig[PCI_CAPABILITY_LIST];
1531 
1532 	/* We can bound our loop, capabilities are dword aligned */
1533 	loops = (PCI_CFG_SPACE_SIZE - PCI_STD_HEADER_SIZEOF) / PCI_CAP_SIZEOF;
1534 	while (pos && loops--) {
1535 		u8 next;
1536 		int i, len = 0;
1537 
1538 		ret = pci_read_config_byte(pdev, pos, &cap);
1539 		if (ret)
1540 			return ret;
1541 
1542 		ret = pci_read_config_byte(pdev,
1543 					   pos + PCI_CAP_LIST_NEXT, &next);
1544 		if (ret)
1545 			return ret;
1546 
1547 		/*
1548 		 * ID 0 is a NULL capability, conflicting with our fake
1549 		 * PCI_CAP_ID_BASIC.  As it has no content, consider it
1550 		 * hidden for now.
1551 		 */
1552 		if (cap && cap <= PCI_CAP_ID_MAX) {
1553 			len = pci_cap_length[cap];
1554 			if (len == 0xFF) { /* Variable length */
1555 				len = vfio_cap_len(vdev, cap, pos);
1556 				if (len < 0)
1557 					return len;
1558 			}
1559 		}
1560 
1561 		if (!len) {
1562 			pci_info(pdev, "%s: hiding cap %#x@%#x\n", __func__,
1563 				 cap, pos);
1564 			*prev = next;
1565 			pos = next;
1566 			continue;
1567 		}
1568 
1569 		/* Sanity check, do we overlap other capabilities? */
1570 		for (i = 0; i < len; i++) {
1571 			if (likely(map[pos + i] == PCI_CAP_ID_INVALID))
1572 				continue;
1573 
1574 			pci_warn(pdev, "%s: PCI config conflict @%#x, was cap %#x now cap %#x\n",
1575 				 __func__, pos + i, map[pos + i], cap);
1576 		}
1577 
1578 		BUILD_BUG_ON(PCI_CAP_ID_MAX >= PCI_CAP_ID_INVALID_VIRT);
1579 
1580 		memset(map + pos, cap, len);
1581 		ret = vfio_fill_vconfig_bytes(vdev, pos, len);
1582 		if (ret)
1583 			return ret;
1584 
1585 		if (cap == PCI_CAP_ID_PM)
1586 			vfio_update_pm_vconfig_bytes(vdev, pos);
1587 
1588 		prev = &vdev->vconfig[pos + PCI_CAP_LIST_NEXT];
1589 		pos = next;
1590 		caps++;
1591 	}
1592 
1593 	/* If we didn't fill any capabilities, clear the status flag */
1594 	if (!caps) {
1595 		__le16 *vstatus = (__le16 *)&vdev->vconfig[PCI_STATUS];
1596 		*vstatus &= ~cpu_to_le16(PCI_STATUS_CAP_LIST);
1597 	}
1598 
1599 	return 0;
1600 }
1601 
1602 static int vfio_ecap_init(struct vfio_pci_core_device *vdev)
1603 {
1604 	struct pci_dev *pdev = vdev->pdev;
1605 	u8 *map = vdev->pci_config_map;
1606 	u16 epos;
1607 	__le32 *prev = NULL;
1608 	int loops, ret, ecaps = 0;
1609 
1610 	if (!vdev->extended_caps)
1611 		return 0;
1612 
1613 	epos = PCI_CFG_SPACE_SIZE;
1614 
1615 	loops = (pdev->cfg_size - PCI_CFG_SPACE_SIZE) / PCI_CAP_SIZEOF;
1616 
1617 	while (loops-- && epos >= PCI_CFG_SPACE_SIZE) {
1618 		u32 header;
1619 		u16 ecap;
1620 		int i, len = 0;
1621 		bool hidden = false;
1622 
1623 		ret = pci_read_config_dword(pdev, epos, &header);
1624 		if (ret)
1625 			return ret;
1626 
1627 		ecap = PCI_EXT_CAP_ID(header);
1628 
1629 		if (ecap <= PCI_EXT_CAP_ID_MAX) {
1630 			len = pci_ext_cap_length[ecap];
1631 			if (len == 0xFF) {
1632 				len = vfio_ext_cap_len(vdev, ecap, epos);
1633 				if (len < 0)
1634 					return len;
1635 			}
1636 		}
1637 
1638 		if (!len) {
1639 			pci_info(pdev, "%s: hiding ecap %#x@%#x\n",
1640 				 __func__, ecap, epos);
1641 
1642 			/* If not the first in the chain, we can skip over it */
1643 			if (prev) {
1644 				u32 val = epos = PCI_EXT_CAP_NEXT(header);
1645 				*prev &= cpu_to_le32(~(0xffcU << 20));
1646 				*prev |= cpu_to_le32(val << 20);
1647 				continue;
1648 			}
1649 
1650 			/*
1651 			 * Otherwise, fill in a placeholder, the direct
1652 			 * readfn will virtualize this automatically
1653 			 */
1654 			len = PCI_CAP_SIZEOF;
1655 			hidden = true;
1656 		}
1657 
1658 		for (i = 0; i < len; i++) {
1659 			if (likely(map[epos + i] == PCI_CAP_ID_INVALID))
1660 				continue;
1661 
1662 			pci_warn(pdev, "%s: PCI config conflict @%#x, was ecap %#x now ecap %#x\n",
1663 				 __func__, epos + i, map[epos + i], ecap);
1664 		}
1665 
1666 		/*
1667 		 * Even though ecap is 2 bytes, we're currently a long way
1668 		 * from exceeding 1 byte capabilities.  If we ever make it
1669 		 * up to 0xFE we'll need to up this to a two-byte, byte map.
1670 		 */
1671 		BUILD_BUG_ON(PCI_EXT_CAP_ID_MAX >= PCI_CAP_ID_INVALID_VIRT);
1672 
1673 		memset(map + epos, ecap, len);
1674 		ret = vfio_fill_vconfig_bytes(vdev, epos, len);
1675 		if (ret)
1676 			return ret;
1677 
1678 		/*
1679 		 * If we're just using this capability to anchor the list,
1680 		 * hide the real ID.  Only count real ecaps.  XXX PCI spec
1681 		 * indicates to use cap id = 0, version = 0, next = 0 if
1682 		 * ecaps are absent, hope users check all the way to next.
1683 		 */
1684 		if (hidden)
1685 			*(__le32 *)&vdev->vconfig[epos] &=
1686 				cpu_to_le32((0xffcU << 20));
1687 		else
1688 			ecaps++;
1689 
1690 		prev = (__le32 *)&vdev->vconfig[epos];
1691 		epos = PCI_EXT_CAP_NEXT(header);
1692 	}
1693 
1694 	if (!ecaps)
1695 		*(u32 *)&vdev->vconfig[PCI_CFG_SPACE_SIZE] = 0;
1696 
1697 	return 0;
1698 }
1699 
1700 /*
1701  * Nag about hardware bugs, hopefully to have vendors fix them, but at least
1702  * to collect a list of dependencies for the VF INTx pin quirk below.
1703  */
1704 static const struct pci_device_id known_bogus_vf_intx_pin[] = {
1705 	{ PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x270c) },
1706 	{}
1707 };
1708 
1709 /*
1710  * For each device we allocate a pci_config_map that indicates the
1711  * capability occupying each dword and thus the struct perm_bits we
1712  * use for read and write.  We also allocate a virtualized config
1713  * space which tracks reads and writes to bits that we emulate for
1714  * the user.  Initial values filled from device.
1715  *
1716  * Using shared struct perm_bits between all vfio-pci devices saves
1717  * us from allocating cfg_size buffers for virt and write for every
1718  * device.  We could remove vconfig and allocate individual buffers
1719  * for each area requiring emulated bits, but the array of pointers
1720  * would be comparable in size (at least for standard config space).
1721  */
1722 int vfio_config_init(struct vfio_pci_core_device *vdev)
1723 {
1724 	struct pci_dev *pdev = vdev->pdev;
1725 	u8 *map, *vconfig;
1726 	int ret;
1727 
1728 	/*
1729 	 * Config space, caps and ecaps are all dword aligned, so we could
1730 	 * use one byte per dword to record the type.  However, there are
1731 	 * no requiremenst on the length of a capability, so the gap between
1732 	 * capabilities needs byte granularity.
1733 	 */
1734 	map = kmalloc(pdev->cfg_size, GFP_KERNEL);
1735 	if (!map)
1736 		return -ENOMEM;
1737 
1738 	vconfig = kmalloc(pdev->cfg_size, GFP_KERNEL);
1739 	if (!vconfig) {
1740 		kfree(map);
1741 		return -ENOMEM;
1742 	}
1743 
1744 	vdev->pci_config_map = map;
1745 	vdev->vconfig = vconfig;
1746 
1747 	memset(map, PCI_CAP_ID_BASIC, PCI_STD_HEADER_SIZEOF);
1748 	memset(map + PCI_STD_HEADER_SIZEOF, PCI_CAP_ID_INVALID,
1749 	       pdev->cfg_size - PCI_STD_HEADER_SIZEOF);
1750 
1751 	ret = vfio_fill_vconfig_bytes(vdev, 0, PCI_STD_HEADER_SIZEOF);
1752 	if (ret)
1753 		goto out;
1754 
1755 	vdev->bardirty = true;
1756 
1757 	/*
1758 	 * XXX can we just pci_load_saved_state/pci_restore_state?
1759 	 * may need to rebuild vconfig after that
1760 	 */
1761 
1762 	/* For restore after reset */
1763 	vdev->rbar[0] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_0]);
1764 	vdev->rbar[1] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_1]);
1765 	vdev->rbar[2] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_2]);
1766 	vdev->rbar[3] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_3]);
1767 	vdev->rbar[4] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_4]);
1768 	vdev->rbar[5] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_5]);
1769 	vdev->rbar[6] = le32_to_cpu(*(__le32 *)&vconfig[PCI_ROM_ADDRESS]);
1770 
1771 	if (pdev->is_virtfn) {
1772 		*(__le16 *)&vconfig[PCI_VENDOR_ID] = cpu_to_le16(pdev->vendor);
1773 		*(__le16 *)&vconfig[PCI_DEVICE_ID] = cpu_to_le16(pdev->device);
1774 
1775 		/*
1776 		 * Per SR-IOV spec rev 1.1, 3.4.1.18 the interrupt pin register
1777 		 * does not apply to VFs and VFs must implement this register
1778 		 * as read-only with value zero.  Userspace is not readily able
1779 		 * to identify whether a device is a VF and thus that the pin
1780 		 * definition on the device is bogus should it violate this
1781 		 * requirement.  We already virtualize the pin register for
1782 		 * other purposes, so we simply need to replace the bogus value
1783 		 * and consider VFs when we determine INTx IRQ count.
1784 		 */
1785 		if (vconfig[PCI_INTERRUPT_PIN] &&
1786 		    !pci_match_id(known_bogus_vf_intx_pin, pdev))
1787 			pci_warn(pdev,
1788 				 "Hardware bug: VF reports bogus INTx pin %d\n",
1789 				 vconfig[PCI_INTERRUPT_PIN]);
1790 
1791 		vconfig[PCI_INTERRUPT_PIN] = 0; /* Gratuitous for good VFs */
1792 	}
1793 	if (pdev->no_command_memory) {
1794 		/*
1795 		 * VFs and devices that set pdev->no_command_memory do not
1796 		 * implement the memory enable bit of the COMMAND register
1797 		 * therefore we'll not have it set in our initial copy of
1798 		 * config space after pci_enable_device().  For consistency
1799 		 * with PFs, set the virtual enable bit here.
1800 		 */
1801 		*(__le16 *)&vconfig[PCI_COMMAND] |=
1802 					cpu_to_le16(PCI_COMMAND_MEMORY);
1803 	}
1804 
1805 	if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) || vdev->nointx)
1806 		vconfig[PCI_INTERRUPT_PIN] = 0;
1807 
1808 	ret = vfio_cap_init(vdev);
1809 	if (ret)
1810 		goto out;
1811 
1812 	ret = vfio_ecap_init(vdev);
1813 	if (ret)
1814 		goto out;
1815 
1816 	return 0;
1817 
1818 out:
1819 	kfree(map);
1820 	vdev->pci_config_map = NULL;
1821 	kfree(vconfig);
1822 	vdev->vconfig = NULL;
1823 	return pcibios_err_to_errno(ret);
1824 }
1825 
1826 void vfio_config_free(struct vfio_pci_core_device *vdev)
1827 {
1828 	kfree(vdev->vconfig);
1829 	vdev->vconfig = NULL;
1830 	kfree(vdev->pci_config_map);
1831 	vdev->pci_config_map = NULL;
1832 	if (vdev->msi_perm) {
1833 		free_perm_bits(vdev->msi_perm);
1834 		kfree(vdev->msi_perm);
1835 		vdev->msi_perm = NULL;
1836 	}
1837 }
1838 
1839 /*
1840  * Find the remaining number of bytes in a dword that match the given
1841  * position.  Stop at either the end of the capability or the dword boundary.
1842  */
1843 static size_t vfio_pci_cap_remaining_dword(struct vfio_pci_core_device *vdev,
1844 					   loff_t pos)
1845 {
1846 	u8 cap = vdev->pci_config_map[pos];
1847 	size_t i;
1848 
1849 	for (i = 1; (pos + i) % 4 && vdev->pci_config_map[pos + i] == cap; i++)
1850 		/* nop */;
1851 
1852 	return i;
1853 }
1854 
1855 static ssize_t vfio_config_do_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1856 				 size_t count, loff_t *ppos, bool iswrite)
1857 {
1858 	struct pci_dev *pdev = vdev->pdev;
1859 	struct perm_bits *perm;
1860 	__le32 val = 0;
1861 	int cap_start = 0, offset;
1862 	u8 cap_id;
1863 	ssize_t ret;
1864 
1865 	if (*ppos < 0 || *ppos >= pdev->cfg_size ||
1866 	    *ppos + count > pdev->cfg_size)
1867 		return -EFAULT;
1868 
1869 	/*
1870 	 * Chop accesses into aligned chunks containing no more than a
1871 	 * single capability.  Caller increments to the next chunk.
1872 	 */
1873 	count = min(count, vfio_pci_cap_remaining_dword(vdev, *ppos));
1874 	if (count >= 4 && !(*ppos % 4))
1875 		count = 4;
1876 	else if (count >= 2 && !(*ppos % 2))
1877 		count = 2;
1878 	else
1879 		count = 1;
1880 
1881 	ret = count;
1882 
1883 	cap_id = vdev->pci_config_map[*ppos];
1884 
1885 	if (cap_id == PCI_CAP_ID_INVALID) {
1886 		perm = &unassigned_perms;
1887 		cap_start = *ppos;
1888 	} else if (cap_id == PCI_CAP_ID_INVALID_VIRT) {
1889 		perm = &virt_perms;
1890 		cap_start = *ppos;
1891 	} else {
1892 		if (*ppos >= PCI_CFG_SPACE_SIZE) {
1893 			WARN_ON(cap_id > PCI_EXT_CAP_ID_MAX);
1894 
1895 			perm = &ecap_perms[cap_id];
1896 			cap_start = vfio_find_cap_start(vdev, *ppos);
1897 		} else {
1898 			WARN_ON(cap_id > PCI_CAP_ID_MAX);
1899 
1900 			perm = &cap_perms[cap_id];
1901 
1902 			if (cap_id == PCI_CAP_ID_MSI)
1903 				perm = vdev->msi_perm;
1904 
1905 			if (cap_id > PCI_CAP_ID_BASIC)
1906 				cap_start = vfio_find_cap_start(vdev, *ppos);
1907 		}
1908 	}
1909 
1910 	WARN_ON(!cap_start && cap_id != PCI_CAP_ID_BASIC);
1911 	WARN_ON(cap_start > *ppos);
1912 
1913 	offset = *ppos - cap_start;
1914 
1915 	if (iswrite) {
1916 		if (!perm->writefn)
1917 			return ret;
1918 
1919 		if (copy_from_user(&val, buf, count))
1920 			return -EFAULT;
1921 
1922 		ret = perm->writefn(vdev, *ppos, count, perm, offset, val);
1923 	} else {
1924 		if (perm->readfn) {
1925 			ret = perm->readfn(vdev, *ppos, count,
1926 					   perm, offset, &val);
1927 			if (ret < 0)
1928 				return ret;
1929 		}
1930 
1931 		if (copy_to_user(buf, &val, count))
1932 			return -EFAULT;
1933 	}
1934 
1935 	return ret;
1936 }
1937 
1938 ssize_t vfio_pci_config_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1939 			   size_t count, loff_t *ppos, bool iswrite)
1940 {
1941 	size_t done = 0;
1942 	int ret = 0;
1943 	loff_t pos = *ppos;
1944 
1945 	pos &= VFIO_PCI_OFFSET_MASK;
1946 
1947 	while (count) {
1948 		ret = vfio_config_do_rw(vdev, buf, count, &pos, iswrite);
1949 		if (ret < 0)
1950 			return ret;
1951 
1952 		count -= ret;
1953 		done += ret;
1954 		buf += ret;
1955 		pos += ret;
1956 	}
1957 
1958 	*ppos += done;
1959 
1960 	return done;
1961 }
1962