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 "vfio_pci_private.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 thru 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_device *vdev, int pos, int count,
112 			  struct perm_bits *perm, int offset, __le32 *val);
113 	int	(*writefn)(struct vfio_pci_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_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_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_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_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_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_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_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 /*
399  * Restore the *real* BARs after we detect a FLR or backdoor reset.
400  * (backdoor = some device specific technique that we didn't catch)
401  */
402 static void vfio_bar_restore(struct vfio_pci_device *vdev)
403 {
404 	struct pci_dev *pdev = vdev->pdev;
405 	u32 *rbar = vdev->rbar;
406 	u16 cmd;
407 	int i;
408 
409 	if (pdev->is_virtfn)
410 		return;
411 
412 	pci_info(pdev, "%s: reset recovery - restoring BARs\n", __func__);
413 
414 	for (i = PCI_BASE_ADDRESS_0; i <= PCI_BASE_ADDRESS_5; i += 4, rbar++)
415 		pci_user_write_config_dword(pdev, i, *rbar);
416 
417 	pci_user_write_config_dword(pdev, PCI_ROM_ADDRESS, *rbar);
418 
419 	if (vdev->nointx) {
420 		pci_user_read_config_word(pdev, PCI_COMMAND, &cmd);
421 		cmd |= PCI_COMMAND_INTX_DISABLE;
422 		pci_user_write_config_word(pdev, PCI_COMMAND, cmd);
423 	}
424 }
425 
426 static __le32 vfio_generate_bar_flags(struct pci_dev *pdev, int bar)
427 {
428 	unsigned long flags = pci_resource_flags(pdev, bar);
429 	u32 val;
430 
431 	if (flags & IORESOURCE_IO)
432 		return cpu_to_le32(PCI_BASE_ADDRESS_SPACE_IO);
433 
434 	val = PCI_BASE_ADDRESS_SPACE_MEMORY;
435 
436 	if (flags & IORESOURCE_PREFETCH)
437 		val |= PCI_BASE_ADDRESS_MEM_PREFETCH;
438 
439 	if (flags & IORESOURCE_MEM_64)
440 		val |= PCI_BASE_ADDRESS_MEM_TYPE_64;
441 
442 	return cpu_to_le32(val);
443 }
444 
445 /*
446  * Pretend we're hardware and tweak the values of the *virtual* PCI BARs
447  * to reflect the hardware capabilities.  This implements BAR sizing.
448  */
449 static void vfio_bar_fixup(struct vfio_pci_device *vdev)
450 {
451 	struct pci_dev *pdev = vdev->pdev;
452 	int i;
453 	__le32 *vbar;
454 	u64 mask;
455 
456 	vbar = (__le32 *)&vdev->vconfig[PCI_BASE_ADDRESS_0];
457 
458 	for (i = 0; i < PCI_STD_NUM_BARS; i++, vbar++) {
459 		int bar = i + PCI_STD_RESOURCES;
460 
461 		if (!pci_resource_start(pdev, bar)) {
462 			*vbar = 0; /* Unmapped by host = unimplemented to user */
463 			continue;
464 		}
465 
466 		mask = ~(pci_resource_len(pdev, bar) - 1);
467 
468 		*vbar &= cpu_to_le32((u32)mask);
469 		*vbar |= vfio_generate_bar_flags(pdev, bar);
470 
471 		if (*vbar & cpu_to_le32(PCI_BASE_ADDRESS_MEM_TYPE_64)) {
472 			vbar++;
473 			*vbar &= cpu_to_le32((u32)(mask >> 32));
474 			i++;
475 		}
476 	}
477 
478 	vbar = (__le32 *)&vdev->vconfig[PCI_ROM_ADDRESS];
479 
480 	/*
481 	 * NB. REGION_INFO will have reported zero size if we weren't able
482 	 * to read the ROM, but we still return the actual BAR size here if
483 	 * it exists (or the shadow ROM space).
484 	 */
485 	if (pci_resource_start(pdev, PCI_ROM_RESOURCE)) {
486 		mask = ~(pci_resource_len(pdev, PCI_ROM_RESOURCE) - 1);
487 		mask |= PCI_ROM_ADDRESS_ENABLE;
488 		*vbar &= cpu_to_le32((u32)mask);
489 	} else if (pdev->resource[PCI_ROM_RESOURCE].flags &
490 					IORESOURCE_ROM_SHADOW) {
491 		mask = ~(0x20000 - 1);
492 		mask |= PCI_ROM_ADDRESS_ENABLE;
493 		*vbar &= cpu_to_le32((u32)mask);
494 	} else
495 		*vbar = 0;
496 
497 	vdev->bardirty = false;
498 }
499 
500 static int vfio_basic_config_read(struct vfio_pci_device *vdev, int pos,
501 				  int count, struct perm_bits *perm,
502 				  int offset, __le32 *val)
503 {
504 	if (is_bar(offset)) /* pos == offset for basic config */
505 		vfio_bar_fixup(vdev);
506 
507 	count = vfio_default_config_read(vdev, pos, count, perm, offset, val);
508 
509 	/* Mask in virtual memory enable for SR-IOV devices */
510 	if (offset == PCI_COMMAND && vdev->pdev->is_virtfn) {
511 		u16 cmd = le16_to_cpu(*(__le16 *)&vdev->vconfig[PCI_COMMAND]);
512 		u32 tmp_val = le32_to_cpu(*val);
513 
514 		tmp_val |= cmd & PCI_COMMAND_MEMORY;
515 		*val = cpu_to_le32(tmp_val);
516 	}
517 
518 	return count;
519 }
520 
521 /* Test whether BARs match the value we think they should contain */
522 static bool vfio_need_bar_restore(struct vfio_pci_device *vdev)
523 {
524 	int i = 0, pos = PCI_BASE_ADDRESS_0, ret;
525 	u32 bar;
526 
527 	for (; pos <= PCI_BASE_ADDRESS_5; i++, pos += 4) {
528 		if (vdev->rbar[i]) {
529 			ret = pci_user_read_config_dword(vdev->pdev, pos, &bar);
530 			if (ret || vdev->rbar[i] != bar)
531 				return true;
532 		}
533 	}
534 
535 	return false;
536 }
537 
538 static int vfio_basic_config_write(struct vfio_pci_device *vdev, int pos,
539 				   int count, struct perm_bits *perm,
540 				   int offset, __le32 val)
541 {
542 	struct pci_dev *pdev = vdev->pdev;
543 	__le16 *virt_cmd;
544 	u16 new_cmd = 0;
545 	int ret;
546 
547 	virt_cmd = (__le16 *)&vdev->vconfig[PCI_COMMAND];
548 
549 	if (offset == PCI_COMMAND) {
550 		bool phys_mem, virt_mem, new_mem, phys_io, virt_io, new_io;
551 		u16 phys_cmd;
552 
553 		ret = pci_user_read_config_word(pdev, PCI_COMMAND, &phys_cmd);
554 		if (ret)
555 			return ret;
556 
557 		new_cmd = le32_to_cpu(val);
558 
559 		phys_mem = !!(phys_cmd & PCI_COMMAND_MEMORY);
560 		virt_mem = !!(le16_to_cpu(*virt_cmd) & PCI_COMMAND_MEMORY);
561 		new_mem = !!(new_cmd & PCI_COMMAND_MEMORY);
562 
563 		phys_io = !!(phys_cmd & PCI_COMMAND_IO);
564 		virt_io = !!(le16_to_cpu(*virt_cmd) & PCI_COMMAND_IO);
565 		new_io = !!(new_cmd & PCI_COMMAND_IO);
566 
567 		/*
568 		 * If the user is writing mem/io enable (new_mem/io) and we
569 		 * think it's already enabled (virt_mem/io), but the hardware
570 		 * shows it disabled (phys_mem/io, then the device has
571 		 * undergone some kind of backdoor reset and needs to be
572 		 * restored before we allow it to enable the bars.
573 		 * SR-IOV devices will trigger this, but we catch them later
574 		 */
575 		if ((new_mem && virt_mem && !phys_mem) ||
576 		    (new_io && virt_io && !phys_io) ||
577 		    vfio_need_bar_restore(vdev))
578 			vfio_bar_restore(vdev);
579 	}
580 
581 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
582 	if (count < 0)
583 		return count;
584 
585 	/*
586 	 * Save current memory/io enable bits in vconfig to allow for
587 	 * the test above next time.
588 	 */
589 	if (offset == PCI_COMMAND) {
590 		u16 mask = PCI_COMMAND_MEMORY | PCI_COMMAND_IO;
591 
592 		*virt_cmd &= cpu_to_le16(~mask);
593 		*virt_cmd |= cpu_to_le16(new_cmd & mask);
594 	}
595 
596 	/* Emulate INTx disable */
597 	if (offset >= PCI_COMMAND && offset <= PCI_COMMAND + 1) {
598 		bool virt_intx_disable;
599 
600 		virt_intx_disable = !!(le16_to_cpu(*virt_cmd) &
601 				       PCI_COMMAND_INTX_DISABLE);
602 
603 		if (virt_intx_disable && !vdev->virq_disabled) {
604 			vdev->virq_disabled = true;
605 			vfio_pci_intx_mask(vdev);
606 		} else if (!virt_intx_disable && vdev->virq_disabled) {
607 			vdev->virq_disabled = false;
608 			vfio_pci_intx_unmask(vdev);
609 		}
610 	}
611 
612 	if (is_bar(offset))
613 		vdev->bardirty = true;
614 
615 	return count;
616 }
617 
618 /* Permissions for the Basic PCI Header */
619 static int __init init_pci_cap_basic_perm(struct perm_bits *perm)
620 {
621 	if (alloc_perm_bits(perm, PCI_STD_HEADER_SIZEOF))
622 		return -ENOMEM;
623 
624 	perm->readfn = vfio_basic_config_read;
625 	perm->writefn = vfio_basic_config_write;
626 
627 	/* Virtualized for SR-IOV functions, which just have FFFF */
628 	p_setw(perm, PCI_VENDOR_ID, (u16)ALL_VIRT, NO_WRITE);
629 	p_setw(perm, PCI_DEVICE_ID, (u16)ALL_VIRT, NO_WRITE);
630 
631 	/*
632 	 * Virtualize INTx disable, we use it internally for interrupt
633 	 * control and can emulate it for non-PCI 2.3 devices.
634 	 */
635 	p_setw(perm, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE, (u16)ALL_WRITE);
636 
637 	/* Virtualize capability list, we might want to skip/disable */
638 	p_setw(perm, PCI_STATUS, PCI_STATUS_CAP_LIST, NO_WRITE);
639 
640 	/* No harm to write */
641 	p_setb(perm, PCI_CACHE_LINE_SIZE, NO_VIRT, (u8)ALL_WRITE);
642 	p_setb(perm, PCI_LATENCY_TIMER, NO_VIRT, (u8)ALL_WRITE);
643 	p_setb(perm, PCI_BIST, NO_VIRT, (u8)ALL_WRITE);
644 
645 	/* Virtualize all bars, can't touch the real ones */
646 	p_setd(perm, PCI_BASE_ADDRESS_0, ALL_VIRT, ALL_WRITE);
647 	p_setd(perm, PCI_BASE_ADDRESS_1, ALL_VIRT, ALL_WRITE);
648 	p_setd(perm, PCI_BASE_ADDRESS_2, ALL_VIRT, ALL_WRITE);
649 	p_setd(perm, PCI_BASE_ADDRESS_3, ALL_VIRT, ALL_WRITE);
650 	p_setd(perm, PCI_BASE_ADDRESS_4, ALL_VIRT, ALL_WRITE);
651 	p_setd(perm, PCI_BASE_ADDRESS_5, ALL_VIRT, ALL_WRITE);
652 	p_setd(perm, PCI_ROM_ADDRESS, ALL_VIRT, ALL_WRITE);
653 
654 	/* Allow us to adjust capability chain */
655 	p_setb(perm, PCI_CAPABILITY_LIST, (u8)ALL_VIRT, NO_WRITE);
656 
657 	/* Sometimes used by sw, just virtualize */
658 	p_setb(perm, PCI_INTERRUPT_LINE, (u8)ALL_VIRT, (u8)ALL_WRITE);
659 
660 	/* Virtualize interrupt pin to allow hiding INTx */
661 	p_setb(perm, PCI_INTERRUPT_PIN, (u8)ALL_VIRT, (u8)NO_WRITE);
662 
663 	return 0;
664 }
665 
666 static int vfio_pm_config_write(struct vfio_pci_device *vdev, int pos,
667 				int count, struct perm_bits *perm,
668 				int offset, __le32 val)
669 {
670 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
671 	if (count < 0)
672 		return count;
673 
674 	if (offset == PCI_PM_CTRL) {
675 		pci_power_t state;
676 
677 		switch (le32_to_cpu(val) & PCI_PM_CTRL_STATE_MASK) {
678 		case 0:
679 			state = PCI_D0;
680 			break;
681 		case 1:
682 			state = PCI_D1;
683 			break;
684 		case 2:
685 			state = PCI_D2;
686 			break;
687 		case 3:
688 			state = PCI_D3hot;
689 			break;
690 		}
691 
692 		vfio_pci_set_power_state(vdev, state);
693 	}
694 
695 	return count;
696 }
697 
698 /* Permissions for the Power Management capability */
699 static int __init init_pci_cap_pm_perm(struct perm_bits *perm)
700 {
701 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_PM]))
702 		return -ENOMEM;
703 
704 	perm->writefn = vfio_pm_config_write;
705 
706 	/*
707 	 * We always virtualize the next field so we can remove
708 	 * capabilities from the chain if we want to.
709 	 */
710 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
711 
712 	/*
713 	 * Power management is defined *per function*, so we can let
714 	 * the user change power state, but we trap and initiate the
715 	 * change ourselves, so the state bits are read-only.
716 	 */
717 	p_setd(perm, PCI_PM_CTRL, NO_VIRT, ~PCI_PM_CTRL_STATE_MASK);
718 	return 0;
719 }
720 
721 static int vfio_vpd_config_write(struct vfio_pci_device *vdev, int pos,
722 				 int count, struct perm_bits *perm,
723 				 int offset, __le32 val)
724 {
725 	struct pci_dev *pdev = vdev->pdev;
726 	__le16 *paddr = (__le16 *)(vdev->vconfig + pos - offset + PCI_VPD_ADDR);
727 	__le32 *pdata = (__le32 *)(vdev->vconfig + pos - offset + PCI_VPD_DATA);
728 	u16 addr;
729 	u32 data;
730 
731 	/*
732 	 * Write through to emulation.  If the write includes the upper byte
733 	 * of PCI_VPD_ADDR, then the PCI_VPD_ADDR_F bit is written and we
734 	 * have work to do.
735 	 */
736 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
737 	if (count < 0 || offset > PCI_VPD_ADDR + 1 ||
738 	    offset + count <= PCI_VPD_ADDR + 1)
739 		return count;
740 
741 	addr = le16_to_cpu(*paddr);
742 
743 	if (addr & PCI_VPD_ADDR_F) {
744 		data = le32_to_cpu(*pdata);
745 		if (pci_write_vpd(pdev, addr & ~PCI_VPD_ADDR_F, 4, &data) != 4)
746 			return count;
747 	} else {
748 		data = 0;
749 		if (pci_read_vpd(pdev, addr, 4, &data) < 0)
750 			return count;
751 		*pdata = cpu_to_le32(data);
752 	}
753 
754 	/*
755 	 * Toggle PCI_VPD_ADDR_F in the emulated PCI_VPD_ADDR register to
756 	 * signal completion.  If an error occurs above, we assume that not
757 	 * toggling this bit will induce a driver timeout.
758 	 */
759 	addr ^= PCI_VPD_ADDR_F;
760 	*paddr = cpu_to_le16(addr);
761 
762 	return count;
763 }
764 
765 /* Permissions for Vital Product Data capability */
766 static int __init init_pci_cap_vpd_perm(struct perm_bits *perm)
767 {
768 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_VPD]))
769 		return -ENOMEM;
770 
771 	perm->writefn = vfio_vpd_config_write;
772 
773 	/*
774 	 * We always virtualize the next field so we can remove
775 	 * capabilities from the chain if we want to.
776 	 */
777 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
778 
779 	/*
780 	 * Both the address and data registers are virtualized to
781 	 * enable access through the pci_vpd_read/write functions
782 	 */
783 	p_setw(perm, PCI_VPD_ADDR, (u16)ALL_VIRT, (u16)ALL_WRITE);
784 	p_setd(perm, PCI_VPD_DATA, ALL_VIRT, ALL_WRITE);
785 
786 	return 0;
787 }
788 
789 /* Permissions for PCI-X capability */
790 static int __init init_pci_cap_pcix_perm(struct perm_bits *perm)
791 {
792 	/* Alloc 24, but only 8 are used in v0 */
793 	if (alloc_perm_bits(perm, PCI_CAP_PCIX_SIZEOF_V2))
794 		return -ENOMEM;
795 
796 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
797 
798 	p_setw(perm, PCI_X_CMD, NO_VIRT, (u16)ALL_WRITE);
799 	p_setd(perm, PCI_X_ECC_CSR, NO_VIRT, ALL_WRITE);
800 	return 0;
801 }
802 
803 static int vfio_exp_config_write(struct vfio_pci_device *vdev, int pos,
804 				 int count, struct perm_bits *perm,
805 				 int offset, __le32 val)
806 {
807 	__le16 *ctrl = (__le16 *)(vdev->vconfig + pos -
808 				  offset + PCI_EXP_DEVCTL);
809 	int readrq = le16_to_cpu(*ctrl) & PCI_EXP_DEVCTL_READRQ;
810 
811 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
812 	if (count < 0)
813 		return count;
814 
815 	/*
816 	 * The FLR bit is virtualized, if set and the device supports PCIe
817 	 * FLR, issue a reset_function.  Regardless, clear the bit, the spec
818 	 * requires it to be always read as zero.  NB, reset_function might
819 	 * not use a PCIe FLR, we don't have that level of granularity.
820 	 */
821 	if (*ctrl & cpu_to_le16(PCI_EXP_DEVCTL_BCR_FLR)) {
822 		u32 cap;
823 		int ret;
824 
825 		*ctrl &= ~cpu_to_le16(PCI_EXP_DEVCTL_BCR_FLR);
826 
827 		ret = pci_user_read_config_dword(vdev->pdev,
828 						 pos - offset + PCI_EXP_DEVCAP,
829 						 &cap);
830 
831 		if (!ret && (cap & PCI_EXP_DEVCAP_FLR))
832 			pci_try_reset_function(vdev->pdev);
833 	}
834 
835 	/*
836 	 * MPS is virtualized to the user, writes do not change the physical
837 	 * register since determining a proper MPS value requires a system wide
838 	 * device view.  The MRRS is largely independent of MPS, but since the
839 	 * user does not have that system-wide view, they might set a safe, but
840 	 * inefficiently low value.  Here we allow writes through to hardware,
841 	 * but we set the floor to the physical device MPS setting, so that
842 	 * we can at least use full TLPs, as defined by the MPS value.
843 	 *
844 	 * NB, if any devices actually depend on an artificially low MRRS
845 	 * setting, this will need to be revisited, perhaps with a quirk
846 	 * though pcie_set_readrq().
847 	 */
848 	if (readrq != (le16_to_cpu(*ctrl) & PCI_EXP_DEVCTL_READRQ)) {
849 		readrq = 128 <<
850 			((le16_to_cpu(*ctrl) & PCI_EXP_DEVCTL_READRQ) >> 12);
851 		readrq = max(readrq, pcie_get_mps(vdev->pdev));
852 
853 		pcie_set_readrq(vdev->pdev, readrq);
854 	}
855 
856 	return count;
857 }
858 
859 /* Permissions for PCI Express capability */
860 static int __init init_pci_cap_exp_perm(struct perm_bits *perm)
861 {
862 	/* Alloc largest of possible sizes */
863 	if (alloc_perm_bits(perm, PCI_CAP_EXP_ENDPOINT_SIZEOF_V2))
864 		return -ENOMEM;
865 
866 	perm->writefn = vfio_exp_config_write;
867 
868 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
869 
870 	/*
871 	 * Allow writes to device control fields, except devctl_phantom,
872 	 * which could confuse IOMMU, MPS, which can break communication
873 	 * with other physical devices, and the ARI bit in devctl2, which
874 	 * is set at probe time.  FLR and MRRS get virtualized via our
875 	 * writefn.
876 	 */
877 	p_setw(perm, PCI_EXP_DEVCTL,
878 	       PCI_EXP_DEVCTL_BCR_FLR | PCI_EXP_DEVCTL_PAYLOAD |
879 	       PCI_EXP_DEVCTL_READRQ, ~PCI_EXP_DEVCTL_PHANTOM);
880 	p_setw(perm, PCI_EXP_DEVCTL2, NO_VIRT, ~PCI_EXP_DEVCTL2_ARI);
881 	return 0;
882 }
883 
884 static int vfio_af_config_write(struct vfio_pci_device *vdev, int pos,
885 				int count, struct perm_bits *perm,
886 				int offset, __le32 val)
887 {
888 	u8 *ctrl = vdev->vconfig + pos - offset + PCI_AF_CTRL;
889 
890 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
891 	if (count < 0)
892 		return count;
893 
894 	/*
895 	 * The FLR bit is virtualized, if set and the device supports AF
896 	 * FLR, issue a reset_function.  Regardless, clear the bit, the spec
897 	 * requires it to be always read as zero.  NB, reset_function might
898 	 * not use an AF FLR, we don't have that level of granularity.
899 	 */
900 	if (*ctrl & PCI_AF_CTRL_FLR) {
901 		u8 cap;
902 		int ret;
903 
904 		*ctrl &= ~PCI_AF_CTRL_FLR;
905 
906 		ret = pci_user_read_config_byte(vdev->pdev,
907 						pos - offset + PCI_AF_CAP,
908 						&cap);
909 
910 		if (!ret && (cap & PCI_AF_CAP_FLR) && (cap & PCI_AF_CAP_TP))
911 			pci_try_reset_function(vdev->pdev);
912 	}
913 
914 	return count;
915 }
916 
917 /* Permissions for Advanced Function capability */
918 static int __init init_pci_cap_af_perm(struct perm_bits *perm)
919 {
920 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_AF]))
921 		return -ENOMEM;
922 
923 	perm->writefn = vfio_af_config_write;
924 
925 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
926 	p_setb(perm, PCI_AF_CTRL, PCI_AF_CTRL_FLR, PCI_AF_CTRL_FLR);
927 	return 0;
928 }
929 
930 /* Permissions for Advanced Error Reporting extended capability */
931 static int __init init_pci_ext_cap_err_perm(struct perm_bits *perm)
932 {
933 	u32 mask;
934 
935 	if (alloc_perm_bits(perm, pci_ext_cap_length[PCI_EXT_CAP_ID_ERR]))
936 		return -ENOMEM;
937 
938 	/*
939 	 * Virtualize the first dword of all express capabilities
940 	 * because it includes the next pointer.  This lets us later
941 	 * remove capabilities from the chain if we need to.
942 	 */
943 	p_setd(perm, 0, ALL_VIRT, NO_WRITE);
944 
945 	/* Writable bits mask */
946 	mask =	PCI_ERR_UNC_UND |		/* Undefined */
947 		PCI_ERR_UNC_DLP |		/* Data Link Protocol */
948 		PCI_ERR_UNC_SURPDN |		/* Surprise Down */
949 		PCI_ERR_UNC_POISON_TLP |	/* Poisoned TLP */
950 		PCI_ERR_UNC_FCP |		/* Flow Control Protocol */
951 		PCI_ERR_UNC_COMP_TIME |		/* Completion Timeout */
952 		PCI_ERR_UNC_COMP_ABORT |	/* Completer Abort */
953 		PCI_ERR_UNC_UNX_COMP |		/* Unexpected Completion */
954 		PCI_ERR_UNC_RX_OVER |		/* Receiver Overflow */
955 		PCI_ERR_UNC_MALF_TLP |		/* Malformed TLP */
956 		PCI_ERR_UNC_ECRC |		/* ECRC Error Status */
957 		PCI_ERR_UNC_UNSUP |		/* Unsupported Request */
958 		PCI_ERR_UNC_ACSV |		/* ACS Violation */
959 		PCI_ERR_UNC_INTN |		/* internal error */
960 		PCI_ERR_UNC_MCBTLP |		/* MC blocked TLP */
961 		PCI_ERR_UNC_ATOMEG |		/* Atomic egress blocked */
962 		PCI_ERR_UNC_TLPPRE;		/* TLP prefix blocked */
963 	p_setd(perm, PCI_ERR_UNCOR_STATUS, NO_VIRT, mask);
964 	p_setd(perm, PCI_ERR_UNCOR_MASK, NO_VIRT, mask);
965 	p_setd(perm, PCI_ERR_UNCOR_SEVER, NO_VIRT, mask);
966 
967 	mask =	PCI_ERR_COR_RCVR |		/* Receiver Error Status */
968 		PCI_ERR_COR_BAD_TLP |		/* Bad TLP Status */
969 		PCI_ERR_COR_BAD_DLLP |		/* Bad DLLP Status */
970 		PCI_ERR_COR_REP_ROLL |		/* REPLAY_NUM Rollover */
971 		PCI_ERR_COR_REP_TIMER |		/* Replay Timer Timeout */
972 		PCI_ERR_COR_ADV_NFAT |		/* Advisory Non-Fatal */
973 		PCI_ERR_COR_INTERNAL |		/* Corrected Internal */
974 		PCI_ERR_COR_LOG_OVER;		/* Header Log Overflow */
975 	p_setd(perm, PCI_ERR_COR_STATUS, NO_VIRT, mask);
976 	p_setd(perm, PCI_ERR_COR_MASK, NO_VIRT, mask);
977 
978 	mask =	PCI_ERR_CAP_ECRC_GENE |		/* ECRC Generation Enable */
979 		PCI_ERR_CAP_ECRC_CHKE;		/* ECRC Check Enable */
980 	p_setd(perm, PCI_ERR_CAP, NO_VIRT, mask);
981 	return 0;
982 }
983 
984 /* Permissions for Power Budgeting extended capability */
985 static int __init init_pci_ext_cap_pwr_perm(struct perm_bits *perm)
986 {
987 	if (alloc_perm_bits(perm, pci_ext_cap_length[PCI_EXT_CAP_ID_PWR]))
988 		return -ENOMEM;
989 
990 	p_setd(perm, 0, ALL_VIRT, NO_WRITE);
991 
992 	/* Writing the data selector is OK, the info is still read-only */
993 	p_setb(perm, PCI_PWR_DATA, NO_VIRT, (u8)ALL_WRITE);
994 	return 0;
995 }
996 
997 /*
998  * Initialize the shared permission tables
999  */
1000 void vfio_pci_uninit_perm_bits(void)
1001 {
1002 	free_perm_bits(&cap_perms[PCI_CAP_ID_BASIC]);
1003 
1004 	free_perm_bits(&cap_perms[PCI_CAP_ID_PM]);
1005 	free_perm_bits(&cap_perms[PCI_CAP_ID_VPD]);
1006 	free_perm_bits(&cap_perms[PCI_CAP_ID_PCIX]);
1007 	free_perm_bits(&cap_perms[PCI_CAP_ID_EXP]);
1008 	free_perm_bits(&cap_perms[PCI_CAP_ID_AF]);
1009 
1010 	free_perm_bits(&ecap_perms[PCI_EXT_CAP_ID_ERR]);
1011 	free_perm_bits(&ecap_perms[PCI_EXT_CAP_ID_PWR]);
1012 }
1013 
1014 int __init vfio_pci_init_perm_bits(void)
1015 {
1016 	int ret;
1017 
1018 	/* Basic config space */
1019 	ret = init_pci_cap_basic_perm(&cap_perms[PCI_CAP_ID_BASIC]);
1020 
1021 	/* Capabilities */
1022 	ret |= init_pci_cap_pm_perm(&cap_perms[PCI_CAP_ID_PM]);
1023 	ret |= init_pci_cap_vpd_perm(&cap_perms[PCI_CAP_ID_VPD]);
1024 	ret |= init_pci_cap_pcix_perm(&cap_perms[PCI_CAP_ID_PCIX]);
1025 	cap_perms[PCI_CAP_ID_VNDR].writefn = vfio_raw_config_write;
1026 	ret |= init_pci_cap_exp_perm(&cap_perms[PCI_CAP_ID_EXP]);
1027 	ret |= init_pci_cap_af_perm(&cap_perms[PCI_CAP_ID_AF]);
1028 
1029 	/* Extended capabilities */
1030 	ret |= init_pci_ext_cap_err_perm(&ecap_perms[PCI_EXT_CAP_ID_ERR]);
1031 	ret |= init_pci_ext_cap_pwr_perm(&ecap_perms[PCI_EXT_CAP_ID_PWR]);
1032 	ecap_perms[PCI_EXT_CAP_ID_VNDR].writefn = vfio_raw_config_write;
1033 
1034 	if (ret)
1035 		vfio_pci_uninit_perm_bits();
1036 
1037 	return ret;
1038 }
1039 
1040 static int vfio_find_cap_start(struct vfio_pci_device *vdev, int pos)
1041 {
1042 	u8 cap;
1043 	int base = (pos >= PCI_CFG_SPACE_SIZE) ? PCI_CFG_SPACE_SIZE :
1044 						 PCI_STD_HEADER_SIZEOF;
1045 	cap = vdev->pci_config_map[pos];
1046 
1047 	if (cap == PCI_CAP_ID_BASIC)
1048 		return 0;
1049 
1050 	/* XXX Can we have to abutting capabilities of the same type? */
1051 	while (pos - 1 >= base && vdev->pci_config_map[pos - 1] == cap)
1052 		pos--;
1053 
1054 	return pos;
1055 }
1056 
1057 static int vfio_msi_config_read(struct vfio_pci_device *vdev, int pos,
1058 				int count, struct perm_bits *perm,
1059 				int offset, __le32 *val)
1060 {
1061 	/* Update max available queue size from msi_qmax */
1062 	if (offset <= PCI_MSI_FLAGS && offset + count >= PCI_MSI_FLAGS) {
1063 		__le16 *flags;
1064 		int start;
1065 
1066 		start = vfio_find_cap_start(vdev, pos);
1067 
1068 		flags = (__le16 *)&vdev->vconfig[start];
1069 
1070 		*flags &= cpu_to_le16(~PCI_MSI_FLAGS_QMASK);
1071 		*flags |= cpu_to_le16(vdev->msi_qmax << 1);
1072 	}
1073 
1074 	return vfio_default_config_read(vdev, pos, count, perm, offset, val);
1075 }
1076 
1077 static int vfio_msi_config_write(struct vfio_pci_device *vdev, int pos,
1078 				 int count, struct perm_bits *perm,
1079 				 int offset, __le32 val)
1080 {
1081 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
1082 	if (count < 0)
1083 		return count;
1084 
1085 	/* Fixup and write configured queue size and enable to hardware */
1086 	if (offset <= PCI_MSI_FLAGS && offset + count >= PCI_MSI_FLAGS) {
1087 		__le16 *pflags;
1088 		u16 flags;
1089 		int start, ret;
1090 
1091 		start = vfio_find_cap_start(vdev, pos);
1092 
1093 		pflags = (__le16 *)&vdev->vconfig[start + PCI_MSI_FLAGS];
1094 
1095 		flags = le16_to_cpu(*pflags);
1096 
1097 		/* MSI is enabled via ioctl */
1098 		if  (!is_msi(vdev))
1099 			flags &= ~PCI_MSI_FLAGS_ENABLE;
1100 
1101 		/* Check queue size */
1102 		if ((flags & PCI_MSI_FLAGS_QSIZE) >> 4 > vdev->msi_qmax) {
1103 			flags &= ~PCI_MSI_FLAGS_QSIZE;
1104 			flags |= vdev->msi_qmax << 4;
1105 		}
1106 
1107 		/* Write back to virt and to hardware */
1108 		*pflags = cpu_to_le16(flags);
1109 		ret = pci_user_write_config_word(vdev->pdev,
1110 						 start + PCI_MSI_FLAGS,
1111 						 flags);
1112 		if (ret)
1113 			return ret;
1114 	}
1115 
1116 	return count;
1117 }
1118 
1119 /*
1120  * MSI determination is per-device, so this routine gets used beyond
1121  * initialization time. Don't add __init
1122  */
1123 static int init_pci_cap_msi_perm(struct perm_bits *perm, int len, u16 flags)
1124 {
1125 	if (alloc_perm_bits(perm, len))
1126 		return -ENOMEM;
1127 
1128 	perm->readfn = vfio_msi_config_read;
1129 	perm->writefn = vfio_msi_config_write;
1130 
1131 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
1132 
1133 	/*
1134 	 * The upper byte of the control register is reserved,
1135 	 * just setup the lower byte.
1136 	 */
1137 	p_setb(perm, PCI_MSI_FLAGS, (u8)ALL_VIRT, (u8)ALL_WRITE);
1138 	p_setd(perm, PCI_MSI_ADDRESS_LO, ALL_VIRT, ALL_WRITE);
1139 	if (flags & PCI_MSI_FLAGS_64BIT) {
1140 		p_setd(perm, PCI_MSI_ADDRESS_HI, ALL_VIRT, ALL_WRITE);
1141 		p_setw(perm, PCI_MSI_DATA_64, (u16)ALL_VIRT, (u16)ALL_WRITE);
1142 		if (flags & PCI_MSI_FLAGS_MASKBIT) {
1143 			p_setd(perm, PCI_MSI_MASK_64, NO_VIRT, ALL_WRITE);
1144 			p_setd(perm, PCI_MSI_PENDING_64, NO_VIRT, ALL_WRITE);
1145 		}
1146 	} else {
1147 		p_setw(perm, PCI_MSI_DATA_32, (u16)ALL_VIRT, (u16)ALL_WRITE);
1148 		if (flags & PCI_MSI_FLAGS_MASKBIT) {
1149 			p_setd(perm, PCI_MSI_MASK_32, NO_VIRT, ALL_WRITE);
1150 			p_setd(perm, PCI_MSI_PENDING_32, NO_VIRT, ALL_WRITE);
1151 		}
1152 	}
1153 	return 0;
1154 }
1155 
1156 /* Determine MSI CAP field length; initialize msi_perms on 1st call per vdev */
1157 static int vfio_msi_cap_len(struct vfio_pci_device *vdev, u8 pos)
1158 {
1159 	struct pci_dev *pdev = vdev->pdev;
1160 	int len, ret;
1161 	u16 flags;
1162 
1163 	ret = pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &flags);
1164 	if (ret)
1165 		return pcibios_err_to_errno(ret);
1166 
1167 	len = 10; /* Minimum size */
1168 	if (flags & PCI_MSI_FLAGS_64BIT)
1169 		len += 4;
1170 	if (flags & PCI_MSI_FLAGS_MASKBIT)
1171 		len += 10;
1172 
1173 	if (vdev->msi_perm)
1174 		return len;
1175 
1176 	vdev->msi_perm = kmalloc(sizeof(struct perm_bits), GFP_KERNEL);
1177 	if (!vdev->msi_perm)
1178 		return -ENOMEM;
1179 
1180 	ret = init_pci_cap_msi_perm(vdev->msi_perm, len, flags);
1181 	if (ret) {
1182 		kfree(vdev->msi_perm);
1183 		return ret;
1184 	}
1185 
1186 	return len;
1187 }
1188 
1189 /* Determine extended capability length for VC (2 & 9) and MFVC */
1190 static int vfio_vc_cap_len(struct vfio_pci_device *vdev, u16 pos)
1191 {
1192 	struct pci_dev *pdev = vdev->pdev;
1193 	u32 tmp;
1194 	int ret, evcc, phases, vc_arb;
1195 	int len = PCI_CAP_VC_BASE_SIZEOF;
1196 
1197 	ret = pci_read_config_dword(pdev, pos + PCI_VC_PORT_CAP1, &tmp);
1198 	if (ret)
1199 		return pcibios_err_to_errno(ret);
1200 
1201 	evcc = tmp & PCI_VC_CAP1_EVCC; /* extended vc count */
1202 	ret = pci_read_config_dword(pdev, pos + PCI_VC_PORT_CAP2, &tmp);
1203 	if (ret)
1204 		return pcibios_err_to_errno(ret);
1205 
1206 	if (tmp & PCI_VC_CAP2_128_PHASE)
1207 		phases = 128;
1208 	else if (tmp & PCI_VC_CAP2_64_PHASE)
1209 		phases = 64;
1210 	else if (tmp & PCI_VC_CAP2_32_PHASE)
1211 		phases = 32;
1212 	else
1213 		phases = 0;
1214 
1215 	vc_arb = phases * 4;
1216 
1217 	/*
1218 	 * Port arbitration tables are root & switch only;
1219 	 * function arbitration tables are function 0 only.
1220 	 * In either case, we'll never let user write them so
1221 	 * we don't care how big they are
1222 	 */
1223 	len += (1 + evcc) * PCI_CAP_VC_PER_VC_SIZEOF;
1224 	if (vc_arb) {
1225 		len = round_up(len, 16);
1226 		len += vc_arb / 8;
1227 	}
1228 	return len;
1229 }
1230 
1231 static int vfio_cap_len(struct vfio_pci_device *vdev, u8 cap, u8 pos)
1232 {
1233 	struct pci_dev *pdev = vdev->pdev;
1234 	u32 dword;
1235 	u16 word;
1236 	u8 byte;
1237 	int ret;
1238 
1239 	switch (cap) {
1240 	case PCI_CAP_ID_MSI:
1241 		return vfio_msi_cap_len(vdev, pos);
1242 	case PCI_CAP_ID_PCIX:
1243 		ret = pci_read_config_word(pdev, pos + PCI_X_CMD, &word);
1244 		if (ret)
1245 			return pcibios_err_to_errno(ret);
1246 
1247 		if (PCI_X_CMD_VERSION(word)) {
1248 			if (pdev->cfg_size > PCI_CFG_SPACE_SIZE) {
1249 				/* Test for extended capabilities */
1250 				pci_read_config_dword(pdev, PCI_CFG_SPACE_SIZE,
1251 						      &dword);
1252 				vdev->extended_caps = (dword != 0);
1253 			}
1254 			return PCI_CAP_PCIX_SIZEOF_V2;
1255 		} else
1256 			return PCI_CAP_PCIX_SIZEOF_V0;
1257 	case PCI_CAP_ID_VNDR:
1258 		/* length follows next field */
1259 		ret = pci_read_config_byte(pdev, pos + PCI_CAP_FLAGS, &byte);
1260 		if (ret)
1261 			return pcibios_err_to_errno(ret);
1262 
1263 		return byte;
1264 	case PCI_CAP_ID_EXP:
1265 		if (pdev->cfg_size > PCI_CFG_SPACE_SIZE) {
1266 			/* Test for extended capabilities */
1267 			pci_read_config_dword(pdev, PCI_CFG_SPACE_SIZE, &dword);
1268 			vdev->extended_caps = (dword != 0);
1269 		}
1270 
1271 		/* length based on version and type */
1272 		if ((pcie_caps_reg(pdev) & PCI_EXP_FLAGS_VERS) == 1) {
1273 			if (pci_pcie_type(pdev) == PCI_EXP_TYPE_RC_END)
1274 				return 0xc; /* "All Devices" only, no link */
1275 			return PCI_CAP_EXP_ENDPOINT_SIZEOF_V1;
1276 		} else {
1277 			if (pci_pcie_type(pdev) == PCI_EXP_TYPE_RC_END)
1278 				return 0x2c; /* No link */
1279 			return PCI_CAP_EXP_ENDPOINT_SIZEOF_V2;
1280 		}
1281 	case PCI_CAP_ID_HT:
1282 		ret = pci_read_config_byte(pdev, pos + 3, &byte);
1283 		if (ret)
1284 			return pcibios_err_to_errno(ret);
1285 
1286 		return (byte & HT_3BIT_CAP_MASK) ?
1287 			HT_CAP_SIZEOF_SHORT : HT_CAP_SIZEOF_LONG;
1288 	case PCI_CAP_ID_SATA:
1289 		ret = pci_read_config_byte(pdev, pos + PCI_SATA_REGS, &byte);
1290 		if (ret)
1291 			return pcibios_err_to_errno(ret);
1292 
1293 		byte &= PCI_SATA_REGS_MASK;
1294 		if (byte == PCI_SATA_REGS_INLINE)
1295 			return PCI_SATA_SIZEOF_LONG;
1296 		else
1297 			return PCI_SATA_SIZEOF_SHORT;
1298 	default:
1299 		pci_warn(pdev, "%s: unknown length for PCI cap %#x@%#x\n",
1300 			 __func__, cap, pos);
1301 	}
1302 
1303 	return 0;
1304 }
1305 
1306 static int vfio_ext_cap_len(struct vfio_pci_device *vdev, u16 ecap, u16 epos)
1307 {
1308 	struct pci_dev *pdev = vdev->pdev;
1309 	u8 byte;
1310 	u32 dword;
1311 	int ret;
1312 
1313 	switch (ecap) {
1314 	case PCI_EXT_CAP_ID_VNDR:
1315 		ret = pci_read_config_dword(pdev, epos + PCI_VSEC_HDR, &dword);
1316 		if (ret)
1317 			return pcibios_err_to_errno(ret);
1318 
1319 		return dword >> PCI_VSEC_HDR_LEN_SHIFT;
1320 	case PCI_EXT_CAP_ID_VC:
1321 	case PCI_EXT_CAP_ID_VC9:
1322 	case PCI_EXT_CAP_ID_MFVC:
1323 		return vfio_vc_cap_len(vdev, epos);
1324 	case PCI_EXT_CAP_ID_ACS:
1325 		ret = pci_read_config_byte(pdev, epos + PCI_ACS_CAP, &byte);
1326 		if (ret)
1327 			return pcibios_err_to_errno(ret);
1328 
1329 		if (byte & PCI_ACS_EC) {
1330 			int bits;
1331 
1332 			ret = pci_read_config_byte(pdev,
1333 						   epos + PCI_ACS_EGRESS_BITS,
1334 						   &byte);
1335 			if (ret)
1336 				return pcibios_err_to_errno(ret);
1337 
1338 			bits = byte ? round_up(byte, 32) : 256;
1339 			return 8 + (bits / 8);
1340 		}
1341 		return 8;
1342 
1343 	case PCI_EXT_CAP_ID_REBAR:
1344 		ret = pci_read_config_byte(pdev, epos + PCI_REBAR_CTRL, &byte);
1345 		if (ret)
1346 			return pcibios_err_to_errno(ret);
1347 
1348 		byte &= PCI_REBAR_CTRL_NBAR_MASK;
1349 		byte >>= PCI_REBAR_CTRL_NBAR_SHIFT;
1350 
1351 		return 4 + (byte * 8);
1352 	case PCI_EXT_CAP_ID_DPA:
1353 		ret = pci_read_config_byte(pdev, epos + PCI_DPA_CAP, &byte);
1354 		if (ret)
1355 			return pcibios_err_to_errno(ret);
1356 
1357 		byte &= PCI_DPA_CAP_SUBSTATE_MASK;
1358 		return PCI_DPA_BASE_SIZEOF + byte + 1;
1359 	case PCI_EXT_CAP_ID_TPH:
1360 		ret = pci_read_config_dword(pdev, epos + PCI_TPH_CAP, &dword);
1361 		if (ret)
1362 			return pcibios_err_to_errno(ret);
1363 
1364 		if ((dword & PCI_TPH_CAP_LOC_MASK) == PCI_TPH_LOC_CAP) {
1365 			int sts;
1366 
1367 			sts = dword & PCI_TPH_CAP_ST_MASK;
1368 			sts >>= PCI_TPH_CAP_ST_SHIFT;
1369 			return PCI_TPH_BASE_SIZEOF + (sts * 2) + 2;
1370 		}
1371 		return PCI_TPH_BASE_SIZEOF;
1372 	default:
1373 		pci_warn(pdev, "%s: unknown length for PCI ecap %#x@%#x\n",
1374 			 __func__, ecap, epos);
1375 	}
1376 
1377 	return 0;
1378 }
1379 
1380 static int vfio_fill_vconfig_bytes(struct vfio_pci_device *vdev,
1381 				   int offset, int size)
1382 {
1383 	struct pci_dev *pdev = vdev->pdev;
1384 	int ret = 0;
1385 
1386 	/*
1387 	 * We try to read physical config space in the largest chunks
1388 	 * we can, assuming that all of the fields support dword access.
1389 	 * pci_save_state() makes this same assumption and seems to do ok.
1390 	 */
1391 	while (size) {
1392 		int filled;
1393 
1394 		if (size >= 4 && !(offset % 4)) {
1395 			__le32 *dwordp = (__le32 *)&vdev->vconfig[offset];
1396 			u32 dword;
1397 
1398 			ret = pci_read_config_dword(pdev, offset, &dword);
1399 			if (ret)
1400 				return ret;
1401 			*dwordp = cpu_to_le32(dword);
1402 			filled = 4;
1403 		} else if (size >= 2 && !(offset % 2)) {
1404 			__le16 *wordp = (__le16 *)&vdev->vconfig[offset];
1405 			u16 word;
1406 
1407 			ret = pci_read_config_word(pdev, offset, &word);
1408 			if (ret)
1409 				return ret;
1410 			*wordp = cpu_to_le16(word);
1411 			filled = 2;
1412 		} else {
1413 			u8 *byte = &vdev->vconfig[offset];
1414 			ret = pci_read_config_byte(pdev, offset, byte);
1415 			if (ret)
1416 				return ret;
1417 			filled = 1;
1418 		}
1419 
1420 		offset += filled;
1421 		size -= filled;
1422 	}
1423 
1424 	return ret;
1425 }
1426 
1427 static int vfio_cap_init(struct vfio_pci_device *vdev)
1428 {
1429 	struct pci_dev *pdev = vdev->pdev;
1430 	u8 *map = vdev->pci_config_map;
1431 	u16 status;
1432 	u8 pos, *prev, cap;
1433 	int loops, ret, caps = 0;
1434 
1435 	/* Any capabilities? */
1436 	ret = pci_read_config_word(pdev, PCI_STATUS, &status);
1437 	if (ret)
1438 		return ret;
1439 
1440 	if (!(status & PCI_STATUS_CAP_LIST))
1441 		return 0; /* Done */
1442 
1443 	ret = pci_read_config_byte(pdev, PCI_CAPABILITY_LIST, &pos);
1444 	if (ret)
1445 		return ret;
1446 
1447 	/* Mark the previous position in case we want to skip a capability */
1448 	prev = &vdev->vconfig[PCI_CAPABILITY_LIST];
1449 
1450 	/* We can bound our loop, capabilities are dword aligned */
1451 	loops = (PCI_CFG_SPACE_SIZE - PCI_STD_HEADER_SIZEOF) / PCI_CAP_SIZEOF;
1452 	while (pos && loops--) {
1453 		u8 next;
1454 		int i, len = 0;
1455 
1456 		ret = pci_read_config_byte(pdev, pos, &cap);
1457 		if (ret)
1458 			return ret;
1459 
1460 		ret = pci_read_config_byte(pdev,
1461 					   pos + PCI_CAP_LIST_NEXT, &next);
1462 		if (ret)
1463 			return ret;
1464 
1465 		if (cap <= PCI_CAP_ID_MAX) {
1466 			len = pci_cap_length[cap];
1467 			if (len == 0xFF) { /* Variable length */
1468 				len = vfio_cap_len(vdev, cap, pos);
1469 				if (len < 0)
1470 					return len;
1471 			}
1472 		}
1473 
1474 		if (!len) {
1475 			pci_info(pdev, "%s: hiding cap %#x@%#x\n", __func__,
1476 				 cap, pos);
1477 			*prev = next;
1478 			pos = next;
1479 			continue;
1480 		}
1481 
1482 		/* Sanity check, do we overlap other capabilities? */
1483 		for (i = 0; i < len; i++) {
1484 			if (likely(map[pos + i] == PCI_CAP_ID_INVALID))
1485 				continue;
1486 
1487 			pci_warn(pdev, "%s: PCI config conflict @%#x, was cap %#x now cap %#x\n",
1488 				 __func__, pos + i, map[pos + i], cap);
1489 		}
1490 
1491 		BUILD_BUG_ON(PCI_CAP_ID_MAX >= PCI_CAP_ID_INVALID_VIRT);
1492 
1493 		memset(map + pos, cap, len);
1494 		ret = vfio_fill_vconfig_bytes(vdev, pos, len);
1495 		if (ret)
1496 			return ret;
1497 
1498 		prev = &vdev->vconfig[pos + PCI_CAP_LIST_NEXT];
1499 		pos = next;
1500 		caps++;
1501 	}
1502 
1503 	/* If we didn't fill any capabilities, clear the status flag */
1504 	if (!caps) {
1505 		__le16 *vstatus = (__le16 *)&vdev->vconfig[PCI_STATUS];
1506 		*vstatus &= ~cpu_to_le16(PCI_STATUS_CAP_LIST);
1507 	}
1508 
1509 	return 0;
1510 }
1511 
1512 static int vfio_ecap_init(struct vfio_pci_device *vdev)
1513 {
1514 	struct pci_dev *pdev = vdev->pdev;
1515 	u8 *map = vdev->pci_config_map;
1516 	u16 epos;
1517 	__le32 *prev = NULL;
1518 	int loops, ret, ecaps = 0;
1519 
1520 	if (!vdev->extended_caps)
1521 		return 0;
1522 
1523 	epos = PCI_CFG_SPACE_SIZE;
1524 
1525 	loops = (pdev->cfg_size - PCI_CFG_SPACE_SIZE) / PCI_CAP_SIZEOF;
1526 
1527 	while (loops-- && epos >= PCI_CFG_SPACE_SIZE) {
1528 		u32 header;
1529 		u16 ecap;
1530 		int i, len = 0;
1531 		bool hidden = false;
1532 
1533 		ret = pci_read_config_dword(pdev, epos, &header);
1534 		if (ret)
1535 			return ret;
1536 
1537 		ecap = PCI_EXT_CAP_ID(header);
1538 
1539 		if (ecap <= PCI_EXT_CAP_ID_MAX) {
1540 			len = pci_ext_cap_length[ecap];
1541 			if (len == 0xFF) {
1542 				len = vfio_ext_cap_len(vdev, ecap, epos);
1543 				if (len < 0)
1544 					return ret;
1545 			}
1546 		}
1547 
1548 		if (!len) {
1549 			pci_info(pdev, "%s: hiding ecap %#x@%#x\n",
1550 				 __func__, ecap, epos);
1551 
1552 			/* If not the first in the chain, we can skip over it */
1553 			if (prev) {
1554 				u32 val = epos = PCI_EXT_CAP_NEXT(header);
1555 				*prev &= cpu_to_le32(~(0xffcU << 20));
1556 				*prev |= cpu_to_le32(val << 20);
1557 				continue;
1558 			}
1559 
1560 			/*
1561 			 * Otherwise, fill in a placeholder, the direct
1562 			 * readfn will virtualize this automatically
1563 			 */
1564 			len = PCI_CAP_SIZEOF;
1565 			hidden = true;
1566 		}
1567 
1568 		for (i = 0; i < len; i++) {
1569 			if (likely(map[epos + i] == PCI_CAP_ID_INVALID))
1570 				continue;
1571 
1572 			pci_warn(pdev, "%s: PCI config conflict @%#x, was ecap %#x now ecap %#x\n",
1573 				 __func__, epos + i, map[epos + i], ecap);
1574 		}
1575 
1576 		/*
1577 		 * Even though ecap is 2 bytes, we're currently a long way
1578 		 * from exceeding 1 byte capabilities.  If we ever make it
1579 		 * up to 0xFE we'll need to up this to a two-byte, byte map.
1580 		 */
1581 		BUILD_BUG_ON(PCI_EXT_CAP_ID_MAX >= PCI_CAP_ID_INVALID_VIRT);
1582 
1583 		memset(map + epos, ecap, len);
1584 		ret = vfio_fill_vconfig_bytes(vdev, epos, len);
1585 		if (ret)
1586 			return ret;
1587 
1588 		/*
1589 		 * If we're just using this capability to anchor the list,
1590 		 * hide the real ID.  Only count real ecaps.  XXX PCI spec
1591 		 * indicates to use cap id = 0, version = 0, next = 0 if
1592 		 * ecaps are absent, hope users check all the way to next.
1593 		 */
1594 		if (hidden)
1595 			*(__le32 *)&vdev->vconfig[epos] &=
1596 				cpu_to_le32((0xffcU << 20));
1597 		else
1598 			ecaps++;
1599 
1600 		prev = (__le32 *)&vdev->vconfig[epos];
1601 		epos = PCI_EXT_CAP_NEXT(header);
1602 	}
1603 
1604 	if (!ecaps)
1605 		*(u32 *)&vdev->vconfig[PCI_CFG_SPACE_SIZE] = 0;
1606 
1607 	return 0;
1608 }
1609 
1610 /*
1611  * Nag about hardware bugs, hopefully to have vendors fix them, but at least
1612  * to collect a list of dependencies for the VF INTx pin quirk below.
1613  */
1614 static const struct pci_device_id known_bogus_vf_intx_pin[] = {
1615 	{ PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x270c) },
1616 	{}
1617 };
1618 
1619 /*
1620  * For each device we allocate a pci_config_map that indicates the
1621  * capability occupying each dword and thus the struct perm_bits we
1622  * use for read and write.  We also allocate a virtualized config
1623  * space which tracks reads and writes to bits that we emulate for
1624  * the user.  Initial values filled from device.
1625  *
1626  * Using shared struct perm_bits between all vfio-pci devices saves
1627  * us from allocating cfg_size buffers for virt and write for every
1628  * device.  We could remove vconfig and allocate individual buffers
1629  * for each area requiring emulated bits, but the array of pointers
1630  * would be comparable in size (at least for standard config space).
1631  */
1632 int vfio_config_init(struct vfio_pci_device *vdev)
1633 {
1634 	struct pci_dev *pdev = vdev->pdev;
1635 	u8 *map, *vconfig;
1636 	int ret;
1637 
1638 	/*
1639 	 * Config space, caps and ecaps are all dword aligned, so we could
1640 	 * use one byte per dword to record the type.  However, there are
1641 	 * no requiremenst on the length of a capability, so the gap between
1642 	 * capabilities needs byte granularity.
1643 	 */
1644 	map = kmalloc(pdev->cfg_size, GFP_KERNEL);
1645 	if (!map)
1646 		return -ENOMEM;
1647 
1648 	vconfig = kmalloc(pdev->cfg_size, GFP_KERNEL);
1649 	if (!vconfig) {
1650 		kfree(map);
1651 		return -ENOMEM;
1652 	}
1653 
1654 	vdev->pci_config_map = map;
1655 	vdev->vconfig = vconfig;
1656 
1657 	memset(map, PCI_CAP_ID_BASIC, PCI_STD_HEADER_SIZEOF);
1658 	memset(map + PCI_STD_HEADER_SIZEOF, PCI_CAP_ID_INVALID,
1659 	       pdev->cfg_size - PCI_STD_HEADER_SIZEOF);
1660 
1661 	ret = vfio_fill_vconfig_bytes(vdev, 0, PCI_STD_HEADER_SIZEOF);
1662 	if (ret)
1663 		goto out;
1664 
1665 	vdev->bardirty = true;
1666 
1667 	/*
1668 	 * XXX can we just pci_load_saved_state/pci_restore_state?
1669 	 * may need to rebuild vconfig after that
1670 	 */
1671 
1672 	/* For restore after reset */
1673 	vdev->rbar[0] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_0]);
1674 	vdev->rbar[1] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_1]);
1675 	vdev->rbar[2] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_2]);
1676 	vdev->rbar[3] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_3]);
1677 	vdev->rbar[4] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_4]);
1678 	vdev->rbar[5] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_5]);
1679 	vdev->rbar[6] = le32_to_cpu(*(__le32 *)&vconfig[PCI_ROM_ADDRESS]);
1680 
1681 	if (pdev->is_virtfn) {
1682 		*(__le16 *)&vconfig[PCI_VENDOR_ID] = cpu_to_le16(pdev->vendor);
1683 		*(__le16 *)&vconfig[PCI_DEVICE_ID] = cpu_to_le16(pdev->device);
1684 
1685 		/*
1686 		 * Per SR-IOV spec rev 1.1, 3.4.1.18 the interrupt pin register
1687 		 * does not apply to VFs and VFs must implement this register
1688 		 * as read-only with value zero.  Userspace is not readily able
1689 		 * to identify whether a device is a VF and thus that the pin
1690 		 * definition on the device is bogus should it violate this
1691 		 * requirement.  We already virtualize the pin register for
1692 		 * other purposes, so we simply need to replace the bogus value
1693 		 * and consider VFs when we determine INTx IRQ count.
1694 		 */
1695 		if (vconfig[PCI_INTERRUPT_PIN] &&
1696 		    !pci_match_id(known_bogus_vf_intx_pin, pdev))
1697 			pci_warn(pdev,
1698 				 "Hardware bug: VF reports bogus INTx pin %d\n",
1699 				 vconfig[PCI_INTERRUPT_PIN]);
1700 
1701 		vconfig[PCI_INTERRUPT_PIN] = 0; /* Gratuitous for good VFs */
1702 	}
1703 
1704 	if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) || vdev->nointx)
1705 		vconfig[PCI_INTERRUPT_PIN] = 0;
1706 
1707 	ret = vfio_cap_init(vdev);
1708 	if (ret)
1709 		goto out;
1710 
1711 	ret = vfio_ecap_init(vdev);
1712 	if (ret)
1713 		goto out;
1714 
1715 	return 0;
1716 
1717 out:
1718 	kfree(map);
1719 	vdev->pci_config_map = NULL;
1720 	kfree(vconfig);
1721 	vdev->vconfig = NULL;
1722 	return pcibios_err_to_errno(ret);
1723 }
1724 
1725 void vfio_config_free(struct vfio_pci_device *vdev)
1726 {
1727 	kfree(vdev->vconfig);
1728 	vdev->vconfig = NULL;
1729 	kfree(vdev->pci_config_map);
1730 	vdev->pci_config_map = NULL;
1731 	kfree(vdev->msi_perm);
1732 	vdev->msi_perm = NULL;
1733 }
1734 
1735 /*
1736  * Find the remaining number of bytes in a dword that match the given
1737  * position.  Stop at either the end of the capability or the dword boundary.
1738  */
1739 static size_t vfio_pci_cap_remaining_dword(struct vfio_pci_device *vdev,
1740 					   loff_t pos)
1741 {
1742 	u8 cap = vdev->pci_config_map[pos];
1743 	size_t i;
1744 
1745 	for (i = 1; (pos + i) % 4 && vdev->pci_config_map[pos + i] == cap; i++)
1746 		/* nop */;
1747 
1748 	return i;
1749 }
1750 
1751 static ssize_t vfio_config_do_rw(struct vfio_pci_device *vdev, char __user *buf,
1752 				 size_t count, loff_t *ppos, bool iswrite)
1753 {
1754 	struct pci_dev *pdev = vdev->pdev;
1755 	struct perm_bits *perm;
1756 	__le32 val = 0;
1757 	int cap_start = 0, offset;
1758 	u8 cap_id;
1759 	ssize_t ret;
1760 
1761 	if (*ppos < 0 || *ppos >= pdev->cfg_size ||
1762 	    *ppos + count > pdev->cfg_size)
1763 		return -EFAULT;
1764 
1765 	/*
1766 	 * Chop accesses into aligned chunks containing no more than a
1767 	 * single capability.  Caller increments to the next chunk.
1768 	 */
1769 	count = min(count, vfio_pci_cap_remaining_dword(vdev, *ppos));
1770 	if (count >= 4 && !(*ppos % 4))
1771 		count = 4;
1772 	else if (count >= 2 && !(*ppos % 2))
1773 		count = 2;
1774 	else
1775 		count = 1;
1776 
1777 	ret = count;
1778 
1779 	cap_id = vdev->pci_config_map[*ppos];
1780 
1781 	if (cap_id == PCI_CAP_ID_INVALID) {
1782 		perm = &unassigned_perms;
1783 		cap_start = *ppos;
1784 	} else if (cap_id == PCI_CAP_ID_INVALID_VIRT) {
1785 		perm = &virt_perms;
1786 		cap_start = *ppos;
1787 	} else {
1788 		if (*ppos >= PCI_CFG_SPACE_SIZE) {
1789 			WARN_ON(cap_id > PCI_EXT_CAP_ID_MAX);
1790 
1791 			perm = &ecap_perms[cap_id];
1792 			cap_start = vfio_find_cap_start(vdev, *ppos);
1793 		} else {
1794 			WARN_ON(cap_id > PCI_CAP_ID_MAX);
1795 
1796 			perm = &cap_perms[cap_id];
1797 
1798 			if (cap_id == PCI_CAP_ID_MSI)
1799 				perm = vdev->msi_perm;
1800 
1801 			if (cap_id > PCI_CAP_ID_BASIC)
1802 				cap_start = vfio_find_cap_start(vdev, *ppos);
1803 		}
1804 	}
1805 
1806 	WARN_ON(!cap_start && cap_id != PCI_CAP_ID_BASIC);
1807 	WARN_ON(cap_start > *ppos);
1808 
1809 	offset = *ppos - cap_start;
1810 
1811 	if (iswrite) {
1812 		if (!perm->writefn)
1813 			return ret;
1814 
1815 		if (copy_from_user(&val, buf, count))
1816 			return -EFAULT;
1817 
1818 		ret = perm->writefn(vdev, *ppos, count, perm, offset, val);
1819 	} else {
1820 		if (perm->readfn) {
1821 			ret = perm->readfn(vdev, *ppos, count,
1822 					   perm, offset, &val);
1823 			if (ret < 0)
1824 				return ret;
1825 		}
1826 
1827 		if (copy_to_user(buf, &val, count))
1828 			return -EFAULT;
1829 	}
1830 
1831 	return ret;
1832 }
1833 
1834 ssize_t vfio_pci_config_rw(struct vfio_pci_device *vdev, char __user *buf,
1835 			   size_t count, loff_t *ppos, bool iswrite)
1836 {
1837 	size_t done = 0;
1838 	int ret = 0;
1839 	loff_t pos = *ppos;
1840 
1841 	pos &= VFIO_PCI_OFFSET_MASK;
1842 
1843 	while (count) {
1844 		ret = vfio_config_do_rw(vdev, buf, count, &pos, iswrite);
1845 		if (ret < 0)
1846 			return ret;
1847 
1848 		count -= ret;
1849 		done += ret;
1850 		buf += ret;
1851 		pos += ret;
1852 	}
1853 
1854 	*ppos += done;
1855 
1856 	return done;
1857 }
1858