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_priv.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 	[PCI_EXT_CAP_ID_DVSEC]	=	0xFF,
100 };
101 
102 /*
103  * Read/Write Permission Bits - one bit for each bit in capability
104  * Any field can be read if it exists, but what is read depends on
105  * whether the field is 'virtualized', or just pass through to the
106  * hardware.  Any virtualized field is also virtualized for writes.
107  * Writes are only permitted if they have a 1 bit here.
108  */
109 struct perm_bits {
110 	u8	*virt;		/* read/write virtual data, not hw */
111 	u8	*write;		/* writeable bits */
112 	int	(*readfn)(struct vfio_pci_core_device *vdev, int pos, int count,
113 			  struct perm_bits *perm, int offset, __le32 *val);
114 	int	(*writefn)(struct vfio_pci_core_device *vdev, int pos, int count,
115 			   struct perm_bits *perm, int offset, __le32 val);
116 };
117 
118 #define	NO_VIRT		0
119 #define	ALL_VIRT	0xFFFFFFFFU
120 #define	NO_WRITE	0
121 #define	ALL_WRITE	0xFFFFFFFFU
122 
123 static int vfio_user_config_read(struct pci_dev *pdev, int offset,
124 				 __le32 *val, int count)
125 {
126 	int ret = -EINVAL;
127 	u32 tmp_val = 0;
128 
129 	switch (count) {
130 	case 1:
131 	{
132 		u8 tmp;
133 		ret = pci_user_read_config_byte(pdev, offset, &tmp);
134 		tmp_val = tmp;
135 		break;
136 	}
137 	case 2:
138 	{
139 		u16 tmp;
140 		ret = pci_user_read_config_word(pdev, offset, &tmp);
141 		tmp_val = tmp;
142 		break;
143 	}
144 	case 4:
145 		ret = pci_user_read_config_dword(pdev, offset, &tmp_val);
146 		break;
147 	}
148 
149 	*val = cpu_to_le32(tmp_val);
150 
151 	return ret;
152 }
153 
154 static int vfio_user_config_write(struct pci_dev *pdev, int offset,
155 				  __le32 val, int count)
156 {
157 	int ret = -EINVAL;
158 	u32 tmp_val = le32_to_cpu(val);
159 
160 	switch (count) {
161 	case 1:
162 		ret = pci_user_write_config_byte(pdev, offset, tmp_val);
163 		break;
164 	case 2:
165 		ret = pci_user_write_config_word(pdev, offset, tmp_val);
166 		break;
167 	case 4:
168 		ret = pci_user_write_config_dword(pdev, offset, tmp_val);
169 		break;
170 	}
171 
172 	return ret;
173 }
174 
175 static int vfio_default_config_read(struct vfio_pci_core_device *vdev, int pos,
176 				    int count, struct perm_bits *perm,
177 				    int offset, __le32 *val)
178 {
179 	__le32 virt = 0;
180 
181 	memcpy(val, vdev->vconfig + pos, count);
182 
183 	memcpy(&virt, perm->virt + offset, count);
184 
185 	/* Any non-virtualized bits? */
186 	if (cpu_to_le32(~0U >> (32 - (count * 8))) != virt) {
187 		struct pci_dev *pdev = vdev->pdev;
188 		__le32 phys_val = 0;
189 		int ret;
190 
191 		ret = vfio_user_config_read(pdev, pos, &phys_val, count);
192 		if (ret)
193 			return ret;
194 
195 		*val = (phys_val & ~virt) | (*val & virt);
196 	}
197 
198 	return count;
199 }
200 
201 static int vfio_default_config_write(struct vfio_pci_core_device *vdev, int pos,
202 				     int count, struct perm_bits *perm,
203 				     int offset, __le32 val)
204 {
205 	__le32 virt = 0, write = 0;
206 
207 	memcpy(&write, perm->write + offset, count);
208 
209 	if (!write)
210 		return count; /* drop, no writable bits */
211 
212 	memcpy(&virt, perm->virt + offset, count);
213 
214 	/* Virtualized and writable bits go to vconfig */
215 	if (write & virt) {
216 		__le32 virt_val = 0;
217 
218 		memcpy(&virt_val, vdev->vconfig + pos, count);
219 
220 		virt_val &= ~(write & virt);
221 		virt_val |= (val & (write & virt));
222 
223 		memcpy(vdev->vconfig + pos, &virt_val, count);
224 	}
225 
226 	/* Non-virtualized and writable bits go to hardware */
227 	if (write & ~virt) {
228 		struct pci_dev *pdev = vdev->pdev;
229 		__le32 phys_val = 0;
230 		int ret;
231 
232 		ret = vfio_user_config_read(pdev, pos, &phys_val, count);
233 		if (ret)
234 			return ret;
235 
236 		phys_val &= ~(write & ~virt);
237 		phys_val |= (val & (write & ~virt));
238 
239 		ret = vfio_user_config_write(pdev, pos, phys_val, count);
240 		if (ret)
241 			return ret;
242 	}
243 
244 	return count;
245 }
246 
247 /* Allow direct read from hardware, except for capability next pointer */
248 static int vfio_direct_config_read(struct vfio_pci_core_device *vdev, int pos,
249 				   int count, struct perm_bits *perm,
250 				   int offset, __le32 *val)
251 {
252 	int ret;
253 
254 	ret = vfio_user_config_read(vdev->pdev, pos, val, count);
255 	if (ret)
256 		return ret;
257 
258 	if (pos >= PCI_CFG_SPACE_SIZE) { /* Extended cap header mangling */
259 		if (offset < 4)
260 			memcpy(val, vdev->vconfig + pos, count);
261 	} else if (pos >= PCI_STD_HEADER_SIZEOF) { /* Std cap mangling */
262 		if (offset == PCI_CAP_LIST_ID && count > 1)
263 			memcpy(val, vdev->vconfig + pos,
264 			       min(PCI_CAP_FLAGS, count));
265 		else if (offset == PCI_CAP_LIST_NEXT)
266 			memcpy(val, vdev->vconfig + pos, 1);
267 	}
268 
269 	return count;
270 }
271 
272 /* Raw access skips any kind of virtualization */
273 static int vfio_raw_config_write(struct vfio_pci_core_device *vdev, int pos,
274 				 int count, struct perm_bits *perm,
275 				 int offset, __le32 val)
276 {
277 	int ret;
278 
279 	ret = vfio_user_config_write(vdev->pdev, pos, val, count);
280 	if (ret)
281 		return ret;
282 
283 	return count;
284 }
285 
286 static int vfio_raw_config_read(struct vfio_pci_core_device *vdev, int pos,
287 				int count, struct perm_bits *perm,
288 				int offset, __le32 *val)
289 {
290 	int ret;
291 
292 	ret = vfio_user_config_read(vdev->pdev, pos, val, count);
293 	if (ret)
294 		return ret;
295 
296 	return count;
297 }
298 
299 /* Virt access uses only virtualization */
300 static int vfio_virt_config_write(struct vfio_pci_core_device *vdev, int pos,
301 				  int count, struct perm_bits *perm,
302 				  int offset, __le32 val)
303 {
304 	memcpy(vdev->vconfig + pos, &val, count);
305 	return count;
306 }
307 
308 static int vfio_virt_config_read(struct vfio_pci_core_device *vdev, int pos,
309 				 int count, struct perm_bits *perm,
310 				 int offset, __le32 *val)
311 {
312 	memcpy(val, vdev->vconfig + pos, count);
313 	return count;
314 }
315 
316 /* Default capability regions to read-only, no-virtualization */
317 static struct perm_bits cap_perms[PCI_CAP_ID_MAX + 1] = {
318 	[0 ... PCI_CAP_ID_MAX] = { .readfn = vfio_direct_config_read }
319 };
320 static struct perm_bits ecap_perms[PCI_EXT_CAP_ID_MAX + 1] = {
321 	[0 ... PCI_EXT_CAP_ID_MAX] = { .readfn = vfio_direct_config_read }
322 };
323 /*
324  * Default unassigned regions to raw read-write access.  Some devices
325  * require this to function as they hide registers between the gaps in
326  * config space (be2net).  Like MMIO and I/O port registers, we have
327  * to trust the hardware isolation.
328  */
329 static struct perm_bits unassigned_perms = {
330 	.readfn = vfio_raw_config_read,
331 	.writefn = vfio_raw_config_write
332 };
333 
334 static struct perm_bits virt_perms = {
335 	.readfn = vfio_virt_config_read,
336 	.writefn = vfio_virt_config_write
337 };
338 
339 static void free_perm_bits(struct perm_bits *perm)
340 {
341 	kfree(perm->virt);
342 	kfree(perm->write);
343 	perm->virt = NULL;
344 	perm->write = NULL;
345 }
346 
347 static int alloc_perm_bits(struct perm_bits *perm, int size)
348 {
349 	/*
350 	 * Round up all permission bits to the next dword, this lets us
351 	 * ignore whether a read/write exceeds the defined capability
352 	 * structure.  We can do this because:
353 	 *  - Standard config space is already dword aligned
354 	 *  - Capabilities are all dword aligned (bits 0:1 of next reserved)
355 	 *  - Express capabilities defined as dword aligned
356 	 */
357 	size = round_up(size, 4);
358 
359 	/*
360 	 * Zero state is
361 	 * - All Readable, None Writeable, None Virtualized
362 	 */
363 	perm->virt = kzalloc(size, GFP_KERNEL);
364 	perm->write = kzalloc(size, GFP_KERNEL);
365 	if (!perm->virt || !perm->write) {
366 		free_perm_bits(perm);
367 		return -ENOMEM;
368 	}
369 
370 	perm->readfn = vfio_default_config_read;
371 	perm->writefn = vfio_default_config_write;
372 
373 	return 0;
374 }
375 
376 /*
377  * Helper functions for filling in permission tables
378  */
379 static inline void p_setb(struct perm_bits *p, int off, u8 virt, u8 write)
380 {
381 	p->virt[off] = virt;
382 	p->write[off] = write;
383 }
384 
385 /* Handle endian-ness - pci and tables are little-endian */
386 static inline void p_setw(struct perm_bits *p, int off, u16 virt, u16 write)
387 {
388 	*(__le16 *)(&p->virt[off]) = cpu_to_le16(virt);
389 	*(__le16 *)(&p->write[off]) = cpu_to_le16(write);
390 }
391 
392 /* Handle endian-ness - pci and tables are little-endian */
393 static inline void p_setd(struct perm_bits *p, int off, u32 virt, u32 write)
394 {
395 	*(__le32 *)(&p->virt[off]) = cpu_to_le32(virt);
396 	*(__le32 *)(&p->write[off]) = cpu_to_le32(write);
397 }
398 
399 /* Caller should hold memory_lock semaphore */
400 bool __vfio_pci_memory_enabled(struct vfio_pci_core_device *vdev)
401 {
402 	struct pci_dev *pdev = vdev->pdev;
403 	u16 cmd = le16_to_cpu(*(__le16 *)&vdev->vconfig[PCI_COMMAND]);
404 
405 	/*
406 	 * Memory region cannot be accessed if device power state is D3.
407 	 *
408 	 * SR-IOV VF memory enable is handled by the MSE bit in the
409 	 * PF SR-IOV capability, there's therefore no need to trigger
410 	 * faults based on the virtual value.
411 	 */
412 	return pdev->current_state < PCI_D3hot &&
413 	       (pdev->no_command_memory || (cmd & PCI_COMMAND_MEMORY));
414 }
415 
416 /*
417  * Restore the *real* BARs after we detect a FLR or backdoor reset.
418  * (backdoor = some device specific technique that we didn't catch)
419  */
420 static void vfio_bar_restore(struct vfio_pci_core_device *vdev)
421 {
422 	struct pci_dev *pdev = vdev->pdev;
423 	u32 *rbar = vdev->rbar;
424 	u16 cmd;
425 	int i;
426 
427 	if (pdev->is_virtfn)
428 		return;
429 
430 	pci_info(pdev, "%s: reset recovery - restoring BARs\n", __func__);
431 
432 	for (i = PCI_BASE_ADDRESS_0; i <= PCI_BASE_ADDRESS_5; i += 4, rbar++)
433 		pci_user_write_config_dword(pdev, i, *rbar);
434 
435 	pci_user_write_config_dword(pdev, PCI_ROM_ADDRESS, *rbar);
436 
437 	if (vdev->nointx) {
438 		pci_user_read_config_word(pdev, PCI_COMMAND, &cmd);
439 		cmd |= PCI_COMMAND_INTX_DISABLE;
440 		pci_user_write_config_word(pdev, PCI_COMMAND, cmd);
441 	}
442 }
443 
444 static __le32 vfio_generate_bar_flags(struct pci_dev *pdev, int bar)
445 {
446 	unsigned long flags = pci_resource_flags(pdev, bar);
447 	u32 val;
448 
449 	if (flags & IORESOURCE_IO)
450 		return cpu_to_le32(PCI_BASE_ADDRESS_SPACE_IO);
451 
452 	val = PCI_BASE_ADDRESS_SPACE_MEMORY;
453 
454 	if (flags & IORESOURCE_PREFETCH)
455 		val |= PCI_BASE_ADDRESS_MEM_PREFETCH;
456 
457 	if (flags & IORESOURCE_MEM_64)
458 		val |= PCI_BASE_ADDRESS_MEM_TYPE_64;
459 
460 	return cpu_to_le32(val);
461 }
462 
463 /*
464  * Pretend we're hardware and tweak the values of the *virtual* PCI BARs
465  * to reflect the hardware capabilities.  This implements BAR sizing.
466  */
467 static void vfio_bar_fixup(struct vfio_pci_core_device *vdev)
468 {
469 	struct pci_dev *pdev = vdev->pdev;
470 	int i;
471 	__le32 *vbar;
472 	u64 mask;
473 
474 	if (!vdev->bardirty)
475 		return;
476 
477 	vbar = (__le32 *)&vdev->vconfig[PCI_BASE_ADDRESS_0];
478 
479 	for (i = 0; i < PCI_STD_NUM_BARS; i++, vbar++) {
480 		int bar = i + PCI_STD_RESOURCES;
481 
482 		if (!pci_resource_start(pdev, bar)) {
483 			*vbar = 0; /* Unmapped by host = unimplemented to user */
484 			continue;
485 		}
486 
487 		mask = ~(pci_resource_len(pdev, bar) - 1);
488 
489 		*vbar &= cpu_to_le32((u32)mask);
490 		*vbar |= vfio_generate_bar_flags(pdev, bar);
491 
492 		if (*vbar & cpu_to_le32(PCI_BASE_ADDRESS_MEM_TYPE_64)) {
493 			vbar++;
494 			*vbar &= cpu_to_le32((u32)(mask >> 32));
495 			i++;
496 		}
497 	}
498 
499 	vbar = (__le32 *)&vdev->vconfig[PCI_ROM_ADDRESS];
500 
501 	/*
502 	 * NB. REGION_INFO will have reported zero size if we weren't able
503 	 * to read the ROM, but we still return the actual BAR size here if
504 	 * it exists (or the shadow ROM space).
505 	 */
506 	if (pci_resource_start(pdev, PCI_ROM_RESOURCE)) {
507 		mask = ~(pci_resource_len(pdev, PCI_ROM_RESOURCE) - 1);
508 		mask |= PCI_ROM_ADDRESS_ENABLE;
509 		*vbar &= cpu_to_le32((u32)mask);
510 	} else if (pdev->resource[PCI_ROM_RESOURCE].flags &
511 					IORESOURCE_ROM_SHADOW) {
512 		mask = ~(0x20000 - 1);
513 		mask |= PCI_ROM_ADDRESS_ENABLE;
514 		*vbar &= cpu_to_le32((u32)mask);
515 	} else
516 		*vbar = 0;
517 
518 	vdev->bardirty = false;
519 }
520 
521 static int vfio_basic_config_read(struct vfio_pci_core_device *vdev, int pos,
522 				  int count, struct perm_bits *perm,
523 				  int offset, __le32 *val)
524 {
525 	if (is_bar(offset)) /* pos == offset for basic config */
526 		vfio_bar_fixup(vdev);
527 
528 	count = vfio_default_config_read(vdev, pos, count, perm, offset, val);
529 
530 	/* Mask in virtual memory enable */
531 	if (offset == PCI_COMMAND && vdev->pdev->no_command_memory) {
532 		u16 cmd = le16_to_cpu(*(__le16 *)&vdev->vconfig[PCI_COMMAND]);
533 		u32 tmp_val = le32_to_cpu(*val);
534 
535 		tmp_val |= cmd & PCI_COMMAND_MEMORY;
536 		*val = cpu_to_le32(tmp_val);
537 	}
538 
539 	return count;
540 }
541 
542 /* Test whether BARs match the value we think they should contain */
543 static bool vfio_need_bar_restore(struct vfio_pci_core_device *vdev)
544 {
545 	int i = 0, pos = PCI_BASE_ADDRESS_0, ret;
546 	u32 bar;
547 
548 	for (; pos <= PCI_BASE_ADDRESS_5; i++, pos += 4) {
549 		if (vdev->rbar[i]) {
550 			ret = pci_user_read_config_dword(vdev->pdev, pos, &bar);
551 			if (ret || vdev->rbar[i] != bar)
552 				return true;
553 		}
554 	}
555 
556 	return false;
557 }
558 
559 static int vfio_basic_config_write(struct vfio_pci_core_device *vdev, int pos,
560 				   int count, struct perm_bits *perm,
561 				   int offset, __le32 val)
562 {
563 	struct pci_dev *pdev = vdev->pdev;
564 	__le16 *virt_cmd;
565 	u16 new_cmd = 0;
566 	int ret;
567 
568 	virt_cmd = (__le16 *)&vdev->vconfig[PCI_COMMAND];
569 
570 	if (offset == PCI_COMMAND) {
571 		bool phys_mem, virt_mem, new_mem, phys_io, virt_io, new_io;
572 		u16 phys_cmd;
573 
574 		ret = pci_user_read_config_word(pdev, PCI_COMMAND, &phys_cmd);
575 		if (ret)
576 			return ret;
577 
578 		new_cmd = le32_to_cpu(val);
579 
580 		phys_io = !!(phys_cmd & PCI_COMMAND_IO);
581 		virt_io = !!(le16_to_cpu(*virt_cmd) & PCI_COMMAND_IO);
582 		new_io = !!(new_cmd & PCI_COMMAND_IO);
583 
584 		phys_mem = !!(phys_cmd & PCI_COMMAND_MEMORY);
585 		virt_mem = !!(le16_to_cpu(*virt_cmd) & PCI_COMMAND_MEMORY);
586 		new_mem = !!(new_cmd & PCI_COMMAND_MEMORY);
587 
588 		if (!new_mem)
589 			vfio_pci_zap_and_down_write_memory_lock(vdev);
590 		else
591 			down_write(&vdev->memory_lock);
592 
593 		/*
594 		 * If the user is writing mem/io enable (new_mem/io) and we
595 		 * think it's already enabled (virt_mem/io), but the hardware
596 		 * shows it disabled (phys_mem/io, then the device has
597 		 * undergone some kind of backdoor reset and needs to be
598 		 * restored before we allow it to enable the bars.
599 		 * SR-IOV devices will trigger this - for mem enable let's
600 		 * catch this now and for io enable it will be caught later
601 		 */
602 		if ((new_mem && virt_mem && !phys_mem &&
603 		     !pdev->no_command_memory) ||
604 		    (new_io && virt_io && !phys_io) ||
605 		    vfio_need_bar_restore(vdev))
606 			vfio_bar_restore(vdev);
607 	}
608 
609 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
610 	if (count < 0) {
611 		if (offset == PCI_COMMAND)
612 			up_write(&vdev->memory_lock);
613 		return count;
614 	}
615 
616 	/*
617 	 * Save current memory/io enable bits in vconfig to allow for
618 	 * the test above next time.
619 	 */
620 	if (offset == PCI_COMMAND) {
621 		u16 mask = PCI_COMMAND_MEMORY | PCI_COMMAND_IO;
622 
623 		*virt_cmd &= cpu_to_le16(~mask);
624 		*virt_cmd |= cpu_to_le16(new_cmd & mask);
625 
626 		up_write(&vdev->memory_lock);
627 	}
628 
629 	/* Emulate INTx disable */
630 	if (offset >= PCI_COMMAND && offset <= PCI_COMMAND + 1) {
631 		bool virt_intx_disable;
632 
633 		virt_intx_disable = !!(le16_to_cpu(*virt_cmd) &
634 				       PCI_COMMAND_INTX_DISABLE);
635 
636 		if (virt_intx_disable && !vdev->virq_disabled) {
637 			vdev->virq_disabled = true;
638 			vfio_pci_intx_mask(vdev);
639 		} else if (!virt_intx_disable && vdev->virq_disabled) {
640 			vdev->virq_disabled = false;
641 			vfio_pci_intx_unmask(vdev);
642 		}
643 	}
644 
645 	if (is_bar(offset))
646 		vdev->bardirty = true;
647 
648 	return count;
649 }
650 
651 /* Permissions for the Basic PCI Header */
652 static int __init init_pci_cap_basic_perm(struct perm_bits *perm)
653 {
654 	if (alloc_perm_bits(perm, PCI_STD_HEADER_SIZEOF))
655 		return -ENOMEM;
656 
657 	perm->readfn = vfio_basic_config_read;
658 	perm->writefn = vfio_basic_config_write;
659 
660 	/* Virtualized for SR-IOV functions, which just have FFFF */
661 	p_setw(perm, PCI_VENDOR_ID, (u16)ALL_VIRT, NO_WRITE);
662 	p_setw(perm, PCI_DEVICE_ID, (u16)ALL_VIRT, NO_WRITE);
663 
664 	/*
665 	 * Virtualize INTx disable, we use it internally for interrupt
666 	 * control and can emulate it for non-PCI 2.3 devices.
667 	 */
668 	p_setw(perm, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE, (u16)ALL_WRITE);
669 
670 	/* Virtualize capability list, we might want to skip/disable */
671 	p_setw(perm, PCI_STATUS, PCI_STATUS_CAP_LIST, NO_WRITE);
672 
673 	/* No harm to write */
674 	p_setb(perm, PCI_CACHE_LINE_SIZE, NO_VIRT, (u8)ALL_WRITE);
675 	p_setb(perm, PCI_LATENCY_TIMER, NO_VIRT, (u8)ALL_WRITE);
676 	p_setb(perm, PCI_BIST, NO_VIRT, (u8)ALL_WRITE);
677 
678 	/* Virtualize all bars, can't touch the real ones */
679 	p_setd(perm, PCI_BASE_ADDRESS_0, ALL_VIRT, ALL_WRITE);
680 	p_setd(perm, PCI_BASE_ADDRESS_1, ALL_VIRT, ALL_WRITE);
681 	p_setd(perm, PCI_BASE_ADDRESS_2, ALL_VIRT, ALL_WRITE);
682 	p_setd(perm, PCI_BASE_ADDRESS_3, ALL_VIRT, ALL_WRITE);
683 	p_setd(perm, PCI_BASE_ADDRESS_4, ALL_VIRT, ALL_WRITE);
684 	p_setd(perm, PCI_BASE_ADDRESS_5, ALL_VIRT, ALL_WRITE);
685 	p_setd(perm, PCI_ROM_ADDRESS, ALL_VIRT, ALL_WRITE);
686 
687 	/* Allow us to adjust capability chain */
688 	p_setb(perm, PCI_CAPABILITY_LIST, (u8)ALL_VIRT, NO_WRITE);
689 
690 	/* Sometimes used by sw, just virtualize */
691 	p_setb(perm, PCI_INTERRUPT_LINE, (u8)ALL_VIRT, (u8)ALL_WRITE);
692 
693 	/* Virtualize interrupt pin to allow hiding INTx */
694 	p_setb(perm, PCI_INTERRUPT_PIN, (u8)ALL_VIRT, (u8)NO_WRITE);
695 
696 	return 0;
697 }
698 
699 /*
700  * It takes all the required locks to protect the access of power related
701  * variables and then invokes vfio_pci_set_power_state().
702  */
703 static void vfio_lock_and_set_power_state(struct vfio_pci_core_device *vdev,
704 					  pci_power_t state)
705 {
706 	if (state >= PCI_D3hot)
707 		vfio_pci_zap_and_down_write_memory_lock(vdev);
708 	else
709 		down_write(&vdev->memory_lock);
710 
711 	vfio_pci_set_power_state(vdev, state);
712 	up_write(&vdev->memory_lock);
713 }
714 
715 static int vfio_pm_config_write(struct vfio_pci_core_device *vdev, int pos,
716 				int count, struct perm_bits *perm,
717 				int offset, __le32 val)
718 {
719 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
720 	if (count < 0)
721 		return count;
722 
723 	if (offset == PCI_PM_CTRL) {
724 		pci_power_t state;
725 
726 		switch (le32_to_cpu(val) & PCI_PM_CTRL_STATE_MASK) {
727 		case 0:
728 			state = PCI_D0;
729 			break;
730 		case 1:
731 			state = PCI_D1;
732 			break;
733 		case 2:
734 			state = PCI_D2;
735 			break;
736 		case 3:
737 			state = PCI_D3hot;
738 			break;
739 		}
740 
741 		vfio_lock_and_set_power_state(vdev, state);
742 	}
743 
744 	return count;
745 }
746 
747 /* Permissions for the Power Management capability */
748 static int __init init_pci_cap_pm_perm(struct perm_bits *perm)
749 {
750 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_PM]))
751 		return -ENOMEM;
752 
753 	perm->writefn = vfio_pm_config_write;
754 
755 	/*
756 	 * We always virtualize the next field so we can remove
757 	 * capabilities from the chain if we want to.
758 	 */
759 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
760 
761 	/*
762 	 * The guests can't process PME events. If any PME event will be
763 	 * generated, then it will be mostly handled in the host and the
764 	 * host will clear the PME_STATUS. So virtualize PME_Support bits.
765 	 * The vconfig bits will be cleared during device capability
766 	 * initialization.
767 	 */
768 	p_setw(perm, PCI_PM_PMC, PCI_PM_CAP_PME_MASK, NO_WRITE);
769 
770 	/*
771 	 * Power management is defined *per function*, so we can let
772 	 * the user change power state, but we trap and initiate the
773 	 * change ourselves, so the state bits are read-only.
774 	 *
775 	 * The guest can't process PME from D3cold so virtualize PME_Status
776 	 * and PME_En bits. The vconfig bits will be cleared during device
777 	 * capability initialization.
778 	 */
779 	p_setd(perm, PCI_PM_CTRL,
780 	       PCI_PM_CTRL_PME_ENABLE | PCI_PM_CTRL_PME_STATUS,
781 	       ~(PCI_PM_CTRL_PME_ENABLE | PCI_PM_CTRL_PME_STATUS |
782 		 PCI_PM_CTRL_STATE_MASK));
783 
784 	return 0;
785 }
786 
787 static int vfio_vpd_config_write(struct vfio_pci_core_device *vdev, int pos,
788 				 int count, struct perm_bits *perm,
789 				 int offset, __le32 val)
790 {
791 	struct pci_dev *pdev = vdev->pdev;
792 	__le16 *paddr = (__le16 *)(vdev->vconfig + pos - offset + PCI_VPD_ADDR);
793 	__le32 *pdata = (__le32 *)(vdev->vconfig + pos - offset + PCI_VPD_DATA);
794 	u16 addr;
795 	u32 data;
796 
797 	/*
798 	 * Write through to emulation.  If the write includes the upper byte
799 	 * of PCI_VPD_ADDR, then the PCI_VPD_ADDR_F bit is written and we
800 	 * have work to do.
801 	 */
802 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
803 	if (count < 0 || offset > PCI_VPD_ADDR + 1 ||
804 	    offset + count <= PCI_VPD_ADDR + 1)
805 		return count;
806 
807 	addr = le16_to_cpu(*paddr);
808 
809 	if (addr & PCI_VPD_ADDR_F) {
810 		data = le32_to_cpu(*pdata);
811 		if (pci_write_vpd(pdev, addr & ~PCI_VPD_ADDR_F, 4, &data) != 4)
812 			return count;
813 	} else {
814 		data = 0;
815 		if (pci_read_vpd(pdev, addr, 4, &data) < 0)
816 			return count;
817 		*pdata = cpu_to_le32(data);
818 	}
819 
820 	/*
821 	 * Toggle PCI_VPD_ADDR_F in the emulated PCI_VPD_ADDR register to
822 	 * signal completion.  If an error occurs above, we assume that not
823 	 * toggling this bit will induce a driver timeout.
824 	 */
825 	addr ^= PCI_VPD_ADDR_F;
826 	*paddr = cpu_to_le16(addr);
827 
828 	return count;
829 }
830 
831 /* Permissions for Vital Product Data capability */
832 static int __init init_pci_cap_vpd_perm(struct perm_bits *perm)
833 {
834 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_VPD]))
835 		return -ENOMEM;
836 
837 	perm->writefn = vfio_vpd_config_write;
838 
839 	/*
840 	 * We always virtualize the next field so we can remove
841 	 * capabilities from the chain if we want to.
842 	 */
843 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
844 
845 	/*
846 	 * Both the address and data registers are virtualized to
847 	 * enable access through the pci_vpd_read/write functions
848 	 */
849 	p_setw(perm, PCI_VPD_ADDR, (u16)ALL_VIRT, (u16)ALL_WRITE);
850 	p_setd(perm, PCI_VPD_DATA, ALL_VIRT, ALL_WRITE);
851 
852 	return 0;
853 }
854 
855 /* Permissions for PCI-X capability */
856 static int __init init_pci_cap_pcix_perm(struct perm_bits *perm)
857 {
858 	/* Alloc 24, but only 8 are used in v0 */
859 	if (alloc_perm_bits(perm, PCI_CAP_PCIX_SIZEOF_V2))
860 		return -ENOMEM;
861 
862 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
863 
864 	p_setw(perm, PCI_X_CMD, NO_VIRT, (u16)ALL_WRITE);
865 	p_setd(perm, PCI_X_ECC_CSR, NO_VIRT, ALL_WRITE);
866 	return 0;
867 }
868 
869 static int vfio_exp_config_write(struct vfio_pci_core_device *vdev, int pos,
870 				 int count, struct perm_bits *perm,
871 				 int offset, __le32 val)
872 {
873 	__le16 *ctrl = (__le16 *)(vdev->vconfig + pos -
874 				  offset + PCI_EXP_DEVCTL);
875 	int readrq = le16_to_cpu(*ctrl) & PCI_EXP_DEVCTL_READRQ;
876 
877 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
878 	if (count < 0)
879 		return count;
880 
881 	/*
882 	 * The FLR bit is virtualized, if set and the device supports PCIe
883 	 * FLR, issue a reset_function.  Regardless, clear the bit, the spec
884 	 * requires it to be always read as zero.  NB, reset_function might
885 	 * not use a PCIe FLR, we don't have that level of granularity.
886 	 */
887 	if (*ctrl & cpu_to_le16(PCI_EXP_DEVCTL_BCR_FLR)) {
888 		u32 cap;
889 		int ret;
890 
891 		*ctrl &= ~cpu_to_le16(PCI_EXP_DEVCTL_BCR_FLR);
892 
893 		ret = pci_user_read_config_dword(vdev->pdev,
894 						 pos - offset + PCI_EXP_DEVCAP,
895 						 &cap);
896 
897 		if (!ret && (cap & PCI_EXP_DEVCAP_FLR)) {
898 			vfio_pci_zap_and_down_write_memory_lock(vdev);
899 			pci_try_reset_function(vdev->pdev);
900 			up_write(&vdev->memory_lock);
901 		}
902 	}
903 
904 	/*
905 	 * MPS is virtualized to the user, writes do not change the physical
906 	 * register since determining a proper MPS value requires a system wide
907 	 * device view.  The MRRS is largely independent of MPS, but since the
908 	 * user does not have that system-wide view, they might set a safe, but
909 	 * inefficiently low value.  Here we allow writes through to hardware,
910 	 * but we set the floor to the physical device MPS setting, so that
911 	 * we can at least use full TLPs, as defined by the MPS value.
912 	 *
913 	 * NB, if any devices actually depend on an artificially low MRRS
914 	 * setting, this will need to be revisited, perhaps with a quirk
915 	 * though pcie_set_readrq().
916 	 */
917 	if (readrq != (le16_to_cpu(*ctrl) & PCI_EXP_DEVCTL_READRQ)) {
918 		readrq = 128 <<
919 			((le16_to_cpu(*ctrl) & PCI_EXP_DEVCTL_READRQ) >> 12);
920 		readrq = max(readrq, pcie_get_mps(vdev->pdev));
921 
922 		pcie_set_readrq(vdev->pdev, readrq);
923 	}
924 
925 	return count;
926 }
927 
928 /* Permissions for PCI Express capability */
929 static int __init init_pci_cap_exp_perm(struct perm_bits *perm)
930 {
931 	/* Alloc largest of possible sizes */
932 	if (alloc_perm_bits(perm, PCI_CAP_EXP_ENDPOINT_SIZEOF_V2))
933 		return -ENOMEM;
934 
935 	perm->writefn = vfio_exp_config_write;
936 
937 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
938 
939 	/*
940 	 * Allow writes to device control fields, except devctl_phantom,
941 	 * which could confuse IOMMU, MPS, which can break communication
942 	 * with other physical devices, and the ARI bit in devctl2, which
943 	 * is set at probe time.  FLR and MRRS get virtualized via our
944 	 * writefn.
945 	 */
946 	p_setw(perm, PCI_EXP_DEVCTL,
947 	       PCI_EXP_DEVCTL_BCR_FLR | PCI_EXP_DEVCTL_PAYLOAD |
948 	       PCI_EXP_DEVCTL_READRQ, ~PCI_EXP_DEVCTL_PHANTOM);
949 	p_setw(perm, PCI_EXP_DEVCTL2, NO_VIRT, ~PCI_EXP_DEVCTL2_ARI);
950 	return 0;
951 }
952 
953 static int vfio_af_config_write(struct vfio_pci_core_device *vdev, int pos,
954 				int count, struct perm_bits *perm,
955 				int offset, __le32 val)
956 {
957 	u8 *ctrl = vdev->vconfig + pos - offset + PCI_AF_CTRL;
958 
959 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
960 	if (count < 0)
961 		return count;
962 
963 	/*
964 	 * The FLR bit is virtualized, if set and the device supports AF
965 	 * FLR, issue a reset_function.  Regardless, clear the bit, the spec
966 	 * requires it to be always read as zero.  NB, reset_function might
967 	 * not use an AF FLR, we don't have that level of granularity.
968 	 */
969 	if (*ctrl & PCI_AF_CTRL_FLR) {
970 		u8 cap;
971 		int ret;
972 
973 		*ctrl &= ~PCI_AF_CTRL_FLR;
974 
975 		ret = pci_user_read_config_byte(vdev->pdev,
976 						pos - offset + PCI_AF_CAP,
977 						&cap);
978 
979 		if (!ret && (cap & PCI_AF_CAP_FLR) && (cap & PCI_AF_CAP_TP)) {
980 			vfio_pci_zap_and_down_write_memory_lock(vdev);
981 			pci_try_reset_function(vdev->pdev);
982 			up_write(&vdev->memory_lock);
983 		}
984 	}
985 
986 	return count;
987 }
988 
989 /* Permissions for Advanced Function capability */
990 static int __init init_pci_cap_af_perm(struct perm_bits *perm)
991 {
992 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_AF]))
993 		return -ENOMEM;
994 
995 	perm->writefn = vfio_af_config_write;
996 
997 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
998 	p_setb(perm, PCI_AF_CTRL, PCI_AF_CTRL_FLR, PCI_AF_CTRL_FLR);
999 	return 0;
1000 }
1001 
1002 /* Permissions for Advanced Error Reporting extended capability */
1003 static int __init init_pci_ext_cap_err_perm(struct perm_bits *perm)
1004 {
1005 	u32 mask;
1006 
1007 	if (alloc_perm_bits(perm, pci_ext_cap_length[PCI_EXT_CAP_ID_ERR]))
1008 		return -ENOMEM;
1009 
1010 	/*
1011 	 * Virtualize the first dword of all express capabilities
1012 	 * because it includes the next pointer.  This lets us later
1013 	 * remove capabilities from the chain if we need to.
1014 	 */
1015 	p_setd(perm, 0, ALL_VIRT, NO_WRITE);
1016 
1017 	/* Writable bits mask */
1018 	mask =	PCI_ERR_UNC_UND |		/* Undefined */
1019 		PCI_ERR_UNC_DLP |		/* Data Link Protocol */
1020 		PCI_ERR_UNC_SURPDN |		/* Surprise Down */
1021 		PCI_ERR_UNC_POISON_TLP |	/* Poisoned TLP */
1022 		PCI_ERR_UNC_FCP |		/* Flow Control Protocol */
1023 		PCI_ERR_UNC_COMP_TIME |		/* Completion Timeout */
1024 		PCI_ERR_UNC_COMP_ABORT |	/* Completer Abort */
1025 		PCI_ERR_UNC_UNX_COMP |		/* Unexpected Completion */
1026 		PCI_ERR_UNC_RX_OVER |		/* Receiver Overflow */
1027 		PCI_ERR_UNC_MALF_TLP |		/* Malformed TLP */
1028 		PCI_ERR_UNC_ECRC |		/* ECRC Error Status */
1029 		PCI_ERR_UNC_UNSUP |		/* Unsupported Request */
1030 		PCI_ERR_UNC_ACSV |		/* ACS Violation */
1031 		PCI_ERR_UNC_INTN |		/* internal error */
1032 		PCI_ERR_UNC_MCBTLP |		/* MC blocked TLP */
1033 		PCI_ERR_UNC_ATOMEG |		/* Atomic egress blocked */
1034 		PCI_ERR_UNC_TLPPRE;		/* TLP prefix blocked */
1035 	p_setd(perm, PCI_ERR_UNCOR_STATUS, NO_VIRT, mask);
1036 	p_setd(perm, PCI_ERR_UNCOR_MASK, NO_VIRT, mask);
1037 	p_setd(perm, PCI_ERR_UNCOR_SEVER, NO_VIRT, mask);
1038 
1039 	mask =	PCI_ERR_COR_RCVR |		/* Receiver Error Status */
1040 		PCI_ERR_COR_BAD_TLP |		/* Bad TLP Status */
1041 		PCI_ERR_COR_BAD_DLLP |		/* Bad DLLP Status */
1042 		PCI_ERR_COR_REP_ROLL |		/* REPLAY_NUM Rollover */
1043 		PCI_ERR_COR_REP_TIMER |		/* Replay Timer Timeout */
1044 		PCI_ERR_COR_ADV_NFAT |		/* Advisory Non-Fatal */
1045 		PCI_ERR_COR_INTERNAL |		/* Corrected Internal */
1046 		PCI_ERR_COR_LOG_OVER;		/* Header Log Overflow */
1047 	p_setd(perm, PCI_ERR_COR_STATUS, NO_VIRT, mask);
1048 	p_setd(perm, PCI_ERR_COR_MASK, NO_VIRT, mask);
1049 
1050 	mask =	PCI_ERR_CAP_ECRC_GENE |		/* ECRC Generation Enable */
1051 		PCI_ERR_CAP_ECRC_CHKE;		/* ECRC Check Enable */
1052 	p_setd(perm, PCI_ERR_CAP, NO_VIRT, mask);
1053 	return 0;
1054 }
1055 
1056 /* Permissions for Power Budgeting extended capability */
1057 static int __init init_pci_ext_cap_pwr_perm(struct perm_bits *perm)
1058 {
1059 	if (alloc_perm_bits(perm, pci_ext_cap_length[PCI_EXT_CAP_ID_PWR]))
1060 		return -ENOMEM;
1061 
1062 	p_setd(perm, 0, ALL_VIRT, NO_WRITE);
1063 
1064 	/* Writing the data selector is OK, the info is still read-only */
1065 	p_setb(perm, PCI_PWR_DATA, NO_VIRT, (u8)ALL_WRITE);
1066 	return 0;
1067 }
1068 
1069 /*
1070  * Initialize the shared permission tables
1071  */
1072 void vfio_pci_uninit_perm_bits(void)
1073 {
1074 	free_perm_bits(&cap_perms[PCI_CAP_ID_BASIC]);
1075 
1076 	free_perm_bits(&cap_perms[PCI_CAP_ID_PM]);
1077 	free_perm_bits(&cap_perms[PCI_CAP_ID_VPD]);
1078 	free_perm_bits(&cap_perms[PCI_CAP_ID_PCIX]);
1079 	free_perm_bits(&cap_perms[PCI_CAP_ID_EXP]);
1080 	free_perm_bits(&cap_perms[PCI_CAP_ID_AF]);
1081 
1082 	free_perm_bits(&ecap_perms[PCI_EXT_CAP_ID_ERR]);
1083 	free_perm_bits(&ecap_perms[PCI_EXT_CAP_ID_PWR]);
1084 }
1085 
1086 int __init vfio_pci_init_perm_bits(void)
1087 {
1088 	int ret;
1089 
1090 	/* Basic config space */
1091 	ret = init_pci_cap_basic_perm(&cap_perms[PCI_CAP_ID_BASIC]);
1092 
1093 	/* Capabilities */
1094 	ret |= init_pci_cap_pm_perm(&cap_perms[PCI_CAP_ID_PM]);
1095 	ret |= init_pci_cap_vpd_perm(&cap_perms[PCI_CAP_ID_VPD]);
1096 	ret |= init_pci_cap_pcix_perm(&cap_perms[PCI_CAP_ID_PCIX]);
1097 	cap_perms[PCI_CAP_ID_VNDR].writefn = vfio_raw_config_write;
1098 	ret |= init_pci_cap_exp_perm(&cap_perms[PCI_CAP_ID_EXP]);
1099 	ret |= init_pci_cap_af_perm(&cap_perms[PCI_CAP_ID_AF]);
1100 
1101 	/* Extended capabilities */
1102 	ret |= init_pci_ext_cap_err_perm(&ecap_perms[PCI_EXT_CAP_ID_ERR]);
1103 	ret |= init_pci_ext_cap_pwr_perm(&ecap_perms[PCI_EXT_CAP_ID_PWR]);
1104 	ecap_perms[PCI_EXT_CAP_ID_VNDR].writefn = vfio_raw_config_write;
1105 	ecap_perms[PCI_EXT_CAP_ID_DVSEC].writefn = vfio_raw_config_write;
1106 
1107 	if (ret)
1108 		vfio_pci_uninit_perm_bits();
1109 
1110 	return ret;
1111 }
1112 
1113 static int vfio_find_cap_start(struct vfio_pci_core_device *vdev, int pos)
1114 {
1115 	u8 cap;
1116 	int base = (pos >= PCI_CFG_SPACE_SIZE) ? PCI_CFG_SPACE_SIZE :
1117 						 PCI_STD_HEADER_SIZEOF;
1118 	cap = vdev->pci_config_map[pos];
1119 
1120 	if (cap == PCI_CAP_ID_BASIC)
1121 		return 0;
1122 
1123 	/* XXX Can we have to abutting capabilities of the same type? */
1124 	while (pos - 1 >= base && vdev->pci_config_map[pos - 1] == cap)
1125 		pos--;
1126 
1127 	return pos;
1128 }
1129 
1130 static int vfio_msi_config_read(struct vfio_pci_core_device *vdev, int pos,
1131 				int count, struct perm_bits *perm,
1132 				int offset, __le32 *val)
1133 {
1134 	/* Update max available queue size from msi_qmax */
1135 	if (offset <= PCI_MSI_FLAGS && offset + count >= PCI_MSI_FLAGS) {
1136 		__le16 *flags;
1137 		int start;
1138 
1139 		start = vfio_find_cap_start(vdev, pos);
1140 
1141 		flags = (__le16 *)&vdev->vconfig[start];
1142 
1143 		*flags &= cpu_to_le16(~PCI_MSI_FLAGS_QMASK);
1144 		*flags |= cpu_to_le16(vdev->msi_qmax << 1);
1145 	}
1146 
1147 	return vfio_default_config_read(vdev, pos, count, perm, offset, val);
1148 }
1149 
1150 static int vfio_msi_config_write(struct vfio_pci_core_device *vdev, int pos,
1151 				 int count, struct perm_bits *perm,
1152 				 int offset, __le32 val)
1153 {
1154 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
1155 	if (count < 0)
1156 		return count;
1157 
1158 	/* Fixup and write configured queue size and enable to hardware */
1159 	if (offset <= PCI_MSI_FLAGS && offset + count >= PCI_MSI_FLAGS) {
1160 		__le16 *pflags;
1161 		u16 flags;
1162 		int start, ret;
1163 
1164 		start = vfio_find_cap_start(vdev, pos);
1165 
1166 		pflags = (__le16 *)&vdev->vconfig[start + PCI_MSI_FLAGS];
1167 
1168 		flags = le16_to_cpu(*pflags);
1169 
1170 		/* MSI is enabled via ioctl */
1171 		if  (vdev->irq_type != VFIO_PCI_MSI_IRQ_INDEX)
1172 			flags &= ~PCI_MSI_FLAGS_ENABLE;
1173 
1174 		/* Check queue size */
1175 		if ((flags & PCI_MSI_FLAGS_QSIZE) >> 4 > vdev->msi_qmax) {
1176 			flags &= ~PCI_MSI_FLAGS_QSIZE;
1177 			flags |= vdev->msi_qmax << 4;
1178 		}
1179 
1180 		/* Write back to virt and to hardware */
1181 		*pflags = cpu_to_le16(flags);
1182 		ret = pci_user_write_config_word(vdev->pdev,
1183 						 start + PCI_MSI_FLAGS,
1184 						 flags);
1185 		if (ret)
1186 			return ret;
1187 	}
1188 
1189 	return count;
1190 }
1191 
1192 /*
1193  * MSI determination is per-device, so this routine gets used beyond
1194  * initialization time. Don't add __init
1195  */
1196 static int init_pci_cap_msi_perm(struct perm_bits *perm, int len, u16 flags)
1197 {
1198 	if (alloc_perm_bits(perm, len))
1199 		return -ENOMEM;
1200 
1201 	perm->readfn = vfio_msi_config_read;
1202 	perm->writefn = vfio_msi_config_write;
1203 
1204 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
1205 
1206 	/*
1207 	 * The upper byte of the control register is reserved,
1208 	 * just setup the lower byte.
1209 	 */
1210 	p_setb(perm, PCI_MSI_FLAGS, (u8)ALL_VIRT, (u8)ALL_WRITE);
1211 	p_setd(perm, PCI_MSI_ADDRESS_LO, ALL_VIRT, ALL_WRITE);
1212 	if (flags & PCI_MSI_FLAGS_64BIT) {
1213 		p_setd(perm, PCI_MSI_ADDRESS_HI, ALL_VIRT, ALL_WRITE);
1214 		p_setw(perm, PCI_MSI_DATA_64, (u16)ALL_VIRT, (u16)ALL_WRITE);
1215 		if (flags & PCI_MSI_FLAGS_MASKBIT) {
1216 			p_setd(perm, PCI_MSI_MASK_64, NO_VIRT, ALL_WRITE);
1217 			p_setd(perm, PCI_MSI_PENDING_64, NO_VIRT, ALL_WRITE);
1218 		}
1219 	} else {
1220 		p_setw(perm, PCI_MSI_DATA_32, (u16)ALL_VIRT, (u16)ALL_WRITE);
1221 		if (flags & PCI_MSI_FLAGS_MASKBIT) {
1222 			p_setd(perm, PCI_MSI_MASK_32, NO_VIRT, ALL_WRITE);
1223 			p_setd(perm, PCI_MSI_PENDING_32, NO_VIRT, ALL_WRITE);
1224 		}
1225 	}
1226 	return 0;
1227 }
1228 
1229 /* Determine MSI CAP field length; initialize msi_perms on 1st call per vdev */
1230 static int vfio_msi_cap_len(struct vfio_pci_core_device *vdev, u8 pos)
1231 {
1232 	struct pci_dev *pdev = vdev->pdev;
1233 	int len, ret;
1234 	u16 flags;
1235 
1236 	ret = pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &flags);
1237 	if (ret)
1238 		return pcibios_err_to_errno(ret);
1239 
1240 	len = 10; /* Minimum size */
1241 	if (flags & PCI_MSI_FLAGS_64BIT)
1242 		len += 4;
1243 	if (flags & PCI_MSI_FLAGS_MASKBIT)
1244 		len += 10;
1245 
1246 	if (vdev->msi_perm)
1247 		return len;
1248 
1249 	vdev->msi_perm = kmalloc(sizeof(struct perm_bits), GFP_KERNEL_ACCOUNT);
1250 	if (!vdev->msi_perm)
1251 		return -ENOMEM;
1252 
1253 	ret = init_pci_cap_msi_perm(vdev->msi_perm, len, flags);
1254 	if (ret) {
1255 		kfree(vdev->msi_perm);
1256 		return ret;
1257 	}
1258 
1259 	return len;
1260 }
1261 
1262 /* Determine extended capability length for VC (2 & 9) and MFVC */
1263 static int vfio_vc_cap_len(struct vfio_pci_core_device *vdev, u16 pos)
1264 {
1265 	struct pci_dev *pdev = vdev->pdev;
1266 	u32 tmp;
1267 	int ret, evcc, phases, vc_arb;
1268 	int len = PCI_CAP_VC_BASE_SIZEOF;
1269 
1270 	ret = pci_read_config_dword(pdev, pos + PCI_VC_PORT_CAP1, &tmp);
1271 	if (ret)
1272 		return pcibios_err_to_errno(ret);
1273 
1274 	evcc = tmp & PCI_VC_CAP1_EVCC; /* extended vc count */
1275 	ret = pci_read_config_dword(pdev, pos + PCI_VC_PORT_CAP2, &tmp);
1276 	if (ret)
1277 		return pcibios_err_to_errno(ret);
1278 
1279 	if (tmp & PCI_VC_CAP2_128_PHASE)
1280 		phases = 128;
1281 	else if (tmp & PCI_VC_CAP2_64_PHASE)
1282 		phases = 64;
1283 	else if (tmp & PCI_VC_CAP2_32_PHASE)
1284 		phases = 32;
1285 	else
1286 		phases = 0;
1287 
1288 	vc_arb = phases * 4;
1289 
1290 	/*
1291 	 * Port arbitration tables are root & switch only;
1292 	 * function arbitration tables are function 0 only.
1293 	 * In either case, we'll never let user write them so
1294 	 * we don't care how big they are
1295 	 */
1296 	len += (1 + evcc) * PCI_CAP_VC_PER_VC_SIZEOF;
1297 	if (vc_arb) {
1298 		len = round_up(len, 16);
1299 		len += vc_arb / 8;
1300 	}
1301 	return len;
1302 }
1303 
1304 static int vfio_cap_len(struct vfio_pci_core_device *vdev, u8 cap, u8 pos)
1305 {
1306 	struct pci_dev *pdev = vdev->pdev;
1307 	u32 dword;
1308 	u16 word;
1309 	u8 byte;
1310 	int ret;
1311 
1312 	switch (cap) {
1313 	case PCI_CAP_ID_MSI:
1314 		return vfio_msi_cap_len(vdev, pos);
1315 	case PCI_CAP_ID_PCIX:
1316 		ret = pci_read_config_word(pdev, pos + PCI_X_CMD, &word);
1317 		if (ret)
1318 			return pcibios_err_to_errno(ret);
1319 
1320 		if (PCI_X_CMD_VERSION(word)) {
1321 			if (pdev->cfg_size > PCI_CFG_SPACE_SIZE) {
1322 				/* Test for extended capabilities */
1323 				pci_read_config_dword(pdev, PCI_CFG_SPACE_SIZE,
1324 						      &dword);
1325 				vdev->extended_caps = (dword != 0);
1326 			}
1327 			return PCI_CAP_PCIX_SIZEOF_V2;
1328 		} else
1329 			return PCI_CAP_PCIX_SIZEOF_V0;
1330 	case PCI_CAP_ID_VNDR:
1331 		/* length follows next field */
1332 		ret = pci_read_config_byte(pdev, pos + PCI_CAP_FLAGS, &byte);
1333 		if (ret)
1334 			return pcibios_err_to_errno(ret);
1335 
1336 		return byte;
1337 	case PCI_CAP_ID_EXP:
1338 		if (pdev->cfg_size > PCI_CFG_SPACE_SIZE) {
1339 			/* Test for extended capabilities */
1340 			pci_read_config_dword(pdev, PCI_CFG_SPACE_SIZE, &dword);
1341 			vdev->extended_caps = (dword != 0);
1342 		}
1343 
1344 		/* length based on version and type */
1345 		if ((pcie_caps_reg(pdev) & PCI_EXP_FLAGS_VERS) == 1) {
1346 			if (pci_pcie_type(pdev) == PCI_EXP_TYPE_RC_END)
1347 				return 0xc; /* "All Devices" only, no link */
1348 			return PCI_CAP_EXP_ENDPOINT_SIZEOF_V1;
1349 		} else {
1350 			if (pci_pcie_type(pdev) == PCI_EXP_TYPE_RC_END)
1351 				return 0x2c; /* No link */
1352 			return PCI_CAP_EXP_ENDPOINT_SIZEOF_V2;
1353 		}
1354 	case PCI_CAP_ID_HT:
1355 		ret = pci_read_config_byte(pdev, pos + 3, &byte);
1356 		if (ret)
1357 			return pcibios_err_to_errno(ret);
1358 
1359 		return (byte & HT_3BIT_CAP_MASK) ?
1360 			HT_CAP_SIZEOF_SHORT : HT_CAP_SIZEOF_LONG;
1361 	case PCI_CAP_ID_SATA:
1362 		ret = pci_read_config_byte(pdev, pos + PCI_SATA_REGS, &byte);
1363 		if (ret)
1364 			return pcibios_err_to_errno(ret);
1365 
1366 		byte &= PCI_SATA_REGS_MASK;
1367 		if (byte == PCI_SATA_REGS_INLINE)
1368 			return PCI_SATA_SIZEOF_LONG;
1369 		else
1370 			return PCI_SATA_SIZEOF_SHORT;
1371 	default:
1372 		pci_warn(pdev, "%s: unknown length for PCI cap %#x@%#x\n",
1373 			 __func__, cap, pos);
1374 	}
1375 
1376 	return 0;
1377 }
1378 
1379 static int vfio_ext_cap_len(struct vfio_pci_core_device *vdev, u16 ecap, u16 epos)
1380 {
1381 	struct pci_dev *pdev = vdev->pdev;
1382 	u8 byte;
1383 	u32 dword;
1384 	int ret;
1385 
1386 	switch (ecap) {
1387 	case PCI_EXT_CAP_ID_VNDR:
1388 		ret = pci_read_config_dword(pdev, epos + PCI_VSEC_HDR, &dword);
1389 		if (ret)
1390 			return pcibios_err_to_errno(ret);
1391 
1392 		return dword >> PCI_VSEC_HDR_LEN_SHIFT;
1393 	case PCI_EXT_CAP_ID_VC:
1394 	case PCI_EXT_CAP_ID_VC9:
1395 	case PCI_EXT_CAP_ID_MFVC:
1396 		return vfio_vc_cap_len(vdev, epos);
1397 	case PCI_EXT_CAP_ID_ACS:
1398 		ret = pci_read_config_byte(pdev, epos + PCI_ACS_CAP, &byte);
1399 		if (ret)
1400 			return pcibios_err_to_errno(ret);
1401 
1402 		if (byte & PCI_ACS_EC) {
1403 			int bits;
1404 
1405 			ret = pci_read_config_byte(pdev,
1406 						   epos + PCI_ACS_EGRESS_BITS,
1407 						   &byte);
1408 			if (ret)
1409 				return pcibios_err_to_errno(ret);
1410 
1411 			bits = byte ? round_up(byte, 32) : 256;
1412 			return 8 + (bits / 8);
1413 		}
1414 		return 8;
1415 
1416 	case PCI_EXT_CAP_ID_REBAR:
1417 		ret = pci_read_config_byte(pdev, epos + PCI_REBAR_CTRL, &byte);
1418 		if (ret)
1419 			return pcibios_err_to_errno(ret);
1420 
1421 		byte &= PCI_REBAR_CTRL_NBAR_MASK;
1422 		byte >>= PCI_REBAR_CTRL_NBAR_SHIFT;
1423 
1424 		return 4 + (byte * 8);
1425 	case PCI_EXT_CAP_ID_DPA:
1426 		ret = pci_read_config_byte(pdev, epos + PCI_DPA_CAP, &byte);
1427 		if (ret)
1428 			return pcibios_err_to_errno(ret);
1429 
1430 		byte &= PCI_DPA_CAP_SUBSTATE_MASK;
1431 		return PCI_DPA_BASE_SIZEOF + byte + 1;
1432 	case PCI_EXT_CAP_ID_TPH:
1433 		ret = pci_read_config_dword(pdev, epos + PCI_TPH_CAP, &dword);
1434 		if (ret)
1435 			return pcibios_err_to_errno(ret);
1436 
1437 		if ((dword & PCI_TPH_CAP_LOC_MASK) == PCI_TPH_LOC_CAP) {
1438 			int sts;
1439 
1440 			sts = dword & PCI_TPH_CAP_ST_MASK;
1441 			sts >>= PCI_TPH_CAP_ST_SHIFT;
1442 			return PCI_TPH_BASE_SIZEOF + (sts * 2) + 2;
1443 		}
1444 		return PCI_TPH_BASE_SIZEOF;
1445 	case PCI_EXT_CAP_ID_DVSEC:
1446 		ret = pci_read_config_dword(pdev, epos + PCI_DVSEC_HEADER1, &dword);
1447 		if (ret)
1448 			return pcibios_err_to_errno(ret);
1449 		return PCI_DVSEC_HEADER1_LEN(dword);
1450 	default:
1451 		pci_warn(pdev, "%s: unknown length for PCI ecap %#x@%#x\n",
1452 			 __func__, ecap, epos);
1453 	}
1454 
1455 	return 0;
1456 }
1457 
1458 static void vfio_update_pm_vconfig_bytes(struct vfio_pci_core_device *vdev,
1459 					 int offset)
1460 {
1461 	__le16 *pmc = (__le16 *)&vdev->vconfig[offset + PCI_PM_PMC];
1462 	__le16 *ctrl = (__le16 *)&vdev->vconfig[offset + PCI_PM_CTRL];
1463 
1464 	/* Clear vconfig PME_Support, PME_Status, and PME_En bits */
1465 	*pmc &= ~cpu_to_le16(PCI_PM_CAP_PME_MASK);
1466 	*ctrl &= ~cpu_to_le16(PCI_PM_CTRL_PME_ENABLE | PCI_PM_CTRL_PME_STATUS);
1467 }
1468 
1469 static int vfio_fill_vconfig_bytes(struct vfio_pci_core_device *vdev,
1470 				   int offset, int size)
1471 {
1472 	struct pci_dev *pdev = vdev->pdev;
1473 	int ret = 0;
1474 
1475 	/*
1476 	 * We try to read physical config space in the largest chunks
1477 	 * we can, assuming that all of the fields support dword access.
1478 	 * pci_save_state() makes this same assumption and seems to do ok.
1479 	 */
1480 	while (size) {
1481 		int filled;
1482 
1483 		if (size >= 4 && !(offset % 4)) {
1484 			__le32 *dwordp = (__le32 *)&vdev->vconfig[offset];
1485 			u32 dword;
1486 
1487 			ret = pci_read_config_dword(pdev, offset, &dword);
1488 			if (ret)
1489 				return ret;
1490 			*dwordp = cpu_to_le32(dword);
1491 			filled = 4;
1492 		} else if (size >= 2 && !(offset % 2)) {
1493 			__le16 *wordp = (__le16 *)&vdev->vconfig[offset];
1494 			u16 word;
1495 
1496 			ret = pci_read_config_word(pdev, offset, &word);
1497 			if (ret)
1498 				return ret;
1499 			*wordp = cpu_to_le16(word);
1500 			filled = 2;
1501 		} else {
1502 			u8 *byte = &vdev->vconfig[offset];
1503 			ret = pci_read_config_byte(pdev, offset, byte);
1504 			if (ret)
1505 				return ret;
1506 			filled = 1;
1507 		}
1508 
1509 		offset += filled;
1510 		size -= filled;
1511 	}
1512 
1513 	return ret;
1514 }
1515 
1516 static int vfio_cap_init(struct vfio_pci_core_device *vdev)
1517 {
1518 	struct pci_dev *pdev = vdev->pdev;
1519 	u8 *map = vdev->pci_config_map;
1520 	u16 status;
1521 	u8 pos, *prev, cap;
1522 	int loops, ret, caps = 0;
1523 
1524 	/* Any capabilities? */
1525 	ret = pci_read_config_word(pdev, PCI_STATUS, &status);
1526 	if (ret)
1527 		return ret;
1528 
1529 	if (!(status & PCI_STATUS_CAP_LIST))
1530 		return 0; /* Done */
1531 
1532 	ret = pci_read_config_byte(pdev, PCI_CAPABILITY_LIST, &pos);
1533 	if (ret)
1534 		return ret;
1535 
1536 	/* Mark the previous position in case we want to skip a capability */
1537 	prev = &vdev->vconfig[PCI_CAPABILITY_LIST];
1538 
1539 	/* We can bound our loop, capabilities are dword aligned */
1540 	loops = (PCI_CFG_SPACE_SIZE - PCI_STD_HEADER_SIZEOF) / PCI_CAP_SIZEOF;
1541 	while (pos && loops--) {
1542 		u8 next;
1543 		int i, len = 0;
1544 
1545 		ret = pci_read_config_byte(pdev, pos, &cap);
1546 		if (ret)
1547 			return ret;
1548 
1549 		ret = pci_read_config_byte(pdev,
1550 					   pos + PCI_CAP_LIST_NEXT, &next);
1551 		if (ret)
1552 			return ret;
1553 
1554 		/*
1555 		 * ID 0 is a NULL capability, conflicting with our fake
1556 		 * PCI_CAP_ID_BASIC.  As it has no content, consider it
1557 		 * hidden for now.
1558 		 */
1559 		if (cap && cap <= PCI_CAP_ID_MAX) {
1560 			len = pci_cap_length[cap];
1561 			if (len == 0xFF) { /* Variable length */
1562 				len = vfio_cap_len(vdev, cap, pos);
1563 				if (len < 0)
1564 					return len;
1565 			}
1566 		}
1567 
1568 		if (!len) {
1569 			pci_info(pdev, "%s: hiding cap %#x@%#x\n", __func__,
1570 				 cap, pos);
1571 			*prev = next;
1572 			pos = next;
1573 			continue;
1574 		}
1575 
1576 		/* Sanity check, do we overlap other capabilities? */
1577 		for (i = 0; i < len; i++) {
1578 			if (likely(map[pos + i] == PCI_CAP_ID_INVALID))
1579 				continue;
1580 
1581 			pci_warn(pdev, "%s: PCI config conflict @%#x, was cap %#x now cap %#x\n",
1582 				 __func__, pos + i, map[pos + i], cap);
1583 		}
1584 
1585 		BUILD_BUG_ON(PCI_CAP_ID_MAX >= PCI_CAP_ID_INVALID_VIRT);
1586 
1587 		memset(map + pos, cap, len);
1588 		ret = vfio_fill_vconfig_bytes(vdev, pos, len);
1589 		if (ret)
1590 			return ret;
1591 
1592 		if (cap == PCI_CAP_ID_PM)
1593 			vfio_update_pm_vconfig_bytes(vdev, pos);
1594 
1595 		prev = &vdev->vconfig[pos + PCI_CAP_LIST_NEXT];
1596 		pos = next;
1597 		caps++;
1598 	}
1599 
1600 	/* If we didn't fill any capabilities, clear the status flag */
1601 	if (!caps) {
1602 		__le16 *vstatus = (__le16 *)&vdev->vconfig[PCI_STATUS];
1603 		*vstatus &= ~cpu_to_le16(PCI_STATUS_CAP_LIST);
1604 	}
1605 
1606 	return 0;
1607 }
1608 
1609 static int vfio_ecap_init(struct vfio_pci_core_device *vdev)
1610 {
1611 	struct pci_dev *pdev = vdev->pdev;
1612 	u8 *map = vdev->pci_config_map;
1613 	u16 epos;
1614 	__le32 *prev = NULL;
1615 	int loops, ret, ecaps = 0;
1616 
1617 	if (!vdev->extended_caps)
1618 		return 0;
1619 
1620 	epos = PCI_CFG_SPACE_SIZE;
1621 
1622 	loops = (pdev->cfg_size - PCI_CFG_SPACE_SIZE) / PCI_CAP_SIZEOF;
1623 
1624 	while (loops-- && epos >= PCI_CFG_SPACE_SIZE) {
1625 		u32 header;
1626 		u16 ecap;
1627 		int i, len = 0;
1628 		bool hidden = false;
1629 
1630 		ret = pci_read_config_dword(pdev, epos, &header);
1631 		if (ret)
1632 			return ret;
1633 
1634 		ecap = PCI_EXT_CAP_ID(header);
1635 
1636 		if (ecap <= PCI_EXT_CAP_ID_MAX) {
1637 			len = pci_ext_cap_length[ecap];
1638 			if (len == 0xFF) {
1639 				len = vfio_ext_cap_len(vdev, ecap, epos);
1640 				if (len < 0)
1641 					return len;
1642 			}
1643 		}
1644 
1645 		if (!len) {
1646 			pci_info(pdev, "%s: hiding ecap %#x@%#x\n",
1647 				 __func__, ecap, epos);
1648 
1649 			/* If not the first in the chain, we can skip over it */
1650 			if (prev) {
1651 				u32 val = epos = PCI_EXT_CAP_NEXT(header);
1652 				*prev &= cpu_to_le32(~(0xffcU << 20));
1653 				*prev |= cpu_to_le32(val << 20);
1654 				continue;
1655 			}
1656 
1657 			/*
1658 			 * Otherwise, fill in a placeholder, the direct
1659 			 * readfn will virtualize this automatically
1660 			 */
1661 			len = PCI_CAP_SIZEOF;
1662 			hidden = true;
1663 		}
1664 
1665 		for (i = 0; i < len; i++) {
1666 			if (likely(map[epos + i] == PCI_CAP_ID_INVALID))
1667 				continue;
1668 
1669 			pci_warn(pdev, "%s: PCI config conflict @%#x, was ecap %#x now ecap %#x\n",
1670 				 __func__, epos + i, map[epos + i], ecap);
1671 		}
1672 
1673 		/*
1674 		 * Even though ecap is 2 bytes, we're currently a long way
1675 		 * from exceeding 1 byte capabilities.  If we ever make it
1676 		 * up to 0xFE we'll need to up this to a two-byte, byte map.
1677 		 */
1678 		BUILD_BUG_ON(PCI_EXT_CAP_ID_MAX >= PCI_CAP_ID_INVALID_VIRT);
1679 
1680 		memset(map + epos, ecap, len);
1681 		ret = vfio_fill_vconfig_bytes(vdev, epos, len);
1682 		if (ret)
1683 			return ret;
1684 
1685 		/*
1686 		 * If we're just using this capability to anchor the list,
1687 		 * hide the real ID.  Only count real ecaps.  XXX PCI spec
1688 		 * indicates to use cap id = 0, version = 0, next = 0 if
1689 		 * ecaps are absent, hope users check all the way to next.
1690 		 */
1691 		if (hidden)
1692 			*(__le32 *)&vdev->vconfig[epos] &=
1693 				cpu_to_le32((0xffcU << 20));
1694 		else
1695 			ecaps++;
1696 
1697 		prev = (__le32 *)&vdev->vconfig[epos];
1698 		epos = PCI_EXT_CAP_NEXT(header);
1699 	}
1700 
1701 	if (!ecaps)
1702 		*(u32 *)&vdev->vconfig[PCI_CFG_SPACE_SIZE] = 0;
1703 
1704 	return 0;
1705 }
1706 
1707 /*
1708  * Nag about hardware bugs, hopefully to have vendors fix them, but at least
1709  * to collect a list of dependencies for the VF INTx pin quirk below.
1710  */
1711 static const struct pci_device_id known_bogus_vf_intx_pin[] = {
1712 	{ PCI_DEVICE(PCI_VENDOR_ID_INTEL, 0x270c) },
1713 	{}
1714 };
1715 
1716 /*
1717  * For each device we allocate a pci_config_map that indicates the
1718  * capability occupying each dword and thus the struct perm_bits we
1719  * use for read and write.  We also allocate a virtualized config
1720  * space which tracks reads and writes to bits that we emulate for
1721  * the user.  Initial values filled from device.
1722  *
1723  * Using shared struct perm_bits between all vfio-pci devices saves
1724  * us from allocating cfg_size buffers for virt and write for every
1725  * device.  We could remove vconfig and allocate individual buffers
1726  * for each area requiring emulated bits, but the array of pointers
1727  * would be comparable in size (at least for standard config space).
1728  */
1729 int vfio_config_init(struct vfio_pci_core_device *vdev)
1730 {
1731 	struct pci_dev *pdev = vdev->pdev;
1732 	u8 *map, *vconfig;
1733 	int ret;
1734 
1735 	/*
1736 	 * Config space, caps and ecaps are all dword aligned, so we could
1737 	 * use one byte per dword to record the type.  However, there are
1738 	 * no requirements on the length of a capability, so the gap between
1739 	 * capabilities needs byte granularity.
1740 	 */
1741 	map = kmalloc(pdev->cfg_size, GFP_KERNEL_ACCOUNT);
1742 	if (!map)
1743 		return -ENOMEM;
1744 
1745 	vconfig = kmalloc(pdev->cfg_size, GFP_KERNEL_ACCOUNT);
1746 	if (!vconfig) {
1747 		kfree(map);
1748 		return -ENOMEM;
1749 	}
1750 
1751 	vdev->pci_config_map = map;
1752 	vdev->vconfig = vconfig;
1753 
1754 	memset(map, PCI_CAP_ID_BASIC, PCI_STD_HEADER_SIZEOF);
1755 	memset(map + PCI_STD_HEADER_SIZEOF, PCI_CAP_ID_INVALID,
1756 	       pdev->cfg_size - PCI_STD_HEADER_SIZEOF);
1757 
1758 	ret = vfio_fill_vconfig_bytes(vdev, 0, PCI_STD_HEADER_SIZEOF);
1759 	if (ret)
1760 		goto out;
1761 
1762 	vdev->bardirty = true;
1763 
1764 	/*
1765 	 * XXX can we just pci_load_saved_state/pci_restore_state?
1766 	 * may need to rebuild vconfig after that
1767 	 */
1768 
1769 	/* For restore after reset */
1770 	vdev->rbar[0] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_0]);
1771 	vdev->rbar[1] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_1]);
1772 	vdev->rbar[2] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_2]);
1773 	vdev->rbar[3] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_3]);
1774 	vdev->rbar[4] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_4]);
1775 	vdev->rbar[5] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_5]);
1776 	vdev->rbar[6] = le32_to_cpu(*(__le32 *)&vconfig[PCI_ROM_ADDRESS]);
1777 
1778 	if (pdev->is_virtfn) {
1779 		*(__le16 *)&vconfig[PCI_VENDOR_ID] = cpu_to_le16(pdev->vendor);
1780 		*(__le16 *)&vconfig[PCI_DEVICE_ID] = cpu_to_le16(pdev->device);
1781 
1782 		/*
1783 		 * Per SR-IOV spec rev 1.1, 3.4.1.18 the interrupt pin register
1784 		 * does not apply to VFs and VFs must implement this register
1785 		 * as read-only with value zero.  Userspace is not readily able
1786 		 * to identify whether a device is a VF and thus that the pin
1787 		 * definition on the device is bogus should it violate this
1788 		 * requirement.  We already virtualize the pin register for
1789 		 * other purposes, so we simply need to replace the bogus value
1790 		 * and consider VFs when we determine INTx IRQ count.
1791 		 */
1792 		if (vconfig[PCI_INTERRUPT_PIN] &&
1793 		    !pci_match_id(known_bogus_vf_intx_pin, pdev))
1794 			pci_warn(pdev,
1795 				 "Hardware bug: VF reports bogus INTx pin %d\n",
1796 				 vconfig[PCI_INTERRUPT_PIN]);
1797 
1798 		vconfig[PCI_INTERRUPT_PIN] = 0; /* Gratuitous for good VFs */
1799 	}
1800 	if (pdev->no_command_memory) {
1801 		/*
1802 		 * VFs and devices that set pdev->no_command_memory do not
1803 		 * implement the memory enable bit of the COMMAND register
1804 		 * therefore we'll not have it set in our initial copy of
1805 		 * config space after pci_enable_device().  For consistency
1806 		 * with PFs, set the virtual enable bit here.
1807 		 */
1808 		*(__le16 *)&vconfig[PCI_COMMAND] |=
1809 					cpu_to_le16(PCI_COMMAND_MEMORY);
1810 	}
1811 
1812 	if (!IS_ENABLED(CONFIG_VFIO_PCI_INTX) || vdev->nointx)
1813 		vconfig[PCI_INTERRUPT_PIN] = 0;
1814 
1815 	ret = vfio_cap_init(vdev);
1816 	if (ret)
1817 		goto out;
1818 
1819 	ret = vfio_ecap_init(vdev);
1820 	if (ret)
1821 		goto out;
1822 
1823 	return 0;
1824 
1825 out:
1826 	kfree(map);
1827 	vdev->pci_config_map = NULL;
1828 	kfree(vconfig);
1829 	vdev->vconfig = NULL;
1830 	return pcibios_err_to_errno(ret);
1831 }
1832 
1833 void vfio_config_free(struct vfio_pci_core_device *vdev)
1834 {
1835 	kfree(vdev->vconfig);
1836 	vdev->vconfig = NULL;
1837 	kfree(vdev->pci_config_map);
1838 	vdev->pci_config_map = NULL;
1839 	if (vdev->msi_perm) {
1840 		free_perm_bits(vdev->msi_perm);
1841 		kfree(vdev->msi_perm);
1842 		vdev->msi_perm = NULL;
1843 	}
1844 }
1845 
1846 /*
1847  * Find the remaining number of bytes in a dword that match the given
1848  * position.  Stop at either the end of the capability or the dword boundary.
1849  */
1850 static size_t vfio_pci_cap_remaining_dword(struct vfio_pci_core_device *vdev,
1851 					   loff_t pos)
1852 {
1853 	u8 cap = vdev->pci_config_map[pos];
1854 	size_t i;
1855 
1856 	for (i = 1; (pos + i) % 4 && vdev->pci_config_map[pos + i] == cap; i++)
1857 		/* nop */;
1858 
1859 	return i;
1860 }
1861 
1862 static ssize_t vfio_config_do_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1863 				 size_t count, loff_t *ppos, bool iswrite)
1864 {
1865 	struct pci_dev *pdev = vdev->pdev;
1866 	struct perm_bits *perm;
1867 	__le32 val = 0;
1868 	int cap_start = 0, offset;
1869 	u8 cap_id;
1870 	ssize_t ret;
1871 
1872 	if (*ppos < 0 || *ppos >= pdev->cfg_size ||
1873 	    *ppos + count > pdev->cfg_size)
1874 		return -EFAULT;
1875 
1876 	/*
1877 	 * Chop accesses into aligned chunks containing no more than a
1878 	 * single capability.  Caller increments to the next chunk.
1879 	 */
1880 	count = min(count, vfio_pci_cap_remaining_dword(vdev, *ppos));
1881 	if (count >= 4 && !(*ppos % 4))
1882 		count = 4;
1883 	else if (count >= 2 && !(*ppos % 2))
1884 		count = 2;
1885 	else
1886 		count = 1;
1887 
1888 	ret = count;
1889 
1890 	cap_id = vdev->pci_config_map[*ppos];
1891 
1892 	if (cap_id == PCI_CAP_ID_INVALID) {
1893 		perm = &unassigned_perms;
1894 		cap_start = *ppos;
1895 	} else if (cap_id == PCI_CAP_ID_INVALID_VIRT) {
1896 		perm = &virt_perms;
1897 		cap_start = *ppos;
1898 	} else {
1899 		if (*ppos >= PCI_CFG_SPACE_SIZE) {
1900 			WARN_ON(cap_id > PCI_EXT_CAP_ID_MAX);
1901 
1902 			perm = &ecap_perms[cap_id];
1903 			cap_start = vfio_find_cap_start(vdev, *ppos);
1904 		} else {
1905 			WARN_ON(cap_id > PCI_CAP_ID_MAX);
1906 
1907 			perm = &cap_perms[cap_id];
1908 
1909 			if (cap_id == PCI_CAP_ID_MSI)
1910 				perm = vdev->msi_perm;
1911 
1912 			if (cap_id > PCI_CAP_ID_BASIC)
1913 				cap_start = vfio_find_cap_start(vdev, *ppos);
1914 		}
1915 	}
1916 
1917 	WARN_ON(!cap_start && cap_id != PCI_CAP_ID_BASIC);
1918 	WARN_ON(cap_start > *ppos);
1919 
1920 	offset = *ppos - cap_start;
1921 
1922 	if (iswrite) {
1923 		if (!perm->writefn)
1924 			return ret;
1925 
1926 		if (copy_from_user(&val, buf, count))
1927 			return -EFAULT;
1928 
1929 		ret = perm->writefn(vdev, *ppos, count, perm, offset, val);
1930 	} else {
1931 		if (perm->readfn) {
1932 			ret = perm->readfn(vdev, *ppos, count,
1933 					   perm, offset, &val);
1934 			if (ret < 0)
1935 				return ret;
1936 		}
1937 
1938 		if (copy_to_user(buf, &val, count))
1939 			return -EFAULT;
1940 	}
1941 
1942 	return ret;
1943 }
1944 
1945 ssize_t vfio_pci_config_rw(struct vfio_pci_core_device *vdev, char __user *buf,
1946 			   size_t count, loff_t *ppos, bool iswrite)
1947 {
1948 	size_t done = 0;
1949 	int ret = 0;
1950 	loff_t pos = *ppos;
1951 
1952 	pos &= VFIO_PCI_OFFSET_MASK;
1953 
1954 	while (count) {
1955 		ret = vfio_config_do_rw(vdev, buf, count, &pos, iswrite);
1956 		if (ret < 0)
1957 			return ret;
1958 
1959 		count -= ret;
1960 		done += ret;
1961 		buf += ret;
1962 		pos += ret;
1963 	}
1964 
1965 	*ppos += done;
1966 
1967 	return done;
1968 }
1969