1 /*
2  * VFIO PCI config space virtualization
3  *
4  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
5  *     Author: Alex Williamson <alex.williamson@redhat.com>
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License version 2 as
9  * published by the Free Software Foundation.
10  *
11  * Derived from original vfio:
12  * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
13  * Author: Tom Lyon, pugs@cisco.com
14  */
15 
16 /*
17  * This code handles reading and writing of PCI configuration registers.
18  * This is hairy because we want to allow a lot of flexibility to the
19  * user driver, but cannot trust it with all of the config fields.
20  * Tables determine which fields can be read and written, as well as
21  * which fields are 'virtualized' - special actions and translations to
22  * make it appear to the user that he has control, when in fact things
23  * must be negotiated with the underlying OS.
24  */
25 
26 #include <linux/fs.h>
27 #include <linux/pci.h>
28 #include <linux/uaccess.h>
29 #include <linux/vfio.h>
30 
31 #include "vfio_pci_private.h"
32 
33 #define PCI_CFG_SPACE_SIZE	256
34 
35 /* Useful "pseudo" capabilities */
36 #define PCI_CAP_ID_BASIC	0
37 #define PCI_CAP_ID_INVALID	0xFF
38 
39 #define is_bar(offset)	\
40 	((offset >= PCI_BASE_ADDRESS_0 && offset < PCI_BASE_ADDRESS_5 + 4) || \
41 	 (offset >= PCI_ROM_ADDRESS && offset < PCI_ROM_ADDRESS + 4))
42 
43 /*
44  * Lengths of PCI Config Capabilities
45  *   0: Removed from the user visible capability list
46  *   FF: Variable length
47  */
48 static u8 pci_cap_length[] = {
49 	[PCI_CAP_ID_BASIC]	= PCI_STD_HEADER_SIZEOF, /* pci config header */
50 	[PCI_CAP_ID_PM]		= PCI_PM_SIZEOF,
51 	[PCI_CAP_ID_AGP]	= PCI_AGP_SIZEOF,
52 	[PCI_CAP_ID_VPD]	= PCI_CAP_VPD_SIZEOF,
53 	[PCI_CAP_ID_SLOTID]	= 0,		/* bridge - don't care */
54 	[PCI_CAP_ID_MSI]	= 0xFF,		/* 10, 14, 20, or 24 */
55 	[PCI_CAP_ID_CHSWP]	= 0,		/* cpci - not yet */
56 	[PCI_CAP_ID_PCIX]	= 0xFF,		/* 8 or 24 */
57 	[PCI_CAP_ID_HT]		= 0xFF,		/* hypertransport */
58 	[PCI_CAP_ID_VNDR]	= 0xFF,		/* variable */
59 	[PCI_CAP_ID_DBG]	= 0,		/* debug - don't care */
60 	[PCI_CAP_ID_CCRC]	= 0,		/* cpci - not yet */
61 	[PCI_CAP_ID_SHPC]	= 0,		/* hotswap - not yet */
62 	[PCI_CAP_ID_SSVID]	= 0,		/* bridge - don't care */
63 	[PCI_CAP_ID_AGP3]	= 0,		/* AGP8x - not yet */
64 	[PCI_CAP_ID_SECDEV]	= 0,		/* secure device not yet */
65 	[PCI_CAP_ID_EXP]	= 0xFF,		/* 20 or 44 */
66 	[PCI_CAP_ID_MSIX]	= PCI_CAP_MSIX_SIZEOF,
67 	[PCI_CAP_ID_SATA]	= 0xFF,
68 	[PCI_CAP_ID_AF]		= PCI_CAP_AF_SIZEOF,
69 };
70 
71 /*
72  * Lengths of PCIe/PCI-X Extended Config Capabilities
73  *   0: Removed or masked from the user visible capabilty list
74  *   FF: Variable length
75  */
76 static u16 pci_ext_cap_length[] = {
77 	[PCI_EXT_CAP_ID_ERR]	=	PCI_ERR_ROOT_COMMAND,
78 	[PCI_EXT_CAP_ID_VC]	=	0xFF,
79 	[PCI_EXT_CAP_ID_DSN]	=	PCI_EXT_CAP_DSN_SIZEOF,
80 	[PCI_EXT_CAP_ID_PWR]	=	PCI_EXT_CAP_PWR_SIZEOF,
81 	[PCI_EXT_CAP_ID_RCLD]	=	0,	/* root only - don't care */
82 	[PCI_EXT_CAP_ID_RCILC]	=	0,	/* root only - don't care */
83 	[PCI_EXT_CAP_ID_RCEC]	=	0,	/* root only - don't care */
84 	[PCI_EXT_CAP_ID_MFVC]	=	0xFF,
85 	[PCI_EXT_CAP_ID_VC9]	=	0xFF,	/* same as CAP_ID_VC */
86 	[PCI_EXT_CAP_ID_RCRB]	=	0,	/* root only - don't care */
87 	[PCI_EXT_CAP_ID_VNDR]	=	0xFF,
88 	[PCI_EXT_CAP_ID_CAC]	=	0,	/* obsolete */
89 	[PCI_EXT_CAP_ID_ACS]	=	0xFF,
90 	[PCI_EXT_CAP_ID_ARI]	=	PCI_EXT_CAP_ARI_SIZEOF,
91 	[PCI_EXT_CAP_ID_ATS]	=	PCI_EXT_CAP_ATS_SIZEOF,
92 	[PCI_EXT_CAP_ID_SRIOV]	=	PCI_EXT_CAP_SRIOV_SIZEOF,
93 	[PCI_EXT_CAP_ID_MRIOV]	=	0,	/* not yet */
94 	[PCI_EXT_CAP_ID_MCAST]	=	PCI_EXT_CAP_MCAST_ENDPOINT_SIZEOF,
95 	[PCI_EXT_CAP_ID_PRI]	=	PCI_EXT_CAP_PRI_SIZEOF,
96 	[PCI_EXT_CAP_ID_AMD_XXX] =	0,	/* not yet */
97 	[PCI_EXT_CAP_ID_REBAR]	=	0xFF,
98 	[PCI_EXT_CAP_ID_DPA]	=	0xFF,
99 	[PCI_EXT_CAP_ID_TPH]	=	0xFF,
100 	[PCI_EXT_CAP_ID_LTR]	=	PCI_EXT_CAP_LTR_SIZEOF,
101 	[PCI_EXT_CAP_ID_SECPCI]	=	0,	/* not yet */
102 	[PCI_EXT_CAP_ID_PMUX]	=	0,	/* not yet */
103 	[PCI_EXT_CAP_ID_PASID]	=	0,	/* not yet */
104 };
105 
106 /*
107  * Read/Write Permission Bits - one bit for each bit in capability
108  * Any field can be read if it exists, but what is read depends on
109  * whether the field is 'virtualized', or just pass thru to the
110  * hardware.  Any virtualized field is also virtualized for writes.
111  * Writes are only permitted if they have a 1 bit here.
112  */
113 struct perm_bits {
114 	u8	*virt;		/* read/write virtual data, not hw */
115 	u8	*write;		/* writeable bits */
116 	int	(*readfn)(struct vfio_pci_device *vdev, int pos, int count,
117 			  struct perm_bits *perm, int offset, __le32 *val);
118 	int	(*writefn)(struct vfio_pci_device *vdev, int pos, int count,
119 			   struct perm_bits *perm, int offset, __le32 val);
120 };
121 
122 #define	NO_VIRT		0
123 #define	ALL_VIRT	0xFFFFFFFFU
124 #define	NO_WRITE	0
125 #define	ALL_WRITE	0xFFFFFFFFU
126 
127 static int vfio_user_config_read(struct pci_dev *pdev, int offset,
128 				 __le32 *val, int count)
129 {
130 	int ret = -EINVAL;
131 	u32 tmp_val = 0;
132 
133 	switch (count) {
134 	case 1:
135 	{
136 		u8 tmp;
137 		ret = pci_user_read_config_byte(pdev, offset, &tmp);
138 		tmp_val = tmp;
139 		break;
140 	}
141 	case 2:
142 	{
143 		u16 tmp;
144 		ret = pci_user_read_config_word(pdev, offset, &tmp);
145 		tmp_val = tmp;
146 		break;
147 	}
148 	case 4:
149 		ret = pci_user_read_config_dword(pdev, offset, &tmp_val);
150 		break;
151 	}
152 
153 	*val = cpu_to_le32(tmp_val);
154 
155 	return pcibios_err_to_errno(ret);
156 }
157 
158 static int vfio_user_config_write(struct pci_dev *pdev, int offset,
159 				  __le32 val, int count)
160 {
161 	int ret = -EINVAL;
162 	u32 tmp_val = le32_to_cpu(val);
163 
164 	switch (count) {
165 	case 1:
166 		ret = pci_user_write_config_byte(pdev, offset, tmp_val);
167 		break;
168 	case 2:
169 		ret = pci_user_write_config_word(pdev, offset, tmp_val);
170 		break;
171 	case 4:
172 		ret = pci_user_write_config_dword(pdev, offset, tmp_val);
173 		break;
174 	}
175 
176 	return pcibios_err_to_errno(ret);
177 }
178 
179 static int vfio_default_config_read(struct vfio_pci_device *vdev, int pos,
180 				    int count, struct perm_bits *perm,
181 				    int offset, __le32 *val)
182 {
183 	__le32 virt = 0;
184 
185 	memcpy(val, vdev->vconfig + pos, count);
186 
187 	memcpy(&virt, perm->virt + offset, count);
188 
189 	/* Any non-virtualized bits? */
190 	if (cpu_to_le32(~0U >> (32 - (count * 8))) != virt) {
191 		struct pci_dev *pdev = vdev->pdev;
192 		__le32 phys_val = 0;
193 		int ret;
194 
195 		ret = vfio_user_config_read(pdev, pos, &phys_val, count);
196 		if (ret)
197 			return ret;
198 
199 		*val = (phys_val & ~virt) | (*val & virt);
200 	}
201 
202 	return count;
203 }
204 
205 static int vfio_default_config_write(struct vfio_pci_device *vdev, int pos,
206 				     int count, struct perm_bits *perm,
207 				     int offset, __le32 val)
208 {
209 	__le32 virt = 0, write = 0;
210 
211 	memcpy(&write, perm->write + offset, count);
212 
213 	if (!write)
214 		return count; /* drop, no writable bits */
215 
216 	memcpy(&virt, perm->virt + offset, count);
217 
218 	/* Virtualized and writable bits go to vconfig */
219 	if (write & virt) {
220 		__le32 virt_val = 0;
221 
222 		memcpy(&virt_val, vdev->vconfig + pos, count);
223 
224 		virt_val &= ~(write & virt);
225 		virt_val |= (val & (write & virt));
226 
227 		memcpy(vdev->vconfig + pos, &virt_val, count);
228 	}
229 
230 	/* Non-virtualzed and writable bits go to hardware */
231 	if (write & ~virt) {
232 		struct pci_dev *pdev = vdev->pdev;
233 		__le32 phys_val = 0;
234 		int ret;
235 
236 		ret = vfio_user_config_read(pdev, pos, &phys_val, count);
237 		if (ret)
238 			return ret;
239 
240 		phys_val &= ~(write & ~virt);
241 		phys_val |= (val & (write & ~virt));
242 
243 		ret = vfio_user_config_write(pdev, pos, phys_val, count);
244 		if (ret)
245 			return ret;
246 	}
247 
248 	return count;
249 }
250 
251 /* Allow direct read from hardware, except for capability next pointer */
252 static int vfio_direct_config_read(struct vfio_pci_device *vdev, int pos,
253 				   int count, struct perm_bits *perm,
254 				   int offset, __le32 *val)
255 {
256 	int ret;
257 
258 	ret = vfio_user_config_read(vdev->pdev, pos, val, count);
259 	if (ret)
260 		return pcibios_err_to_errno(ret);
261 
262 	if (pos >= PCI_CFG_SPACE_SIZE) { /* Extended cap header mangling */
263 		if (offset < 4)
264 			memcpy(val, vdev->vconfig + pos, count);
265 	} else if (pos >= PCI_STD_HEADER_SIZEOF) { /* Std cap mangling */
266 		if (offset == PCI_CAP_LIST_ID && count > 1)
267 			memcpy(val, vdev->vconfig + pos,
268 			       min(PCI_CAP_FLAGS, count));
269 		else if (offset == PCI_CAP_LIST_NEXT)
270 			memcpy(val, vdev->vconfig + pos, 1);
271 	}
272 
273 	return count;
274 }
275 
276 static int vfio_direct_config_write(struct vfio_pci_device *vdev, int pos,
277 				    int count, struct perm_bits *perm,
278 				    int offset, __le32 val)
279 {
280 	int ret;
281 
282 	ret = vfio_user_config_write(vdev->pdev, pos, val, count);
283 	if (ret)
284 		return ret;
285 
286 	return count;
287 }
288 
289 /* Default all regions to read-only, no-virtualization */
290 static struct perm_bits cap_perms[PCI_CAP_ID_MAX + 1] = {
291 	[0 ... PCI_CAP_ID_MAX] = { .readfn = vfio_direct_config_read }
292 };
293 static struct perm_bits ecap_perms[PCI_EXT_CAP_ID_MAX + 1] = {
294 	[0 ... PCI_EXT_CAP_ID_MAX] = { .readfn = vfio_direct_config_read }
295 };
296 
297 static void free_perm_bits(struct perm_bits *perm)
298 {
299 	kfree(perm->virt);
300 	kfree(perm->write);
301 	perm->virt = NULL;
302 	perm->write = NULL;
303 }
304 
305 static int alloc_perm_bits(struct perm_bits *perm, int size)
306 {
307 	/*
308 	 * Round up all permission bits to the next dword, this lets us
309 	 * ignore whether a read/write exceeds the defined capability
310 	 * structure.  We can do this because:
311 	 *  - Standard config space is already dword aligned
312 	 *  - Capabilities are all dword alinged (bits 0:1 of next reserved)
313 	 *  - Express capabilities defined as dword aligned
314 	 */
315 	size = round_up(size, 4);
316 
317 	/*
318 	 * Zero state is
319 	 * - All Readable, None Writeable, None Virtualized
320 	 */
321 	perm->virt = kzalloc(size, GFP_KERNEL);
322 	perm->write = kzalloc(size, GFP_KERNEL);
323 	if (!perm->virt || !perm->write) {
324 		free_perm_bits(perm);
325 		return -ENOMEM;
326 	}
327 
328 	perm->readfn = vfio_default_config_read;
329 	perm->writefn = vfio_default_config_write;
330 
331 	return 0;
332 }
333 
334 /*
335  * Helper functions for filling in permission tables
336  */
337 static inline void p_setb(struct perm_bits *p, int off, u8 virt, u8 write)
338 {
339 	p->virt[off] = virt;
340 	p->write[off] = write;
341 }
342 
343 /* Handle endian-ness - pci and tables are little-endian */
344 static inline void p_setw(struct perm_bits *p, int off, u16 virt, u16 write)
345 {
346 	*(__le16 *)(&p->virt[off]) = cpu_to_le16(virt);
347 	*(__le16 *)(&p->write[off]) = cpu_to_le16(write);
348 }
349 
350 /* Handle endian-ness - pci and tables are little-endian */
351 static inline void p_setd(struct perm_bits *p, int off, u32 virt, u32 write)
352 {
353 	*(__le32 *)(&p->virt[off]) = cpu_to_le32(virt);
354 	*(__le32 *)(&p->write[off]) = cpu_to_le32(write);
355 }
356 
357 /*
358  * Restore the *real* BARs after we detect a FLR or backdoor reset.
359  * (backdoor = some device specific technique that we didn't catch)
360  */
361 static void vfio_bar_restore(struct vfio_pci_device *vdev)
362 {
363 	struct pci_dev *pdev = vdev->pdev;
364 	u32 *rbar = vdev->rbar;
365 	int i;
366 
367 	if (pdev->is_virtfn)
368 		return;
369 
370 	pr_info("%s: %s reset recovery - restoring bars\n",
371 		__func__, dev_name(&pdev->dev));
372 
373 	for (i = PCI_BASE_ADDRESS_0; i <= PCI_BASE_ADDRESS_5; i += 4, rbar++)
374 		pci_user_write_config_dword(pdev, i, *rbar);
375 
376 	pci_user_write_config_dword(pdev, PCI_ROM_ADDRESS, *rbar);
377 }
378 
379 static __le32 vfio_generate_bar_flags(struct pci_dev *pdev, int bar)
380 {
381 	unsigned long flags = pci_resource_flags(pdev, bar);
382 	u32 val;
383 
384 	if (flags & IORESOURCE_IO)
385 		return cpu_to_le32(PCI_BASE_ADDRESS_SPACE_IO);
386 
387 	val = PCI_BASE_ADDRESS_SPACE_MEMORY;
388 
389 	if (flags & IORESOURCE_PREFETCH)
390 		val |= PCI_BASE_ADDRESS_MEM_PREFETCH;
391 
392 	if (flags & IORESOURCE_MEM_64)
393 		val |= PCI_BASE_ADDRESS_MEM_TYPE_64;
394 
395 	return cpu_to_le32(val);
396 }
397 
398 /*
399  * Pretend we're hardware and tweak the values of the *virtual* PCI BARs
400  * to reflect the hardware capabilities.  This implements BAR sizing.
401  */
402 static void vfio_bar_fixup(struct vfio_pci_device *vdev)
403 {
404 	struct pci_dev *pdev = vdev->pdev;
405 	int i;
406 	__le32 *bar;
407 	u64 mask;
408 
409 	bar = (__le32 *)&vdev->vconfig[PCI_BASE_ADDRESS_0];
410 
411 	for (i = PCI_STD_RESOURCES; i <= PCI_STD_RESOURCE_END; i++, bar++) {
412 		if (!pci_resource_start(pdev, i)) {
413 			*bar = 0; /* Unmapped by host = unimplemented to user */
414 			continue;
415 		}
416 
417 		mask = ~(pci_resource_len(pdev, i) - 1);
418 
419 		*bar &= cpu_to_le32((u32)mask);
420 		*bar |= vfio_generate_bar_flags(pdev, i);
421 
422 		if (*bar & cpu_to_le32(PCI_BASE_ADDRESS_MEM_TYPE_64)) {
423 			bar++;
424 			*bar &= cpu_to_le32((u32)(mask >> 32));
425 			i++;
426 		}
427 	}
428 
429 	bar = (__le32 *)&vdev->vconfig[PCI_ROM_ADDRESS];
430 
431 	/*
432 	 * NB. we expose the actual BAR size here, regardless of whether
433 	 * we can read it.  When we report the REGION_INFO for the ROM
434 	 * we report what PCI tells us is the actual ROM size.
435 	 */
436 	if (pci_resource_start(pdev, PCI_ROM_RESOURCE)) {
437 		mask = ~(pci_resource_len(pdev, PCI_ROM_RESOURCE) - 1);
438 		mask |= PCI_ROM_ADDRESS_ENABLE;
439 		*bar &= cpu_to_le32((u32)mask);
440 	} else
441 		*bar = 0;
442 
443 	vdev->bardirty = false;
444 }
445 
446 static int vfio_basic_config_read(struct vfio_pci_device *vdev, int pos,
447 				  int count, struct perm_bits *perm,
448 				  int offset, __le32 *val)
449 {
450 	if (is_bar(offset)) /* pos == offset for basic config */
451 		vfio_bar_fixup(vdev);
452 
453 	count = vfio_default_config_read(vdev, pos, count, perm, offset, val);
454 
455 	/* Mask in virtual memory enable for SR-IOV devices */
456 	if (offset == PCI_COMMAND && vdev->pdev->is_virtfn) {
457 		u16 cmd = le16_to_cpu(*(__le16 *)&vdev->vconfig[PCI_COMMAND]);
458 		u32 tmp_val = le32_to_cpu(*val);
459 
460 		tmp_val |= cmd & PCI_COMMAND_MEMORY;
461 		*val = cpu_to_le32(tmp_val);
462 	}
463 
464 	return count;
465 }
466 
467 static int vfio_basic_config_write(struct vfio_pci_device *vdev, int pos,
468 				   int count, struct perm_bits *perm,
469 				   int offset, __le32 val)
470 {
471 	struct pci_dev *pdev = vdev->pdev;
472 	__le16 *virt_cmd;
473 	u16 new_cmd = 0;
474 	int ret;
475 
476 	virt_cmd = (__le16 *)&vdev->vconfig[PCI_COMMAND];
477 
478 	if (offset == PCI_COMMAND) {
479 		bool phys_mem, virt_mem, new_mem, phys_io, virt_io, new_io;
480 		u16 phys_cmd;
481 
482 		ret = pci_user_read_config_word(pdev, PCI_COMMAND, &phys_cmd);
483 		if (ret)
484 			return ret;
485 
486 		new_cmd = le32_to_cpu(val);
487 
488 		phys_mem = !!(phys_cmd & PCI_COMMAND_MEMORY);
489 		virt_mem = !!(le16_to_cpu(*virt_cmd) & PCI_COMMAND_MEMORY);
490 		new_mem = !!(new_cmd & PCI_COMMAND_MEMORY);
491 
492 		phys_io = !!(phys_cmd & PCI_COMMAND_IO);
493 		virt_io = !!(le16_to_cpu(*virt_cmd) & PCI_COMMAND_IO);
494 		new_io = !!(new_cmd & PCI_COMMAND_IO);
495 
496 		/*
497 		 * If the user is writing mem/io enable (new_mem/io) and we
498 		 * think it's already enabled (virt_mem/io), but the hardware
499 		 * shows it disabled (phys_mem/io, then the device has
500 		 * undergone some kind of backdoor reset and needs to be
501 		 * restored before we allow it to enable the bars.
502 		 * SR-IOV devices will trigger this, but we catch them later
503 		 */
504 		if ((new_mem && virt_mem && !phys_mem) ||
505 		    (new_io && virt_io && !phys_io))
506 			vfio_bar_restore(vdev);
507 	}
508 
509 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
510 	if (count < 0)
511 		return count;
512 
513 	/*
514 	 * Save current memory/io enable bits in vconfig to allow for
515 	 * the test above next time.
516 	 */
517 	if (offset == PCI_COMMAND) {
518 		u16 mask = PCI_COMMAND_MEMORY | PCI_COMMAND_IO;
519 
520 		*virt_cmd &= cpu_to_le16(~mask);
521 		*virt_cmd |= cpu_to_le16(new_cmd & mask);
522 	}
523 
524 	/* Emulate INTx disable */
525 	if (offset >= PCI_COMMAND && offset <= PCI_COMMAND + 1) {
526 		bool virt_intx_disable;
527 
528 		virt_intx_disable = !!(le16_to_cpu(*virt_cmd) &
529 				       PCI_COMMAND_INTX_DISABLE);
530 
531 		if (virt_intx_disable && !vdev->virq_disabled) {
532 			vdev->virq_disabled = true;
533 			vfio_pci_intx_mask(vdev);
534 		} else if (!virt_intx_disable && vdev->virq_disabled) {
535 			vdev->virq_disabled = false;
536 			vfio_pci_intx_unmask(vdev);
537 		}
538 	}
539 
540 	if (is_bar(offset))
541 		vdev->bardirty = true;
542 
543 	return count;
544 }
545 
546 /* Permissions for the Basic PCI Header */
547 static int __init init_pci_cap_basic_perm(struct perm_bits *perm)
548 {
549 	if (alloc_perm_bits(perm, PCI_STD_HEADER_SIZEOF))
550 		return -ENOMEM;
551 
552 	perm->readfn = vfio_basic_config_read;
553 	perm->writefn = vfio_basic_config_write;
554 
555 	/* Virtualized for SR-IOV functions, which just have FFFF */
556 	p_setw(perm, PCI_VENDOR_ID, (u16)ALL_VIRT, NO_WRITE);
557 	p_setw(perm, PCI_DEVICE_ID, (u16)ALL_VIRT, NO_WRITE);
558 
559 	/*
560 	 * Virtualize INTx disable, we use it internally for interrupt
561 	 * control and can emulate it for non-PCI 2.3 devices.
562 	 */
563 	p_setw(perm, PCI_COMMAND, PCI_COMMAND_INTX_DISABLE, (u16)ALL_WRITE);
564 
565 	/* Virtualize capability list, we might want to skip/disable */
566 	p_setw(perm, PCI_STATUS, PCI_STATUS_CAP_LIST, NO_WRITE);
567 
568 	/* No harm to write */
569 	p_setb(perm, PCI_CACHE_LINE_SIZE, NO_VIRT, (u8)ALL_WRITE);
570 	p_setb(perm, PCI_LATENCY_TIMER, NO_VIRT, (u8)ALL_WRITE);
571 	p_setb(perm, PCI_BIST, NO_VIRT, (u8)ALL_WRITE);
572 
573 	/* Virtualize all bars, can't touch the real ones */
574 	p_setd(perm, PCI_BASE_ADDRESS_0, ALL_VIRT, ALL_WRITE);
575 	p_setd(perm, PCI_BASE_ADDRESS_1, ALL_VIRT, ALL_WRITE);
576 	p_setd(perm, PCI_BASE_ADDRESS_2, ALL_VIRT, ALL_WRITE);
577 	p_setd(perm, PCI_BASE_ADDRESS_3, ALL_VIRT, ALL_WRITE);
578 	p_setd(perm, PCI_BASE_ADDRESS_4, ALL_VIRT, ALL_WRITE);
579 	p_setd(perm, PCI_BASE_ADDRESS_5, ALL_VIRT, ALL_WRITE);
580 	p_setd(perm, PCI_ROM_ADDRESS, ALL_VIRT, ALL_WRITE);
581 
582 	/* Allow us to adjust capability chain */
583 	p_setb(perm, PCI_CAPABILITY_LIST, (u8)ALL_VIRT, NO_WRITE);
584 
585 	/* Sometimes used by sw, just virtualize */
586 	p_setb(perm, PCI_INTERRUPT_LINE, (u8)ALL_VIRT, (u8)ALL_WRITE);
587 	return 0;
588 }
589 
590 /* Permissions for the Power Management capability */
591 static int __init init_pci_cap_pm_perm(struct perm_bits *perm)
592 {
593 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_PM]))
594 		return -ENOMEM;
595 
596 	/*
597 	 * We always virtualize the next field so we can remove
598 	 * capabilities from the chain if we want to.
599 	 */
600 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
601 
602 	/*
603 	 * Power management is defined *per function*,
604 	 * so we let the user write this
605 	 */
606 	p_setd(perm, PCI_PM_CTRL, NO_VIRT, ALL_WRITE);
607 	return 0;
608 }
609 
610 /* Permissions for PCI-X capability */
611 static int __init init_pci_cap_pcix_perm(struct perm_bits *perm)
612 {
613 	/* Alloc 24, but only 8 are used in v0 */
614 	if (alloc_perm_bits(perm, PCI_CAP_PCIX_SIZEOF_V2))
615 		return -ENOMEM;
616 
617 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
618 
619 	p_setw(perm, PCI_X_CMD, NO_VIRT, (u16)ALL_WRITE);
620 	p_setd(perm, PCI_X_ECC_CSR, NO_VIRT, ALL_WRITE);
621 	return 0;
622 }
623 
624 /* Permissions for PCI Express capability */
625 static int __init init_pci_cap_exp_perm(struct perm_bits *perm)
626 {
627 	/* Alloc larger of two possible sizes */
628 	if (alloc_perm_bits(perm, PCI_CAP_EXP_ENDPOINT_SIZEOF_V2))
629 		return -ENOMEM;
630 
631 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
632 
633 	/*
634 	 * Allow writes to device control fields (includes FLR!)
635 	 * but not to devctl_phantom which could confuse IOMMU
636 	 * or to the ARI bit in devctl2 which is set at probe time
637 	 */
638 	p_setw(perm, PCI_EXP_DEVCTL, NO_VIRT, ~PCI_EXP_DEVCTL_PHANTOM);
639 	p_setw(perm, PCI_EXP_DEVCTL2, NO_VIRT, ~PCI_EXP_DEVCTL2_ARI);
640 	return 0;
641 }
642 
643 /* Permissions for Advanced Function capability */
644 static int __init init_pci_cap_af_perm(struct perm_bits *perm)
645 {
646 	if (alloc_perm_bits(perm, pci_cap_length[PCI_CAP_ID_AF]))
647 		return -ENOMEM;
648 
649 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
650 	p_setb(perm, PCI_AF_CTRL, NO_VIRT, PCI_AF_CTRL_FLR);
651 	return 0;
652 }
653 
654 /* Permissions for Advanced Error Reporting extended capability */
655 static int __init init_pci_ext_cap_err_perm(struct perm_bits *perm)
656 {
657 	u32 mask;
658 
659 	if (alloc_perm_bits(perm, pci_ext_cap_length[PCI_EXT_CAP_ID_ERR]))
660 		return -ENOMEM;
661 
662 	/*
663 	 * Virtualize the first dword of all express capabilities
664 	 * because it includes the next pointer.  This lets us later
665 	 * remove capabilities from the chain if we need to.
666 	 */
667 	p_setd(perm, 0, ALL_VIRT, NO_WRITE);
668 
669 	/* Writable bits mask */
670 	mask =	PCI_ERR_UNC_TRAIN |		/* Training */
671 		PCI_ERR_UNC_DLP |		/* Data Link Protocol */
672 		PCI_ERR_UNC_SURPDN |		/* Surprise Down */
673 		PCI_ERR_UNC_POISON_TLP |	/* Poisoned TLP */
674 		PCI_ERR_UNC_FCP |		/* Flow Control Protocol */
675 		PCI_ERR_UNC_COMP_TIME |		/* Completion Timeout */
676 		PCI_ERR_UNC_COMP_ABORT |	/* Completer Abort */
677 		PCI_ERR_UNC_UNX_COMP |		/* Unexpected Completion */
678 		PCI_ERR_UNC_RX_OVER |		/* Receiver Overflow */
679 		PCI_ERR_UNC_MALF_TLP |		/* Malformed TLP */
680 		PCI_ERR_UNC_ECRC |		/* ECRC Error Status */
681 		PCI_ERR_UNC_UNSUP |		/* Unsupported Request */
682 		PCI_ERR_UNC_ACSV |		/* ACS Violation */
683 		PCI_ERR_UNC_INTN |		/* internal error */
684 		PCI_ERR_UNC_MCBTLP |		/* MC blocked TLP */
685 		PCI_ERR_UNC_ATOMEG |		/* Atomic egress blocked */
686 		PCI_ERR_UNC_TLPPRE;		/* TLP prefix blocked */
687 	p_setd(perm, PCI_ERR_UNCOR_STATUS, NO_VIRT, mask);
688 	p_setd(perm, PCI_ERR_UNCOR_MASK, NO_VIRT, mask);
689 	p_setd(perm, PCI_ERR_UNCOR_SEVER, NO_VIRT, mask);
690 
691 	mask =	PCI_ERR_COR_RCVR |		/* Receiver Error Status */
692 		PCI_ERR_COR_BAD_TLP |		/* Bad TLP Status */
693 		PCI_ERR_COR_BAD_DLLP |		/* Bad DLLP Status */
694 		PCI_ERR_COR_REP_ROLL |		/* REPLAY_NUM Rollover */
695 		PCI_ERR_COR_REP_TIMER |		/* Replay Timer Timeout */
696 		PCI_ERR_COR_ADV_NFAT |		/* Advisory Non-Fatal */
697 		PCI_ERR_COR_INTERNAL |		/* Corrected Internal */
698 		PCI_ERR_COR_LOG_OVER;		/* Header Log Overflow */
699 	p_setd(perm, PCI_ERR_COR_STATUS, NO_VIRT, mask);
700 	p_setd(perm, PCI_ERR_COR_MASK, NO_VIRT, mask);
701 
702 	mask =	PCI_ERR_CAP_ECRC_GENE |		/* ECRC Generation Enable */
703 		PCI_ERR_CAP_ECRC_CHKE;		/* ECRC Check Enable */
704 	p_setd(perm, PCI_ERR_CAP, NO_VIRT, mask);
705 	return 0;
706 }
707 
708 /* Permissions for Power Budgeting extended capability */
709 static int __init init_pci_ext_cap_pwr_perm(struct perm_bits *perm)
710 {
711 	if (alloc_perm_bits(perm, pci_ext_cap_length[PCI_EXT_CAP_ID_PWR]))
712 		return -ENOMEM;
713 
714 	p_setd(perm, 0, ALL_VIRT, NO_WRITE);
715 
716 	/* Writing the data selector is OK, the info is still read-only */
717 	p_setb(perm, PCI_PWR_DATA, NO_VIRT, (u8)ALL_WRITE);
718 	return 0;
719 }
720 
721 /*
722  * Initialize the shared permission tables
723  */
724 void vfio_pci_uninit_perm_bits(void)
725 {
726 	free_perm_bits(&cap_perms[PCI_CAP_ID_BASIC]);
727 
728 	free_perm_bits(&cap_perms[PCI_CAP_ID_PM]);
729 	free_perm_bits(&cap_perms[PCI_CAP_ID_PCIX]);
730 	free_perm_bits(&cap_perms[PCI_CAP_ID_EXP]);
731 	free_perm_bits(&cap_perms[PCI_CAP_ID_AF]);
732 
733 	free_perm_bits(&ecap_perms[PCI_EXT_CAP_ID_ERR]);
734 	free_perm_bits(&ecap_perms[PCI_EXT_CAP_ID_PWR]);
735 }
736 
737 int __init vfio_pci_init_perm_bits(void)
738 {
739 	int ret;
740 
741 	/* Basic config space */
742 	ret = init_pci_cap_basic_perm(&cap_perms[PCI_CAP_ID_BASIC]);
743 
744 	/* Capabilities */
745 	ret |= init_pci_cap_pm_perm(&cap_perms[PCI_CAP_ID_PM]);
746 	cap_perms[PCI_CAP_ID_VPD].writefn = vfio_direct_config_write;
747 	ret |= init_pci_cap_pcix_perm(&cap_perms[PCI_CAP_ID_PCIX]);
748 	cap_perms[PCI_CAP_ID_VNDR].writefn = vfio_direct_config_write;
749 	ret |= init_pci_cap_exp_perm(&cap_perms[PCI_CAP_ID_EXP]);
750 	ret |= init_pci_cap_af_perm(&cap_perms[PCI_CAP_ID_AF]);
751 
752 	/* Extended capabilities */
753 	ret |= init_pci_ext_cap_err_perm(&ecap_perms[PCI_EXT_CAP_ID_ERR]);
754 	ret |= init_pci_ext_cap_pwr_perm(&ecap_perms[PCI_EXT_CAP_ID_PWR]);
755 	ecap_perms[PCI_EXT_CAP_ID_VNDR].writefn = vfio_direct_config_write;
756 
757 	if (ret)
758 		vfio_pci_uninit_perm_bits();
759 
760 	return ret;
761 }
762 
763 static int vfio_find_cap_start(struct vfio_pci_device *vdev, int pos)
764 {
765 	u8 cap;
766 	int base = (pos >= PCI_CFG_SPACE_SIZE) ? PCI_CFG_SPACE_SIZE :
767 						 PCI_STD_HEADER_SIZEOF;
768 	base /= 4;
769 	pos /= 4;
770 
771 	cap = vdev->pci_config_map[pos];
772 
773 	if (cap == PCI_CAP_ID_BASIC)
774 		return 0;
775 
776 	/* XXX Can we have to abutting capabilities of the same type? */
777 	while (pos - 1 >= base && vdev->pci_config_map[pos - 1] == cap)
778 		pos--;
779 
780 	return pos * 4;
781 }
782 
783 static int vfio_msi_config_read(struct vfio_pci_device *vdev, int pos,
784 				int count, struct perm_bits *perm,
785 				int offset, __le32 *val)
786 {
787 	/* Update max available queue size from msi_qmax */
788 	if (offset <= PCI_MSI_FLAGS && offset + count >= PCI_MSI_FLAGS) {
789 		__le16 *flags;
790 		int start;
791 
792 		start = vfio_find_cap_start(vdev, pos);
793 
794 		flags = (__le16 *)&vdev->vconfig[start];
795 
796 		*flags &= cpu_to_le16(~PCI_MSI_FLAGS_QMASK);
797 		*flags |= cpu_to_le16(vdev->msi_qmax << 1);
798 	}
799 
800 	return vfio_default_config_read(vdev, pos, count, perm, offset, val);
801 }
802 
803 static int vfio_msi_config_write(struct vfio_pci_device *vdev, int pos,
804 				 int count, struct perm_bits *perm,
805 				 int offset, __le32 val)
806 {
807 	count = vfio_default_config_write(vdev, pos, count, perm, offset, val);
808 	if (count < 0)
809 		return count;
810 
811 	/* Fixup and write configured queue size and enable to hardware */
812 	if (offset <= PCI_MSI_FLAGS && offset + count >= PCI_MSI_FLAGS) {
813 		__le16 *pflags;
814 		u16 flags;
815 		int start, ret;
816 
817 		start = vfio_find_cap_start(vdev, pos);
818 
819 		pflags = (__le16 *)&vdev->vconfig[start + PCI_MSI_FLAGS];
820 
821 		flags = le16_to_cpu(*pflags);
822 
823 		/* MSI is enabled via ioctl */
824 		if  (!is_msi(vdev))
825 			flags &= ~PCI_MSI_FLAGS_ENABLE;
826 
827 		/* Check queue size */
828 		if ((flags & PCI_MSI_FLAGS_QSIZE) >> 4 > vdev->msi_qmax) {
829 			flags &= ~PCI_MSI_FLAGS_QSIZE;
830 			flags |= vdev->msi_qmax << 4;
831 		}
832 
833 		/* Write back to virt and to hardware */
834 		*pflags = cpu_to_le16(flags);
835 		ret = pci_user_write_config_word(vdev->pdev,
836 						 start + PCI_MSI_FLAGS,
837 						 flags);
838 		if (ret)
839 			return pcibios_err_to_errno(ret);
840 	}
841 
842 	return count;
843 }
844 
845 /*
846  * MSI determination is per-device, so this routine gets used beyond
847  * initialization time. Don't add __init
848  */
849 static int init_pci_cap_msi_perm(struct perm_bits *perm, int len, u16 flags)
850 {
851 	if (alloc_perm_bits(perm, len))
852 		return -ENOMEM;
853 
854 	perm->readfn = vfio_msi_config_read;
855 	perm->writefn = vfio_msi_config_write;
856 
857 	p_setb(perm, PCI_CAP_LIST_NEXT, (u8)ALL_VIRT, NO_WRITE);
858 
859 	/*
860 	 * The upper byte of the control register is reserved,
861 	 * just setup the lower byte.
862 	 */
863 	p_setb(perm, PCI_MSI_FLAGS, (u8)ALL_VIRT, (u8)ALL_WRITE);
864 	p_setd(perm, PCI_MSI_ADDRESS_LO, ALL_VIRT, ALL_WRITE);
865 	if (flags & PCI_MSI_FLAGS_64BIT) {
866 		p_setd(perm, PCI_MSI_ADDRESS_HI, ALL_VIRT, ALL_WRITE);
867 		p_setw(perm, PCI_MSI_DATA_64, (u16)ALL_VIRT, (u16)ALL_WRITE);
868 		if (flags & PCI_MSI_FLAGS_MASKBIT) {
869 			p_setd(perm, PCI_MSI_MASK_64, NO_VIRT, ALL_WRITE);
870 			p_setd(perm, PCI_MSI_PENDING_64, NO_VIRT, ALL_WRITE);
871 		}
872 	} else {
873 		p_setw(perm, PCI_MSI_DATA_32, (u16)ALL_VIRT, (u16)ALL_WRITE);
874 		if (flags & PCI_MSI_FLAGS_MASKBIT) {
875 			p_setd(perm, PCI_MSI_MASK_32, NO_VIRT, ALL_WRITE);
876 			p_setd(perm, PCI_MSI_PENDING_32, NO_VIRT, ALL_WRITE);
877 		}
878 	}
879 	return 0;
880 }
881 
882 /* Determine MSI CAP field length; initialize msi_perms on 1st call per vdev */
883 static int vfio_msi_cap_len(struct vfio_pci_device *vdev, u8 pos)
884 {
885 	struct pci_dev *pdev = vdev->pdev;
886 	int len, ret;
887 	u16 flags;
888 
889 	ret = pci_read_config_word(pdev, pos + PCI_MSI_FLAGS, &flags);
890 	if (ret)
891 		return pcibios_err_to_errno(ret);
892 
893 	len = 10; /* Minimum size */
894 	if (flags & PCI_MSI_FLAGS_64BIT)
895 		len += 4;
896 	if (flags & PCI_MSI_FLAGS_MASKBIT)
897 		len += 10;
898 
899 	if (vdev->msi_perm)
900 		return len;
901 
902 	vdev->msi_perm = kmalloc(sizeof(struct perm_bits), GFP_KERNEL);
903 	if (!vdev->msi_perm)
904 		return -ENOMEM;
905 
906 	ret = init_pci_cap_msi_perm(vdev->msi_perm, len, flags);
907 	if (ret)
908 		return ret;
909 
910 	return len;
911 }
912 
913 /* Determine extended capability length for VC (2 & 9) and MFVC */
914 static int vfio_vc_cap_len(struct vfio_pci_device *vdev, u16 pos)
915 {
916 	struct pci_dev *pdev = vdev->pdev;
917 	u32 tmp;
918 	int ret, evcc, phases, vc_arb;
919 	int len = PCI_CAP_VC_BASE_SIZEOF;
920 
921 	ret = pci_read_config_dword(pdev, pos + PCI_VC_PORT_REG1, &tmp);
922 	if (ret)
923 		return pcibios_err_to_errno(ret);
924 
925 	evcc = tmp & PCI_VC_REG1_EVCC; /* extended vc count */
926 	ret = pci_read_config_dword(pdev, pos + PCI_VC_PORT_REG2, &tmp);
927 	if (ret)
928 		return pcibios_err_to_errno(ret);
929 
930 	if (tmp & PCI_VC_REG2_128_PHASE)
931 		phases = 128;
932 	else if (tmp & PCI_VC_REG2_64_PHASE)
933 		phases = 64;
934 	else if (tmp & PCI_VC_REG2_32_PHASE)
935 		phases = 32;
936 	else
937 		phases = 0;
938 
939 	vc_arb = phases * 4;
940 
941 	/*
942 	 * Port arbitration tables are root & switch only;
943 	 * function arbitration tables are function 0 only.
944 	 * In either case, we'll never let user write them so
945 	 * we don't care how big they are
946 	 */
947 	len += (1 + evcc) * PCI_CAP_VC_PER_VC_SIZEOF;
948 	if (vc_arb) {
949 		len = round_up(len, 16);
950 		len += vc_arb / 8;
951 	}
952 	return len;
953 }
954 
955 static int vfio_cap_len(struct vfio_pci_device *vdev, u8 cap, u8 pos)
956 {
957 	struct pci_dev *pdev = vdev->pdev;
958 	u16 word;
959 	u8 byte;
960 	int ret;
961 
962 	switch (cap) {
963 	case PCI_CAP_ID_MSI:
964 		return vfio_msi_cap_len(vdev, pos);
965 	case PCI_CAP_ID_PCIX:
966 		ret = pci_read_config_word(pdev, pos + PCI_X_CMD, &word);
967 		if (ret)
968 			return pcibios_err_to_errno(ret);
969 
970 		if (PCI_X_CMD_VERSION(word)) {
971 			vdev->extended_caps = true;
972 			return PCI_CAP_PCIX_SIZEOF_V2;
973 		} else
974 			return PCI_CAP_PCIX_SIZEOF_V0;
975 	case PCI_CAP_ID_VNDR:
976 		/* length follows next field */
977 		ret = pci_read_config_byte(pdev, pos + PCI_CAP_FLAGS, &byte);
978 		if (ret)
979 			return pcibios_err_to_errno(ret);
980 
981 		return byte;
982 	case PCI_CAP_ID_EXP:
983 		/* length based on version */
984 		ret = pci_read_config_word(pdev, pos + PCI_EXP_FLAGS, &word);
985 		if (ret)
986 			return pcibios_err_to_errno(ret);
987 
988 		if ((word & PCI_EXP_FLAGS_VERS) == 1)
989 			return PCI_CAP_EXP_ENDPOINT_SIZEOF_V1;
990 		else {
991 			vdev->extended_caps = true;
992 			return PCI_CAP_EXP_ENDPOINT_SIZEOF_V2;
993 		}
994 	case PCI_CAP_ID_HT:
995 		ret = pci_read_config_byte(pdev, pos + 3, &byte);
996 		if (ret)
997 			return pcibios_err_to_errno(ret);
998 
999 		return (byte & HT_3BIT_CAP_MASK) ?
1000 			HT_CAP_SIZEOF_SHORT : HT_CAP_SIZEOF_LONG;
1001 	case PCI_CAP_ID_SATA:
1002 		ret = pci_read_config_byte(pdev, pos + PCI_SATA_REGS, &byte);
1003 		if (ret)
1004 			return pcibios_err_to_errno(ret);
1005 
1006 		byte &= PCI_SATA_REGS_MASK;
1007 		if (byte == PCI_SATA_REGS_INLINE)
1008 			return PCI_SATA_SIZEOF_LONG;
1009 		else
1010 			return PCI_SATA_SIZEOF_SHORT;
1011 	default:
1012 		pr_warn("%s: %s unknown length for pci cap 0x%x@0x%x\n",
1013 			dev_name(&pdev->dev), __func__, cap, pos);
1014 	}
1015 
1016 	return 0;
1017 }
1018 
1019 static int vfio_ext_cap_len(struct vfio_pci_device *vdev, u16 ecap, u16 epos)
1020 {
1021 	struct pci_dev *pdev = vdev->pdev;
1022 	u8 byte;
1023 	u32 dword;
1024 	int ret;
1025 
1026 	switch (ecap) {
1027 	case PCI_EXT_CAP_ID_VNDR:
1028 		ret = pci_read_config_dword(pdev, epos + PCI_VSEC_HDR, &dword);
1029 		if (ret)
1030 			return pcibios_err_to_errno(ret);
1031 
1032 		return dword >> PCI_VSEC_HDR_LEN_SHIFT;
1033 	case PCI_EXT_CAP_ID_VC:
1034 	case PCI_EXT_CAP_ID_VC9:
1035 	case PCI_EXT_CAP_ID_MFVC:
1036 		return vfio_vc_cap_len(vdev, epos);
1037 	case PCI_EXT_CAP_ID_ACS:
1038 		ret = pci_read_config_byte(pdev, epos + PCI_ACS_CAP, &byte);
1039 		if (ret)
1040 			return pcibios_err_to_errno(ret);
1041 
1042 		if (byte & PCI_ACS_EC) {
1043 			int bits;
1044 
1045 			ret = pci_read_config_byte(pdev,
1046 						   epos + PCI_ACS_EGRESS_BITS,
1047 						   &byte);
1048 			if (ret)
1049 				return pcibios_err_to_errno(ret);
1050 
1051 			bits = byte ? round_up(byte, 32) : 256;
1052 			return 8 + (bits / 8);
1053 		}
1054 		return 8;
1055 
1056 	case PCI_EXT_CAP_ID_REBAR:
1057 		ret = pci_read_config_byte(pdev, epos + PCI_REBAR_CTRL, &byte);
1058 		if (ret)
1059 			return pcibios_err_to_errno(ret);
1060 
1061 		byte &= PCI_REBAR_CTRL_NBAR_MASK;
1062 		byte >>= PCI_REBAR_CTRL_NBAR_SHIFT;
1063 
1064 		return 4 + (byte * 8);
1065 	case PCI_EXT_CAP_ID_DPA:
1066 		ret = pci_read_config_byte(pdev, epos + PCI_DPA_CAP, &byte);
1067 		if (ret)
1068 			return pcibios_err_to_errno(ret);
1069 
1070 		byte &= PCI_DPA_CAP_SUBSTATE_MASK;
1071 		byte = round_up(byte + 1, 4);
1072 		return PCI_DPA_BASE_SIZEOF + byte;
1073 	case PCI_EXT_CAP_ID_TPH:
1074 		ret = pci_read_config_dword(pdev, epos + PCI_TPH_CAP, &dword);
1075 		if (ret)
1076 			return pcibios_err_to_errno(ret);
1077 
1078 		if ((dword & PCI_TPH_CAP_LOC_MASK) == PCI_TPH_LOC_CAP) {
1079 			int sts;
1080 
1081 			sts = byte & PCI_TPH_CAP_ST_MASK;
1082 			sts >>= PCI_TPH_CAP_ST_SHIFT;
1083 			return PCI_TPH_BASE_SIZEOF + round_up(sts * 2, 4);
1084 		}
1085 		return PCI_TPH_BASE_SIZEOF;
1086 	default:
1087 		pr_warn("%s: %s unknown length for pci ecap 0x%x@0x%x\n",
1088 			dev_name(&pdev->dev), __func__, ecap, epos);
1089 	}
1090 
1091 	return 0;
1092 }
1093 
1094 static int vfio_fill_vconfig_bytes(struct vfio_pci_device *vdev,
1095 				   int offset, int size)
1096 {
1097 	struct pci_dev *pdev = vdev->pdev;
1098 	int ret = 0;
1099 
1100 	/*
1101 	 * We try to read physical config space in the largest chunks
1102 	 * we can, assuming that all of the fields support dword access.
1103 	 * pci_save_state() makes this same assumption and seems to do ok.
1104 	 */
1105 	while (size) {
1106 		int filled;
1107 
1108 		if (size >= 4 && !(offset % 4)) {
1109 			__le32 *dwordp = (__le32 *)&vdev->vconfig[offset];
1110 			u32 dword;
1111 
1112 			ret = pci_read_config_dword(pdev, offset, &dword);
1113 			if (ret)
1114 				return ret;
1115 			*dwordp = cpu_to_le32(dword);
1116 			filled = 4;
1117 		} else if (size >= 2 && !(offset % 2)) {
1118 			__le16 *wordp = (__le16 *)&vdev->vconfig[offset];
1119 			u16 word;
1120 
1121 			ret = pci_read_config_word(pdev, offset, &word);
1122 			if (ret)
1123 				return ret;
1124 			*wordp = cpu_to_le16(word);
1125 			filled = 2;
1126 		} else {
1127 			u8 *byte = &vdev->vconfig[offset];
1128 			ret = pci_read_config_byte(pdev, offset, byte);
1129 			if (ret)
1130 				return ret;
1131 			filled = 1;
1132 		}
1133 
1134 		offset += filled;
1135 		size -= filled;
1136 	}
1137 
1138 	return ret;
1139 }
1140 
1141 static int vfio_cap_init(struct vfio_pci_device *vdev)
1142 {
1143 	struct pci_dev *pdev = vdev->pdev;
1144 	u8 *map = vdev->pci_config_map;
1145 	u16 status;
1146 	u8 pos, *prev, cap;
1147 	int loops, ret, caps = 0;
1148 
1149 	/* Any capabilities? */
1150 	ret = pci_read_config_word(pdev, PCI_STATUS, &status);
1151 	if (ret)
1152 		return ret;
1153 
1154 	if (!(status & PCI_STATUS_CAP_LIST))
1155 		return 0; /* Done */
1156 
1157 	ret = pci_read_config_byte(pdev, PCI_CAPABILITY_LIST, &pos);
1158 	if (ret)
1159 		return ret;
1160 
1161 	/* Mark the previous position in case we want to skip a capability */
1162 	prev = &vdev->vconfig[PCI_CAPABILITY_LIST];
1163 
1164 	/* We can bound our loop, capabilities are dword aligned */
1165 	loops = (PCI_CFG_SPACE_SIZE - PCI_STD_HEADER_SIZEOF) / PCI_CAP_SIZEOF;
1166 	while (pos && loops--) {
1167 		u8 next;
1168 		int i, len = 0;
1169 
1170 		ret = pci_read_config_byte(pdev, pos, &cap);
1171 		if (ret)
1172 			return ret;
1173 
1174 		ret = pci_read_config_byte(pdev,
1175 					   pos + PCI_CAP_LIST_NEXT, &next);
1176 		if (ret)
1177 			return ret;
1178 
1179 		if (cap <= PCI_CAP_ID_MAX) {
1180 			len = pci_cap_length[cap];
1181 			if (len == 0xFF) { /* Variable length */
1182 				len = vfio_cap_len(vdev, cap, pos);
1183 				if (len < 0)
1184 					return len;
1185 			}
1186 		}
1187 
1188 		if (!len) {
1189 			pr_info("%s: %s hiding cap 0x%x\n",
1190 				__func__, dev_name(&pdev->dev), cap);
1191 			*prev = next;
1192 			pos = next;
1193 			continue;
1194 		}
1195 
1196 		/* Sanity check, do we overlap other capabilities? */
1197 		for (i = 0; i < len; i += 4) {
1198 			if (likely(map[(pos + i) / 4] == PCI_CAP_ID_INVALID))
1199 				continue;
1200 
1201 			pr_warn("%s: %s pci config conflict @0x%x, was cap 0x%x now cap 0x%x\n",
1202 				__func__, dev_name(&pdev->dev),
1203 				pos + i, map[pos + i], cap);
1204 		}
1205 
1206 		memset(map + (pos / 4), cap, len / 4);
1207 		ret = vfio_fill_vconfig_bytes(vdev, pos, len);
1208 		if (ret)
1209 			return ret;
1210 
1211 		prev = &vdev->vconfig[pos + PCI_CAP_LIST_NEXT];
1212 		pos = next;
1213 		caps++;
1214 	}
1215 
1216 	/* If we didn't fill any capabilities, clear the status flag */
1217 	if (!caps) {
1218 		__le16 *vstatus = (__le16 *)&vdev->vconfig[PCI_STATUS];
1219 		*vstatus &= ~cpu_to_le16(PCI_STATUS_CAP_LIST);
1220 	}
1221 
1222 	return 0;
1223 }
1224 
1225 static int vfio_ecap_init(struct vfio_pci_device *vdev)
1226 {
1227 	struct pci_dev *pdev = vdev->pdev;
1228 	u8 *map = vdev->pci_config_map;
1229 	u16 epos;
1230 	__le32 *prev = NULL;
1231 	int loops, ret, ecaps = 0;
1232 
1233 	if (!vdev->extended_caps)
1234 		return 0;
1235 
1236 	epos = PCI_CFG_SPACE_SIZE;
1237 
1238 	loops = (pdev->cfg_size - PCI_CFG_SPACE_SIZE) / PCI_CAP_SIZEOF;
1239 
1240 	while (loops-- && epos >= PCI_CFG_SPACE_SIZE) {
1241 		u32 header;
1242 		u16 ecap;
1243 		int i, len = 0;
1244 		bool hidden = false;
1245 
1246 		ret = pci_read_config_dword(pdev, epos, &header);
1247 		if (ret)
1248 			return ret;
1249 
1250 		ecap = PCI_EXT_CAP_ID(header);
1251 
1252 		if (ecap <= PCI_EXT_CAP_ID_MAX) {
1253 			len = pci_ext_cap_length[ecap];
1254 			if (len == 0xFF) {
1255 				len = vfio_ext_cap_len(vdev, ecap, epos);
1256 				if (len < 0)
1257 					return ret;
1258 			}
1259 		}
1260 
1261 		if (!len) {
1262 			pr_info("%s: %s hiding ecap 0x%x@0x%x\n",
1263 				__func__, dev_name(&pdev->dev), ecap, epos);
1264 
1265 			/* If not the first in the chain, we can skip over it */
1266 			if (prev) {
1267 				u32 val = epos = PCI_EXT_CAP_NEXT(header);
1268 				*prev &= cpu_to_le32(~(0xffcU << 20));
1269 				*prev |= cpu_to_le32(val << 20);
1270 				continue;
1271 			}
1272 
1273 			/*
1274 			 * Otherwise, fill in a placeholder, the direct
1275 			 * readfn will virtualize this automatically
1276 			 */
1277 			len = PCI_CAP_SIZEOF;
1278 			hidden = true;
1279 		}
1280 
1281 		for (i = 0; i < len; i += 4) {
1282 			if (likely(map[(epos + i) / 4] == PCI_CAP_ID_INVALID))
1283 				continue;
1284 
1285 			pr_warn("%s: %s pci config conflict @0x%x, was ecap 0x%x now ecap 0x%x\n",
1286 				__func__, dev_name(&pdev->dev),
1287 				epos + i, map[epos + i], ecap);
1288 		}
1289 
1290 		/*
1291 		 * Even though ecap is 2 bytes, we're currently a long way
1292 		 * from exceeding 1 byte capabilities.  If we ever make it
1293 		 * up to 0xFF we'll need to up this to a two-byte, byte map.
1294 		 */
1295 		BUILD_BUG_ON(PCI_EXT_CAP_ID_MAX >= PCI_CAP_ID_INVALID);
1296 
1297 		memset(map + (epos / 4), ecap, len / 4);
1298 		ret = vfio_fill_vconfig_bytes(vdev, epos, len);
1299 		if (ret)
1300 			return ret;
1301 
1302 		/*
1303 		 * If we're just using this capability to anchor the list,
1304 		 * hide the real ID.  Only count real ecaps.  XXX PCI spec
1305 		 * indicates to use cap id = 0, version = 0, next = 0 if
1306 		 * ecaps are absent, hope users check all the way to next.
1307 		 */
1308 		if (hidden)
1309 			*(__le32 *)&vdev->vconfig[epos] &=
1310 				cpu_to_le32((0xffcU << 20));
1311 		else
1312 			ecaps++;
1313 
1314 		prev = (__le32 *)&vdev->vconfig[epos];
1315 		epos = PCI_EXT_CAP_NEXT(header);
1316 	}
1317 
1318 	if (!ecaps)
1319 		*(u32 *)&vdev->vconfig[PCI_CFG_SPACE_SIZE] = 0;
1320 
1321 	return 0;
1322 }
1323 
1324 /*
1325  * For each device we allocate a pci_config_map that indicates the
1326  * capability occupying each dword and thus the struct perm_bits we
1327  * use for read and write.  We also allocate a virtualized config
1328  * space which tracks reads and writes to bits that we emulate for
1329  * the user.  Initial values filled from device.
1330  *
1331  * Using shared stuct perm_bits between all vfio-pci devices saves
1332  * us from allocating cfg_size buffers for virt and write for every
1333  * device.  We could remove vconfig and allocate individual buffers
1334  * for each area requring emulated bits, but the array of pointers
1335  * would be comparable in size (at least for standard config space).
1336  */
1337 int vfio_config_init(struct vfio_pci_device *vdev)
1338 {
1339 	struct pci_dev *pdev = vdev->pdev;
1340 	u8 *map, *vconfig;
1341 	int ret;
1342 
1343 	/*
1344 	 * Config space, caps and ecaps are all dword aligned, so we can
1345 	 * use one byte per dword to record the type.
1346 	 */
1347 	map = kmalloc(pdev->cfg_size / 4, GFP_KERNEL);
1348 	if (!map)
1349 		return -ENOMEM;
1350 
1351 	vconfig = kmalloc(pdev->cfg_size, GFP_KERNEL);
1352 	if (!vconfig) {
1353 		kfree(map);
1354 		return -ENOMEM;
1355 	}
1356 
1357 	vdev->pci_config_map = map;
1358 	vdev->vconfig = vconfig;
1359 
1360 	memset(map, PCI_CAP_ID_BASIC, PCI_STD_HEADER_SIZEOF / 4);
1361 	memset(map + (PCI_STD_HEADER_SIZEOF / 4), PCI_CAP_ID_INVALID,
1362 	       (pdev->cfg_size - PCI_STD_HEADER_SIZEOF) / 4);
1363 
1364 	ret = vfio_fill_vconfig_bytes(vdev, 0, PCI_STD_HEADER_SIZEOF);
1365 	if (ret)
1366 		goto out;
1367 
1368 	vdev->bardirty = true;
1369 
1370 	/*
1371 	 * XXX can we just pci_load_saved_state/pci_restore_state?
1372 	 * may need to rebuild vconfig after that
1373 	 */
1374 
1375 	/* For restore after reset */
1376 	vdev->rbar[0] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_0]);
1377 	vdev->rbar[1] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_1]);
1378 	vdev->rbar[2] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_2]);
1379 	vdev->rbar[3] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_3]);
1380 	vdev->rbar[4] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_4]);
1381 	vdev->rbar[5] = le32_to_cpu(*(__le32 *)&vconfig[PCI_BASE_ADDRESS_5]);
1382 	vdev->rbar[6] = le32_to_cpu(*(__le32 *)&vconfig[PCI_ROM_ADDRESS]);
1383 
1384 	if (pdev->is_virtfn) {
1385 		*(__le16 *)&vconfig[PCI_VENDOR_ID] = cpu_to_le16(pdev->vendor);
1386 		*(__le16 *)&vconfig[PCI_DEVICE_ID] = cpu_to_le16(pdev->device);
1387 	}
1388 
1389 	ret = vfio_cap_init(vdev);
1390 	if (ret)
1391 		goto out;
1392 
1393 	ret = vfio_ecap_init(vdev);
1394 	if (ret)
1395 		goto out;
1396 
1397 	return 0;
1398 
1399 out:
1400 	kfree(map);
1401 	vdev->pci_config_map = NULL;
1402 	kfree(vconfig);
1403 	vdev->vconfig = NULL;
1404 	return pcibios_err_to_errno(ret);
1405 }
1406 
1407 void vfio_config_free(struct vfio_pci_device *vdev)
1408 {
1409 	kfree(vdev->vconfig);
1410 	vdev->vconfig = NULL;
1411 	kfree(vdev->pci_config_map);
1412 	vdev->pci_config_map = NULL;
1413 	kfree(vdev->msi_perm);
1414 	vdev->msi_perm = NULL;
1415 }
1416 
1417 static ssize_t vfio_config_do_rw(struct vfio_pci_device *vdev, char __user *buf,
1418 				 size_t count, loff_t *ppos, bool iswrite)
1419 {
1420 	struct pci_dev *pdev = vdev->pdev;
1421 	struct perm_bits *perm;
1422 	__le32 val = 0;
1423 	int cap_start = 0, offset;
1424 	u8 cap_id;
1425 	ssize_t ret = count;
1426 
1427 	if (*ppos < 0 || *ppos + count > pdev->cfg_size)
1428 		return -EFAULT;
1429 
1430 	/*
1431 	 * gcc can't seem to figure out we're a static function, only called
1432 	 * with count of 1/2/4 and hits copy_from_user_overflow without this.
1433 	 */
1434 	if (count > sizeof(val))
1435 		return -EINVAL;
1436 
1437 	cap_id = vdev->pci_config_map[*ppos / 4];
1438 
1439 	if (cap_id == PCI_CAP_ID_INVALID) {
1440 		if (iswrite)
1441 			return ret; /* drop */
1442 
1443 		/*
1444 		 * Per PCI spec 3.0, section 6.1, reads from reserved and
1445 		 * unimplemented registers return 0
1446 		 */
1447 		if (copy_to_user(buf, &val, count))
1448 			return -EFAULT;
1449 
1450 		return ret;
1451 	}
1452 
1453 	/*
1454 	 * All capabilities are minimum 4 bytes and aligned on dword
1455 	 * boundaries.  Since we don't support unaligned accesses, we're
1456 	 * only ever accessing a single capability.
1457 	 */
1458 	if (*ppos >= PCI_CFG_SPACE_SIZE) {
1459 		WARN_ON(cap_id > PCI_EXT_CAP_ID_MAX);
1460 
1461 		perm = &ecap_perms[cap_id];
1462 		cap_start = vfio_find_cap_start(vdev, *ppos);
1463 
1464 	} else {
1465 		WARN_ON(cap_id > PCI_CAP_ID_MAX);
1466 
1467 		perm = &cap_perms[cap_id];
1468 
1469 		if (cap_id == PCI_CAP_ID_MSI)
1470 			perm = vdev->msi_perm;
1471 
1472 		if (cap_id > PCI_CAP_ID_BASIC)
1473 			cap_start = vfio_find_cap_start(vdev, *ppos);
1474 	}
1475 
1476 	WARN_ON(!cap_start && cap_id != PCI_CAP_ID_BASIC);
1477 	WARN_ON(cap_start > *ppos);
1478 
1479 	offset = *ppos - cap_start;
1480 
1481 	if (iswrite) {
1482 		if (!perm->writefn)
1483 			return ret;
1484 
1485 		if (copy_from_user(&val, buf, count))
1486 			return -EFAULT;
1487 
1488 		ret = perm->writefn(vdev, *ppos, count, perm, offset, val);
1489 	} else {
1490 		if (perm->readfn) {
1491 			ret = perm->readfn(vdev, *ppos, count,
1492 					   perm, offset, &val);
1493 			if (ret < 0)
1494 				return ret;
1495 		}
1496 
1497 		if (copy_to_user(buf, &val, count))
1498 			return -EFAULT;
1499 	}
1500 
1501 	return ret;
1502 }
1503 
1504 ssize_t vfio_pci_config_readwrite(struct vfio_pci_device *vdev,
1505 				  char __user *buf, size_t count,
1506 				  loff_t *ppos, bool iswrite)
1507 {
1508 	size_t done = 0;
1509 	int ret = 0;
1510 	loff_t pos = *ppos;
1511 
1512 	pos &= VFIO_PCI_OFFSET_MASK;
1513 
1514 	/*
1515 	 * We want to both keep the access size the caller users as well as
1516 	 * support reading large chunks of config space in a single call.
1517 	 * PCI doesn't support unaligned accesses, so we can safely break
1518 	 * those apart.
1519 	 */
1520 	while (count) {
1521 		if (count >= 4 && !(pos % 4))
1522 			ret = vfio_config_do_rw(vdev, buf, 4, &pos, iswrite);
1523 		else if (count >= 2 && !(pos % 2))
1524 			ret = vfio_config_do_rw(vdev, buf, 2, &pos, iswrite);
1525 		else
1526 			ret = vfio_config_do_rw(vdev, buf, 1, &pos, iswrite);
1527 
1528 		if (ret < 0)
1529 			return ret;
1530 
1531 		count -= ret;
1532 		done += ret;
1533 		buf += ret;
1534 		pos += ret;
1535 	}
1536 
1537 	*ppos += done;
1538 
1539 	return done;
1540 }
1541