1 /*
2  * Support PCI/PCIe on PowerNV platforms
3  *
4  * Copyright 2011 Benjamin Herrenschmidt, IBM Corp.
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; either version
9  * 2 of the License, or (at your option) any later version.
10  */
11 
12 #undef DEBUG
13 
14 #include <linux/kernel.h>
15 #include <linux/pci.h>
16 #include <linux/crash_dump.h>
17 #include <linux/debugfs.h>
18 #include <linux/delay.h>
19 #include <linux/string.h>
20 #include <linux/init.h>
21 #include <linux/bootmem.h>
22 #include <linux/irq.h>
23 #include <linux/io.h>
24 #include <linux/msi.h>
25 #include <linux/memblock.h>
26 #include <linux/iommu.h>
27 #include <linux/rculist.h>
28 #include <linux/sizes.h>
29 
30 #include <asm/sections.h>
31 #include <asm/io.h>
32 #include <asm/prom.h>
33 #include <asm/pci-bridge.h>
34 #include <asm/machdep.h>
35 #include <asm/msi_bitmap.h>
36 #include <asm/ppc-pci.h>
37 #include <asm/opal.h>
38 #include <asm/iommu.h>
39 #include <asm/tce.h>
40 #include <asm/xics.h>
41 #include <asm/debug.h>
42 #include <asm/firmware.h>
43 #include <asm/pnv-pci.h>
44 #include <asm/mmzone.h>
45 
46 #include <misc/cxl-base.h>
47 
48 #include "powernv.h"
49 #include "pci.h"
50 
51 #define PNV_IODA1_M64_NUM	16	/* Number of M64 BARs	*/
52 #define PNV_IODA1_M64_SEGS	8	/* Segments per M64 BAR	*/
53 #define PNV_IODA1_DMA32_SEGSIZE	0x10000000
54 
55 #define POWERNV_IOMMU_DEFAULT_LEVELS	1
56 #define POWERNV_IOMMU_MAX_LEVELS	5
57 
58 static const char * const pnv_phb_names[] = { "IODA1", "IODA2", "NPU" };
59 static void pnv_pci_ioda2_table_free_pages(struct iommu_table *tbl);
60 
61 void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
62 			    const char *fmt, ...)
63 {
64 	struct va_format vaf;
65 	va_list args;
66 	char pfix[32];
67 
68 	va_start(args, fmt);
69 
70 	vaf.fmt = fmt;
71 	vaf.va = &args;
72 
73 	if (pe->flags & PNV_IODA_PE_DEV)
74 		strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix));
75 	else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
76 		sprintf(pfix, "%04x:%02x     ",
77 			pci_domain_nr(pe->pbus), pe->pbus->number);
78 #ifdef CONFIG_PCI_IOV
79 	else if (pe->flags & PNV_IODA_PE_VF)
80 		sprintf(pfix, "%04x:%02x:%2x.%d",
81 			pci_domain_nr(pe->parent_dev->bus),
82 			(pe->rid & 0xff00) >> 8,
83 			PCI_SLOT(pe->rid), PCI_FUNC(pe->rid));
84 #endif /* CONFIG_PCI_IOV*/
85 
86 	printk("%spci %s: [PE# %.3d] %pV",
87 	       level, pfix, pe->pe_number, &vaf);
88 
89 	va_end(args);
90 }
91 
92 static bool pnv_iommu_bypass_disabled __read_mostly;
93 
94 static int __init iommu_setup(char *str)
95 {
96 	if (!str)
97 		return -EINVAL;
98 
99 	while (*str) {
100 		if (!strncmp(str, "nobypass", 8)) {
101 			pnv_iommu_bypass_disabled = true;
102 			pr_info("PowerNV: IOMMU bypass window disabled.\n");
103 			break;
104 		}
105 		str += strcspn(str, ",");
106 		if (*str == ',')
107 			str++;
108 	}
109 
110 	return 0;
111 }
112 early_param("iommu", iommu_setup);
113 
114 static inline bool pnv_pci_is_mem_pref_64(unsigned long flags)
115 {
116 	return ((flags & (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH)) ==
117 		(IORESOURCE_MEM_64 | IORESOURCE_PREFETCH));
118 }
119 
120 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
121 {
122 	phb->ioda.pe_array[pe_no].phb = phb;
123 	phb->ioda.pe_array[pe_no].pe_number = pe_no;
124 
125 	return &phb->ioda.pe_array[pe_no];
126 }
127 
128 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
129 {
130 	if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) {
131 		pr_warn("%s: Invalid PE %d on PHB#%x\n",
132 			__func__, pe_no, phb->hose->global_number);
133 		return;
134 	}
135 
136 	if (test_and_set_bit(pe_no, phb->ioda.pe_alloc))
137 		pr_debug("%s: PE %d was reserved on PHB#%x\n",
138 			 __func__, pe_no, phb->hose->global_number);
139 
140 	pnv_ioda_init_pe(phb, pe_no);
141 }
142 
143 static struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb)
144 {
145 	unsigned long pe = phb->ioda.total_pe_num - 1;
146 
147 	for (pe = phb->ioda.total_pe_num - 1; pe >= 0; pe--) {
148 		if (!test_and_set_bit(pe, phb->ioda.pe_alloc))
149 			return pnv_ioda_init_pe(phb, pe);
150 	}
151 
152 	return NULL;
153 }
154 
155 static void pnv_ioda_free_pe(struct pnv_ioda_pe *pe)
156 {
157 	struct pnv_phb *phb = pe->phb;
158 
159 	WARN_ON(pe->pdev);
160 
161 	memset(pe, 0, sizeof(struct pnv_ioda_pe));
162 	clear_bit(pe->pe_number, phb->ioda.pe_alloc);
163 }
164 
165 /* The default M64 BAR is shared by all PEs */
166 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
167 {
168 	const char *desc;
169 	struct resource *r;
170 	s64 rc;
171 
172 	/* Configure the default M64 BAR */
173 	rc = opal_pci_set_phb_mem_window(phb->opal_id,
174 					 OPAL_M64_WINDOW_TYPE,
175 					 phb->ioda.m64_bar_idx,
176 					 phb->ioda.m64_base,
177 					 0, /* unused */
178 					 phb->ioda.m64_size);
179 	if (rc != OPAL_SUCCESS) {
180 		desc = "configuring";
181 		goto fail;
182 	}
183 
184 	/* Enable the default M64 BAR */
185 	rc = opal_pci_phb_mmio_enable(phb->opal_id,
186 				      OPAL_M64_WINDOW_TYPE,
187 				      phb->ioda.m64_bar_idx,
188 				      OPAL_ENABLE_M64_SPLIT);
189 	if (rc != OPAL_SUCCESS) {
190 		desc = "enabling";
191 		goto fail;
192 	}
193 
194 	/*
195 	 * Exclude the segments for reserved and root bus PE, which
196 	 * are first or last two PEs.
197 	 */
198 	r = &phb->hose->mem_resources[1];
199 	if (phb->ioda.reserved_pe_idx == 0)
200 		r->start += (2 * phb->ioda.m64_segsize);
201 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
202 		r->end -= (2 * phb->ioda.m64_segsize);
203 	else
204 		pr_warn("  Cannot strip M64 segment for reserved PE#%d\n",
205 			phb->ioda.reserved_pe_idx);
206 
207 	return 0;
208 
209 fail:
210 	pr_warn("  Failure %lld %s M64 BAR#%d\n",
211 		rc, desc, phb->ioda.m64_bar_idx);
212 	opal_pci_phb_mmio_enable(phb->opal_id,
213 				 OPAL_M64_WINDOW_TYPE,
214 				 phb->ioda.m64_bar_idx,
215 				 OPAL_DISABLE_M64);
216 	return -EIO;
217 }
218 
219 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev,
220 					 unsigned long *pe_bitmap)
221 {
222 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
223 	struct pnv_phb *phb = hose->private_data;
224 	struct resource *r;
225 	resource_size_t base, sgsz, start, end;
226 	int segno, i;
227 
228 	base = phb->ioda.m64_base;
229 	sgsz = phb->ioda.m64_segsize;
230 	for (i = 0; i <= PCI_ROM_RESOURCE; i++) {
231 		r = &pdev->resource[i];
232 		if (!r->parent || !pnv_pci_is_mem_pref_64(r->flags))
233 			continue;
234 
235 		start = _ALIGN_DOWN(r->start - base, sgsz);
236 		end = _ALIGN_UP(r->end - base, sgsz);
237 		for (segno = start / sgsz; segno < end / sgsz; segno++) {
238 			if (pe_bitmap)
239 				set_bit(segno, pe_bitmap);
240 			else
241 				pnv_ioda_reserve_pe(phb, segno);
242 		}
243 	}
244 }
245 
246 static int pnv_ioda1_init_m64(struct pnv_phb *phb)
247 {
248 	struct resource *r;
249 	int index;
250 
251 	/*
252 	 * There are 16 M64 BARs, each of which has 8 segments. So
253 	 * there are as many M64 segments as the maximum number of
254 	 * PEs, which is 128.
255 	 */
256 	for (index = 0; index < PNV_IODA1_M64_NUM; index++) {
257 		unsigned long base, segsz = phb->ioda.m64_segsize;
258 		int64_t rc;
259 
260 		base = phb->ioda.m64_base +
261 		       index * PNV_IODA1_M64_SEGS * segsz;
262 		rc = opal_pci_set_phb_mem_window(phb->opal_id,
263 				OPAL_M64_WINDOW_TYPE, index, base, 0,
264 				PNV_IODA1_M64_SEGS * segsz);
265 		if (rc != OPAL_SUCCESS) {
266 			pr_warn("  Error %lld setting M64 PHB#%d-BAR#%d\n",
267 				rc, phb->hose->global_number, index);
268 			goto fail;
269 		}
270 
271 		rc = opal_pci_phb_mmio_enable(phb->opal_id,
272 				OPAL_M64_WINDOW_TYPE, index,
273 				OPAL_ENABLE_M64_SPLIT);
274 		if (rc != OPAL_SUCCESS) {
275 			pr_warn("  Error %lld enabling M64 PHB#%d-BAR#%d\n",
276 				rc, phb->hose->global_number, index);
277 			goto fail;
278 		}
279 	}
280 
281 	/*
282 	 * Exclude the segments for reserved and root bus PE, which
283 	 * are first or last two PEs.
284 	 */
285 	r = &phb->hose->mem_resources[1];
286 	if (phb->ioda.reserved_pe_idx == 0)
287 		r->start += (2 * phb->ioda.m64_segsize);
288 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
289 		r->end -= (2 * phb->ioda.m64_segsize);
290 	else
291 		WARN(1, "Wrong reserved PE#%d on PHB#%d\n",
292 		     phb->ioda.reserved_pe_idx, phb->hose->global_number);
293 
294 	return 0;
295 
296 fail:
297 	for ( ; index >= 0; index--)
298 		opal_pci_phb_mmio_enable(phb->opal_id,
299 			OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64);
300 
301 	return -EIO;
302 }
303 
304 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus,
305 				    unsigned long *pe_bitmap,
306 				    bool all)
307 {
308 	struct pci_dev *pdev;
309 
310 	list_for_each_entry(pdev, &bus->devices, bus_list) {
311 		pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap);
312 
313 		if (all && pdev->subordinate)
314 			pnv_ioda_reserve_m64_pe(pdev->subordinate,
315 						pe_bitmap, all);
316 	}
317 }
318 
319 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all)
320 {
321 	struct pci_controller *hose = pci_bus_to_host(bus);
322 	struct pnv_phb *phb = hose->private_data;
323 	struct pnv_ioda_pe *master_pe, *pe;
324 	unsigned long size, *pe_alloc;
325 	int i;
326 
327 	/* Root bus shouldn't use M64 */
328 	if (pci_is_root_bus(bus))
329 		return NULL;
330 
331 	/* Allocate bitmap */
332 	size = _ALIGN_UP(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
333 	pe_alloc = kzalloc(size, GFP_KERNEL);
334 	if (!pe_alloc) {
335 		pr_warn("%s: Out of memory !\n",
336 			__func__);
337 		return NULL;
338 	}
339 
340 	/* Figure out reserved PE numbers by the PE */
341 	pnv_ioda_reserve_m64_pe(bus, pe_alloc, all);
342 
343 	/*
344 	 * the current bus might not own M64 window and that's all
345 	 * contributed by its child buses. For the case, we needn't
346 	 * pick M64 dependent PE#.
347 	 */
348 	if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) {
349 		kfree(pe_alloc);
350 		return NULL;
351 	}
352 
353 	/*
354 	 * Figure out the master PE and put all slave PEs to master
355 	 * PE's list to form compound PE.
356 	 */
357 	master_pe = NULL;
358 	i = -1;
359 	while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) <
360 		phb->ioda.total_pe_num) {
361 		pe = &phb->ioda.pe_array[i];
362 
363 		phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number;
364 		if (!master_pe) {
365 			pe->flags |= PNV_IODA_PE_MASTER;
366 			INIT_LIST_HEAD(&pe->slaves);
367 			master_pe = pe;
368 		} else {
369 			pe->flags |= PNV_IODA_PE_SLAVE;
370 			pe->master = master_pe;
371 			list_add_tail(&pe->list, &master_pe->slaves);
372 		}
373 
374 		/*
375 		 * P7IOC supports M64DT, which helps mapping M64 segment
376 		 * to one particular PE#. However, PHB3 has fixed mapping
377 		 * between M64 segment and PE#. In order to have same logic
378 		 * for P7IOC and PHB3, we enforce fixed mapping between M64
379 		 * segment and PE# on P7IOC.
380 		 */
381 		if (phb->type == PNV_PHB_IODA1) {
382 			int64_t rc;
383 
384 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
385 					pe->pe_number, OPAL_M64_WINDOW_TYPE,
386 					pe->pe_number / PNV_IODA1_M64_SEGS,
387 					pe->pe_number % PNV_IODA1_M64_SEGS);
388 			if (rc != OPAL_SUCCESS)
389 				pr_warn("%s: Error %lld mapping M64 for PHB#%d-PE#%d\n",
390 					__func__, rc, phb->hose->global_number,
391 					pe->pe_number);
392 		}
393 	}
394 
395 	kfree(pe_alloc);
396 	return master_pe;
397 }
398 
399 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
400 {
401 	struct pci_controller *hose = phb->hose;
402 	struct device_node *dn = hose->dn;
403 	struct resource *res;
404 	u32 m64_range[2], i;
405 	const u32 *r;
406 	u64 pci_addr;
407 
408 	if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) {
409 		pr_info("  Not support M64 window\n");
410 		return;
411 	}
412 
413 	if (!firmware_has_feature(FW_FEATURE_OPAL)) {
414 		pr_info("  Firmware too old to support M64 window\n");
415 		return;
416 	}
417 
418 	r = of_get_property(dn, "ibm,opal-m64-window", NULL);
419 	if (!r) {
420 		pr_info("  No <ibm,opal-m64-window> on %s\n",
421 			dn->full_name);
422 		return;
423 	}
424 
425 	/*
426 	 * Find the available M64 BAR range and pickup the last one for
427 	 * covering the whole 64-bits space. We support only one range.
428 	 */
429 	if (of_property_read_u32_array(dn, "ibm,opal-available-m64-ranges",
430 				       m64_range, 2)) {
431 		/* In absence of the property, assume 0..15 */
432 		m64_range[0] = 0;
433 		m64_range[1] = 16;
434 	}
435 	/* We only support 64 bits in our allocator */
436 	if (m64_range[1] > 63) {
437 		pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n",
438 			__func__, m64_range[1], phb->hose->global_number);
439 		m64_range[1] = 63;
440 	}
441 	/* Empty range, no m64 */
442 	if (m64_range[1] <= m64_range[0]) {
443 		pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n",
444 			__func__, phb->hose->global_number);
445 		return;
446 	}
447 
448 	/* Configure M64 informations */
449 	res = &hose->mem_resources[1];
450 	res->name = dn->full_name;
451 	res->start = of_translate_address(dn, r + 2);
452 	res->end = res->start + of_read_number(r + 4, 2) - 1;
453 	res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
454 	pci_addr = of_read_number(r, 2);
455 	hose->mem_offset[1] = res->start - pci_addr;
456 
457 	phb->ioda.m64_size = resource_size(res);
458 	phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num;
459 	phb->ioda.m64_base = pci_addr;
460 
461 	/* This lines up nicely with the display from processing OF ranges */
462 	pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n",
463 		res->start, res->end, pci_addr, m64_range[0],
464 		m64_range[0] + m64_range[1] - 1);
465 
466 	/* Mark all M64 used up by default */
467 	phb->ioda.m64_bar_alloc = (unsigned long)-1;
468 
469 	/* Use last M64 BAR to cover M64 window */
470 	m64_range[1]--;
471 	phb->ioda.m64_bar_idx = m64_range[0] + m64_range[1];
472 
473 	pr_info(" Using M64 #%d as default window\n", phb->ioda.m64_bar_idx);
474 
475 	/* Mark remaining ones free */
476 	for (i = m64_range[0]; i < m64_range[1]; i++)
477 		clear_bit(i, &phb->ioda.m64_bar_alloc);
478 
479 	/*
480 	 * Setup init functions for M64 based on IODA version, IODA3 uses
481 	 * the IODA2 code.
482 	 */
483 	if (phb->type == PNV_PHB_IODA1)
484 		phb->init_m64 = pnv_ioda1_init_m64;
485 	else
486 		phb->init_m64 = pnv_ioda2_init_m64;
487 	phb->reserve_m64_pe = pnv_ioda_reserve_m64_pe;
488 	phb->pick_m64_pe = pnv_ioda_pick_m64_pe;
489 }
490 
491 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
492 {
493 	struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
494 	struct pnv_ioda_pe *slave;
495 	s64 rc;
496 
497 	/* Fetch master PE */
498 	if (pe->flags & PNV_IODA_PE_SLAVE) {
499 		pe = pe->master;
500 		if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
501 			return;
502 
503 		pe_no = pe->pe_number;
504 	}
505 
506 	/* Freeze master PE */
507 	rc = opal_pci_eeh_freeze_set(phb->opal_id,
508 				     pe_no,
509 				     OPAL_EEH_ACTION_SET_FREEZE_ALL);
510 	if (rc != OPAL_SUCCESS) {
511 		pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
512 			__func__, rc, phb->hose->global_number, pe_no);
513 		return;
514 	}
515 
516 	/* Freeze slave PEs */
517 	if (!(pe->flags & PNV_IODA_PE_MASTER))
518 		return;
519 
520 	list_for_each_entry(slave, &pe->slaves, list) {
521 		rc = opal_pci_eeh_freeze_set(phb->opal_id,
522 					     slave->pe_number,
523 					     OPAL_EEH_ACTION_SET_FREEZE_ALL);
524 		if (rc != OPAL_SUCCESS)
525 			pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
526 				__func__, rc, phb->hose->global_number,
527 				slave->pe_number);
528 	}
529 }
530 
531 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
532 {
533 	struct pnv_ioda_pe *pe, *slave;
534 	s64 rc;
535 
536 	/* Find master PE */
537 	pe = &phb->ioda.pe_array[pe_no];
538 	if (pe->flags & PNV_IODA_PE_SLAVE) {
539 		pe = pe->master;
540 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
541 		pe_no = pe->pe_number;
542 	}
543 
544 	/* Clear frozen state for master PE */
545 	rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
546 	if (rc != OPAL_SUCCESS) {
547 		pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
548 			__func__, rc, opt, phb->hose->global_number, pe_no);
549 		return -EIO;
550 	}
551 
552 	if (!(pe->flags & PNV_IODA_PE_MASTER))
553 		return 0;
554 
555 	/* Clear frozen state for slave PEs */
556 	list_for_each_entry(slave, &pe->slaves, list) {
557 		rc = opal_pci_eeh_freeze_clear(phb->opal_id,
558 					     slave->pe_number,
559 					     opt);
560 		if (rc != OPAL_SUCCESS) {
561 			pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
562 				__func__, rc, opt, phb->hose->global_number,
563 				slave->pe_number);
564 			return -EIO;
565 		}
566 	}
567 
568 	return 0;
569 }
570 
571 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
572 {
573 	struct pnv_ioda_pe *slave, *pe;
574 	u8 fstate, state;
575 	__be16 pcierr;
576 	s64 rc;
577 
578 	/* Sanity check on PE number */
579 	if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num)
580 		return OPAL_EEH_STOPPED_PERM_UNAVAIL;
581 
582 	/*
583 	 * Fetch the master PE and the PE instance might be
584 	 * not initialized yet.
585 	 */
586 	pe = &phb->ioda.pe_array[pe_no];
587 	if (pe->flags & PNV_IODA_PE_SLAVE) {
588 		pe = pe->master;
589 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
590 		pe_no = pe->pe_number;
591 	}
592 
593 	/* Check the master PE */
594 	rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
595 					&state, &pcierr, NULL);
596 	if (rc != OPAL_SUCCESS) {
597 		pr_warn("%s: Failure %lld getting "
598 			"PHB#%x-PE#%x state\n",
599 			__func__, rc,
600 			phb->hose->global_number, pe_no);
601 		return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
602 	}
603 
604 	/* Check the slave PE */
605 	if (!(pe->flags & PNV_IODA_PE_MASTER))
606 		return state;
607 
608 	list_for_each_entry(slave, &pe->slaves, list) {
609 		rc = opal_pci_eeh_freeze_status(phb->opal_id,
610 						slave->pe_number,
611 						&fstate,
612 						&pcierr,
613 						NULL);
614 		if (rc != OPAL_SUCCESS) {
615 			pr_warn("%s: Failure %lld getting "
616 				"PHB#%x-PE#%x state\n",
617 				__func__, rc,
618 				phb->hose->global_number, slave->pe_number);
619 			return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
620 		}
621 
622 		/*
623 		 * Override the result based on the ascending
624 		 * priority.
625 		 */
626 		if (fstate > state)
627 			state = fstate;
628 	}
629 
630 	return state;
631 }
632 
633 /* Currently those 2 are only used when MSIs are enabled, this will change
634  * but in the meantime, we need to protect them to avoid warnings
635  */
636 #ifdef CONFIG_PCI_MSI
637 struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
638 {
639 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
640 	struct pnv_phb *phb = hose->private_data;
641 	struct pci_dn *pdn = pci_get_pdn(dev);
642 
643 	if (!pdn)
644 		return NULL;
645 	if (pdn->pe_number == IODA_INVALID_PE)
646 		return NULL;
647 	return &phb->ioda.pe_array[pdn->pe_number];
648 }
649 #endif /* CONFIG_PCI_MSI */
650 
651 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
652 				  struct pnv_ioda_pe *parent,
653 				  struct pnv_ioda_pe *child,
654 				  bool is_add)
655 {
656 	const char *desc = is_add ? "adding" : "removing";
657 	uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
658 			      OPAL_REMOVE_PE_FROM_DOMAIN;
659 	struct pnv_ioda_pe *slave;
660 	long rc;
661 
662 	/* Parent PE affects child PE */
663 	rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
664 				child->pe_number, op);
665 	if (rc != OPAL_SUCCESS) {
666 		pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
667 			rc, desc);
668 		return -ENXIO;
669 	}
670 
671 	if (!(child->flags & PNV_IODA_PE_MASTER))
672 		return 0;
673 
674 	/* Compound case: parent PE affects slave PEs */
675 	list_for_each_entry(slave, &child->slaves, list) {
676 		rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
677 					slave->pe_number, op);
678 		if (rc != OPAL_SUCCESS) {
679 			pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
680 				rc, desc);
681 			return -ENXIO;
682 		}
683 	}
684 
685 	return 0;
686 }
687 
688 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
689 			      struct pnv_ioda_pe *pe,
690 			      bool is_add)
691 {
692 	struct pnv_ioda_pe *slave;
693 	struct pci_dev *pdev = NULL;
694 	int ret;
695 
696 	/*
697 	 * Clear PE frozen state. If it's master PE, we need
698 	 * clear slave PE frozen state as well.
699 	 */
700 	if (is_add) {
701 		opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
702 					  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
703 		if (pe->flags & PNV_IODA_PE_MASTER) {
704 			list_for_each_entry(slave, &pe->slaves, list)
705 				opal_pci_eeh_freeze_clear(phb->opal_id,
706 							  slave->pe_number,
707 							  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
708 		}
709 	}
710 
711 	/*
712 	 * Associate PE in PELT. We need add the PE into the
713 	 * corresponding PELT-V as well. Otherwise, the error
714 	 * originated from the PE might contribute to other
715 	 * PEs.
716 	 */
717 	ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
718 	if (ret)
719 		return ret;
720 
721 	/* For compound PEs, any one affects all of them */
722 	if (pe->flags & PNV_IODA_PE_MASTER) {
723 		list_for_each_entry(slave, &pe->slaves, list) {
724 			ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
725 			if (ret)
726 				return ret;
727 		}
728 	}
729 
730 	if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
731 		pdev = pe->pbus->self;
732 	else if (pe->flags & PNV_IODA_PE_DEV)
733 		pdev = pe->pdev->bus->self;
734 #ifdef CONFIG_PCI_IOV
735 	else if (pe->flags & PNV_IODA_PE_VF)
736 		pdev = pe->parent_dev;
737 #endif /* CONFIG_PCI_IOV */
738 	while (pdev) {
739 		struct pci_dn *pdn = pci_get_pdn(pdev);
740 		struct pnv_ioda_pe *parent;
741 
742 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
743 			parent = &phb->ioda.pe_array[pdn->pe_number];
744 			ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
745 			if (ret)
746 				return ret;
747 		}
748 
749 		pdev = pdev->bus->self;
750 	}
751 
752 	return 0;
753 }
754 
755 static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
756 {
757 	struct pci_dev *parent;
758 	uint8_t bcomp, dcomp, fcomp;
759 	int64_t rc;
760 	long rid_end, rid;
761 
762 	/* Currently, we just deconfigure VF PE. Bus PE will always there.*/
763 	if (pe->pbus) {
764 		int count;
765 
766 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
767 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
768 		parent = pe->pbus->self;
769 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
770 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
771 		else
772 			count = 1;
773 
774 		switch(count) {
775 		case  1: bcomp = OpalPciBusAll;         break;
776 		case  2: bcomp = OpalPciBus7Bits;       break;
777 		case  4: bcomp = OpalPciBus6Bits;       break;
778 		case  8: bcomp = OpalPciBus5Bits;       break;
779 		case 16: bcomp = OpalPciBus4Bits;       break;
780 		case 32: bcomp = OpalPciBus3Bits;       break;
781 		default:
782 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
783 			        count);
784 			/* Do an exact match only */
785 			bcomp = OpalPciBusAll;
786 		}
787 		rid_end = pe->rid + (count << 8);
788 	} else {
789 #ifdef CONFIG_PCI_IOV
790 		if (pe->flags & PNV_IODA_PE_VF)
791 			parent = pe->parent_dev;
792 		else
793 #endif
794 			parent = pe->pdev->bus->self;
795 		bcomp = OpalPciBusAll;
796 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
797 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
798 		rid_end = pe->rid + 1;
799 	}
800 
801 	/* Clear the reverse map */
802 	for (rid = pe->rid; rid < rid_end; rid++)
803 		phb->ioda.pe_rmap[rid] = IODA_INVALID_PE;
804 
805 	/* Release from all parents PELT-V */
806 	while (parent) {
807 		struct pci_dn *pdn = pci_get_pdn(parent);
808 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
809 			rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
810 						pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
811 			/* XXX What to do in case of error ? */
812 		}
813 		parent = parent->bus->self;
814 	}
815 
816 	opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
817 				  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
818 
819 	/* Disassociate PE in PELT */
820 	rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
821 				pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
822 	if (rc)
823 		pe_warn(pe, "OPAL error %ld remove self from PELTV\n", rc);
824 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
825 			     bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
826 	if (rc)
827 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
828 
829 	pe->pbus = NULL;
830 	pe->pdev = NULL;
831 #ifdef CONFIG_PCI_IOV
832 	pe->parent_dev = NULL;
833 #endif
834 
835 	return 0;
836 }
837 
838 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
839 {
840 	struct pci_dev *parent;
841 	uint8_t bcomp, dcomp, fcomp;
842 	long rc, rid_end, rid;
843 
844 	/* Bus validation ? */
845 	if (pe->pbus) {
846 		int count;
847 
848 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
849 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
850 		parent = pe->pbus->self;
851 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
852 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
853 		else
854 			count = 1;
855 
856 		switch(count) {
857 		case  1: bcomp = OpalPciBusAll;		break;
858 		case  2: bcomp = OpalPciBus7Bits;	break;
859 		case  4: bcomp = OpalPciBus6Bits;	break;
860 		case  8: bcomp = OpalPciBus5Bits;	break;
861 		case 16: bcomp = OpalPciBus4Bits;	break;
862 		case 32: bcomp = OpalPciBus3Bits;	break;
863 		default:
864 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
865 			        count);
866 			/* Do an exact match only */
867 			bcomp = OpalPciBusAll;
868 		}
869 		rid_end = pe->rid + (count << 8);
870 	} else {
871 #ifdef CONFIG_PCI_IOV
872 		if (pe->flags & PNV_IODA_PE_VF)
873 			parent = pe->parent_dev;
874 		else
875 #endif /* CONFIG_PCI_IOV */
876 			parent = pe->pdev->bus->self;
877 		bcomp = OpalPciBusAll;
878 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
879 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
880 		rid_end = pe->rid + 1;
881 	}
882 
883 	/*
884 	 * Associate PE in PELT. We need add the PE into the
885 	 * corresponding PELT-V as well. Otherwise, the error
886 	 * originated from the PE might contribute to other
887 	 * PEs.
888 	 */
889 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
890 			     bcomp, dcomp, fcomp, OPAL_MAP_PE);
891 	if (rc) {
892 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
893 		return -ENXIO;
894 	}
895 
896 	/*
897 	 * Configure PELTV. NPUs don't have a PELTV table so skip
898 	 * configuration on them.
899 	 */
900 	if (phb->type != PNV_PHB_NPU)
901 		pnv_ioda_set_peltv(phb, pe, true);
902 
903 	/* Setup reverse map */
904 	for (rid = pe->rid; rid < rid_end; rid++)
905 		phb->ioda.pe_rmap[rid] = pe->pe_number;
906 
907 	/* Setup one MVTs on IODA1 */
908 	if (phb->type != PNV_PHB_IODA1) {
909 		pe->mve_number = 0;
910 		goto out;
911 	}
912 
913 	pe->mve_number = pe->pe_number;
914 	rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
915 	if (rc != OPAL_SUCCESS) {
916 		pe_err(pe, "OPAL error %ld setting up MVE %d\n",
917 		       rc, pe->mve_number);
918 		pe->mve_number = -1;
919 	} else {
920 		rc = opal_pci_set_mve_enable(phb->opal_id,
921 					     pe->mve_number, OPAL_ENABLE_MVE);
922 		if (rc) {
923 			pe_err(pe, "OPAL error %ld enabling MVE %d\n",
924 			       rc, pe->mve_number);
925 			pe->mve_number = -1;
926 		}
927 	}
928 
929 out:
930 	return 0;
931 }
932 
933 #ifdef CONFIG_PCI_IOV
934 static int pnv_pci_vf_resource_shift(struct pci_dev *dev, int offset)
935 {
936 	struct pci_dn *pdn = pci_get_pdn(dev);
937 	int i;
938 	struct resource *res, res2;
939 	resource_size_t size;
940 	u16 num_vfs;
941 
942 	if (!dev->is_physfn)
943 		return -EINVAL;
944 
945 	/*
946 	 * "offset" is in VFs.  The M64 windows are sized so that when they
947 	 * are segmented, each segment is the same size as the IOV BAR.
948 	 * Each segment is in a separate PE, and the high order bits of the
949 	 * address are the PE number.  Therefore, each VF's BAR is in a
950 	 * separate PE, and changing the IOV BAR start address changes the
951 	 * range of PEs the VFs are in.
952 	 */
953 	num_vfs = pdn->num_vfs;
954 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
955 		res = &dev->resource[i + PCI_IOV_RESOURCES];
956 		if (!res->flags || !res->parent)
957 			continue;
958 
959 		/*
960 		 * The actual IOV BAR range is determined by the start address
961 		 * and the actual size for num_vfs VFs BAR.  This check is to
962 		 * make sure that after shifting, the range will not overlap
963 		 * with another device.
964 		 */
965 		size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
966 		res2.flags = res->flags;
967 		res2.start = res->start + (size * offset);
968 		res2.end = res2.start + (size * num_vfs) - 1;
969 
970 		if (res2.end > res->end) {
971 			dev_err(&dev->dev, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
972 				i, &res2, res, num_vfs, offset);
973 			return -EBUSY;
974 		}
975 	}
976 
977 	/*
978 	 * After doing so, there would be a "hole" in the /proc/iomem when
979 	 * offset is a positive value. It looks like the device return some
980 	 * mmio back to the system, which actually no one could use it.
981 	 */
982 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
983 		res = &dev->resource[i + PCI_IOV_RESOURCES];
984 		if (!res->flags || !res->parent)
985 			continue;
986 
987 		size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
988 		res2 = *res;
989 		res->start += size * offset;
990 
991 		dev_info(&dev->dev, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n",
992 			 i, &res2, res, (offset > 0) ? "En" : "Dis",
993 			 num_vfs, offset);
994 		pci_update_resource(dev, i + PCI_IOV_RESOURCES);
995 	}
996 	return 0;
997 }
998 #endif /* CONFIG_PCI_IOV */
999 
1000 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
1001 {
1002 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
1003 	struct pnv_phb *phb = hose->private_data;
1004 	struct pci_dn *pdn = pci_get_pdn(dev);
1005 	struct pnv_ioda_pe *pe;
1006 
1007 	if (!pdn) {
1008 		pr_err("%s: Device tree node not associated properly\n",
1009 			   pci_name(dev));
1010 		return NULL;
1011 	}
1012 	if (pdn->pe_number != IODA_INVALID_PE)
1013 		return NULL;
1014 
1015 	pe = pnv_ioda_alloc_pe(phb);
1016 	if (!pe) {
1017 		pr_warning("%s: Not enough PE# available, disabling device\n",
1018 			   pci_name(dev));
1019 		return NULL;
1020 	}
1021 
1022 	/* NOTE: We get only one ref to the pci_dev for the pdn, not for the
1023 	 * pointer in the PE data structure, both should be destroyed at the
1024 	 * same time. However, this needs to be looked at more closely again
1025 	 * once we actually start removing things (Hotplug, SR-IOV, ...)
1026 	 *
1027 	 * At some point we want to remove the PDN completely anyways
1028 	 */
1029 	pci_dev_get(dev);
1030 	pdn->pcidev = dev;
1031 	pdn->pe_number = pe->pe_number;
1032 	pe->flags = PNV_IODA_PE_DEV;
1033 	pe->pdev = dev;
1034 	pe->pbus = NULL;
1035 	pe->mve_number = -1;
1036 	pe->rid = dev->bus->number << 8 | pdn->devfn;
1037 
1038 	pe_info(pe, "Associated device to PE\n");
1039 
1040 	if (pnv_ioda_configure_pe(phb, pe)) {
1041 		/* XXX What do we do here ? */
1042 		pnv_ioda_free_pe(pe);
1043 		pdn->pe_number = IODA_INVALID_PE;
1044 		pe->pdev = NULL;
1045 		pci_dev_put(dev);
1046 		return NULL;
1047 	}
1048 
1049 	/* Put PE to the list */
1050 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1051 
1052 	return pe;
1053 }
1054 
1055 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
1056 {
1057 	struct pci_dev *dev;
1058 
1059 	list_for_each_entry(dev, &bus->devices, bus_list) {
1060 		struct pci_dn *pdn = pci_get_pdn(dev);
1061 
1062 		if (pdn == NULL) {
1063 			pr_warn("%s: No device node associated with device !\n",
1064 				pci_name(dev));
1065 			continue;
1066 		}
1067 
1068 		/*
1069 		 * In partial hotplug case, the PCI device might be still
1070 		 * associated with the PE and needn't attach it to the PE
1071 		 * again.
1072 		 */
1073 		if (pdn->pe_number != IODA_INVALID_PE)
1074 			continue;
1075 
1076 		pe->device_count++;
1077 		pdn->pcidev = dev;
1078 		pdn->pe_number = pe->pe_number;
1079 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1080 			pnv_ioda_setup_same_PE(dev->subordinate, pe);
1081 	}
1082 }
1083 
1084 /*
1085  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1086  * single PCI bus. Another one that contains the primary PCI bus and its
1087  * subordinate PCI devices and buses. The second type of PE is normally
1088  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1089  */
1090 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
1091 {
1092 	struct pci_controller *hose = pci_bus_to_host(bus);
1093 	struct pnv_phb *phb = hose->private_data;
1094 	struct pnv_ioda_pe *pe = NULL;
1095 	unsigned int pe_num;
1096 
1097 	/*
1098 	 * In partial hotplug case, the PE instance might be still alive.
1099 	 * We should reuse it instead of allocating a new one.
1100 	 */
1101 	pe_num = phb->ioda.pe_rmap[bus->number << 8];
1102 	if (pe_num != IODA_INVALID_PE) {
1103 		pe = &phb->ioda.pe_array[pe_num];
1104 		pnv_ioda_setup_same_PE(bus, pe);
1105 		return NULL;
1106 	}
1107 
1108 	/* PE number for root bus should have been reserved */
1109 	if (pci_is_root_bus(bus) &&
1110 	    phb->ioda.root_pe_idx != IODA_INVALID_PE)
1111 		pe = &phb->ioda.pe_array[phb->ioda.root_pe_idx];
1112 
1113 	/* Check if PE is determined by M64 */
1114 	if (!pe && phb->pick_m64_pe)
1115 		pe = phb->pick_m64_pe(bus, all);
1116 
1117 	/* The PE number isn't pinned by M64 */
1118 	if (!pe)
1119 		pe = pnv_ioda_alloc_pe(phb);
1120 
1121 	if (!pe) {
1122 		pr_warning("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1123 			__func__, pci_domain_nr(bus), bus->number);
1124 		return NULL;
1125 	}
1126 
1127 	pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1128 	pe->pbus = bus;
1129 	pe->pdev = NULL;
1130 	pe->mve_number = -1;
1131 	pe->rid = bus->busn_res.start << 8;
1132 
1133 	if (all)
1134 		pe_info(pe, "Secondary bus %d..%d associated with PE#%d\n",
1135 			bus->busn_res.start, bus->busn_res.end, pe->pe_number);
1136 	else
1137 		pe_info(pe, "Secondary bus %d associated with PE#%d\n",
1138 			bus->busn_res.start, pe->pe_number);
1139 
1140 	if (pnv_ioda_configure_pe(phb, pe)) {
1141 		/* XXX What do we do here ? */
1142 		pnv_ioda_free_pe(pe);
1143 		pe->pbus = NULL;
1144 		return NULL;
1145 	}
1146 
1147 	/* Associate it with all child devices */
1148 	pnv_ioda_setup_same_PE(bus, pe);
1149 
1150 	/* Put PE to the list */
1151 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1152 
1153 	return pe;
1154 }
1155 
1156 static struct pnv_ioda_pe *pnv_ioda_setup_npu_PE(struct pci_dev *npu_pdev)
1157 {
1158 	int pe_num, found_pe = false, rc;
1159 	long rid;
1160 	struct pnv_ioda_pe *pe;
1161 	struct pci_dev *gpu_pdev;
1162 	struct pci_dn *npu_pdn;
1163 	struct pci_controller *hose = pci_bus_to_host(npu_pdev->bus);
1164 	struct pnv_phb *phb = hose->private_data;
1165 
1166 	/*
1167 	 * Due to a hardware errata PE#0 on the NPU is reserved for
1168 	 * error handling. This means we only have three PEs remaining
1169 	 * which need to be assigned to four links, implying some
1170 	 * links must share PEs.
1171 	 *
1172 	 * To achieve this we assign PEs such that NPUs linking the
1173 	 * same GPU get assigned the same PE.
1174 	 */
1175 	gpu_pdev = pnv_pci_get_gpu_dev(npu_pdev);
1176 	for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) {
1177 		pe = &phb->ioda.pe_array[pe_num];
1178 		if (!pe->pdev)
1179 			continue;
1180 
1181 		if (pnv_pci_get_gpu_dev(pe->pdev) == gpu_pdev) {
1182 			/*
1183 			 * This device has the same peer GPU so should
1184 			 * be assigned the same PE as the existing
1185 			 * peer NPU.
1186 			 */
1187 			dev_info(&npu_pdev->dev,
1188 				"Associating to existing PE %d\n", pe_num);
1189 			pci_dev_get(npu_pdev);
1190 			npu_pdn = pci_get_pdn(npu_pdev);
1191 			rid = npu_pdev->bus->number << 8 | npu_pdn->devfn;
1192 			npu_pdn->pcidev = npu_pdev;
1193 			npu_pdn->pe_number = pe_num;
1194 			phb->ioda.pe_rmap[rid] = pe->pe_number;
1195 
1196 			/* Map the PE to this link */
1197 			rc = opal_pci_set_pe(phb->opal_id, pe_num, rid,
1198 					OpalPciBusAll,
1199 					OPAL_COMPARE_RID_DEVICE_NUMBER,
1200 					OPAL_COMPARE_RID_FUNCTION_NUMBER,
1201 					OPAL_MAP_PE);
1202 			WARN_ON(rc != OPAL_SUCCESS);
1203 			found_pe = true;
1204 			break;
1205 		}
1206 	}
1207 
1208 	if (!found_pe)
1209 		/*
1210 		 * Could not find an existing PE so allocate a new
1211 		 * one.
1212 		 */
1213 		return pnv_ioda_setup_dev_PE(npu_pdev);
1214 	else
1215 		return pe;
1216 }
1217 
1218 static void pnv_ioda_setup_npu_PEs(struct pci_bus *bus)
1219 {
1220 	struct pci_dev *pdev;
1221 
1222 	list_for_each_entry(pdev, &bus->devices, bus_list)
1223 		pnv_ioda_setup_npu_PE(pdev);
1224 }
1225 
1226 static void pnv_pci_ioda_setup_PEs(void)
1227 {
1228 	struct pci_controller *hose, *tmp;
1229 	struct pnv_phb *phb;
1230 
1231 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1232 		phb = hose->private_data;
1233 		if (phb->type == PNV_PHB_NPU) {
1234 			/* PE#0 is needed for error reporting */
1235 			pnv_ioda_reserve_pe(phb, 0);
1236 			pnv_ioda_setup_npu_PEs(hose->bus);
1237 		}
1238 	}
1239 }
1240 
1241 #ifdef CONFIG_PCI_IOV
1242 static int pnv_pci_vf_release_m64(struct pci_dev *pdev, u16 num_vfs)
1243 {
1244 	struct pci_bus        *bus;
1245 	struct pci_controller *hose;
1246 	struct pnv_phb        *phb;
1247 	struct pci_dn         *pdn;
1248 	int                    i, j;
1249 	int                    m64_bars;
1250 
1251 	bus = pdev->bus;
1252 	hose = pci_bus_to_host(bus);
1253 	phb = hose->private_data;
1254 	pdn = pci_get_pdn(pdev);
1255 
1256 	if (pdn->m64_single_mode)
1257 		m64_bars = num_vfs;
1258 	else
1259 		m64_bars = 1;
1260 
1261 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
1262 		for (j = 0; j < m64_bars; j++) {
1263 			if (pdn->m64_map[j][i] == IODA_INVALID_M64)
1264 				continue;
1265 			opal_pci_phb_mmio_enable(phb->opal_id,
1266 				OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 0);
1267 			clear_bit(pdn->m64_map[j][i], &phb->ioda.m64_bar_alloc);
1268 			pdn->m64_map[j][i] = IODA_INVALID_M64;
1269 		}
1270 
1271 	kfree(pdn->m64_map);
1272 	return 0;
1273 }
1274 
1275 static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)
1276 {
1277 	struct pci_bus        *bus;
1278 	struct pci_controller *hose;
1279 	struct pnv_phb        *phb;
1280 	struct pci_dn         *pdn;
1281 	unsigned int           win;
1282 	struct resource       *res;
1283 	int                    i, j;
1284 	int64_t                rc;
1285 	int                    total_vfs;
1286 	resource_size_t        size, start;
1287 	int                    pe_num;
1288 	int                    m64_bars;
1289 
1290 	bus = pdev->bus;
1291 	hose = pci_bus_to_host(bus);
1292 	phb = hose->private_data;
1293 	pdn = pci_get_pdn(pdev);
1294 	total_vfs = pci_sriov_get_totalvfs(pdev);
1295 
1296 	if (pdn->m64_single_mode)
1297 		m64_bars = num_vfs;
1298 	else
1299 		m64_bars = 1;
1300 
1301 	pdn->m64_map = kmalloc(sizeof(*pdn->m64_map) * m64_bars, GFP_KERNEL);
1302 	if (!pdn->m64_map)
1303 		return -ENOMEM;
1304 	/* Initialize the m64_map to IODA_INVALID_M64 */
1305 	for (i = 0; i < m64_bars ; i++)
1306 		for (j = 0; j < PCI_SRIOV_NUM_BARS; j++)
1307 			pdn->m64_map[i][j] = IODA_INVALID_M64;
1308 
1309 
1310 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1311 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
1312 		if (!res->flags || !res->parent)
1313 			continue;
1314 
1315 		for (j = 0; j < m64_bars; j++) {
1316 			do {
1317 				win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
1318 						phb->ioda.m64_bar_idx + 1, 0);
1319 
1320 				if (win >= phb->ioda.m64_bar_idx + 1)
1321 					goto m64_failed;
1322 			} while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc));
1323 
1324 			pdn->m64_map[j][i] = win;
1325 
1326 			if (pdn->m64_single_mode) {
1327 				size = pci_iov_resource_size(pdev,
1328 							PCI_IOV_RESOURCES + i);
1329 				start = res->start + size * j;
1330 			} else {
1331 				size = resource_size(res);
1332 				start = res->start;
1333 			}
1334 
1335 			/* Map the M64 here */
1336 			if (pdn->m64_single_mode) {
1337 				pe_num = pdn->pe_num_map[j];
1338 				rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1339 						pe_num, OPAL_M64_WINDOW_TYPE,
1340 						pdn->m64_map[j][i], 0);
1341 			}
1342 
1343 			rc = opal_pci_set_phb_mem_window(phb->opal_id,
1344 						 OPAL_M64_WINDOW_TYPE,
1345 						 pdn->m64_map[j][i],
1346 						 start,
1347 						 0, /* unused */
1348 						 size);
1349 
1350 
1351 			if (rc != OPAL_SUCCESS) {
1352 				dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n",
1353 					win, rc);
1354 				goto m64_failed;
1355 			}
1356 
1357 			if (pdn->m64_single_mode)
1358 				rc = opal_pci_phb_mmio_enable(phb->opal_id,
1359 				     OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 2);
1360 			else
1361 				rc = opal_pci_phb_mmio_enable(phb->opal_id,
1362 				     OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 1);
1363 
1364 			if (rc != OPAL_SUCCESS) {
1365 				dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n",
1366 					win, rc);
1367 				goto m64_failed;
1368 			}
1369 		}
1370 	}
1371 	return 0;
1372 
1373 m64_failed:
1374 	pnv_pci_vf_release_m64(pdev, num_vfs);
1375 	return -EBUSY;
1376 }
1377 
1378 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
1379 		int num);
1380 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable);
1381 
1382 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe)
1383 {
1384 	struct iommu_table    *tbl;
1385 	int64_t               rc;
1386 
1387 	tbl = pe->table_group.tables[0];
1388 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
1389 	if (rc)
1390 		pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
1391 
1392 	pnv_pci_ioda2_set_bypass(pe, false);
1393 	if (pe->table_group.group) {
1394 		iommu_group_put(pe->table_group.group);
1395 		BUG_ON(pe->table_group.group);
1396 	}
1397 	pnv_pci_ioda2_table_free_pages(tbl);
1398 	iommu_free_table(tbl, of_node_full_name(dev->dev.of_node));
1399 }
1400 
1401 static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
1402 {
1403 	struct pci_bus        *bus;
1404 	struct pci_controller *hose;
1405 	struct pnv_phb        *phb;
1406 	struct pnv_ioda_pe    *pe, *pe_n;
1407 	struct pci_dn         *pdn;
1408 
1409 	bus = pdev->bus;
1410 	hose = pci_bus_to_host(bus);
1411 	phb = hose->private_data;
1412 	pdn = pci_get_pdn(pdev);
1413 
1414 	if (!pdev->is_physfn)
1415 		return;
1416 
1417 	list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) {
1418 		if (pe->parent_dev != pdev)
1419 			continue;
1420 
1421 		pnv_pci_ioda2_release_dma_pe(pdev, pe);
1422 
1423 		/* Remove from list */
1424 		mutex_lock(&phb->ioda.pe_list_mutex);
1425 		list_del(&pe->list);
1426 		mutex_unlock(&phb->ioda.pe_list_mutex);
1427 
1428 		pnv_ioda_deconfigure_pe(phb, pe);
1429 
1430 		pnv_ioda_free_pe(pe);
1431 	}
1432 }
1433 
1434 void pnv_pci_sriov_disable(struct pci_dev *pdev)
1435 {
1436 	struct pci_bus        *bus;
1437 	struct pci_controller *hose;
1438 	struct pnv_phb        *phb;
1439 	struct pnv_ioda_pe    *pe;
1440 	struct pci_dn         *pdn;
1441 	struct pci_sriov      *iov;
1442 	u16                    num_vfs, i;
1443 
1444 	bus = pdev->bus;
1445 	hose = pci_bus_to_host(bus);
1446 	phb = hose->private_data;
1447 	pdn = pci_get_pdn(pdev);
1448 	iov = pdev->sriov;
1449 	num_vfs = pdn->num_vfs;
1450 
1451 	/* Release VF PEs */
1452 	pnv_ioda_release_vf_PE(pdev);
1453 
1454 	if (phb->type == PNV_PHB_IODA2) {
1455 		if (!pdn->m64_single_mode)
1456 			pnv_pci_vf_resource_shift(pdev, -*pdn->pe_num_map);
1457 
1458 		/* Release M64 windows */
1459 		pnv_pci_vf_release_m64(pdev, num_vfs);
1460 
1461 		/* Release PE numbers */
1462 		if (pdn->m64_single_mode) {
1463 			for (i = 0; i < num_vfs; i++) {
1464 				if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1465 					continue;
1466 
1467 				pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1468 				pnv_ioda_free_pe(pe);
1469 			}
1470 		} else
1471 			bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1472 		/* Releasing pe_num_map */
1473 		kfree(pdn->pe_num_map);
1474 	}
1475 }
1476 
1477 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1478 				       struct pnv_ioda_pe *pe);
1479 static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs)
1480 {
1481 	struct pci_bus        *bus;
1482 	struct pci_controller *hose;
1483 	struct pnv_phb        *phb;
1484 	struct pnv_ioda_pe    *pe;
1485 	int                    pe_num;
1486 	u16                    vf_index;
1487 	struct pci_dn         *pdn;
1488 
1489 	bus = pdev->bus;
1490 	hose = pci_bus_to_host(bus);
1491 	phb = hose->private_data;
1492 	pdn = pci_get_pdn(pdev);
1493 
1494 	if (!pdev->is_physfn)
1495 		return;
1496 
1497 	/* Reserve PE for each VF */
1498 	for (vf_index = 0; vf_index < num_vfs; vf_index++) {
1499 		if (pdn->m64_single_mode)
1500 			pe_num = pdn->pe_num_map[vf_index];
1501 		else
1502 			pe_num = *pdn->pe_num_map + vf_index;
1503 
1504 		pe = &phb->ioda.pe_array[pe_num];
1505 		pe->pe_number = pe_num;
1506 		pe->phb = phb;
1507 		pe->flags = PNV_IODA_PE_VF;
1508 		pe->pbus = NULL;
1509 		pe->parent_dev = pdev;
1510 		pe->mve_number = -1;
1511 		pe->rid = (pci_iov_virtfn_bus(pdev, vf_index) << 8) |
1512 			   pci_iov_virtfn_devfn(pdev, vf_index);
1513 
1514 		pe_info(pe, "VF %04d:%02d:%02d.%d associated with PE#%d\n",
1515 			hose->global_number, pdev->bus->number,
1516 			PCI_SLOT(pci_iov_virtfn_devfn(pdev, vf_index)),
1517 			PCI_FUNC(pci_iov_virtfn_devfn(pdev, vf_index)), pe_num);
1518 
1519 		if (pnv_ioda_configure_pe(phb, pe)) {
1520 			/* XXX What do we do here ? */
1521 			pnv_ioda_free_pe(pe);
1522 			pe->pdev = NULL;
1523 			continue;
1524 		}
1525 
1526 		/* Put PE to the list */
1527 		mutex_lock(&phb->ioda.pe_list_mutex);
1528 		list_add_tail(&pe->list, &phb->ioda.pe_list);
1529 		mutex_unlock(&phb->ioda.pe_list_mutex);
1530 
1531 		pnv_pci_ioda2_setup_dma_pe(phb, pe);
1532 	}
1533 }
1534 
1535 int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1536 {
1537 	struct pci_bus        *bus;
1538 	struct pci_controller *hose;
1539 	struct pnv_phb        *phb;
1540 	struct pnv_ioda_pe    *pe;
1541 	struct pci_dn         *pdn;
1542 	int                    ret;
1543 	u16                    i;
1544 
1545 	bus = pdev->bus;
1546 	hose = pci_bus_to_host(bus);
1547 	phb = hose->private_data;
1548 	pdn = pci_get_pdn(pdev);
1549 
1550 	if (phb->type == PNV_PHB_IODA2) {
1551 		if (!pdn->vfs_expanded) {
1552 			dev_info(&pdev->dev, "don't support this SRIOV device"
1553 				" with non 64bit-prefetchable IOV BAR\n");
1554 			return -ENOSPC;
1555 		}
1556 
1557 		/*
1558 		 * When M64 BARs functions in Single PE mode, the number of VFs
1559 		 * could be enabled must be less than the number of M64 BARs.
1560 		 */
1561 		if (pdn->m64_single_mode && num_vfs > phb->ioda.m64_bar_idx) {
1562 			dev_info(&pdev->dev, "Not enough M64 BAR for VFs\n");
1563 			return -EBUSY;
1564 		}
1565 
1566 		/* Allocating pe_num_map */
1567 		if (pdn->m64_single_mode)
1568 			pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map) * num_vfs,
1569 					GFP_KERNEL);
1570 		else
1571 			pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map), GFP_KERNEL);
1572 
1573 		if (!pdn->pe_num_map)
1574 			return -ENOMEM;
1575 
1576 		if (pdn->m64_single_mode)
1577 			for (i = 0; i < num_vfs; i++)
1578 				pdn->pe_num_map[i] = IODA_INVALID_PE;
1579 
1580 		/* Calculate available PE for required VFs */
1581 		if (pdn->m64_single_mode) {
1582 			for (i = 0; i < num_vfs; i++) {
1583 				pe = pnv_ioda_alloc_pe(phb);
1584 				if (!pe) {
1585 					ret = -EBUSY;
1586 					goto m64_failed;
1587 				}
1588 
1589 				pdn->pe_num_map[i] = pe->pe_number;
1590 			}
1591 		} else {
1592 			mutex_lock(&phb->ioda.pe_alloc_mutex);
1593 			*pdn->pe_num_map = bitmap_find_next_zero_area(
1594 				phb->ioda.pe_alloc, phb->ioda.total_pe_num,
1595 				0, num_vfs, 0);
1596 			if (*pdn->pe_num_map >= phb->ioda.total_pe_num) {
1597 				mutex_unlock(&phb->ioda.pe_alloc_mutex);
1598 				dev_info(&pdev->dev, "Failed to enable VF%d\n", num_vfs);
1599 				kfree(pdn->pe_num_map);
1600 				return -EBUSY;
1601 			}
1602 			bitmap_set(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1603 			mutex_unlock(&phb->ioda.pe_alloc_mutex);
1604 		}
1605 		pdn->num_vfs = num_vfs;
1606 
1607 		/* Assign M64 window accordingly */
1608 		ret = pnv_pci_vf_assign_m64(pdev, num_vfs);
1609 		if (ret) {
1610 			dev_info(&pdev->dev, "Not enough M64 window resources\n");
1611 			goto m64_failed;
1612 		}
1613 
1614 		/*
1615 		 * When using one M64 BAR to map one IOV BAR, we need to shift
1616 		 * the IOV BAR according to the PE# allocated to the VFs.
1617 		 * Otherwise, the PE# for the VF will conflict with others.
1618 		 */
1619 		if (!pdn->m64_single_mode) {
1620 			ret = pnv_pci_vf_resource_shift(pdev, *pdn->pe_num_map);
1621 			if (ret)
1622 				goto m64_failed;
1623 		}
1624 	}
1625 
1626 	/* Setup VF PEs */
1627 	pnv_ioda_setup_vf_PE(pdev, num_vfs);
1628 
1629 	return 0;
1630 
1631 m64_failed:
1632 	if (pdn->m64_single_mode) {
1633 		for (i = 0; i < num_vfs; i++) {
1634 			if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1635 				continue;
1636 
1637 			pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1638 			pnv_ioda_free_pe(pe);
1639 		}
1640 	} else
1641 		bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1642 
1643 	/* Releasing pe_num_map */
1644 	kfree(pdn->pe_num_map);
1645 
1646 	return ret;
1647 }
1648 
1649 int pcibios_sriov_disable(struct pci_dev *pdev)
1650 {
1651 	pnv_pci_sriov_disable(pdev);
1652 
1653 	/* Release PCI data */
1654 	remove_dev_pci_data(pdev);
1655 	return 0;
1656 }
1657 
1658 int pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1659 {
1660 	/* Allocate PCI data */
1661 	add_dev_pci_data(pdev);
1662 
1663 	return pnv_pci_sriov_enable(pdev, num_vfs);
1664 }
1665 #endif /* CONFIG_PCI_IOV */
1666 
1667 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
1668 {
1669 	struct pci_dn *pdn = pci_get_pdn(pdev);
1670 	struct pnv_ioda_pe *pe;
1671 
1672 	/*
1673 	 * The function can be called while the PE#
1674 	 * hasn't been assigned. Do nothing for the
1675 	 * case.
1676 	 */
1677 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1678 		return;
1679 
1680 	pe = &phb->ioda.pe_array[pdn->pe_number];
1681 	WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1682 	set_dma_offset(&pdev->dev, pe->tce_bypass_base);
1683 	set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]);
1684 	/*
1685 	 * Note: iommu_add_device() will fail here as
1686 	 * for physical PE: the device is already added by now;
1687 	 * for virtual PE: sysfs entries are not ready yet and
1688 	 * tce_iommu_bus_notifier will add the device to a group later.
1689 	 */
1690 }
1691 
1692 static int pnv_pci_ioda_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
1693 {
1694 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1695 	struct pnv_phb *phb = hose->private_data;
1696 	struct pci_dn *pdn = pci_get_pdn(pdev);
1697 	struct pnv_ioda_pe *pe;
1698 	uint64_t top;
1699 	bool bypass = false;
1700 
1701 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1702 		return -ENODEV;;
1703 
1704 	pe = &phb->ioda.pe_array[pdn->pe_number];
1705 	if (pe->tce_bypass_enabled) {
1706 		top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1707 		bypass = (dma_mask >= top);
1708 	}
1709 
1710 	if (bypass) {
1711 		dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n");
1712 		set_dma_ops(&pdev->dev, &dma_direct_ops);
1713 	} else {
1714 		dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n");
1715 		set_dma_ops(&pdev->dev, &dma_iommu_ops);
1716 	}
1717 	*pdev->dev.dma_mask = dma_mask;
1718 
1719 	/* Update peer npu devices */
1720 	pnv_npu_try_dma_set_bypass(pdev, bypass);
1721 
1722 	return 0;
1723 }
1724 
1725 static u64 pnv_pci_ioda_dma_get_required_mask(struct pci_dev *pdev)
1726 {
1727 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1728 	struct pnv_phb *phb = hose->private_data;
1729 	struct pci_dn *pdn = pci_get_pdn(pdev);
1730 	struct pnv_ioda_pe *pe;
1731 	u64 end, mask;
1732 
1733 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1734 		return 0;
1735 
1736 	pe = &phb->ioda.pe_array[pdn->pe_number];
1737 	if (!pe->tce_bypass_enabled)
1738 		return __dma_get_required_mask(&pdev->dev);
1739 
1740 
1741 	end = pe->tce_bypass_base + memblock_end_of_DRAM();
1742 	mask = 1ULL << (fls64(end) - 1);
1743 	mask += mask - 1;
1744 
1745 	return mask;
1746 }
1747 
1748 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe,
1749 				   struct pci_bus *bus)
1750 {
1751 	struct pci_dev *dev;
1752 
1753 	list_for_each_entry(dev, &bus->devices, bus_list) {
1754 		set_iommu_table_base(&dev->dev, pe->table_group.tables[0]);
1755 		set_dma_offset(&dev->dev, pe->tce_bypass_base);
1756 		iommu_add_device(&dev->dev);
1757 
1758 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1759 			pnv_ioda_setup_bus_dma(pe, dev->subordinate);
1760 	}
1761 }
1762 
1763 static inline __be64 __iomem *pnv_ioda_get_inval_reg(struct pnv_phb *phb,
1764 						     bool real_mode)
1765 {
1766 	return real_mode ? (__be64 __iomem *)(phb->regs_phys + 0x210) :
1767 		(phb->regs + 0x210);
1768 }
1769 
1770 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table *tbl,
1771 		unsigned long index, unsigned long npages, bool rm)
1772 {
1773 	struct iommu_table_group_link *tgl = list_first_entry_or_null(
1774 			&tbl->it_group_list, struct iommu_table_group_link,
1775 			next);
1776 	struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1777 			struct pnv_ioda_pe, table_group);
1778 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
1779 	unsigned long start, end, inc;
1780 
1781 	start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset);
1782 	end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset +
1783 			npages - 1);
1784 
1785 	/* p7ioc-style invalidation, 2 TCEs per write */
1786 	start |= (1ull << 63);
1787 	end |= (1ull << 63);
1788 	inc = 16;
1789         end |= inc - 1;	/* round up end to be different than start */
1790 
1791         mb(); /* Ensure above stores are visible */
1792         while (start <= end) {
1793 		if (rm)
1794 			__raw_rm_writeq(cpu_to_be64(start), invalidate);
1795 		else
1796 			__raw_writeq(cpu_to_be64(start), invalidate);
1797                 start += inc;
1798         }
1799 
1800 	/*
1801 	 * The iommu layer will do another mb() for us on build()
1802 	 * and we don't care on free()
1803 	 */
1804 }
1805 
1806 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index,
1807 		long npages, unsigned long uaddr,
1808 		enum dma_data_direction direction,
1809 		struct dma_attrs *attrs)
1810 {
1811 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1812 			attrs);
1813 
1814 	if (!ret)
1815 		pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1816 
1817 	return ret;
1818 }
1819 
1820 #ifdef CONFIG_IOMMU_API
1821 static int pnv_ioda1_tce_xchg(struct iommu_table *tbl, long index,
1822 		unsigned long *hpa, enum dma_data_direction *direction)
1823 {
1824 	long ret = pnv_tce_xchg(tbl, index, hpa, direction);
1825 
1826 	if (!ret)
1827 		pnv_pci_p7ioc_tce_invalidate(tbl, index, 1, false);
1828 
1829 	return ret;
1830 }
1831 #endif
1832 
1833 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index,
1834 		long npages)
1835 {
1836 	pnv_tce_free(tbl, index, npages);
1837 
1838 	pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1839 }
1840 
1841 static struct iommu_table_ops pnv_ioda1_iommu_ops = {
1842 	.set = pnv_ioda1_tce_build,
1843 #ifdef CONFIG_IOMMU_API
1844 	.exchange = pnv_ioda1_tce_xchg,
1845 #endif
1846 	.clear = pnv_ioda1_tce_free,
1847 	.get = pnv_tce_get,
1848 };
1849 
1850 #define PHB3_TCE_KILL_INVAL_ALL		PPC_BIT(0)
1851 #define PHB3_TCE_KILL_INVAL_PE		PPC_BIT(1)
1852 #define PHB3_TCE_KILL_INVAL_ONE		PPC_BIT(2)
1853 
1854 void pnv_pci_phb3_tce_invalidate_entire(struct pnv_phb *phb, bool rm)
1855 {
1856 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(phb, rm);
1857 	const unsigned long val = PHB3_TCE_KILL_INVAL_ALL;
1858 
1859 	mb(); /* Ensure previous TCE table stores are visible */
1860 	if (rm)
1861 		__raw_rm_writeq(cpu_to_be64(val), invalidate);
1862 	else
1863 		__raw_writeq(cpu_to_be64(val), invalidate);
1864 }
1865 
1866 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1867 {
1868 	/* 01xb - invalidate TCEs that match the specified PE# */
1869 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, false);
1870 	unsigned long val = PHB3_TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
1871 
1872 	mb(); /* Ensure above stores are visible */
1873 	__raw_writeq(cpu_to_be64(val), invalidate);
1874 }
1875 
1876 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe *pe, bool rm,
1877 					unsigned shift, unsigned long index,
1878 					unsigned long npages)
1879 {
1880 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, false);
1881 	unsigned long start, end, inc;
1882 
1883 	/* We'll invalidate DMA address in PE scope */
1884 	start = PHB3_TCE_KILL_INVAL_ONE;
1885 	start |= (pe->pe_number & 0xFF);
1886 	end = start;
1887 
1888 	/* Figure out the start, end and step */
1889 	start |= (index << shift);
1890 	end |= ((index + npages - 1) << shift);
1891 	inc = (0x1ull << shift);
1892 	mb();
1893 
1894 	while (start <= end) {
1895 		if (rm)
1896 			__raw_rm_writeq(cpu_to_be64(start), invalidate);
1897 		else
1898 			__raw_writeq(cpu_to_be64(start), invalidate);
1899 		start += inc;
1900 	}
1901 }
1902 
1903 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1904 {
1905 	struct pnv_phb *phb = pe->phb;
1906 
1907 	if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
1908 		pnv_pci_phb3_tce_invalidate_pe(pe);
1909 	else
1910 		opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL_PE,
1911 				  pe->pe_number, 0, 0, 0);
1912 }
1913 
1914 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
1915 		unsigned long index, unsigned long npages, bool rm)
1916 {
1917 	struct iommu_table_group_link *tgl;
1918 
1919 	list_for_each_entry_rcu(tgl, &tbl->it_group_list, next) {
1920 		struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1921 				struct pnv_ioda_pe, table_group);
1922 		struct pnv_phb *phb = pe->phb;
1923 		unsigned int shift = tbl->it_page_shift;
1924 
1925 		if (phb->type == PNV_PHB_NPU) {
1926 			/*
1927 			 * The NVLink hardware does not support TCE kill
1928 			 * per TCE entry so we have to invalidate
1929 			 * the entire cache for it.
1930 			 */
1931 			pnv_pci_phb3_tce_invalidate_entire(phb, rm);
1932 			continue;
1933 		}
1934 		if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
1935 			pnv_pci_phb3_tce_invalidate(pe, rm, shift,
1936 						    index, npages);
1937 		else if (rm)
1938 			opal_rm_pci_tce_kill(phb->opal_id,
1939 					     OPAL_PCI_TCE_KILL_PAGES,
1940 					     pe->pe_number, 1u << shift,
1941 					     index << shift, npages);
1942 		else
1943 			opal_pci_tce_kill(phb->opal_id,
1944 					  OPAL_PCI_TCE_KILL_PAGES,
1945 					  pe->pe_number, 1u << shift,
1946 					  index << shift, npages);
1947 	}
1948 }
1949 
1950 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index,
1951 		long npages, unsigned long uaddr,
1952 		enum dma_data_direction direction,
1953 		struct dma_attrs *attrs)
1954 {
1955 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1956 			attrs);
1957 
1958 	if (!ret)
1959 		pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1960 
1961 	return ret;
1962 }
1963 
1964 #ifdef CONFIG_IOMMU_API
1965 static int pnv_ioda2_tce_xchg(struct iommu_table *tbl, long index,
1966 		unsigned long *hpa, enum dma_data_direction *direction)
1967 {
1968 	long ret = pnv_tce_xchg(tbl, index, hpa, direction);
1969 
1970 	if (!ret)
1971 		pnv_pci_ioda2_tce_invalidate(tbl, index, 1, false);
1972 
1973 	return ret;
1974 }
1975 #endif
1976 
1977 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index,
1978 		long npages)
1979 {
1980 	pnv_tce_free(tbl, index, npages);
1981 
1982 	pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1983 }
1984 
1985 static void pnv_ioda2_table_free(struct iommu_table *tbl)
1986 {
1987 	pnv_pci_ioda2_table_free_pages(tbl);
1988 	iommu_free_table(tbl, "pnv");
1989 }
1990 
1991 static struct iommu_table_ops pnv_ioda2_iommu_ops = {
1992 	.set = pnv_ioda2_tce_build,
1993 #ifdef CONFIG_IOMMU_API
1994 	.exchange = pnv_ioda2_tce_xchg,
1995 #endif
1996 	.clear = pnv_ioda2_tce_free,
1997 	.get = pnv_tce_get,
1998 	.free = pnv_ioda2_table_free,
1999 };
2000 
2001 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data)
2002 {
2003 	unsigned int *weight = (unsigned int *)data;
2004 
2005 	/* This is quite simplistic. The "base" weight of a device
2006 	 * is 10. 0 means no DMA is to be accounted for it.
2007 	 */
2008 	if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
2009 		return 0;
2010 
2011 	if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
2012 	    dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
2013 	    dev->class == PCI_CLASS_SERIAL_USB_EHCI)
2014 		*weight += 3;
2015 	else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
2016 		*weight += 15;
2017 	else
2018 		*weight += 10;
2019 
2020 	return 0;
2021 }
2022 
2023 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe)
2024 {
2025 	unsigned int weight = 0;
2026 
2027 	/* SRIOV VF has same DMA32 weight as its PF */
2028 #ifdef CONFIG_PCI_IOV
2029 	if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) {
2030 		pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight);
2031 		return weight;
2032 	}
2033 #endif
2034 
2035 	if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) {
2036 		pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight);
2037 	} else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) {
2038 		struct pci_dev *pdev;
2039 
2040 		list_for_each_entry(pdev, &pe->pbus->devices, bus_list)
2041 			pnv_pci_ioda_dev_dma_weight(pdev, &weight);
2042 	} else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) {
2043 		pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight);
2044 	}
2045 
2046 	return weight;
2047 }
2048 
2049 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
2050 				       struct pnv_ioda_pe *pe)
2051 {
2052 
2053 	struct page *tce_mem = NULL;
2054 	struct iommu_table *tbl;
2055 	unsigned int weight, total_weight = 0;
2056 	unsigned int tce32_segsz, base, segs, avail, i;
2057 	int64_t rc;
2058 	void *addr;
2059 
2060 	/* XXX FIXME: Handle 64-bit only DMA devices */
2061 	/* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
2062 	/* XXX FIXME: Allocate multi-level tables on PHB3 */
2063 	weight = pnv_pci_ioda_pe_dma_weight(pe);
2064 	if (!weight)
2065 		return;
2066 
2067 	pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight,
2068 		     &total_weight);
2069 	segs = (weight * phb->ioda.dma32_count) / total_weight;
2070 	if (!segs)
2071 		segs = 1;
2072 
2073 	/*
2074 	 * Allocate contiguous DMA32 segments. We begin with the expected
2075 	 * number of segments. With one more attempt, the number of DMA32
2076 	 * segments to be allocated is decreased by one until one segment
2077 	 * is allocated successfully.
2078 	 */
2079 	do {
2080 		for (base = 0; base <= phb->ioda.dma32_count - segs; base++) {
2081 			for (avail = 0, i = base; i < base + segs; i++) {
2082 				if (phb->ioda.dma32_segmap[i] ==
2083 				    IODA_INVALID_PE)
2084 					avail++;
2085 			}
2086 
2087 			if (avail == segs)
2088 				goto found;
2089 		}
2090 	} while (--segs);
2091 
2092 	if (!segs) {
2093 		pe_warn(pe, "No available DMA32 segments\n");
2094 		return;
2095 	}
2096 
2097 found:
2098 	tbl = pnv_pci_table_alloc(phb->hose->node);
2099 	iommu_register_group(&pe->table_group, phb->hose->global_number,
2100 			pe->pe_number);
2101 	pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group);
2102 
2103 	/* Grab a 32-bit TCE table */
2104 	pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
2105 		weight, total_weight, base, segs);
2106 	pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
2107 		base * PNV_IODA1_DMA32_SEGSIZE,
2108 		(base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1);
2109 
2110 	/* XXX Currently, we allocate one big contiguous table for the
2111 	 * TCEs. We only really need one chunk per 256M of TCE space
2112 	 * (ie per segment) but that's an optimization for later, it
2113 	 * requires some added smarts with our get/put_tce implementation
2114 	 *
2115 	 * Each TCE page is 4KB in size and each TCE entry occupies 8
2116 	 * bytes
2117 	 */
2118 	tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3);
2119 	tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
2120 				   get_order(tce32_segsz * segs));
2121 	if (!tce_mem) {
2122 		pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
2123 		goto fail;
2124 	}
2125 	addr = page_address(tce_mem);
2126 	memset(addr, 0, tce32_segsz * segs);
2127 
2128 	/* Configure HW */
2129 	for (i = 0; i < segs; i++) {
2130 		rc = opal_pci_map_pe_dma_window(phb->opal_id,
2131 					      pe->pe_number,
2132 					      base + i, 1,
2133 					      __pa(addr) + tce32_segsz * i,
2134 					      tce32_segsz, IOMMU_PAGE_SIZE_4K);
2135 		if (rc) {
2136 			pe_err(pe, " Failed to configure 32-bit TCE table,"
2137 			       " err %ld\n", rc);
2138 			goto fail;
2139 		}
2140 	}
2141 
2142 	/* Setup DMA32 segment mapping */
2143 	for (i = base; i < base + segs; i++)
2144 		phb->ioda.dma32_segmap[i] = pe->pe_number;
2145 
2146 	/* Setup linux iommu table */
2147 	pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs,
2148 				  base * PNV_IODA1_DMA32_SEGSIZE,
2149 				  IOMMU_PAGE_SHIFT_4K);
2150 
2151 	tbl->it_ops = &pnv_ioda1_iommu_ops;
2152 	pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
2153 	pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
2154 	iommu_init_table(tbl, phb->hose->node);
2155 
2156 	if (pe->flags & PNV_IODA_PE_DEV) {
2157 		/*
2158 		 * Setting table base here only for carrying iommu_group
2159 		 * further down to let iommu_add_device() do the job.
2160 		 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2161 		 */
2162 		set_iommu_table_base(&pe->pdev->dev, tbl);
2163 		iommu_add_device(&pe->pdev->dev);
2164 	} else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2165 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
2166 
2167 	return;
2168  fail:
2169 	/* XXX Failure: Try to fallback to 64-bit only ? */
2170 	if (tce_mem)
2171 		__free_pages(tce_mem, get_order(tce32_segsz * segs));
2172 	if (tbl) {
2173 		pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
2174 		iommu_free_table(tbl, "pnv");
2175 	}
2176 }
2177 
2178 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
2179 		int num, struct iommu_table *tbl)
2180 {
2181 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2182 			table_group);
2183 	struct pnv_phb *phb = pe->phb;
2184 	int64_t rc;
2185 	const unsigned long size = tbl->it_indirect_levels ?
2186 			tbl->it_level_size : tbl->it_size;
2187 	const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
2188 	const __u64 win_size = tbl->it_size << tbl->it_page_shift;
2189 
2190 	pe_info(pe, "Setting up window#%d %llx..%llx pg=%x\n", num,
2191 			start_addr, start_addr + win_size - 1,
2192 			IOMMU_PAGE_SIZE(tbl));
2193 
2194 	/*
2195 	 * Map TCE table through TVT. The TVE index is the PE number
2196 	 * shifted by 1 bit for 32-bits DMA space.
2197 	 */
2198 	rc = opal_pci_map_pe_dma_window(phb->opal_id,
2199 			pe->pe_number,
2200 			(pe->pe_number << 1) + num,
2201 			tbl->it_indirect_levels + 1,
2202 			__pa(tbl->it_base),
2203 			size << 3,
2204 			IOMMU_PAGE_SIZE(tbl));
2205 	if (rc) {
2206 		pe_err(pe, "Failed to configure TCE table, err %ld\n", rc);
2207 		return rc;
2208 	}
2209 
2210 	pnv_pci_link_table_and_group(phb->hose->node, num,
2211 			tbl, &pe->table_group);
2212 	pnv_pci_phb3_tce_invalidate_pe(pe);
2213 
2214 	return 0;
2215 }
2216 
2217 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
2218 {
2219 	uint16_t window_id = (pe->pe_number << 1 ) + 1;
2220 	int64_t rc;
2221 
2222 	pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
2223 	if (enable) {
2224 		phys_addr_t top = memblock_end_of_DRAM();
2225 
2226 		top = roundup_pow_of_two(top);
2227 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2228 						     pe->pe_number,
2229 						     window_id,
2230 						     pe->tce_bypass_base,
2231 						     top);
2232 	} else {
2233 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2234 						     pe->pe_number,
2235 						     window_id,
2236 						     pe->tce_bypass_base,
2237 						     0);
2238 	}
2239 	if (rc)
2240 		pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
2241 	else
2242 		pe->tce_bypass_enabled = enable;
2243 }
2244 
2245 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset,
2246 		__u32 page_shift, __u64 window_size, __u32 levels,
2247 		struct iommu_table *tbl);
2248 
2249 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group,
2250 		int num, __u32 page_shift, __u64 window_size, __u32 levels,
2251 		struct iommu_table **ptbl)
2252 {
2253 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2254 			table_group);
2255 	int nid = pe->phb->hose->node;
2256 	__u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start;
2257 	long ret;
2258 	struct iommu_table *tbl;
2259 
2260 	tbl = pnv_pci_table_alloc(nid);
2261 	if (!tbl)
2262 		return -ENOMEM;
2263 
2264 	ret = pnv_pci_ioda2_table_alloc_pages(nid,
2265 			bus_offset, page_shift, window_size,
2266 			levels, tbl);
2267 	if (ret) {
2268 		iommu_free_table(tbl, "pnv");
2269 		return ret;
2270 	}
2271 
2272 	tbl->it_ops = &pnv_ioda2_iommu_ops;
2273 
2274 	*ptbl = tbl;
2275 
2276 	return 0;
2277 }
2278 
2279 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe)
2280 {
2281 	struct iommu_table *tbl = NULL;
2282 	long rc;
2283 
2284 	/*
2285 	 * crashkernel= specifies the kdump kernel's maximum memory at
2286 	 * some offset and there is no guaranteed the result is a power
2287 	 * of 2, which will cause errors later.
2288 	 */
2289 	const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max());
2290 
2291 	/*
2292 	 * In memory constrained environments, e.g. kdump kernel, the
2293 	 * DMA window can be larger than available memory, which will
2294 	 * cause errors later.
2295 	 */
2296 	const u64 window_size = min((u64)pe->table_group.tce32_size, max_memory);
2297 
2298 	rc = pnv_pci_ioda2_create_table(&pe->table_group, 0,
2299 			IOMMU_PAGE_SHIFT_4K,
2300 			window_size,
2301 			POWERNV_IOMMU_DEFAULT_LEVELS, &tbl);
2302 	if (rc) {
2303 		pe_err(pe, "Failed to create 32-bit TCE table, err %ld",
2304 				rc);
2305 		return rc;
2306 	}
2307 
2308 	iommu_init_table(tbl, pe->phb->hose->node);
2309 
2310 	rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl);
2311 	if (rc) {
2312 		pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n",
2313 				rc);
2314 		pnv_ioda2_table_free(tbl);
2315 		return rc;
2316 	}
2317 
2318 	if (!pnv_iommu_bypass_disabled)
2319 		pnv_pci_ioda2_set_bypass(pe, true);
2320 
2321 	/*
2322 	 * Setting table base here only for carrying iommu_group
2323 	 * further down to let iommu_add_device() do the job.
2324 	 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2325 	 */
2326 	if (pe->flags & PNV_IODA_PE_DEV)
2327 		set_iommu_table_base(&pe->pdev->dev, tbl);
2328 
2329 	return 0;
2330 }
2331 
2332 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV)
2333 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
2334 		int num)
2335 {
2336 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2337 			table_group);
2338 	struct pnv_phb *phb = pe->phb;
2339 	long ret;
2340 
2341 	pe_info(pe, "Removing DMA window #%d\n", num);
2342 
2343 	ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
2344 			(pe->pe_number << 1) + num,
2345 			0/* levels */, 0/* table address */,
2346 			0/* table size */, 0/* page size */);
2347 	if (ret)
2348 		pe_warn(pe, "Unmapping failed, ret = %ld\n", ret);
2349 	else
2350 		pnv_pci_phb3_tce_invalidate_pe(pe);
2351 
2352 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
2353 
2354 	return ret;
2355 }
2356 #endif
2357 
2358 #ifdef CONFIG_IOMMU_API
2359 static unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift,
2360 		__u64 window_size, __u32 levels)
2361 {
2362 	unsigned long bytes = 0;
2363 	const unsigned window_shift = ilog2(window_size);
2364 	unsigned entries_shift = window_shift - page_shift;
2365 	unsigned table_shift = entries_shift + 3;
2366 	unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift);
2367 	unsigned long direct_table_size;
2368 
2369 	if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) ||
2370 			(window_size > memory_hotplug_max()) ||
2371 			!is_power_of_2(window_size))
2372 		return 0;
2373 
2374 	/* Calculate a direct table size from window_size and levels */
2375 	entries_shift = (entries_shift + levels - 1) / levels;
2376 	table_shift = entries_shift + 3;
2377 	table_shift = max_t(unsigned, table_shift, PAGE_SHIFT);
2378 	direct_table_size =  1UL << table_shift;
2379 
2380 	for ( ; levels; --levels) {
2381 		bytes += _ALIGN_UP(tce_table_size, direct_table_size);
2382 
2383 		tce_table_size /= direct_table_size;
2384 		tce_table_size <<= 3;
2385 		tce_table_size = _ALIGN_UP(tce_table_size, direct_table_size);
2386 	}
2387 
2388 	return bytes;
2389 }
2390 
2391 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group)
2392 {
2393 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2394 						table_group);
2395 	/* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
2396 	struct iommu_table *tbl = pe->table_group.tables[0];
2397 
2398 	pnv_pci_ioda2_set_bypass(pe, false);
2399 	pnv_pci_ioda2_unset_window(&pe->table_group, 0);
2400 	pnv_ioda2_table_free(tbl);
2401 }
2402 
2403 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group)
2404 {
2405 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2406 						table_group);
2407 
2408 	pnv_pci_ioda2_setup_default_config(pe);
2409 }
2410 
2411 static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
2412 	.get_table_size = pnv_pci_ioda2_get_table_size,
2413 	.create_table = pnv_pci_ioda2_create_table,
2414 	.set_window = pnv_pci_ioda2_set_window,
2415 	.unset_window = pnv_pci_ioda2_unset_window,
2416 	.take_ownership = pnv_ioda2_take_ownership,
2417 	.release_ownership = pnv_ioda2_release_ownership,
2418 };
2419 
2420 static int gpe_table_group_to_npe_cb(struct device *dev, void *opaque)
2421 {
2422 	struct pci_controller *hose;
2423 	struct pnv_phb *phb;
2424 	struct pnv_ioda_pe **ptmppe = opaque;
2425 	struct pci_dev *pdev = container_of(dev, struct pci_dev, dev);
2426 	struct pci_dn *pdn = pci_get_pdn(pdev);
2427 
2428 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
2429 		return 0;
2430 
2431 	hose = pci_bus_to_host(pdev->bus);
2432 	phb = hose->private_data;
2433 	if (phb->type != PNV_PHB_NPU)
2434 		return 0;
2435 
2436 	*ptmppe = &phb->ioda.pe_array[pdn->pe_number];
2437 
2438 	return 1;
2439 }
2440 
2441 /*
2442  * This returns PE of associated NPU.
2443  * This assumes that NPU is in the same IOMMU group with GPU and there is
2444  * no other PEs.
2445  */
2446 static struct pnv_ioda_pe *gpe_table_group_to_npe(
2447 		struct iommu_table_group *table_group)
2448 {
2449 	struct pnv_ioda_pe *npe = NULL;
2450 	int ret = iommu_group_for_each_dev(table_group->group, &npe,
2451 			gpe_table_group_to_npe_cb);
2452 
2453 	BUG_ON(!ret || !npe);
2454 
2455 	return npe;
2456 }
2457 
2458 static long pnv_pci_ioda2_npu_set_window(struct iommu_table_group *table_group,
2459 		int num, struct iommu_table *tbl)
2460 {
2461 	long ret = pnv_pci_ioda2_set_window(table_group, num, tbl);
2462 
2463 	if (ret)
2464 		return ret;
2465 
2466 	ret = pnv_npu_set_window(gpe_table_group_to_npe(table_group), num, tbl);
2467 	if (ret)
2468 		pnv_pci_ioda2_unset_window(table_group, num);
2469 
2470 	return ret;
2471 }
2472 
2473 static long pnv_pci_ioda2_npu_unset_window(
2474 		struct iommu_table_group *table_group,
2475 		int num)
2476 {
2477 	long ret = pnv_pci_ioda2_unset_window(table_group, num);
2478 
2479 	if (ret)
2480 		return ret;
2481 
2482 	return pnv_npu_unset_window(gpe_table_group_to_npe(table_group), num);
2483 }
2484 
2485 static void pnv_ioda2_npu_take_ownership(struct iommu_table_group *table_group)
2486 {
2487 	/*
2488 	 * Detach NPU first as pnv_ioda2_take_ownership() will destroy
2489 	 * the iommu_table if 32bit DMA is enabled.
2490 	 */
2491 	pnv_npu_take_ownership(gpe_table_group_to_npe(table_group));
2492 	pnv_ioda2_take_ownership(table_group);
2493 }
2494 
2495 static struct iommu_table_group_ops pnv_pci_ioda2_npu_ops = {
2496 	.get_table_size = pnv_pci_ioda2_get_table_size,
2497 	.create_table = pnv_pci_ioda2_create_table,
2498 	.set_window = pnv_pci_ioda2_npu_set_window,
2499 	.unset_window = pnv_pci_ioda2_npu_unset_window,
2500 	.take_ownership = pnv_ioda2_npu_take_ownership,
2501 	.release_ownership = pnv_ioda2_release_ownership,
2502 };
2503 
2504 static void pnv_pci_ioda_setup_iommu_api(void)
2505 {
2506 	struct pci_controller *hose, *tmp;
2507 	struct pnv_phb *phb;
2508 	struct pnv_ioda_pe *pe, *gpe;
2509 
2510 	/*
2511 	 * Now we have all PHBs discovered, time to add NPU devices to
2512 	 * the corresponding IOMMU groups.
2513 	 */
2514 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2515 		phb = hose->private_data;
2516 
2517 		if (phb->type != PNV_PHB_NPU)
2518 			continue;
2519 
2520 		list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2521 			gpe = pnv_pci_npu_setup_iommu(pe);
2522 			if (gpe)
2523 				gpe->table_group.ops = &pnv_pci_ioda2_npu_ops;
2524 		}
2525 	}
2526 }
2527 #else /* !CONFIG_IOMMU_API */
2528 static void pnv_pci_ioda_setup_iommu_api(void) { };
2529 #endif
2530 
2531 static __be64 *pnv_pci_ioda2_table_do_alloc_pages(int nid, unsigned shift,
2532 		unsigned levels, unsigned long limit,
2533 		unsigned long *current_offset, unsigned long *total_allocated)
2534 {
2535 	struct page *tce_mem = NULL;
2536 	__be64 *addr, *tmp;
2537 	unsigned order = max_t(unsigned, shift, PAGE_SHIFT) - PAGE_SHIFT;
2538 	unsigned long allocated = 1UL << (order + PAGE_SHIFT);
2539 	unsigned entries = 1UL << (shift - 3);
2540 	long i;
2541 
2542 	tce_mem = alloc_pages_node(nid, GFP_KERNEL, order);
2543 	if (!tce_mem) {
2544 		pr_err("Failed to allocate a TCE memory, order=%d\n", order);
2545 		return NULL;
2546 	}
2547 	addr = page_address(tce_mem);
2548 	memset(addr, 0, allocated);
2549 	*total_allocated += allocated;
2550 
2551 	--levels;
2552 	if (!levels) {
2553 		*current_offset += allocated;
2554 		return addr;
2555 	}
2556 
2557 	for (i = 0; i < entries; ++i) {
2558 		tmp = pnv_pci_ioda2_table_do_alloc_pages(nid, shift,
2559 				levels, limit, current_offset, total_allocated);
2560 		if (!tmp)
2561 			break;
2562 
2563 		addr[i] = cpu_to_be64(__pa(tmp) |
2564 				TCE_PCI_READ | TCE_PCI_WRITE);
2565 
2566 		if (*current_offset >= limit)
2567 			break;
2568 	}
2569 
2570 	return addr;
2571 }
2572 
2573 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr,
2574 		unsigned long size, unsigned level);
2575 
2576 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset,
2577 		__u32 page_shift, __u64 window_size, __u32 levels,
2578 		struct iommu_table *tbl)
2579 {
2580 	void *addr;
2581 	unsigned long offset = 0, level_shift, total_allocated = 0;
2582 	const unsigned window_shift = ilog2(window_size);
2583 	unsigned entries_shift = window_shift - page_shift;
2584 	unsigned table_shift = max_t(unsigned, entries_shift + 3, PAGE_SHIFT);
2585 	const unsigned long tce_table_size = 1UL << table_shift;
2586 
2587 	if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS))
2588 		return -EINVAL;
2589 
2590 	if ((window_size > memory_hotplug_max()) || !is_power_of_2(window_size))
2591 		return -EINVAL;
2592 
2593 	/* Adjust direct table size from window_size and levels */
2594 	entries_shift = (entries_shift + levels - 1) / levels;
2595 	level_shift = entries_shift + 3;
2596 	level_shift = max_t(unsigned, level_shift, PAGE_SHIFT);
2597 
2598 	/* Allocate TCE table */
2599 	addr = pnv_pci_ioda2_table_do_alloc_pages(nid, level_shift,
2600 			levels, tce_table_size, &offset, &total_allocated);
2601 
2602 	/* addr==NULL means that the first level allocation failed */
2603 	if (!addr)
2604 		return -ENOMEM;
2605 
2606 	/*
2607 	 * First level was allocated but some lower level failed as
2608 	 * we did not allocate as much as we wanted,
2609 	 * release partially allocated table.
2610 	 */
2611 	if (offset < tce_table_size) {
2612 		pnv_pci_ioda2_table_do_free_pages(addr,
2613 				1ULL << (level_shift - 3), levels - 1);
2614 		return -ENOMEM;
2615 	}
2616 
2617 	/* Setup linux iommu table */
2618 	pnv_pci_setup_iommu_table(tbl, addr, tce_table_size, bus_offset,
2619 			page_shift);
2620 	tbl->it_level_size = 1ULL << (level_shift - 3);
2621 	tbl->it_indirect_levels = levels - 1;
2622 	tbl->it_allocated_size = total_allocated;
2623 
2624 	pr_devel("Created TCE table: ws=%08llx ts=%lx @%08llx\n",
2625 			window_size, tce_table_size, bus_offset);
2626 
2627 	return 0;
2628 }
2629 
2630 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr,
2631 		unsigned long size, unsigned level)
2632 {
2633 	const unsigned long addr_ul = (unsigned long) addr &
2634 			~(TCE_PCI_READ | TCE_PCI_WRITE);
2635 
2636 	if (level) {
2637 		long i;
2638 		u64 *tmp = (u64 *) addr_ul;
2639 
2640 		for (i = 0; i < size; ++i) {
2641 			unsigned long hpa = be64_to_cpu(tmp[i]);
2642 
2643 			if (!(hpa & (TCE_PCI_READ | TCE_PCI_WRITE)))
2644 				continue;
2645 
2646 			pnv_pci_ioda2_table_do_free_pages(__va(hpa), size,
2647 					level - 1);
2648 		}
2649 	}
2650 
2651 	free_pages(addr_ul, get_order(size << 3));
2652 }
2653 
2654 static void pnv_pci_ioda2_table_free_pages(struct iommu_table *tbl)
2655 {
2656 	const unsigned long size = tbl->it_indirect_levels ?
2657 			tbl->it_level_size : tbl->it_size;
2658 
2659 	if (!tbl->it_size)
2660 		return;
2661 
2662 	pnv_pci_ioda2_table_do_free_pages((__be64 *)tbl->it_base, size,
2663 			tbl->it_indirect_levels);
2664 }
2665 
2666 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
2667 				       struct pnv_ioda_pe *pe)
2668 {
2669 	int64_t rc;
2670 
2671 	if (!pnv_pci_ioda_pe_dma_weight(pe))
2672 		return;
2673 
2674 	/* TVE #1 is selected by PCI address bit 59 */
2675 	pe->tce_bypass_base = 1ull << 59;
2676 
2677 	iommu_register_group(&pe->table_group, phb->hose->global_number,
2678 			pe->pe_number);
2679 
2680 	/* The PE will reserve all possible 32-bits space */
2681 	pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
2682 		phb->ioda.m32_pci_base);
2683 
2684 	/* Setup linux iommu table */
2685 	pe->table_group.tce32_start = 0;
2686 	pe->table_group.tce32_size = phb->ioda.m32_pci_base;
2687 	pe->table_group.max_dynamic_windows_supported =
2688 			IOMMU_TABLE_GROUP_MAX_TABLES;
2689 	pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS;
2690 	pe->table_group.pgsizes = SZ_4K | SZ_64K | SZ_16M;
2691 #ifdef CONFIG_IOMMU_API
2692 	pe->table_group.ops = &pnv_pci_ioda2_ops;
2693 #endif
2694 
2695 	rc = pnv_pci_ioda2_setup_default_config(pe);
2696 	if (rc)
2697 		return;
2698 
2699 	if (pe->flags & PNV_IODA_PE_DEV)
2700 		iommu_add_device(&pe->pdev->dev);
2701 	else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2702 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
2703 }
2704 
2705 #ifdef CONFIG_PCI_MSI
2706 static void pnv_ioda2_msi_eoi(struct irq_data *d)
2707 {
2708 	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
2709 	struct irq_chip *chip = irq_data_get_irq_chip(d);
2710 	struct pnv_phb *phb = container_of(chip, struct pnv_phb,
2711 					   ioda.irq_chip);
2712 	int64_t rc;
2713 
2714 	rc = opal_pci_msi_eoi(phb->opal_id, hw_irq);
2715 	WARN_ON_ONCE(rc);
2716 
2717 	icp_native_eoi(d);
2718 }
2719 
2720 
2721 void pnv_set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2722 {
2723 	struct irq_data *idata;
2724 	struct irq_chip *ichip;
2725 
2726 	/* The MSI EOI OPAL call is only needed on PHB3 */
2727 	if (phb->model != PNV_PHB_MODEL_PHB3)
2728 		return;
2729 
2730 	if (!phb->ioda.irq_chip_init) {
2731 		/*
2732 		 * First time we setup an MSI IRQ, we need to setup the
2733 		 * corresponding IRQ chip to route correctly.
2734 		 */
2735 		idata = irq_get_irq_data(virq);
2736 		ichip = irq_data_get_irq_chip(idata);
2737 		phb->ioda.irq_chip_init = 1;
2738 		phb->ioda.irq_chip = *ichip;
2739 		phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2740 	}
2741 	irq_set_chip(virq, &phb->ioda.irq_chip);
2742 }
2743 
2744 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2745 				  unsigned int hwirq, unsigned int virq,
2746 				  unsigned int is_64, struct msi_msg *msg)
2747 {
2748 	struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2749 	unsigned int xive_num = hwirq - phb->msi_base;
2750 	__be32 data;
2751 	int rc;
2752 
2753 	/* No PE assigned ? bail out ... no MSI for you ! */
2754 	if (pe == NULL)
2755 		return -ENXIO;
2756 
2757 	/* Check if we have an MVE */
2758 	if (pe->mve_number < 0)
2759 		return -ENXIO;
2760 
2761 	/* Force 32-bit MSI on some broken devices */
2762 	if (dev->no_64bit_msi)
2763 		is_64 = 0;
2764 
2765 	/* Assign XIVE to PE */
2766 	rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2767 	if (rc) {
2768 		pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2769 			pci_name(dev), rc, xive_num);
2770 		return -EIO;
2771 	}
2772 
2773 	if (is_64) {
2774 		__be64 addr64;
2775 
2776 		rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
2777 				     &addr64, &data);
2778 		if (rc) {
2779 			pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2780 				pci_name(dev), rc);
2781 			return -EIO;
2782 		}
2783 		msg->address_hi = be64_to_cpu(addr64) >> 32;
2784 		msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
2785 	} else {
2786 		__be32 addr32;
2787 
2788 		rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
2789 				     &addr32, &data);
2790 		if (rc) {
2791 			pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2792 				pci_name(dev), rc);
2793 			return -EIO;
2794 		}
2795 		msg->address_hi = 0;
2796 		msg->address_lo = be32_to_cpu(addr32);
2797 	}
2798 	msg->data = be32_to_cpu(data);
2799 
2800 	pnv_set_msi_irq_chip(phb, virq);
2801 
2802 	pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
2803 		 " address=%x_%08x data=%x PE# %d\n",
2804 		 pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
2805 		 msg->address_hi, msg->address_lo, data, pe->pe_number);
2806 
2807 	return 0;
2808 }
2809 
2810 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
2811 {
2812 	unsigned int count;
2813 	const __be32 *prop = of_get_property(phb->hose->dn,
2814 					     "ibm,opal-msi-ranges", NULL);
2815 	if (!prop) {
2816 		/* BML Fallback */
2817 		prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
2818 	}
2819 	if (!prop)
2820 		return;
2821 
2822 	phb->msi_base = be32_to_cpup(prop);
2823 	count = be32_to_cpup(prop + 1);
2824 	if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
2825 		pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2826 		       phb->hose->global_number);
2827 		return;
2828 	}
2829 
2830 	phb->msi_setup = pnv_pci_ioda_msi_setup;
2831 	phb->msi32_support = 1;
2832 	pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2833 		count, phb->msi_base);
2834 }
2835 #else
2836 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { }
2837 #endif /* CONFIG_PCI_MSI */
2838 
2839 #ifdef CONFIG_PCI_IOV
2840 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev)
2841 {
2842 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
2843 	struct pnv_phb *phb = hose->private_data;
2844 	const resource_size_t gate = phb->ioda.m64_segsize >> 2;
2845 	struct resource *res;
2846 	int i;
2847 	resource_size_t size, total_vf_bar_sz;
2848 	struct pci_dn *pdn;
2849 	int mul, total_vfs;
2850 
2851 	if (!pdev->is_physfn || pdev->is_added)
2852 		return;
2853 
2854 	pdn = pci_get_pdn(pdev);
2855 	pdn->vfs_expanded = 0;
2856 	pdn->m64_single_mode = false;
2857 
2858 	total_vfs = pci_sriov_get_totalvfs(pdev);
2859 	mul = phb->ioda.total_pe_num;
2860 	total_vf_bar_sz = 0;
2861 
2862 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2863 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
2864 		if (!res->flags || res->parent)
2865 			continue;
2866 		if (!pnv_pci_is_mem_pref_64(res->flags)) {
2867 			dev_warn(&pdev->dev, "Don't support SR-IOV with"
2868 					" non M64 VF BAR%d: %pR. \n",
2869 				 i, res);
2870 			goto truncate_iov;
2871 		}
2872 
2873 		total_vf_bar_sz += pci_iov_resource_size(pdev,
2874 				i + PCI_IOV_RESOURCES);
2875 
2876 		/*
2877 		 * If bigger than quarter of M64 segment size, just round up
2878 		 * power of two.
2879 		 *
2880 		 * Generally, one M64 BAR maps one IOV BAR. To avoid conflict
2881 		 * with other devices, IOV BAR size is expanded to be
2882 		 * (total_pe * VF_BAR_size).  When VF_BAR_size is half of M64
2883 		 * segment size , the expanded size would equal to half of the
2884 		 * whole M64 space size, which will exhaust the M64 Space and
2885 		 * limit the system flexibility.  This is a design decision to
2886 		 * set the boundary to quarter of the M64 segment size.
2887 		 */
2888 		if (total_vf_bar_sz > gate) {
2889 			mul = roundup_pow_of_two(total_vfs);
2890 			dev_info(&pdev->dev,
2891 				"VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n",
2892 				total_vf_bar_sz, gate, mul);
2893 			pdn->m64_single_mode = true;
2894 			break;
2895 		}
2896 	}
2897 
2898 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2899 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
2900 		if (!res->flags || res->parent)
2901 			continue;
2902 
2903 		size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES);
2904 		/*
2905 		 * On PHB3, the minimum size alignment of M64 BAR in single
2906 		 * mode is 32MB.
2907 		 */
2908 		if (pdn->m64_single_mode && (size < SZ_32M))
2909 			goto truncate_iov;
2910 		dev_dbg(&pdev->dev, " Fixing VF BAR%d: %pR to\n", i, res);
2911 		res->end = res->start + size * mul - 1;
2912 		dev_dbg(&pdev->dev, "                       %pR\n", res);
2913 		dev_info(&pdev->dev, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
2914 			 i, res, mul);
2915 	}
2916 	pdn->vfs_expanded = mul;
2917 
2918 	return;
2919 
2920 truncate_iov:
2921 	/* To save MMIO space, IOV BAR is truncated. */
2922 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2923 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
2924 		res->flags = 0;
2925 		res->end = res->start - 1;
2926 	}
2927 }
2928 #endif /* CONFIG_PCI_IOV */
2929 
2930 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe,
2931 				  struct resource *res)
2932 {
2933 	struct pnv_phb *phb = pe->phb;
2934 	struct pci_bus_region region;
2935 	int index;
2936 	int64_t rc;
2937 
2938 	if (!res || !res->flags || res->start > res->end)
2939 		return;
2940 
2941 	if (res->flags & IORESOURCE_IO) {
2942 		region.start = res->start - phb->ioda.io_pci_base;
2943 		region.end   = res->end - phb->ioda.io_pci_base;
2944 		index = region.start / phb->ioda.io_segsize;
2945 
2946 		while (index < phb->ioda.total_pe_num &&
2947 		       region.start <= region.end) {
2948 			phb->ioda.io_segmap[index] = pe->pe_number;
2949 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2950 				pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
2951 			if (rc != OPAL_SUCCESS) {
2952 				pr_err("%s: Error %lld mapping IO segment#%d to PE#%d\n",
2953 				       __func__, rc, index, pe->pe_number);
2954 				break;
2955 			}
2956 
2957 			region.start += phb->ioda.io_segsize;
2958 			index++;
2959 		}
2960 	} else if ((res->flags & IORESOURCE_MEM) &&
2961 		   !pnv_pci_is_mem_pref_64(res->flags)) {
2962 		region.start = res->start -
2963 			       phb->hose->mem_offset[0] -
2964 			       phb->ioda.m32_pci_base;
2965 		region.end   = res->end -
2966 			       phb->hose->mem_offset[0] -
2967 			       phb->ioda.m32_pci_base;
2968 		index = region.start / phb->ioda.m32_segsize;
2969 
2970 		while (index < phb->ioda.total_pe_num &&
2971 		       region.start <= region.end) {
2972 			phb->ioda.m32_segmap[index] = pe->pe_number;
2973 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2974 				pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
2975 			if (rc != OPAL_SUCCESS) {
2976 				pr_err("%s: Error %lld mapping M32 segment#%d to PE#%d",
2977 				       __func__, rc, index, pe->pe_number);
2978 				break;
2979 			}
2980 
2981 			region.start += phb->ioda.m32_segsize;
2982 			index++;
2983 		}
2984 	}
2985 }
2986 
2987 /*
2988  * This function is supposed to be called on basis of PE from top
2989  * to bottom style. So the the I/O or MMIO segment assigned to
2990  * parent PE could be overrided by its child PEs if necessary.
2991  */
2992 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe)
2993 {
2994 	struct pci_dev *pdev;
2995 	int i;
2996 
2997 	/*
2998 	 * NOTE: We only care PCI bus based PE for now. For PCI
2999 	 * device based PE, for example SRIOV sensitive VF should
3000 	 * be figured out later.
3001 	 */
3002 	BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
3003 
3004 	list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
3005 		for (i = 0; i <= PCI_ROM_RESOURCE; i++)
3006 			pnv_ioda_setup_pe_res(pe, &pdev->resource[i]);
3007 
3008 		/*
3009 		 * If the PE contains all subordinate PCI buses, the
3010 		 * windows of the child bridges should be mapped to
3011 		 * the PE as well.
3012 		 */
3013 		if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev))
3014 			continue;
3015 		for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++)
3016 			pnv_ioda_setup_pe_res(pe,
3017 				&pdev->resource[PCI_BRIDGE_RESOURCES + i]);
3018 	}
3019 }
3020 
3021 static void pnv_pci_ioda_create_dbgfs(void)
3022 {
3023 #ifdef CONFIG_DEBUG_FS
3024 	struct pci_controller *hose, *tmp;
3025 	struct pnv_phb *phb;
3026 	char name[16];
3027 
3028 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
3029 		phb = hose->private_data;
3030 
3031 		/* Notify initialization of PHB done */
3032 		phb->initialized = 1;
3033 
3034 		sprintf(name, "PCI%04x", hose->global_number);
3035 		phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
3036 		if (!phb->dbgfs)
3037 			pr_warning("%s: Error on creating debugfs on PHB#%x\n",
3038 				__func__, hose->global_number);
3039 	}
3040 #endif /* CONFIG_DEBUG_FS */
3041 }
3042 
3043 static void pnv_pci_ioda_fixup(void)
3044 {
3045 	pnv_pci_ioda_setup_PEs();
3046 	pnv_pci_ioda_setup_iommu_api();
3047 	pnv_pci_ioda_create_dbgfs();
3048 
3049 #ifdef CONFIG_EEH
3050 	eeh_init();
3051 	eeh_addr_cache_build();
3052 #endif
3053 }
3054 
3055 /*
3056  * Returns the alignment for I/O or memory windows for P2P
3057  * bridges. That actually depends on how PEs are segmented.
3058  * For now, we return I/O or M32 segment size for PE sensitive
3059  * P2P bridges. Otherwise, the default values (4KiB for I/O,
3060  * 1MiB for memory) will be returned.
3061  *
3062  * The current PCI bus might be put into one PE, which was
3063  * create against the parent PCI bridge. For that case, we
3064  * needn't enlarge the alignment so that we can save some
3065  * resources.
3066  */
3067 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
3068 						unsigned long type)
3069 {
3070 	struct pci_dev *bridge;
3071 	struct pci_controller *hose = pci_bus_to_host(bus);
3072 	struct pnv_phb *phb = hose->private_data;
3073 	int num_pci_bridges = 0;
3074 
3075 	bridge = bus->self;
3076 	while (bridge) {
3077 		if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
3078 			num_pci_bridges++;
3079 			if (num_pci_bridges >= 2)
3080 				return 1;
3081 		}
3082 
3083 		bridge = bridge->bus->self;
3084 	}
3085 
3086 	/* We fail back to M32 if M64 isn't supported */
3087 	if (phb->ioda.m64_segsize &&
3088 	    pnv_pci_is_mem_pref_64(type))
3089 		return phb->ioda.m64_segsize;
3090 	if (type & IORESOURCE_MEM)
3091 		return phb->ioda.m32_segsize;
3092 
3093 	return phb->ioda.io_segsize;
3094 }
3095 
3096 /*
3097  * We are updating root port or the upstream port of the
3098  * bridge behind the root port with PHB's windows in order
3099  * to accommodate the changes on required resources during
3100  * PCI (slot) hotplug, which is connected to either root
3101  * port or the downstream ports of PCIe switch behind the
3102  * root port.
3103  */
3104 static void pnv_pci_fixup_bridge_resources(struct pci_bus *bus,
3105 					   unsigned long type)
3106 {
3107 	struct pci_controller *hose = pci_bus_to_host(bus);
3108 	struct pnv_phb *phb = hose->private_data;
3109 	struct pci_dev *bridge = bus->self;
3110 	struct resource *r, *w;
3111 	bool msi_region = false;
3112 	int i;
3113 
3114 	/* Check if we need apply fixup to the bridge's windows */
3115 	if (!pci_is_root_bus(bridge->bus) &&
3116 	    !pci_is_root_bus(bridge->bus->self->bus))
3117 		return;
3118 
3119 	/* Fixup the resources */
3120 	for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
3121 		r = &bridge->resource[PCI_BRIDGE_RESOURCES + i];
3122 		if (!r->flags || !r->parent)
3123 			continue;
3124 
3125 		w = NULL;
3126 		if (r->flags & type & IORESOURCE_IO)
3127 			w = &hose->io_resource;
3128 		else if (pnv_pci_is_mem_pref_64(r->flags) &&
3129 			 (type & IORESOURCE_PREFETCH) &&
3130 			 phb->ioda.m64_segsize)
3131 			w = &hose->mem_resources[1];
3132 		else if (r->flags & type & IORESOURCE_MEM) {
3133 			w = &hose->mem_resources[0];
3134 			msi_region = true;
3135 		}
3136 
3137 		r->start = w->start;
3138 		r->end = w->end;
3139 
3140 		/* The 64KB 32-bits MSI region shouldn't be included in
3141 		 * the 32-bits bridge window. Otherwise, we can see strange
3142 		 * issues. One of them is EEH error observed on Garrison.
3143 		 *
3144 		 * Exclude top 1MB region which is the minimal alignment of
3145 		 * 32-bits bridge window.
3146 		 */
3147 		if (msi_region) {
3148 			r->end += 0x10000;
3149 			r->end -= 0x100000;
3150 		}
3151 	}
3152 }
3153 
3154 static void pnv_pci_setup_bridge(struct pci_bus *bus, unsigned long type)
3155 {
3156 	struct pci_controller *hose = pci_bus_to_host(bus);
3157 	struct pnv_phb *phb = hose->private_data;
3158 	struct pci_dev *bridge = bus->self;
3159 	struct pnv_ioda_pe *pe;
3160 	bool all = (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE);
3161 
3162 	/* Extend bridge's windows if necessary */
3163 	pnv_pci_fixup_bridge_resources(bus, type);
3164 
3165 	/* The PE for root bus should be realized before any one else */
3166 	if (!phb->ioda.root_pe_populated) {
3167 		pe = pnv_ioda_setup_bus_PE(phb->hose->bus, false);
3168 		if (pe) {
3169 			phb->ioda.root_pe_idx = pe->pe_number;
3170 			phb->ioda.root_pe_populated = true;
3171 		}
3172 	}
3173 
3174 	/* Don't assign PE to PCI bus, which doesn't have subordinate devices */
3175 	if (list_empty(&bus->devices))
3176 		return;
3177 
3178 	/* Reserve PEs according to used M64 resources */
3179 	if (phb->reserve_m64_pe)
3180 		phb->reserve_m64_pe(bus, NULL, all);
3181 
3182 	/*
3183 	 * Assign PE. We might run here because of partial hotplug.
3184 	 * For the case, we just pick up the existing PE and should
3185 	 * not allocate resources again.
3186 	 */
3187 	pe = pnv_ioda_setup_bus_PE(bus, all);
3188 	if (!pe)
3189 		return;
3190 
3191 	pnv_ioda_setup_pe_seg(pe);
3192 	switch (phb->type) {
3193 	case PNV_PHB_IODA1:
3194 		pnv_pci_ioda1_setup_dma_pe(phb, pe);
3195 		break;
3196 	case PNV_PHB_IODA2:
3197 		pnv_pci_ioda2_setup_dma_pe(phb, pe);
3198 		break;
3199 	default:
3200 		pr_warn("%s: No DMA for PHB#%d (type %d)\n",
3201 			__func__, phb->hose->global_number, phb->type);
3202 	}
3203 }
3204 
3205 #ifdef CONFIG_PCI_IOV
3206 static resource_size_t pnv_pci_iov_resource_alignment(struct pci_dev *pdev,
3207 						      int resno)
3208 {
3209 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3210 	struct pnv_phb *phb = hose->private_data;
3211 	struct pci_dn *pdn = pci_get_pdn(pdev);
3212 	resource_size_t align;
3213 
3214 	/*
3215 	 * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the
3216 	 * SR-IOV. While from hardware perspective, the range mapped by M64
3217 	 * BAR should be size aligned.
3218 	 *
3219 	 * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra
3220 	 * powernv-specific hardware restriction is gone. But if just use the
3221 	 * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with
3222 	 * in one segment of M64 #15, which introduces the PE conflict between
3223 	 * PF and VF. Based on this, the minimum alignment of an IOV BAR is
3224 	 * m64_segsize.
3225 	 *
3226 	 * This function returns the total IOV BAR size if M64 BAR is in
3227 	 * Shared PE mode or just VF BAR size if not.
3228 	 * If the M64 BAR is in Single PE mode, return the VF BAR size or
3229 	 * M64 segment size if IOV BAR size is less.
3230 	 */
3231 	align = pci_iov_resource_size(pdev, resno);
3232 	if (!pdn->vfs_expanded)
3233 		return align;
3234 	if (pdn->m64_single_mode)
3235 		return max(align, (resource_size_t)phb->ioda.m64_segsize);
3236 
3237 	return pdn->vfs_expanded * align;
3238 }
3239 #endif /* CONFIG_PCI_IOV */
3240 
3241 /* Prevent enabling devices for which we couldn't properly
3242  * assign a PE
3243  */
3244 bool pnv_pci_enable_device_hook(struct pci_dev *dev)
3245 {
3246 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
3247 	struct pnv_phb *phb = hose->private_data;
3248 	struct pci_dn *pdn;
3249 
3250 	/* The function is probably called while the PEs have
3251 	 * not be created yet. For example, resource reassignment
3252 	 * during PCI probe period. We just skip the check if
3253 	 * PEs isn't ready.
3254 	 */
3255 	if (!phb->initialized)
3256 		return true;
3257 
3258 	pdn = pci_get_pdn(dev);
3259 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3260 		return false;
3261 
3262 	return true;
3263 }
3264 
3265 static long pnv_pci_ioda1_unset_window(struct iommu_table_group *table_group,
3266 				       int num)
3267 {
3268 	struct pnv_ioda_pe *pe = container_of(table_group,
3269 					      struct pnv_ioda_pe, table_group);
3270 	struct pnv_phb *phb = pe->phb;
3271 	unsigned int idx;
3272 	long rc;
3273 
3274 	pe_info(pe, "Removing DMA window #%d\n", num);
3275 	for (idx = 0; idx < phb->ioda.dma32_count; idx++) {
3276 		if (phb->ioda.dma32_segmap[idx] != pe->pe_number)
3277 			continue;
3278 
3279 		rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
3280 						idx, 0, 0ul, 0ul, 0ul);
3281 		if (rc != OPAL_SUCCESS) {
3282 			pe_warn(pe, "Failure %ld unmapping DMA32 segment#%d\n",
3283 				rc, idx);
3284 			return rc;
3285 		}
3286 
3287 		phb->ioda.dma32_segmap[idx] = IODA_INVALID_PE;
3288 	}
3289 
3290 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
3291 	return OPAL_SUCCESS;
3292 }
3293 
3294 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe *pe)
3295 {
3296 	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3297 	struct iommu_table *tbl = pe->table_group.tables[0];
3298 	int64_t rc;
3299 
3300 	if (!weight)
3301 		return;
3302 
3303 	rc = pnv_pci_ioda1_unset_window(&pe->table_group, 0);
3304 	if (rc != OPAL_SUCCESS)
3305 		return;
3306 
3307 	pnv_pci_p7ioc_tce_invalidate(tbl, tbl->it_offset, tbl->it_size, false);
3308 	if (pe->table_group.group) {
3309 		iommu_group_put(pe->table_group.group);
3310 		WARN_ON(pe->table_group.group);
3311 	}
3312 
3313 	free_pages(tbl->it_base, get_order(tbl->it_size << 3));
3314 	iommu_free_table(tbl, "pnv");
3315 }
3316 
3317 static void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe *pe)
3318 {
3319 	struct iommu_table *tbl = pe->table_group.tables[0];
3320 	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3321 #ifdef CONFIG_IOMMU_API
3322 	int64_t rc;
3323 #endif
3324 
3325 	if (!weight)
3326 		return;
3327 
3328 #ifdef CONFIG_IOMMU_API
3329 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
3330 	if (rc)
3331 		pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
3332 #endif
3333 
3334 	pnv_pci_ioda2_set_bypass(pe, false);
3335 	if (pe->table_group.group) {
3336 		iommu_group_put(pe->table_group.group);
3337 		WARN_ON(pe->table_group.group);
3338 	}
3339 
3340 	pnv_pci_ioda2_table_free_pages(tbl);
3341 	iommu_free_table(tbl, "pnv");
3342 }
3343 
3344 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe *pe,
3345 				 unsigned short win,
3346 				 unsigned int *map)
3347 {
3348 	struct pnv_phb *phb = pe->phb;
3349 	int idx;
3350 	int64_t rc;
3351 
3352 	for (idx = 0; idx < phb->ioda.total_pe_num; idx++) {
3353 		if (map[idx] != pe->pe_number)
3354 			continue;
3355 
3356 		if (win == OPAL_M64_WINDOW_TYPE)
3357 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3358 					phb->ioda.reserved_pe_idx, win,
3359 					idx / PNV_IODA1_M64_SEGS,
3360 					idx % PNV_IODA1_M64_SEGS);
3361 		else
3362 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3363 					phb->ioda.reserved_pe_idx, win, 0, idx);
3364 
3365 		if (rc != OPAL_SUCCESS)
3366 			pe_warn(pe, "Error %ld unmapping (%d) segment#%d\n",
3367 				rc, win, idx);
3368 
3369 		map[idx] = IODA_INVALID_PE;
3370 	}
3371 }
3372 
3373 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
3374 {
3375 	struct pnv_phb *phb = pe->phb;
3376 
3377 	if (phb->type == PNV_PHB_IODA1) {
3378 		pnv_ioda_free_pe_seg(pe, OPAL_IO_WINDOW_TYPE,
3379 				     phb->ioda.io_segmap);
3380 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3381 				     phb->ioda.m32_segmap);
3382 		pnv_ioda_free_pe_seg(pe, OPAL_M64_WINDOW_TYPE,
3383 				     phb->ioda.m64_segmap);
3384 	} else if (phb->type == PNV_PHB_IODA2) {
3385 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3386 				     phb->ioda.m32_segmap);
3387 	}
3388 }
3389 
3390 static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe)
3391 {
3392 	struct pnv_phb *phb = pe->phb;
3393 	struct pnv_ioda_pe *slave, *tmp;
3394 
3395 	/* Release slave PEs in compound PE */
3396 	if (pe->flags & PNV_IODA_PE_MASTER) {
3397 		list_for_each_entry_safe(slave, tmp, &pe->slaves, list)
3398 			pnv_ioda_release_pe(slave);
3399 	}
3400 
3401 	list_del(&pe->list);
3402 	switch (phb->type) {
3403 	case PNV_PHB_IODA1:
3404 		pnv_pci_ioda1_release_pe_dma(pe);
3405 		break;
3406 	case PNV_PHB_IODA2:
3407 		pnv_pci_ioda2_release_pe_dma(pe);
3408 		break;
3409 	default:
3410 		WARN_ON(1);
3411 	}
3412 
3413 	pnv_ioda_release_pe_seg(pe);
3414 	pnv_ioda_deconfigure_pe(pe->phb, pe);
3415 	pnv_ioda_free_pe(pe);
3416 }
3417 
3418 static void pnv_pci_release_device(struct pci_dev *pdev)
3419 {
3420 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3421 	struct pnv_phb *phb = hose->private_data;
3422 	struct pci_dn *pdn = pci_get_pdn(pdev);
3423 	struct pnv_ioda_pe *pe;
3424 
3425 	if (pdev->is_virtfn)
3426 		return;
3427 
3428 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3429 		return;
3430 
3431 	pe = &phb->ioda.pe_array[pdn->pe_number];
3432 	WARN_ON(--pe->device_count < 0);
3433 	if (pe->device_count == 0)
3434 		pnv_ioda_release_pe(pe);
3435 }
3436 
3437 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
3438 {
3439 	struct pnv_phb *phb = hose->private_data;
3440 
3441 	opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
3442 		       OPAL_ASSERT_RESET);
3443 }
3444 
3445 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
3446 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3447 	.dma_bus_setup		= pnv_pci_dma_bus_setup,
3448 #ifdef CONFIG_PCI_MSI
3449 	.setup_msi_irqs		= pnv_setup_msi_irqs,
3450 	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
3451 #endif
3452 	.enable_device_hook	= pnv_pci_enable_device_hook,
3453 	.release_device		= pnv_pci_release_device,
3454 	.window_alignment	= pnv_pci_window_alignment,
3455 	.setup_bridge		= pnv_pci_setup_bridge,
3456 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3457 	.dma_set_mask		= pnv_pci_ioda_dma_set_mask,
3458 	.dma_get_required_mask	= pnv_pci_ioda_dma_get_required_mask,
3459 	.shutdown		= pnv_pci_ioda_shutdown,
3460 };
3461 
3462 static int pnv_npu_dma_set_mask(struct pci_dev *npdev, u64 dma_mask)
3463 {
3464 	dev_err_once(&npdev->dev,
3465 			"%s operation unsupported for NVLink devices\n",
3466 			__func__);
3467 	return -EPERM;
3468 }
3469 
3470 static const struct pci_controller_ops pnv_npu_ioda_controller_ops = {
3471 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3472 #ifdef CONFIG_PCI_MSI
3473 	.setup_msi_irqs		= pnv_setup_msi_irqs,
3474 	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
3475 #endif
3476 	.enable_device_hook	= pnv_pci_enable_device_hook,
3477 	.window_alignment	= pnv_pci_window_alignment,
3478 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3479 	.dma_set_mask		= pnv_npu_dma_set_mask,
3480 	.shutdown		= pnv_pci_ioda_shutdown,
3481 };
3482 
3483 #ifdef CONFIG_CXL_BASE
3484 const struct pci_controller_ops pnv_cxl_cx4_ioda_controller_ops = {
3485 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3486 	.dma_bus_setup		= pnv_pci_dma_bus_setup,
3487 #ifdef CONFIG_PCI_MSI
3488 	.setup_msi_irqs		= pnv_cxl_cx4_setup_msi_irqs,
3489 	.teardown_msi_irqs	= pnv_cxl_cx4_teardown_msi_irqs,
3490 #endif
3491 	.enable_device_hook	= pnv_cxl_enable_device_hook,
3492 	.disable_device		= pnv_cxl_disable_device,
3493 	.release_device		= pnv_pci_release_device,
3494 	.window_alignment	= pnv_pci_window_alignment,
3495 	.setup_bridge		= pnv_pci_setup_bridge,
3496 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3497 	.dma_set_mask		= pnv_pci_ioda_dma_set_mask,
3498 	.dma_get_required_mask	= pnv_pci_ioda_dma_get_required_mask,
3499 	.shutdown		= pnv_pci_ioda_shutdown,
3500 };
3501 #endif
3502 
3503 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
3504 					 u64 hub_id, int ioda_type)
3505 {
3506 	struct pci_controller *hose;
3507 	struct pnv_phb *phb;
3508 	unsigned long size, m64map_off, m32map_off, pemap_off;
3509 	unsigned long iomap_off = 0, dma32map_off = 0;
3510 	struct resource r;
3511 	const __be64 *prop64;
3512 	const __be32 *prop32;
3513 	int len;
3514 	unsigned int segno;
3515 	u64 phb_id;
3516 	void *aux;
3517 	long rc;
3518 
3519 	if (!of_device_is_available(np))
3520 		return;
3521 
3522 	pr_info("Initializing %s PHB (%s)\n",
3523 		pnv_phb_names[ioda_type], of_node_full_name(np));
3524 
3525 	prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
3526 	if (!prop64) {
3527 		pr_err("  Missing \"ibm,opal-phbid\" property !\n");
3528 		return;
3529 	}
3530 	phb_id = be64_to_cpup(prop64);
3531 	pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
3532 
3533 	phb = memblock_virt_alloc(sizeof(struct pnv_phb), 0);
3534 
3535 	/* Allocate PCI controller */
3536 	phb->hose = hose = pcibios_alloc_controller(np);
3537 	if (!phb->hose) {
3538 		pr_err("  Can't allocate PCI controller for %s\n",
3539 		       np->full_name);
3540 		memblock_free(__pa(phb), sizeof(struct pnv_phb));
3541 		return;
3542 	}
3543 
3544 	spin_lock_init(&phb->lock);
3545 	prop32 = of_get_property(np, "bus-range", &len);
3546 	if (prop32 && len == 8) {
3547 		hose->first_busno = be32_to_cpu(prop32[0]);
3548 		hose->last_busno = be32_to_cpu(prop32[1]);
3549 	} else {
3550 		pr_warn("  Broken <bus-range> on %s\n", np->full_name);
3551 		hose->first_busno = 0;
3552 		hose->last_busno = 0xff;
3553 	}
3554 	hose->private_data = phb;
3555 	phb->hub_id = hub_id;
3556 	phb->opal_id = phb_id;
3557 	phb->type = ioda_type;
3558 	mutex_init(&phb->ioda.pe_alloc_mutex);
3559 
3560 	/* Detect specific models for error handling */
3561 	if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
3562 		phb->model = PNV_PHB_MODEL_P7IOC;
3563 	else if (of_device_is_compatible(np, "ibm,power8-pciex"))
3564 		phb->model = PNV_PHB_MODEL_PHB3;
3565 	else if (of_device_is_compatible(np, "ibm,power8-npu-pciex"))
3566 		phb->model = PNV_PHB_MODEL_NPU;
3567 	else
3568 		phb->model = PNV_PHB_MODEL_UNKNOWN;
3569 
3570 	/* Parse 32-bit and IO ranges (if any) */
3571 	pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
3572 
3573 	/* Get registers */
3574 	if (!of_address_to_resource(np, 0, &r)) {
3575 		phb->regs_phys = r.start;
3576 		phb->regs = ioremap(r.start, resource_size(&r));
3577 		if (phb->regs == NULL)
3578 			pr_err("  Failed to map registers !\n");
3579 	}
3580 
3581 	/* Initialize more IODA stuff */
3582 	phb->ioda.total_pe_num = 1;
3583 	prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
3584 	if (prop32)
3585 		phb->ioda.total_pe_num = be32_to_cpup(prop32);
3586 	prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
3587 	if (prop32)
3588 		phb->ioda.reserved_pe_idx = be32_to_cpup(prop32);
3589 
3590 	/* Invalidate RID to PE# mapping */
3591 	for (segno = 0; segno < ARRAY_SIZE(phb->ioda.pe_rmap); segno++)
3592 		phb->ioda.pe_rmap[segno] = IODA_INVALID_PE;
3593 
3594 	/* Parse 64-bit MMIO range */
3595 	pnv_ioda_parse_m64_window(phb);
3596 
3597 	phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
3598 	/* FW Has already off top 64k of M32 space (MSI space) */
3599 	phb->ioda.m32_size += 0x10000;
3600 
3601 	phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num;
3602 	phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
3603 	phb->ioda.io_size = hose->pci_io_size;
3604 	phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num;
3605 	phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
3606 
3607 	/* Calculate how many 32-bit TCE segments we have */
3608 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3609 				PNV_IODA1_DMA32_SEGSIZE;
3610 
3611 	/* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3612 	size = _ALIGN_UP(max_t(unsigned, phb->ioda.total_pe_num, 8) / 8,
3613 			sizeof(unsigned long));
3614 	m64map_off = size;
3615 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]);
3616 	m32map_off = size;
3617 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]);
3618 	if (phb->type == PNV_PHB_IODA1) {
3619 		iomap_off = size;
3620 		size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]);
3621 		dma32map_off = size;
3622 		size += phb->ioda.dma32_count *
3623 			sizeof(phb->ioda.dma32_segmap[0]);
3624 	}
3625 	pemap_off = size;
3626 	size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe);
3627 	aux = memblock_virt_alloc(size, 0);
3628 	phb->ioda.pe_alloc = aux;
3629 	phb->ioda.m64_segmap = aux + m64map_off;
3630 	phb->ioda.m32_segmap = aux + m32map_off;
3631 	for (segno = 0; segno < phb->ioda.total_pe_num; segno++) {
3632 		phb->ioda.m64_segmap[segno] = IODA_INVALID_PE;
3633 		phb->ioda.m32_segmap[segno] = IODA_INVALID_PE;
3634 	}
3635 	if (phb->type == PNV_PHB_IODA1) {
3636 		phb->ioda.io_segmap = aux + iomap_off;
3637 		for (segno = 0; segno < phb->ioda.total_pe_num; segno++)
3638 			phb->ioda.io_segmap[segno] = IODA_INVALID_PE;
3639 
3640 		phb->ioda.dma32_segmap = aux + dma32map_off;
3641 		for (segno = 0; segno < phb->ioda.dma32_count; segno++)
3642 			phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE;
3643 	}
3644 	phb->ioda.pe_array = aux + pemap_off;
3645 
3646 	/*
3647 	 * Choose PE number for root bus, which shouldn't have
3648 	 * M64 resources consumed by its child devices. To pick
3649 	 * the PE number adjacent to the reserved one if possible.
3650 	 */
3651 	pnv_ioda_reserve_pe(phb, phb->ioda.reserved_pe_idx);
3652 	if (phb->ioda.reserved_pe_idx == 0) {
3653 		phb->ioda.root_pe_idx = 1;
3654 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3655 	} else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) {
3656 		phb->ioda.root_pe_idx = phb->ioda.reserved_pe_idx - 1;
3657 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3658 	} else {
3659 		phb->ioda.root_pe_idx = IODA_INVALID_PE;
3660 	}
3661 
3662 	INIT_LIST_HEAD(&phb->ioda.pe_list);
3663 	mutex_init(&phb->ioda.pe_list_mutex);
3664 
3665 	/* Calculate how many 32-bit TCE segments we have */
3666 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3667 				PNV_IODA1_DMA32_SEGSIZE;
3668 
3669 #if 0 /* We should really do that ... */
3670 	rc = opal_pci_set_phb_mem_window(opal->phb_id,
3671 					 window_type,
3672 					 window_num,
3673 					 starting_real_address,
3674 					 starting_pci_address,
3675 					 segment_size);
3676 #endif
3677 
3678 	pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3679 		phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx,
3680 		phb->ioda.m32_size, phb->ioda.m32_segsize);
3681 	if (phb->ioda.m64_size)
3682 		pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
3683 			phb->ioda.m64_size, phb->ioda.m64_segsize);
3684 	if (phb->ioda.io_size)
3685 		pr_info("                  IO: 0x%x [segment=0x%x]\n",
3686 			phb->ioda.io_size, phb->ioda.io_segsize);
3687 
3688 
3689 	phb->hose->ops = &pnv_pci_ops;
3690 	phb->get_pe_state = pnv_ioda_get_pe_state;
3691 	phb->freeze_pe = pnv_ioda_freeze_pe;
3692 	phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
3693 
3694 	/* Setup MSI support */
3695 	pnv_pci_init_ioda_msis(phb);
3696 
3697 	/*
3698 	 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3699 	 * to let the PCI core do resource assignment. It's supposed
3700 	 * that the PCI core will do correct I/O and MMIO alignment
3701 	 * for the P2P bridge bars so that each PCI bus (excluding
3702 	 * the child P2P bridges) can form individual PE.
3703 	 */
3704 	ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
3705 
3706 	if (phb->type == PNV_PHB_NPU) {
3707 		hose->controller_ops = pnv_npu_ioda_controller_ops;
3708 	} else {
3709 		phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
3710 		hose->controller_ops = pnv_pci_ioda_controller_ops;
3711 	}
3712 
3713 #ifdef CONFIG_PCI_IOV
3714 	ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov_resources;
3715 	ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
3716 #endif
3717 
3718 	pci_add_flags(PCI_REASSIGN_ALL_RSRC);
3719 
3720 	/* Reset IODA tables to a clean state */
3721 	rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
3722 	if (rc)
3723 		pr_warning("  OPAL Error %ld performing IODA table reset !\n", rc);
3724 
3725 	/* If we're running in kdump kerenl, the previous kerenl never
3726 	 * shutdown PCI devices correctly. We already got IODA table
3727 	 * cleaned out. So we have to issue PHB reset to stop all PCI
3728 	 * transactions from previous kerenl.
3729 	 */
3730 	if (is_kdump_kernel()) {
3731 		pr_info("  Issue PHB reset ...\n");
3732 		pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
3733 		pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
3734 	}
3735 
3736 	/* Remove M64 resource if we can't configure it successfully */
3737 	if (!phb->init_m64 || phb->init_m64(phb))
3738 		hose->mem_resources[1].flags = 0;
3739 }
3740 
3741 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
3742 {
3743 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
3744 }
3745 
3746 void __init pnv_pci_init_npu_phb(struct device_node *np)
3747 {
3748 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU);
3749 }
3750 
3751 void __init pnv_pci_init_ioda_hub(struct device_node *np)
3752 {
3753 	struct device_node *phbn;
3754 	const __be64 *prop64;
3755 	u64 hub_id;
3756 
3757 	pr_info("Probing IODA IO-Hub %s\n", np->full_name);
3758 
3759 	prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
3760 	if (!prop64) {
3761 		pr_err(" Missing \"ibm,opal-hubid\" property !\n");
3762 		return;
3763 	}
3764 	hub_id = be64_to_cpup(prop64);
3765 	pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
3766 
3767 	/* Count child PHBs */
3768 	for_each_child_of_node(np, phbn) {
3769 		/* Look for IODA1 PHBs */
3770 		if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
3771 			pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
3772 	}
3773 }
3774