xref: /linux/arch/powerpc/platforms/powernv/pci-ioda.c (revision 59024954a1e7e26b62680e1f2b5725249a6c09f7)
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_m64(struct pnv_phb *phb, struct resource *r)
115 {
116 	/*
117 	 * WARNING: We cannot rely on the resource flags. The Linux PCI
118 	 * allocation code sometimes decides to put a 64-bit prefetchable
119 	 * BAR in the 32-bit window, so we have to compare the addresses.
120 	 *
121 	 * For simplicity we only test resource start.
122 	 */
123 	return (r->start >= phb->ioda.m64_base &&
124 		r->start < (phb->ioda.m64_base + phb->ioda.m64_size));
125 }
126 
127 static inline bool pnv_pci_is_m64_flags(unsigned long resource_flags)
128 {
129 	unsigned long flags = (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
130 
131 	return (resource_flags & flags) == flags;
132 }
133 
134 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
135 {
136 	phb->ioda.pe_array[pe_no].phb = phb;
137 	phb->ioda.pe_array[pe_no].pe_number = pe_no;
138 
139 	return &phb->ioda.pe_array[pe_no];
140 }
141 
142 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
143 {
144 	if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) {
145 		pr_warn("%s: Invalid PE %d on PHB#%x\n",
146 			__func__, pe_no, phb->hose->global_number);
147 		return;
148 	}
149 
150 	if (test_and_set_bit(pe_no, phb->ioda.pe_alloc))
151 		pr_debug("%s: PE %d was reserved on PHB#%x\n",
152 			 __func__, pe_no, phb->hose->global_number);
153 
154 	pnv_ioda_init_pe(phb, pe_no);
155 }
156 
157 static struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb)
158 {
159 	long pe;
160 
161 	for (pe = phb->ioda.total_pe_num - 1; pe >= 0; pe--) {
162 		if (!test_and_set_bit(pe, phb->ioda.pe_alloc))
163 			return pnv_ioda_init_pe(phb, pe);
164 	}
165 
166 	return NULL;
167 }
168 
169 static void pnv_ioda_free_pe(struct pnv_ioda_pe *pe)
170 {
171 	struct pnv_phb *phb = pe->phb;
172 	unsigned int pe_num = pe->pe_number;
173 
174 	WARN_ON(pe->pdev);
175 
176 	memset(pe, 0, sizeof(struct pnv_ioda_pe));
177 	clear_bit(pe_num, phb->ioda.pe_alloc);
178 }
179 
180 /* The default M64 BAR is shared by all PEs */
181 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
182 {
183 	const char *desc;
184 	struct resource *r;
185 	s64 rc;
186 
187 	/* Configure the default M64 BAR */
188 	rc = opal_pci_set_phb_mem_window(phb->opal_id,
189 					 OPAL_M64_WINDOW_TYPE,
190 					 phb->ioda.m64_bar_idx,
191 					 phb->ioda.m64_base,
192 					 0, /* unused */
193 					 phb->ioda.m64_size);
194 	if (rc != OPAL_SUCCESS) {
195 		desc = "configuring";
196 		goto fail;
197 	}
198 
199 	/* Enable the default M64 BAR */
200 	rc = opal_pci_phb_mmio_enable(phb->opal_id,
201 				      OPAL_M64_WINDOW_TYPE,
202 				      phb->ioda.m64_bar_idx,
203 				      OPAL_ENABLE_M64_SPLIT);
204 	if (rc != OPAL_SUCCESS) {
205 		desc = "enabling";
206 		goto fail;
207 	}
208 
209 	/*
210 	 * Exclude the segments for reserved and root bus PE, which
211 	 * are first or last two PEs.
212 	 */
213 	r = &phb->hose->mem_resources[1];
214 	if (phb->ioda.reserved_pe_idx == 0)
215 		r->start += (2 * phb->ioda.m64_segsize);
216 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
217 		r->end -= (2 * phb->ioda.m64_segsize);
218 	else
219 		pr_warn("  Cannot strip M64 segment for reserved PE#%d\n",
220 			phb->ioda.reserved_pe_idx);
221 
222 	return 0;
223 
224 fail:
225 	pr_warn("  Failure %lld %s M64 BAR#%d\n",
226 		rc, desc, phb->ioda.m64_bar_idx);
227 	opal_pci_phb_mmio_enable(phb->opal_id,
228 				 OPAL_M64_WINDOW_TYPE,
229 				 phb->ioda.m64_bar_idx,
230 				 OPAL_DISABLE_M64);
231 	return -EIO;
232 }
233 
234 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev,
235 					 unsigned long *pe_bitmap)
236 {
237 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
238 	struct pnv_phb *phb = hose->private_data;
239 	struct resource *r;
240 	resource_size_t base, sgsz, start, end;
241 	int segno, i;
242 
243 	base = phb->ioda.m64_base;
244 	sgsz = phb->ioda.m64_segsize;
245 	for (i = 0; i <= PCI_ROM_RESOURCE; i++) {
246 		r = &pdev->resource[i];
247 		if (!r->parent || !pnv_pci_is_m64(phb, r))
248 			continue;
249 
250 		start = _ALIGN_DOWN(r->start - base, sgsz);
251 		end = _ALIGN_UP(r->end - base, sgsz);
252 		for (segno = start / sgsz; segno < end / sgsz; segno++) {
253 			if (pe_bitmap)
254 				set_bit(segno, pe_bitmap);
255 			else
256 				pnv_ioda_reserve_pe(phb, segno);
257 		}
258 	}
259 }
260 
261 static int pnv_ioda1_init_m64(struct pnv_phb *phb)
262 {
263 	struct resource *r;
264 	int index;
265 
266 	/*
267 	 * There are 16 M64 BARs, each of which has 8 segments. So
268 	 * there are as many M64 segments as the maximum number of
269 	 * PEs, which is 128.
270 	 */
271 	for (index = 0; index < PNV_IODA1_M64_NUM; index++) {
272 		unsigned long base, segsz = phb->ioda.m64_segsize;
273 		int64_t rc;
274 
275 		base = phb->ioda.m64_base +
276 		       index * PNV_IODA1_M64_SEGS * segsz;
277 		rc = opal_pci_set_phb_mem_window(phb->opal_id,
278 				OPAL_M64_WINDOW_TYPE, index, base, 0,
279 				PNV_IODA1_M64_SEGS * segsz);
280 		if (rc != OPAL_SUCCESS) {
281 			pr_warn("  Error %lld setting M64 PHB#%d-BAR#%d\n",
282 				rc, phb->hose->global_number, index);
283 			goto fail;
284 		}
285 
286 		rc = opal_pci_phb_mmio_enable(phb->opal_id,
287 				OPAL_M64_WINDOW_TYPE, index,
288 				OPAL_ENABLE_M64_SPLIT);
289 		if (rc != OPAL_SUCCESS) {
290 			pr_warn("  Error %lld enabling M64 PHB#%d-BAR#%d\n",
291 				rc, phb->hose->global_number, index);
292 			goto fail;
293 		}
294 	}
295 
296 	/*
297 	 * Exclude the segments for reserved and root bus PE, which
298 	 * are first or last two PEs.
299 	 */
300 	r = &phb->hose->mem_resources[1];
301 	if (phb->ioda.reserved_pe_idx == 0)
302 		r->start += (2 * phb->ioda.m64_segsize);
303 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
304 		r->end -= (2 * phb->ioda.m64_segsize);
305 	else
306 		WARN(1, "Wrong reserved PE#%d on PHB#%d\n",
307 		     phb->ioda.reserved_pe_idx, phb->hose->global_number);
308 
309 	return 0;
310 
311 fail:
312 	for ( ; index >= 0; index--)
313 		opal_pci_phb_mmio_enable(phb->opal_id,
314 			OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64);
315 
316 	return -EIO;
317 }
318 
319 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus,
320 				    unsigned long *pe_bitmap,
321 				    bool all)
322 {
323 	struct pci_dev *pdev;
324 
325 	list_for_each_entry(pdev, &bus->devices, bus_list) {
326 		pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap);
327 
328 		if (all && pdev->subordinate)
329 			pnv_ioda_reserve_m64_pe(pdev->subordinate,
330 						pe_bitmap, all);
331 	}
332 }
333 
334 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all)
335 {
336 	struct pci_controller *hose = pci_bus_to_host(bus);
337 	struct pnv_phb *phb = hose->private_data;
338 	struct pnv_ioda_pe *master_pe, *pe;
339 	unsigned long size, *pe_alloc;
340 	int i;
341 
342 	/* Root bus shouldn't use M64 */
343 	if (pci_is_root_bus(bus))
344 		return NULL;
345 
346 	/* Allocate bitmap */
347 	size = _ALIGN_UP(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
348 	pe_alloc = kzalloc(size, GFP_KERNEL);
349 	if (!pe_alloc) {
350 		pr_warn("%s: Out of memory !\n",
351 			__func__);
352 		return NULL;
353 	}
354 
355 	/* Figure out reserved PE numbers by the PE */
356 	pnv_ioda_reserve_m64_pe(bus, pe_alloc, all);
357 
358 	/*
359 	 * the current bus might not own M64 window and that's all
360 	 * contributed by its child buses. For the case, we needn't
361 	 * pick M64 dependent PE#.
362 	 */
363 	if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) {
364 		kfree(pe_alloc);
365 		return NULL;
366 	}
367 
368 	/*
369 	 * Figure out the master PE and put all slave PEs to master
370 	 * PE's list to form compound PE.
371 	 */
372 	master_pe = NULL;
373 	i = -1;
374 	while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) <
375 		phb->ioda.total_pe_num) {
376 		pe = &phb->ioda.pe_array[i];
377 
378 		phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number;
379 		if (!master_pe) {
380 			pe->flags |= PNV_IODA_PE_MASTER;
381 			INIT_LIST_HEAD(&pe->slaves);
382 			master_pe = pe;
383 		} else {
384 			pe->flags |= PNV_IODA_PE_SLAVE;
385 			pe->master = master_pe;
386 			list_add_tail(&pe->list, &master_pe->slaves);
387 		}
388 
389 		/*
390 		 * P7IOC supports M64DT, which helps mapping M64 segment
391 		 * to one particular PE#. However, PHB3 has fixed mapping
392 		 * between M64 segment and PE#. In order to have same logic
393 		 * for P7IOC and PHB3, we enforce fixed mapping between M64
394 		 * segment and PE# on P7IOC.
395 		 */
396 		if (phb->type == PNV_PHB_IODA1) {
397 			int64_t rc;
398 
399 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
400 					pe->pe_number, OPAL_M64_WINDOW_TYPE,
401 					pe->pe_number / PNV_IODA1_M64_SEGS,
402 					pe->pe_number % PNV_IODA1_M64_SEGS);
403 			if (rc != OPAL_SUCCESS)
404 				pr_warn("%s: Error %lld mapping M64 for PHB#%d-PE#%d\n",
405 					__func__, rc, phb->hose->global_number,
406 					pe->pe_number);
407 		}
408 	}
409 
410 	kfree(pe_alloc);
411 	return master_pe;
412 }
413 
414 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
415 {
416 	struct pci_controller *hose = phb->hose;
417 	struct device_node *dn = hose->dn;
418 	struct resource *res;
419 	u32 m64_range[2], i;
420 	const u32 *r;
421 	u64 pci_addr;
422 
423 	if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) {
424 		pr_info("  Not support M64 window\n");
425 		return;
426 	}
427 
428 	if (!firmware_has_feature(FW_FEATURE_OPAL)) {
429 		pr_info("  Firmware too old to support M64 window\n");
430 		return;
431 	}
432 
433 	r = of_get_property(dn, "ibm,opal-m64-window", NULL);
434 	if (!r) {
435 		pr_info("  No <ibm,opal-m64-window> on %s\n",
436 			dn->full_name);
437 		return;
438 	}
439 
440 	/*
441 	 * Find the available M64 BAR range and pickup the last one for
442 	 * covering the whole 64-bits space. We support only one range.
443 	 */
444 	if (of_property_read_u32_array(dn, "ibm,opal-available-m64-ranges",
445 				       m64_range, 2)) {
446 		/* In absence of the property, assume 0..15 */
447 		m64_range[0] = 0;
448 		m64_range[1] = 16;
449 	}
450 	/* We only support 64 bits in our allocator */
451 	if (m64_range[1] > 63) {
452 		pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n",
453 			__func__, m64_range[1], phb->hose->global_number);
454 		m64_range[1] = 63;
455 	}
456 	/* Empty range, no m64 */
457 	if (m64_range[1] <= m64_range[0]) {
458 		pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n",
459 			__func__, phb->hose->global_number);
460 		return;
461 	}
462 
463 	/* Configure M64 informations */
464 	res = &hose->mem_resources[1];
465 	res->name = dn->full_name;
466 	res->start = of_translate_address(dn, r + 2);
467 	res->end = res->start + of_read_number(r + 4, 2) - 1;
468 	res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
469 	pci_addr = of_read_number(r, 2);
470 	hose->mem_offset[1] = res->start - pci_addr;
471 
472 	phb->ioda.m64_size = resource_size(res);
473 	phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num;
474 	phb->ioda.m64_base = pci_addr;
475 
476 	/* This lines up nicely with the display from processing OF ranges */
477 	pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n",
478 		res->start, res->end, pci_addr, m64_range[0],
479 		m64_range[0] + m64_range[1] - 1);
480 
481 	/* Mark all M64 used up by default */
482 	phb->ioda.m64_bar_alloc = (unsigned long)-1;
483 
484 	/* Use last M64 BAR to cover M64 window */
485 	m64_range[1]--;
486 	phb->ioda.m64_bar_idx = m64_range[0] + m64_range[1];
487 
488 	pr_info(" Using M64 #%d as default window\n", phb->ioda.m64_bar_idx);
489 
490 	/* Mark remaining ones free */
491 	for (i = m64_range[0]; i < m64_range[1]; i++)
492 		clear_bit(i, &phb->ioda.m64_bar_alloc);
493 
494 	/*
495 	 * Setup init functions for M64 based on IODA version, IODA3 uses
496 	 * the IODA2 code.
497 	 */
498 	if (phb->type == PNV_PHB_IODA1)
499 		phb->init_m64 = pnv_ioda1_init_m64;
500 	else
501 		phb->init_m64 = pnv_ioda2_init_m64;
502 	phb->reserve_m64_pe = pnv_ioda_reserve_m64_pe;
503 	phb->pick_m64_pe = pnv_ioda_pick_m64_pe;
504 }
505 
506 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
507 {
508 	struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
509 	struct pnv_ioda_pe *slave;
510 	s64 rc;
511 
512 	/* Fetch master PE */
513 	if (pe->flags & PNV_IODA_PE_SLAVE) {
514 		pe = pe->master;
515 		if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
516 			return;
517 
518 		pe_no = pe->pe_number;
519 	}
520 
521 	/* Freeze master PE */
522 	rc = opal_pci_eeh_freeze_set(phb->opal_id,
523 				     pe_no,
524 				     OPAL_EEH_ACTION_SET_FREEZE_ALL);
525 	if (rc != OPAL_SUCCESS) {
526 		pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
527 			__func__, rc, phb->hose->global_number, pe_no);
528 		return;
529 	}
530 
531 	/* Freeze slave PEs */
532 	if (!(pe->flags & PNV_IODA_PE_MASTER))
533 		return;
534 
535 	list_for_each_entry(slave, &pe->slaves, list) {
536 		rc = opal_pci_eeh_freeze_set(phb->opal_id,
537 					     slave->pe_number,
538 					     OPAL_EEH_ACTION_SET_FREEZE_ALL);
539 		if (rc != OPAL_SUCCESS)
540 			pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
541 				__func__, rc, phb->hose->global_number,
542 				slave->pe_number);
543 	}
544 }
545 
546 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
547 {
548 	struct pnv_ioda_pe *pe, *slave;
549 	s64 rc;
550 
551 	/* Find master PE */
552 	pe = &phb->ioda.pe_array[pe_no];
553 	if (pe->flags & PNV_IODA_PE_SLAVE) {
554 		pe = pe->master;
555 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
556 		pe_no = pe->pe_number;
557 	}
558 
559 	/* Clear frozen state for master PE */
560 	rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
561 	if (rc != OPAL_SUCCESS) {
562 		pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
563 			__func__, rc, opt, phb->hose->global_number, pe_no);
564 		return -EIO;
565 	}
566 
567 	if (!(pe->flags & PNV_IODA_PE_MASTER))
568 		return 0;
569 
570 	/* Clear frozen state for slave PEs */
571 	list_for_each_entry(slave, &pe->slaves, list) {
572 		rc = opal_pci_eeh_freeze_clear(phb->opal_id,
573 					     slave->pe_number,
574 					     opt);
575 		if (rc != OPAL_SUCCESS) {
576 			pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
577 				__func__, rc, opt, phb->hose->global_number,
578 				slave->pe_number);
579 			return -EIO;
580 		}
581 	}
582 
583 	return 0;
584 }
585 
586 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
587 {
588 	struct pnv_ioda_pe *slave, *pe;
589 	u8 fstate, state;
590 	__be16 pcierr;
591 	s64 rc;
592 
593 	/* Sanity check on PE number */
594 	if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num)
595 		return OPAL_EEH_STOPPED_PERM_UNAVAIL;
596 
597 	/*
598 	 * Fetch the master PE and the PE instance might be
599 	 * not initialized yet.
600 	 */
601 	pe = &phb->ioda.pe_array[pe_no];
602 	if (pe->flags & PNV_IODA_PE_SLAVE) {
603 		pe = pe->master;
604 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
605 		pe_no = pe->pe_number;
606 	}
607 
608 	/* Check the master PE */
609 	rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
610 					&state, &pcierr, NULL);
611 	if (rc != OPAL_SUCCESS) {
612 		pr_warn("%s: Failure %lld getting "
613 			"PHB#%x-PE#%x state\n",
614 			__func__, rc,
615 			phb->hose->global_number, pe_no);
616 		return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
617 	}
618 
619 	/* Check the slave PE */
620 	if (!(pe->flags & PNV_IODA_PE_MASTER))
621 		return state;
622 
623 	list_for_each_entry(slave, &pe->slaves, list) {
624 		rc = opal_pci_eeh_freeze_status(phb->opal_id,
625 						slave->pe_number,
626 						&fstate,
627 						&pcierr,
628 						NULL);
629 		if (rc != OPAL_SUCCESS) {
630 			pr_warn("%s: Failure %lld getting "
631 				"PHB#%x-PE#%x state\n",
632 				__func__, rc,
633 				phb->hose->global_number, slave->pe_number);
634 			return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
635 		}
636 
637 		/*
638 		 * Override the result based on the ascending
639 		 * priority.
640 		 */
641 		if (fstate > state)
642 			state = fstate;
643 	}
644 
645 	return state;
646 }
647 
648 /* Currently those 2 are only used when MSIs are enabled, this will change
649  * but in the meantime, we need to protect them to avoid warnings
650  */
651 #ifdef CONFIG_PCI_MSI
652 struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
653 {
654 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
655 	struct pnv_phb *phb = hose->private_data;
656 	struct pci_dn *pdn = pci_get_pdn(dev);
657 
658 	if (!pdn)
659 		return NULL;
660 	if (pdn->pe_number == IODA_INVALID_PE)
661 		return NULL;
662 	return &phb->ioda.pe_array[pdn->pe_number];
663 }
664 #endif /* CONFIG_PCI_MSI */
665 
666 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
667 				  struct pnv_ioda_pe *parent,
668 				  struct pnv_ioda_pe *child,
669 				  bool is_add)
670 {
671 	const char *desc = is_add ? "adding" : "removing";
672 	uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
673 			      OPAL_REMOVE_PE_FROM_DOMAIN;
674 	struct pnv_ioda_pe *slave;
675 	long rc;
676 
677 	/* Parent PE affects child PE */
678 	rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
679 				child->pe_number, op);
680 	if (rc != OPAL_SUCCESS) {
681 		pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
682 			rc, desc);
683 		return -ENXIO;
684 	}
685 
686 	if (!(child->flags & PNV_IODA_PE_MASTER))
687 		return 0;
688 
689 	/* Compound case: parent PE affects slave PEs */
690 	list_for_each_entry(slave, &child->slaves, list) {
691 		rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
692 					slave->pe_number, op);
693 		if (rc != OPAL_SUCCESS) {
694 			pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
695 				rc, desc);
696 			return -ENXIO;
697 		}
698 	}
699 
700 	return 0;
701 }
702 
703 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
704 			      struct pnv_ioda_pe *pe,
705 			      bool is_add)
706 {
707 	struct pnv_ioda_pe *slave;
708 	struct pci_dev *pdev = NULL;
709 	int ret;
710 
711 	/*
712 	 * Clear PE frozen state. If it's master PE, we need
713 	 * clear slave PE frozen state as well.
714 	 */
715 	if (is_add) {
716 		opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
717 					  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
718 		if (pe->flags & PNV_IODA_PE_MASTER) {
719 			list_for_each_entry(slave, &pe->slaves, list)
720 				opal_pci_eeh_freeze_clear(phb->opal_id,
721 							  slave->pe_number,
722 							  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
723 		}
724 	}
725 
726 	/*
727 	 * Associate PE in PELT. We need add the PE into the
728 	 * corresponding PELT-V as well. Otherwise, the error
729 	 * originated from the PE might contribute to other
730 	 * PEs.
731 	 */
732 	ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
733 	if (ret)
734 		return ret;
735 
736 	/* For compound PEs, any one affects all of them */
737 	if (pe->flags & PNV_IODA_PE_MASTER) {
738 		list_for_each_entry(slave, &pe->slaves, list) {
739 			ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
740 			if (ret)
741 				return ret;
742 		}
743 	}
744 
745 	if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
746 		pdev = pe->pbus->self;
747 	else if (pe->flags & PNV_IODA_PE_DEV)
748 		pdev = pe->pdev->bus->self;
749 #ifdef CONFIG_PCI_IOV
750 	else if (pe->flags & PNV_IODA_PE_VF)
751 		pdev = pe->parent_dev;
752 #endif /* CONFIG_PCI_IOV */
753 	while (pdev) {
754 		struct pci_dn *pdn = pci_get_pdn(pdev);
755 		struct pnv_ioda_pe *parent;
756 
757 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
758 			parent = &phb->ioda.pe_array[pdn->pe_number];
759 			ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
760 			if (ret)
761 				return ret;
762 		}
763 
764 		pdev = pdev->bus->self;
765 	}
766 
767 	return 0;
768 }
769 
770 static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
771 {
772 	struct pci_dev *parent;
773 	uint8_t bcomp, dcomp, fcomp;
774 	int64_t rc;
775 	long rid_end, rid;
776 
777 	/* Currently, we just deconfigure VF PE. Bus PE will always there.*/
778 	if (pe->pbus) {
779 		int count;
780 
781 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
782 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
783 		parent = pe->pbus->self;
784 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
785 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
786 		else
787 			count = 1;
788 
789 		switch(count) {
790 		case  1: bcomp = OpalPciBusAll;         break;
791 		case  2: bcomp = OpalPciBus7Bits;       break;
792 		case  4: bcomp = OpalPciBus6Bits;       break;
793 		case  8: bcomp = OpalPciBus5Bits;       break;
794 		case 16: bcomp = OpalPciBus4Bits;       break;
795 		case 32: bcomp = OpalPciBus3Bits;       break;
796 		default:
797 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
798 			        count);
799 			/* Do an exact match only */
800 			bcomp = OpalPciBusAll;
801 		}
802 		rid_end = pe->rid + (count << 8);
803 	} else {
804 #ifdef CONFIG_PCI_IOV
805 		if (pe->flags & PNV_IODA_PE_VF)
806 			parent = pe->parent_dev;
807 		else
808 #endif
809 			parent = pe->pdev->bus->self;
810 		bcomp = OpalPciBusAll;
811 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
812 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
813 		rid_end = pe->rid + 1;
814 	}
815 
816 	/* Clear the reverse map */
817 	for (rid = pe->rid; rid < rid_end; rid++)
818 		phb->ioda.pe_rmap[rid] = IODA_INVALID_PE;
819 
820 	/* Release from all parents PELT-V */
821 	while (parent) {
822 		struct pci_dn *pdn = pci_get_pdn(parent);
823 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
824 			rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
825 						pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
826 			/* XXX What to do in case of error ? */
827 		}
828 		parent = parent->bus->self;
829 	}
830 
831 	opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
832 				  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
833 
834 	/* Disassociate PE in PELT */
835 	rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
836 				pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
837 	if (rc)
838 		pe_warn(pe, "OPAL error %ld remove self from PELTV\n", rc);
839 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
840 			     bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
841 	if (rc)
842 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
843 
844 	pe->pbus = NULL;
845 	pe->pdev = NULL;
846 #ifdef CONFIG_PCI_IOV
847 	pe->parent_dev = NULL;
848 #endif
849 
850 	return 0;
851 }
852 
853 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
854 {
855 	struct pci_dev *parent;
856 	uint8_t bcomp, dcomp, fcomp;
857 	long rc, rid_end, rid;
858 
859 	/* Bus validation ? */
860 	if (pe->pbus) {
861 		int count;
862 
863 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
864 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
865 		parent = pe->pbus->self;
866 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
867 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
868 		else
869 			count = 1;
870 
871 		switch(count) {
872 		case  1: bcomp = OpalPciBusAll;		break;
873 		case  2: bcomp = OpalPciBus7Bits;	break;
874 		case  4: bcomp = OpalPciBus6Bits;	break;
875 		case  8: bcomp = OpalPciBus5Bits;	break;
876 		case 16: bcomp = OpalPciBus4Bits;	break;
877 		case 32: bcomp = OpalPciBus3Bits;	break;
878 		default:
879 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
880 			        count);
881 			/* Do an exact match only */
882 			bcomp = OpalPciBusAll;
883 		}
884 		rid_end = pe->rid + (count << 8);
885 	} else {
886 #ifdef CONFIG_PCI_IOV
887 		if (pe->flags & PNV_IODA_PE_VF)
888 			parent = pe->parent_dev;
889 		else
890 #endif /* CONFIG_PCI_IOV */
891 			parent = pe->pdev->bus->self;
892 		bcomp = OpalPciBusAll;
893 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
894 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
895 		rid_end = pe->rid + 1;
896 	}
897 
898 	/*
899 	 * Associate PE in PELT. We need add the PE into the
900 	 * corresponding PELT-V as well. Otherwise, the error
901 	 * originated from the PE might contribute to other
902 	 * PEs.
903 	 */
904 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
905 			     bcomp, dcomp, fcomp, OPAL_MAP_PE);
906 	if (rc) {
907 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
908 		return -ENXIO;
909 	}
910 
911 	/*
912 	 * Configure PELTV. NPUs don't have a PELTV table so skip
913 	 * configuration on them.
914 	 */
915 	if (phb->type != PNV_PHB_NPU)
916 		pnv_ioda_set_peltv(phb, pe, true);
917 
918 	/* Setup reverse map */
919 	for (rid = pe->rid; rid < rid_end; rid++)
920 		phb->ioda.pe_rmap[rid] = pe->pe_number;
921 
922 	/* Setup one MVTs on IODA1 */
923 	if (phb->type != PNV_PHB_IODA1) {
924 		pe->mve_number = 0;
925 		goto out;
926 	}
927 
928 	pe->mve_number = pe->pe_number;
929 	rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
930 	if (rc != OPAL_SUCCESS) {
931 		pe_err(pe, "OPAL error %ld setting up MVE %d\n",
932 		       rc, pe->mve_number);
933 		pe->mve_number = -1;
934 	} else {
935 		rc = opal_pci_set_mve_enable(phb->opal_id,
936 					     pe->mve_number, OPAL_ENABLE_MVE);
937 		if (rc) {
938 			pe_err(pe, "OPAL error %ld enabling MVE %d\n",
939 			       rc, pe->mve_number);
940 			pe->mve_number = -1;
941 		}
942 	}
943 
944 out:
945 	return 0;
946 }
947 
948 #ifdef CONFIG_PCI_IOV
949 static int pnv_pci_vf_resource_shift(struct pci_dev *dev, int offset)
950 {
951 	struct pci_dn *pdn = pci_get_pdn(dev);
952 	int i;
953 	struct resource *res, res2;
954 	resource_size_t size;
955 	u16 num_vfs;
956 
957 	if (!dev->is_physfn)
958 		return -EINVAL;
959 
960 	/*
961 	 * "offset" is in VFs.  The M64 windows are sized so that when they
962 	 * are segmented, each segment is the same size as the IOV BAR.
963 	 * Each segment is in a separate PE, and the high order bits of the
964 	 * address are the PE number.  Therefore, each VF's BAR is in a
965 	 * separate PE, and changing the IOV BAR start address changes the
966 	 * range of PEs the VFs are in.
967 	 */
968 	num_vfs = pdn->num_vfs;
969 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
970 		res = &dev->resource[i + PCI_IOV_RESOURCES];
971 		if (!res->flags || !res->parent)
972 			continue;
973 
974 		/*
975 		 * The actual IOV BAR range is determined by the start address
976 		 * and the actual size for num_vfs VFs BAR.  This check is to
977 		 * make sure that after shifting, the range will not overlap
978 		 * with another device.
979 		 */
980 		size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
981 		res2.flags = res->flags;
982 		res2.start = res->start + (size * offset);
983 		res2.end = res2.start + (size * num_vfs) - 1;
984 
985 		if (res2.end > res->end) {
986 			dev_err(&dev->dev, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
987 				i, &res2, res, num_vfs, offset);
988 			return -EBUSY;
989 		}
990 	}
991 
992 	/*
993 	 * After doing so, there would be a "hole" in the /proc/iomem when
994 	 * offset is a positive value. It looks like the device return some
995 	 * mmio back to the system, which actually no one could use it.
996 	 */
997 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
998 		res = &dev->resource[i + PCI_IOV_RESOURCES];
999 		if (!res->flags || !res->parent)
1000 			continue;
1001 
1002 		size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
1003 		res2 = *res;
1004 		res->start += size * offset;
1005 
1006 		dev_info(&dev->dev, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n",
1007 			 i, &res2, res, (offset > 0) ? "En" : "Dis",
1008 			 num_vfs, offset);
1009 		pci_update_resource(dev, i + PCI_IOV_RESOURCES);
1010 	}
1011 	return 0;
1012 }
1013 #endif /* CONFIG_PCI_IOV */
1014 
1015 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
1016 {
1017 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
1018 	struct pnv_phb *phb = hose->private_data;
1019 	struct pci_dn *pdn = pci_get_pdn(dev);
1020 	struct pnv_ioda_pe *pe;
1021 
1022 	if (!pdn) {
1023 		pr_err("%s: Device tree node not associated properly\n",
1024 			   pci_name(dev));
1025 		return NULL;
1026 	}
1027 	if (pdn->pe_number != IODA_INVALID_PE)
1028 		return NULL;
1029 
1030 	pe = pnv_ioda_alloc_pe(phb);
1031 	if (!pe) {
1032 		pr_warning("%s: Not enough PE# available, disabling device\n",
1033 			   pci_name(dev));
1034 		return NULL;
1035 	}
1036 
1037 	/* NOTE: We get only one ref to the pci_dev for the pdn, not for the
1038 	 * pointer in the PE data structure, both should be destroyed at the
1039 	 * same time. However, this needs to be looked at more closely again
1040 	 * once we actually start removing things (Hotplug, SR-IOV, ...)
1041 	 *
1042 	 * At some point we want to remove the PDN completely anyways
1043 	 */
1044 	pci_dev_get(dev);
1045 	pdn->pcidev = dev;
1046 	pdn->pe_number = pe->pe_number;
1047 	pe->flags = PNV_IODA_PE_DEV;
1048 	pe->pdev = dev;
1049 	pe->pbus = NULL;
1050 	pe->mve_number = -1;
1051 	pe->rid = dev->bus->number << 8 | pdn->devfn;
1052 
1053 	pe_info(pe, "Associated device to PE\n");
1054 
1055 	if (pnv_ioda_configure_pe(phb, pe)) {
1056 		/* XXX What do we do here ? */
1057 		pnv_ioda_free_pe(pe);
1058 		pdn->pe_number = IODA_INVALID_PE;
1059 		pe->pdev = NULL;
1060 		pci_dev_put(dev);
1061 		return NULL;
1062 	}
1063 
1064 	/* Put PE to the list */
1065 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1066 
1067 	return pe;
1068 }
1069 
1070 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
1071 {
1072 	struct pci_dev *dev;
1073 
1074 	list_for_each_entry(dev, &bus->devices, bus_list) {
1075 		struct pci_dn *pdn = pci_get_pdn(dev);
1076 
1077 		if (pdn == NULL) {
1078 			pr_warn("%s: No device node associated with device !\n",
1079 				pci_name(dev));
1080 			continue;
1081 		}
1082 
1083 		/*
1084 		 * In partial hotplug case, the PCI device might be still
1085 		 * associated with the PE and needn't attach it to the PE
1086 		 * again.
1087 		 */
1088 		if (pdn->pe_number != IODA_INVALID_PE)
1089 			continue;
1090 
1091 		pe->device_count++;
1092 		pdn->pcidev = dev;
1093 		pdn->pe_number = pe->pe_number;
1094 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1095 			pnv_ioda_setup_same_PE(dev->subordinate, pe);
1096 	}
1097 }
1098 
1099 /*
1100  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1101  * single PCI bus. Another one that contains the primary PCI bus and its
1102  * subordinate PCI devices and buses. The second type of PE is normally
1103  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1104  */
1105 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
1106 {
1107 	struct pci_controller *hose = pci_bus_to_host(bus);
1108 	struct pnv_phb *phb = hose->private_data;
1109 	struct pnv_ioda_pe *pe = NULL;
1110 	unsigned int pe_num;
1111 
1112 	/*
1113 	 * In partial hotplug case, the PE instance might be still alive.
1114 	 * We should reuse it instead of allocating a new one.
1115 	 */
1116 	pe_num = phb->ioda.pe_rmap[bus->number << 8];
1117 	if (pe_num != IODA_INVALID_PE) {
1118 		pe = &phb->ioda.pe_array[pe_num];
1119 		pnv_ioda_setup_same_PE(bus, pe);
1120 		return NULL;
1121 	}
1122 
1123 	/* PE number for root bus should have been reserved */
1124 	if (pci_is_root_bus(bus) &&
1125 	    phb->ioda.root_pe_idx != IODA_INVALID_PE)
1126 		pe = &phb->ioda.pe_array[phb->ioda.root_pe_idx];
1127 
1128 	/* Check if PE is determined by M64 */
1129 	if (!pe && phb->pick_m64_pe)
1130 		pe = phb->pick_m64_pe(bus, all);
1131 
1132 	/* The PE number isn't pinned by M64 */
1133 	if (!pe)
1134 		pe = pnv_ioda_alloc_pe(phb);
1135 
1136 	if (!pe) {
1137 		pr_warning("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1138 			__func__, pci_domain_nr(bus), bus->number);
1139 		return NULL;
1140 	}
1141 
1142 	pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1143 	pe->pbus = bus;
1144 	pe->pdev = NULL;
1145 	pe->mve_number = -1;
1146 	pe->rid = bus->busn_res.start << 8;
1147 
1148 	if (all)
1149 		pe_info(pe, "Secondary bus %d..%d associated with PE#%d\n",
1150 			bus->busn_res.start, bus->busn_res.end, pe->pe_number);
1151 	else
1152 		pe_info(pe, "Secondary bus %d associated with PE#%d\n",
1153 			bus->busn_res.start, pe->pe_number);
1154 
1155 	if (pnv_ioda_configure_pe(phb, pe)) {
1156 		/* XXX What do we do here ? */
1157 		pnv_ioda_free_pe(pe);
1158 		pe->pbus = NULL;
1159 		return NULL;
1160 	}
1161 
1162 	/* Associate it with all child devices */
1163 	pnv_ioda_setup_same_PE(bus, pe);
1164 
1165 	/* Put PE to the list */
1166 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1167 
1168 	return pe;
1169 }
1170 
1171 static struct pnv_ioda_pe *pnv_ioda_setup_npu_PE(struct pci_dev *npu_pdev)
1172 {
1173 	int pe_num, found_pe = false, rc;
1174 	long rid;
1175 	struct pnv_ioda_pe *pe;
1176 	struct pci_dev *gpu_pdev;
1177 	struct pci_dn *npu_pdn;
1178 	struct pci_controller *hose = pci_bus_to_host(npu_pdev->bus);
1179 	struct pnv_phb *phb = hose->private_data;
1180 
1181 	/*
1182 	 * Due to a hardware errata PE#0 on the NPU is reserved for
1183 	 * error handling. This means we only have three PEs remaining
1184 	 * which need to be assigned to four links, implying some
1185 	 * links must share PEs.
1186 	 *
1187 	 * To achieve this we assign PEs such that NPUs linking the
1188 	 * same GPU get assigned the same PE.
1189 	 */
1190 	gpu_pdev = pnv_pci_get_gpu_dev(npu_pdev);
1191 	for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) {
1192 		pe = &phb->ioda.pe_array[pe_num];
1193 		if (!pe->pdev)
1194 			continue;
1195 
1196 		if (pnv_pci_get_gpu_dev(pe->pdev) == gpu_pdev) {
1197 			/*
1198 			 * This device has the same peer GPU so should
1199 			 * be assigned the same PE as the existing
1200 			 * peer NPU.
1201 			 */
1202 			dev_info(&npu_pdev->dev,
1203 				"Associating to existing PE %d\n", pe_num);
1204 			pci_dev_get(npu_pdev);
1205 			npu_pdn = pci_get_pdn(npu_pdev);
1206 			rid = npu_pdev->bus->number << 8 | npu_pdn->devfn;
1207 			npu_pdn->pcidev = npu_pdev;
1208 			npu_pdn->pe_number = pe_num;
1209 			phb->ioda.pe_rmap[rid] = pe->pe_number;
1210 
1211 			/* Map the PE to this link */
1212 			rc = opal_pci_set_pe(phb->opal_id, pe_num, rid,
1213 					OpalPciBusAll,
1214 					OPAL_COMPARE_RID_DEVICE_NUMBER,
1215 					OPAL_COMPARE_RID_FUNCTION_NUMBER,
1216 					OPAL_MAP_PE);
1217 			WARN_ON(rc != OPAL_SUCCESS);
1218 			found_pe = true;
1219 			break;
1220 		}
1221 	}
1222 
1223 	if (!found_pe)
1224 		/*
1225 		 * Could not find an existing PE so allocate a new
1226 		 * one.
1227 		 */
1228 		return pnv_ioda_setup_dev_PE(npu_pdev);
1229 	else
1230 		return pe;
1231 }
1232 
1233 static void pnv_ioda_setup_npu_PEs(struct pci_bus *bus)
1234 {
1235 	struct pci_dev *pdev;
1236 
1237 	list_for_each_entry(pdev, &bus->devices, bus_list)
1238 		pnv_ioda_setup_npu_PE(pdev);
1239 }
1240 
1241 static void pnv_pci_ioda_setup_PEs(void)
1242 {
1243 	struct pci_controller *hose, *tmp;
1244 	struct pnv_phb *phb;
1245 
1246 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
1247 		phb = hose->private_data;
1248 		if (phb->type == PNV_PHB_NPU) {
1249 			/* PE#0 is needed for error reporting */
1250 			pnv_ioda_reserve_pe(phb, 0);
1251 			pnv_ioda_setup_npu_PEs(hose->bus);
1252 		}
1253 	}
1254 }
1255 
1256 #ifdef CONFIG_PCI_IOV
1257 static int pnv_pci_vf_release_m64(struct pci_dev *pdev, u16 num_vfs)
1258 {
1259 	struct pci_bus        *bus;
1260 	struct pci_controller *hose;
1261 	struct pnv_phb        *phb;
1262 	struct pci_dn         *pdn;
1263 	int                    i, j;
1264 	int                    m64_bars;
1265 
1266 	bus = pdev->bus;
1267 	hose = pci_bus_to_host(bus);
1268 	phb = hose->private_data;
1269 	pdn = pci_get_pdn(pdev);
1270 
1271 	if (pdn->m64_single_mode)
1272 		m64_bars = num_vfs;
1273 	else
1274 		m64_bars = 1;
1275 
1276 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
1277 		for (j = 0; j < m64_bars; j++) {
1278 			if (pdn->m64_map[j][i] == IODA_INVALID_M64)
1279 				continue;
1280 			opal_pci_phb_mmio_enable(phb->opal_id,
1281 				OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 0);
1282 			clear_bit(pdn->m64_map[j][i], &phb->ioda.m64_bar_alloc);
1283 			pdn->m64_map[j][i] = IODA_INVALID_M64;
1284 		}
1285 
1286 	kfree(pdn->m64_map);
1287 	return 0;
1288 }
1289 
1290 static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)
1291 {
1292 	struct pci_bus        *bus;
1293 	struct pci_controller *hose;
1294 	struct pnv_phb        *phb;
1295 	struct pci_dn         *pdn;
1296 	unsigned int           win;
1297 	struct resource       *res;
1298 	int                    i, j;
1299 	int64_t                rc;
1300 	int                    total_vfs;
1301 	resource_size_t        size, start;
1302 	int                    pe_num;
1303 	int                    m64_bars;
1304 
1305 	bus = pdev->bus;
1306 	hose = pci_bus_to_host(bus);
1307 	phb = hose->private_data;
1308 	pdn = pci_get_pdn(pdev);
1309 	total_vfs = pci_sriov_get_totalvfs(pdev);
1310 
1311 	if (pdn->m64_single_mode)
1312 		m64_bars = num_vfs;
1313 	else
1314 		m64_bars = 1;
1315 
1316 	pdn->m64_map = kmalloc(sizeof(*pdn->m64_map) * m64_bars, GFP_KERNEL);
1317 	if (!pdn->m64_map)
1318 		return -ENOMEM;
1319 	/* Initialize the m64_map to IODA_INVALID_M64 */
1320 	for (i = 0; i < m64_bars ; i++)
1321 		for (j = 0; j < PCI_SRIOV_NUM_BARS; j++)
1322 			pdn->m64_map[i][j] = IODA_INVALID_M64;
1323 
1324 
1325 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1326 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
1327 		if (!res->flags || !res->parent)
1328 			continue;
1329 
1330 		for (j = 0; j < m64_bars; j++) {
1331 			do {
1332 				win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
1333 						phb->ioda.m64_bar_idx + 1, 0);
1334 
1335 				if (win >= phb->ioda.m64_bar_idx + 1)
1336 					goto m64_failed;
1337 			} while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc));
1338 
1339 			pdn->m64_map[j][i] = win;
1340 
1341 			if (pdn->m64_single_mode) {
1342 				size = pci_iov_resource_size(pdev,
1343 							PCI_IOV_RESOURCES + i);
1344 				start = res->start + size * j;
1345 			} else {
1346 				size = resource_size(res);
1347 				start = res->start;
1348 			}
1349 
1350 			/* Map the M64 here */
1351 			if (pdn->m64_single_mode) {
1352 				pe_num = pdn->pe_num_map[j];
1353 				rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1354 						pe_num, OPAL_M64_WINDOW_TYPE,
1355 						pdn->m64_map[j][i], 0);
1356 			}
1357 
1358 			rc = opal_pci_set_phb_mem_window(phb->opal_id,
1359 						 OPAL_M64_WINDOW_TYPE,
1360 						 pdn->m64_map[j][i],
1361 						 start,
1362 						 0, /* unused */
1363 						 size);
1364 
1365 
1366 			if (rc != OPAL_SUCCESS) {
1367 				dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n",
1368 					win, rc);
1369 				goto m64_failed;
1370 			}
1371 
1372 			if (pdn->m64_single_mode)
1373 				rc = opal_pci_phb_mmio_enable(phb->opal_id,
1374 				     OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 2);
1375 			else
1376 				rc = opal_pci_phb_mmio_enable(phb->opal_id,
1377 				     OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 1);
1378 
1379 			if (rc != OPAL_SUCCESS) {
1380 				dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n",
1381 					win, rc);
1382 				goto m64_failed;
1383 			}
1384 		}
1385 	}
1386 	return 0;
1387 
1388 m64_failed:
1389 	pnv_pci_vf_release_m64(pdev, num_vfs);
1390 	return -EBUSY;
1391 }
1392 
1393 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
1394 		int num);
1395 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable);
1396 
1397 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe)
1398 {
1399 	struct iommu_table    *tbl;
1400 	int64_t               rc;
1401 
1402 	tbl = pe->table_group.tables[0];
1403 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
1404 	if (rc)
1405 		pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
1406 
1407 	pnv_pci_ioda2_set_bypass(pe, false);
1408 	if (pe->table_group.group) {
1409 		iommu_group_put(pe->table_group.group);
1410 		BUG_ON(pe->table_group.group);
1411 	}
1412 	pnv_pci_ioda2_table_free_pages(tbl);
1413 	iommu_free_table(tbl, of_node_full_name(dev->dev.of_node));
1414 }
1415 
1416 static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
1417 {
1418 	struct pci_bus        *bus;
1419 	struct pci_controller *hose;
1420 	struct pnv_phb        *phb;
1421 	struct pnv_ioda_pe    *pe, *pe_n;
1422 	struct pci_dn         *pdn;
1423 
1424 	bus = pdev->bus;
1425 	hose = pci_bus_to_host(bus);
1426 	phb = hose->private_data;
1427 	pdn = pci_get_pdn(pdev);
1428 
1429 	if (!pdev->is_physfn)
1430 		return;
1431 
1432 	list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) {
1433 		if (pe->parent_dev != pdev)
1434 			continue;
1435 
1436 		pnv_pci_ioda2_release_dma_pe(pdev, pe);
1437 
1438 		/* Remove from list */
1439 		mutex_lock(&phb->ioda.pe_list_mutex);
1440 		list_del(&pe->list);
1441 		mutex_unlock(&phb->ioda.pe_list_mutex);
1442 
1443 		pnv_ioda_deconfigure_pe(phb, pe);
1444 
1445 		pnv_ioda_free_pe(pe);
1446 	}
1447 }
1448 
1449 void pnv_pci_sriov_disable(struct pci_dev *pdev)
1450 {
1451 	struct pci_bus        *bus;
1452 	struct pci_controller *hose;
1453 	struct pnv_phb        *phb;
1454 	struct pnv_ioda_pe    *pe;
1455 	struct pci_dn         *pdn;
1456 	struct pci_sriov      *iov;
1457 	u16                    num_vfs, i;
1458 
1459 	bus = pdev->bus;
1460 	hose = pci_bus_to_host(bus);
1461 	phb = hose->private_data;
1462 	pdn = pci_get_pdn(pdev);
1463 	iov = pdev->sriov;
1464 	num_vfs = pdn->num_vfs;
1465 
1466 	/* Release VF PEs */
1467 	pnv_ioda_release_vf_PE(pdev);
1468 
1469 	if (phb->type == PNV_PHB_IODA2) {
1470 		if (!pdn->m64_single_mode)
1471 			pnv_pci_vf_resource_shift(pdev, -*pdn->pe_num_map);
1472 
1473 		/* Release M64 windows */
1474 		pnv_pci_vf_release_m64(pdev, num_vfs);
1475 
1476 		/* Release PE numbers */
1477 		if (pdn->m64_single_mode) {
1478 			for (i = 0; i < num_vfs; i++) {
1479 				if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1480 					continue;
1481 
1482 				pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1483 				pnv_ioda_free_pe(pe);
1484 			}
1485 		} else
1486 			bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1487 		/* Releasing pe_num_map */
1488 		kfree(pdn->pe_num_map);
1489 	}
1490 }
1491 
1492 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1493 				       struct pnv_ioda_pe *pe);
1494 static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs)
1495 {
1496 	struct pci_bus        *bus;
1497 	struct pci_controller *hose;
1498 	struct pnv_phb        *phb;
1499 	struct pnv_ioda_pe    *pe;
1500 	int                    pe_num;
1501 	u16                    vf_index;
1502 	struct pci_dn         *pdn;
1503 
1504 	bus = pdev->bus;
1505 	hose = pci_bus_to_host(bus);
1506 	phb = hose->private_data;
1507 	pdn = pci_get_pdn(pdev);
1508 
1509 	if (!pdev->is_physfn)
1510 		return;
1511 
1512 	/* Reserve PE for each VF */
1513 	for (vf_index = 0; vf_index < num_vfs; vf_index++) {
1514 		if (pdn->m64_single_mode)
1515 			pe_num = pdn->pe_num_map[vf_index];
1516 		else
1517 			pe_num = *pdn->pe_num_map + vf_index;
1518 
1519 		pe = &phb->ioda.pe_array[pe_num];
1520 		pe->pe_number = pe_num;
1521 		pe->phb = phb;
1522 		pe->flags = PNV_IODA_PE_VF;
1523 		pe->pbus = NULL;
1524 		pe->parent_dev = pdev;
1525 		pe->mve_number = -1;
1526 		pe->rid = (pci_iov_virtfn_bus(pdev, vf_index) << 8) |
1527 			   pci_iov_virtfn_devfn(pdev, vf_index);
1528 
1529 		pe_info(pe, "VF %04d:%02d:%02d.%d associated with PE#%d\n",
1530 			hose->global_number, pdev->bus->number,
1531 			PCI_SLOT(pci_iov_virtfn_devfn(pdev, vf_index)),
1532 			PCI_FUNC(pci_iov_virtfn_devfn(pdev, vf_index)), pe_num);
1533 
1534 		if (pnv_ioda_configure_pe(phb, pe)) {
1535 			/* XXX What do we do here ? */
1536 			pnv_ioda_free_pe(pe);
1537 			pe->pdev = NULL;
1538 			continue;
1539 		}
1540 
1541 		/* Put PE to the list */
1542 		mutex_lock(&phb->ioda.pe_list_mutex);
1543 		list_add_tail(&pe->list, &phb->ioda.pe_list);
1544 		mutex_unlock(&phb->ioda.pe_list_mutex);
1545 
1546 		pnv_pci_ioda2_setup_dma_pe(phb, pe);
1547 	}
1548 }
1549 
1550 int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1551 {
1552 	struct pci_bus        *bus;
1553 	struct pci_controller *hose;
1554 	struct pnv_phb        *phb;
1555 	struct pnv_ioda_pe    *pe;
1556 	struct pci_dn         *pdn;
1557 	int                    ret;
1558 	u16                    i;
1559 
1560 	bus = pdev->bus;
1561 	hose = pci_bus_to_host(bus);
1562 	phb = hose->private_data;
1563 	pdn = pci_get_pdn(pdev);
1564 
1565 	if (phb->type == PNV_PHB_IODA2) {
1566 		if (!pdn->vfs_expanded) {
1567 			dev_info(&pdev->dev, "don't support this SRIOV device"
1568 				" with non 64bit-prefetchable IOV BAR\n");
1569 			return -ENOSPC;
1570 		}
1571 
1572 		/*
1573 		 * When M64 BARs functions in Single PE mode, the number of VFs
1574 		 * could be enabled must be less than the number of M64 BARs.
1575 		 */
1576 		if (pdn->m64_single_mode && num_vfs > phb->ioda.m64_bar_idx) {
1577 			dev_info(&pdev->dev, "Not enough M64 BAR for VFs\n");
1578 			return -EBUSY;
1579 		}
1580 
1581 		/* Allocating pe_num_map */
1582 		if (pdn->m64_single_mode)
1583 			pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map) * num_vfs,
1584 					GFP_KERNEL);
1585 		else
1586 			pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map), GFP_KERNEL);
1587 
1588 		if (!pdn->pe_num_map)
1589 			return -ENOMEM;
1590 
1591 		if (pdn->m64_single_mode)
1592 			for (i = 0; i < num_vfs; i++)
1593 				pdn->pe_num_map[i] = IODA_INVALID_PE;
1594 
1595 		/* Calculate available PE for required VFs */
1596 		if (pdn->m64_single_mode) {
1597 			for (i = 0; i < num_vfs; i++) {
1598 				pe = pnv_ioda_alloc_pe(phb);
1599 				if (!pe) {
1600 					ret = -EBUSY;
1601 					goto m64_failed;
1602 				}
1603 
1604 				pdn->pe_num_map[i] = pe->pe_number;
1605 			}
1606 		} else {
1607 			mutex_lock(&phb->ioda.pe_alloc_mutex);
1608 			*pdn->pe_num_map = bitmap_find_next_zero_area(
1609 				phb->ioda.pe_alloc, phb->ioda.total_pe_num,
1610 				0, num_vfs, 0);
1611 			if (*pdn->pe_num_map >= phb->ioda.total_pe_num) {
1612 				mutex_unlock(&phb->ioda.pe_alloc_mutex);
1613 				dev_info(&pdev->dev, "Failed to enable VF%d\n", num_vfs);
1614 				kfree(pdn->pe_num_map);
1615 				return -EBUSY;
1616 			}
1617 			bitmap_set(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1618 			mutex_unlock(&phb->ioda.pe_alloc_mutex);
1619 		}
1620 		pdn->num_vfs = num_vfs;
1621 
1622 		/* Assign M64 window accordingly */
1623 		ret = pnv_pci_vf_assign_m64(pdev, num_vfs);
1624 		if (ret) {
1625 			dev_info(&pdev->dev, "Not enough M64 window resources\n");
1626 			goto m64_failed;
1627 		}
1628 
1629 		/*
1630 		 * When using one M64 BAR to map one IOV BAR, we need to shift
1631 		 * the IOV BAR according to the PE# allocated to the VFs.
1632 		 * Otherwise, the PE# for the VF will conflict with others.
1633 		 */
1634 		if (!pdn->m64_single_mode) {
1635 			ret = pnv_pci_vf_resource_shift(pdev, *pdn->pe_num_map);
1636 			if (ret)
1637 				goto m64_failed;
1638 		}
1639 	}
1640 
1641 	/* Setup VF PEs */
1642 	pnv_ioda_setup_vf_PE(pdev, num_vfs);
1643 
1644 	return 0;
1645 
1646 m64_failed:
1647 	if (pdn->m64_single_mode) {
1648 		for (i = 0; i < num_vfs; i++) {
1649 			if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1650 				continue;
1651 
1652 			pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1653 			pnv_ioda_free_pe(pe);
1654 		}
1655 	} else
1656 		bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1657 
1658 	/* Releasing pe_num_map */
1659 	kfree(pdn->pe_num_map);
1660 
1661 	return ret;
1662 }
1663 
1664 int pcibios_sriov_disable(struct pci_dev *pdev)
1665 {
1666 	pnv_pci_sriov_disable(pdev);
1667 
1668 	/* Release PCI data */
1669 	remove_dev_pci_data(pdev);
1670 	return 0;
1671 }
1672 
1673 int pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1674 {
1675 	/* Allocate PCI data */
1676 	add_dev_pci_data(pdev);
1677 
1678 	return pnv_pci_sriov_enable(pdev, num_vfs);
1679 }
1680 #endif /* CONFIG_PCI_IOV */
1681 
1682 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
1683 {
1684 	struct pci_dn *pdn = pci_get_pdn(pdev);
1685 	struct pnv_ioda_pe *pe;
1686 
1687 	/*
1688 	 * The function can be called while the PE#
1689 	 * hasn't been assigned. Do nothing for the
1690 	 * case.
1691 	 */
1692 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1693 		return;
1694 
1695 	pe = &phb->ioda.pe_array[pdn->pe_number];
1696 	WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1697 	set_dma_offset(&pdev->dev, pe->tce_bypass_base);
1698 	set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]);
1699 	/*
1700 	 * Note: iommu_add_device() will fail here as
1701 	 * for physical PE: the device is already added by now;
1702 	 * for virtual PE: sysfs entries are not ready yet and
1703 	 * tce_iommu_bus_notifier will add the device to a group later.
1704 	 */
1705 }
1706 
1707 static int pnv_pci_ioda_dma_set_mask(struct pci_dev *pdev, u64 dma_mask)
1708 {
1709 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1710 	struct pnv_phb *phb = hose->private_data;
1711 	struct pci_dn *pdn = pci_get_pdn(pdev);
1712 	struct pnv_ioda_pe *pe;
1713 	uint64_t top;
1714 	bool bypass = false;
1715 
1716 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1717 		return -ENODEV;;
1718 
1719 	pe = &phb->ioda.pe_array[pdn->pe_number];
1720 	if (pe->tce_bypass_enabled) {
1721 		top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1722 		bypass = (dma_mask >= top);
1723 	}
1724 
1725 	if (bypass) {
1726 		dev_info(&pdev->dev, "Using 64-bit DMA iommu bypass\n");
1727 		set_dma_ops(&pdev->dev, &dma_direct_ops);
1728 	} else {
1729 		dev_info(&pdev->dev, "Using 32-bit DMA via iommu\n");
1730 		set_dma_ops(&pdev->dev, &dma_iommu_ops);
1731 	}
1732 	*pdev->dev.dma_mask = dma_mask;
1733 
1734 	/* Update peer npu devices */
1735 	pnv_npu_try_dma_set_bypass(pdev, bypass);
1736 
1737 	return 0;
1738 }
1739 
1740 static u64 pnv_pci_ioda_dma_get_required_mask(struct pci_dev *pdev)
1741 {
1742 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1743 	struct pnv_phb *phb = hose->private_data;
1744 	struct pci_dn *pdn = pci_get_pdn(pdev);
1745 	struct pnv_ioda_pe *pe;
1746 	u64 end, mask;
1747 
1748 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1749 		return 0;
1750 
1751 	pe = &phb->ioda.pe_array[pdn->pe_number];
1752 	if (!pe->tce_bypass_enabled)
1753 		return __dma_get_required_mask(&pdev->dev);
1754 
1755 
1756 	end = pe->tce_bypass_base + memblock_end_of_DRAM();
1757 	mask = 1ULL << (fls64(end) - 1);
1758 	mask += mask - 1;
1759 
1760 	return mask;
1761 }
1762 
1763 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe,
1764 				   struct pci_bus *bus)
1765 {
1766 	struct pci_dev *dev;
1767 
1768 	list_for_each_entry(dev, &bus->devices, bus_list) {
1769 		set_iommu_table_base(&dev->dev, pe->table_group.tables[0]);
1770 		set_dma_offset(&dev->dev, pe->tce_bypass_base);
1771 		iommu_add_device(&dev->dev);
1772 
1773 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1774 			pnv_ioda_setup_bus_dma(pe, dev->subordinate);
1775 	}
1776 }
1777 
1778 static inline __be64 __iomem *pnv_ioda_get_inval_reg(struct pnv_phb *phb,
1779 						     bool real_mode)
1780 {
1781 	return real_mode ? (__be64 __iomem *)(phb->regs_phys + 0x210) :
1782 		(phb->regs + 0x210);
1783 }
1784 
1785 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table *tbl,
1786 		unsigned long index, unsigned long npages, bool rm)
1787 {
1788 	struct iommu_table_group_link *tgl = list_first_entry_or_null(
1789 			&tbl->it_group_list, struct iommu_table_group_link,
1790 			next);
1791 	struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1792 			struct pnv_ioda_pe, table_group);
1793 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
1794 	unsigned long start, end, inc;
1795 
1796 	start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset);
1797 	end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset +
1798 			npages - 1);
1799 
1800 	/* p7ioc-style invalidation, 2 TCEs per write */
1801 	start |= (1ull << 63);
1802 	end |= (1ull << 63);
1803 	inc = 16;
1804         end |= inc - 1;	/* round up end to be different than start */
1805 
1806         mb(); /* Ensure above stores are visible */
1807         while (start <= end) {
1808 		if (rm)
1809 			__raw_rm_writeq(cpu_to_be64(start), invalidate);
1810 		else
1811 			__raw_writeq(cpu_to_be64(start), invalidate);
1812                 start += inc;
1813         }
1814 
1815 	/*
1816 	 * The iommu layer will do another mb() for us on build()
1817 	 * and we don't care on free()
1818 	 */
1819 }
1820 
1821 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index,
1822 		long npages, unsigned long uaddr,
1823 		enum dma_data_direction direction,
1824 		unsigned long attrs)
1825 {
1826 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1827 			attrs);
1828 
1829 	if (!ret)
1830 		pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1831 
1832 	return ret;
1833 }
1834 
1835 #ifdef CONFIG_IOMMU_API
1836 static int pnv_ioda1_tce_xchg(struct iommu_table *tbl, long index,
1837 		unsigned long *hpa, enum dma_data_direction *direction)
1838 {
1839 	long ret = pnv_tce_xchg(tbl, index, hpa, direction);
1840 
1841 	if (!ret)
1842 		pnv_pci_p7ioc_tce_invalidate(tbl, index, 1, false);
1843 
1844 	return ret;
1845 }
1846 #endif
1847 
1848 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index,
1849 		long npages)
1850 {
1851 	pnv_tce_free(tbl, index, npages);
1852 
1853 	pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1854 }
1855 
1856 static struct iommu_table_ops pnv_ioda1_iommu_ops = {
1857 	.set = pnv_ioda1_tce_build,
1858 #ifdef CONFIG_IOMMU_API
1859 	.exchange = pnv_ioda1_tce_xchg,
1860 #endif
1861 	.clear = pnv_ioda1_tce_free,
1862 	.get = pnv_tce_get,
1863 };
1864 
1865 #define PHB3_TCE_KILL_INVAL_ALL		PPC_BIT(0)
1866 #define PHB3_TCE_KILL_INVAL_PE		PPC_BIT(1)
1867 #define PHB3_TCE_KILL_INVAL_ONE		PPC_BIT(2)
1868 
1869 void pnv_pci_phb3_tce_invalidate_entire(struct pnv_phb *phb, bool rm)
1870 {
1871 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(phb, rm);
1872 	const unsigned long val = PHB3_TCE_KILL_INVAL_ALL;
1873 
1874 	mb(); /* Ensure previous TCE table stores are visible */
1875 	if (rm)
1876 		__raw_rm_writeq(cpu_to_be64(val), invalidate);
1877 	else
1878 		__raw_writeq(cpu_to_be64(val), invalidate);
1879 }
1880 
1881 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1882 {
1883 	/* 01xb - invalidate TCEs that match the specified PE# */
1884 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, false);
1885 	unsigned long val = PHB3_TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
1886 
1887 	mb(); /* Ensure above stores are visible */
1888 	__raw_writeq(cpu_to_be64(val), invalidate);
1889 }
1890 
1891 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe *pe, bool rm,
1892 					unsigned shift, unsigned long index,
1893 					unsigned long npages)
1894 {
1895 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
1896 	unsigned long start, end, inc;
1897 
1898 	/* We'll invalidate DMA address in PE scope */
1899 	start = PHB3_TCE_KILL_INVAL_ONE;
1900 	start |= (pe->pe_number & 0xFF);
1901 	end = start;
1902 
1903 	/* Figure out the start, end and step */
1904 	start |= (index << shift);
1905 	end |= ((index + npages - 1) << shift);
1906 	inc = (0x1ull << shift);
1907 	mb();
1908 
1909 	while (start <= end) {
1910 		if (rm)
1911 			__raw_rm_writeq(cpu_to_be64(start), invalidate);
1912 		else
1913 			__raw_writeq(cpu_to_be64(start), invalidate);
1914 		start += inc;
1915 	}
1916 }
1917 
1918 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1919 {
1920 	struct pnv_phb *phb = pe->phb;
1921 
1922 	if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
1923 		pnv_pci_phb3_tce_invalidate_pe(pe);
1924 	else
1925 		opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL_PE,
1926 				  pe->pe_number, 0, 0, 0);
1927 }
1928 
1929 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
1930 		unsigned long index, unsigned long npages, bool rm)
1931 {
1932 	struct iommu_table_group_link *tgl;
1933 
1934 	list_for_each_entry_rcu(tgl, &tbl->it_group_list, next) {
1935 		struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1936 				struct pnv_ioda_pe, table_group);
1937 		struct pnv_phb *phb = pe->phb;
1938 		unsigned int shift = tbl->it_page_shift;
1939 
1940 		if (phb->type == PNV_PHB_NPU) {
1941 			/*
1942 			 * The NVLink hardware does not support TCE kill
1943 			 * per TCE entry so we have to invalidate
1944 			 * the entire cache for it.
1945 			 */
1946 			pnv_pci_phb3_tce_invalidate_entire(phb, rm);
1947 			continue;
1948 		}
1949 		if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
1950 			pnv_pci_phb3_tce_invalidate(pe, rm, shift,
1951 						    index, npages);
1952 		else if (rm)
1953 			opal_rm_pci_tce_kill(phb->opal_id,
1954 					     OPAL_PCI_TCE_KILL_PAGES,
1955 					     pe->pe_number, 1u << shift,
1956 					     index << shift, npages);
1957 		else
1958 			opal_pci_tce_kill(phb->opal_id,
1959 					  OPAL_PCI_TCE_KILL_PAGES,
1960 					  pe->pe_number, 1u << shift,
1961 					  index << shift, npages);
1962 	}
1963 }
1964 
1965 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index,
1966 		long npages, unsigned long uaddr,
1967 		enum dma_data_direction direction,
1968 		unsigned long attrs)
1969 {
1970 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1971 			attrs);
1972 
1973 	if (!ret)
1974 		pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1975 
1976 	return ret;
1977 }
1978 
1979 #ifdef CONFIG_IOMMU_API
1980 static int pnv_ioda2_tce_xchg(struct iommu_table *tbl, long index,
1981 		unsigned long *hpa, enum dma_data_direction *direction)
1982 {
1983 	long ret = pnv_tce_xchg(tbl, index, hpa, direction);
1984 
1985 	if (!ret)
1986 		pnv_pci_ioda2_tce_invalidate(tbl, index, 1, false);
1987 
1988 	return ret;
1989 }
1990 #endif
1991 
1992 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index,
1993 		long npages)
1994 {
1995 	pnv_tce_free(tbl, index, npages);
1996 
1997 	pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1998 }
1999 
2000 static void pnv_ioda2_table_free(struct iommu_table *tbl)
2001 {
2002 	pnv_pci_ioda2_table_free_pages(tbl);
2003 	iommu_free_table(tbl, "pnv");
2004 }
2005 
2006 static struct iommu_table_ops pnv_ioda2_iommu_ops = {
2007 	.set = pnv_ioda2_tce_build,
2008 #ifdef CONFIG_IOMMU_API
2009 	.exchange = pnv_ioda2_tce_xchg,
2010 #endif
2011 	.clear = pnv_ioda2_tce_free,
2012 	.get = pnv_tce_get,
2013 	.free = pnv_ioda2_table_free,
2014 };
2015 
2016 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data)
2017 {
2018 	unsigned int *weight = (unsigned int *)data;
2019 
2020 	/* This is quite simplistic. The "base" weight of a device
2021 	 * is 10. 0 means no DMA is to be accounted for it.
2022 	 */
2023 	if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
2024 		return 0;
2025 
2026 	if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
2027 	    dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
2028 	    dev->class == PCI_CLASS_SERIAL_USB_EHCI)
2029 		*weight += 3;
2030 	else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
2031 		*weight += 15;
2032 	else
2033 		*weight += 10;
2034 
2035 	return 0;
2036 }
2037 
2038 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe)
2039 {
2040 	unsigned int weight = 0;
2041 
2042 	/* SRIOV VF has same DMA32 weight as its PF */
2043 #ifdef CONFIG_PCI_IOV
2044 	if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) {
2045 		pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight);
2046 		return weight;
2047 	}
2048 #endif
2049 
2050 	if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) {
2051 		pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight);
2052 	} else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) {
2053 		struct pci_dev *pdev;
2054 
2055 		list_for_each_entry(pdev, &pe->pbus->devices, bus_list)
2056 			pnv_pci_ioda_dev_dma_weight(pdev, &weight);
2057 	} else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) {
2058 		pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight);
2059 	}
2060 
2061 	return weight;
2062 }
2063 
2064 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
2065 				       struct pnv_ioda_pe *pe)
2066 {
2067 
2068 	struct page *tce_mem = NULL;
2069 	struct iommu_table *tbl;
2070 	unsigned int weight, total_weight = 0;
2071 	unsigned int tce32_segsz, base, segs, avail, i;
2072 	int64_t rc;
2073 	void *addr;
2074 
2075 	/* XXX FIXME: Handle 64-bit only DMA devices */
2076 	/* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
2077 	/* XXX FIXME: Allocate multi-level tables on PHB3 */
2078 	weight = pnv_pci_ioda_pe_dma_weight(pe);
2079 	if (!weight)
2080 		return;
2081 
2082 	pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight,
2083 		     &total_weight);
2084 	segs = (weight * phb->ioda.dma32_count) / total_weight;
2085 	if (!segs)
2086 		segs = 1;
2087 
2088 	/*
2089 	 * Allocate contiguous DMA32 segments. We begin with the expected
2090 	 * number of segments. With one more attempt, the number of DMA32
2091 	 * segments to be allocated is decreased by one until one segment
2092 	 * is allocated successfully.
2093 	 */
2094 	do {
2095 		for (base = 0; base <= phb->ioda.dma32_count - segs; base++) {
2096 			for (avail = 0, i = base; i < base + segs; i++) {
2097 				if (phb->ioda.dma32_segmap[i] ==
2098 				    IODA_INVALID_PE)
2099 					avail++;
2100 			}
2101 
2102 			if (avail == segs)
2103 				goto found;
2104 		}
2105 	} while (--segs);
2106 
2107 	if (!segs) {
2108 		pe_warn(pe, "No available DMA32 segments\n");
2109 		return;
2110 	}
2111 
2112 found:
2113 	tbl = pnv_pci_table_alloc(phb->hose->node);
2114 	iommu_register_group(&pe->table_group, phb->hose->global_number,
2115 			pe->pe_number);
2116 	pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group);
2117 
2118 	/* Grab a 32-bit TCE table */
2119 	pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
2120 		weight, total_weight, base, segs);
2121 	pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
2122 		base * PNV_IODA1_DMA32_SEGSIZE,
2123 		(base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1);
2124 
2125 	/* XXX Currently, we allocate one big contiguous table for the
2126 	 * TCEs. We only really need one chunk per 256M of TCE space
2127 	 * (ie per segment) but that's an optimization for later, it
2128 	 * requires some added smarts with our get/put_tce implementation
2129 	 *
2130 	 * Each TCE page is 4KB in size and each TCE entry occupies 8
2131 	 * bytes
2132 	 */
2133 	tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3);
2134 	tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
2135 				   get_order(tce32_segsz * segs));
2136 	if (!tce_mem) {
2137 		pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
2138 		goto fail;
2139 	}
2140 	addr = page_address(tce_mem);
2141 	memset(addr, 0, tce32_segsz * segs);
2142 
2143 	/* Configure HW */
2144 	for (i = 0; i < segs; i++) {
2145 		rc = opal_pci_map_pe_dma_window(phb->opal_id,
2146 					      pe->pe_number,
2147 					      base + i, 1,
2148 					      __pa(addr) + tce32_segsz * i,
2149 					      tce32_segsz, IOMMU_PAGE_SIZE_4K);
2150 		if (rc) {
2151 			pe_err(pe, " Failed to configure 32-bit TCE table,"
2152 			       " err %ld\n", rc);
2153 			goto fail;
2154 		}
2155 	}
2156 
2157 	/* Setup DMA32 segment mapping */
2158 	for (i = base; i < base + segs; i++)
2159 		phb->ioda.dma32_segmap[i] = pe->pe_number;
2160 
2161 	/* Setup linux iommu table */
2162 	pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs,
2163 				  base * PNV_IODA1_DMA32_SEGSIZE,
2164 				  IOMMU_PAGE_SHIFT_4K);
2165 
2166 	tbl->it_ops = &pnv_ioda1_iommu_ops;
2167 	pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
2168 	pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
2169 	iommu_init_table(tbl, phb->hose->node);
2170 
2171 	if (pe->flags & PNV_IODA_PE_DEV) {
2172 		/*
2173 		 * Setting table base here only for carrying iommu_group
2174 		 * further down to let iommu_add_device() do the job.
2175 		 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2176 		 */
2177 		set_iommu_table_base(&pe->pdev->dev, tbl);
2178 		iommu_add_device(&pe->pdev->dev);
2179 	} else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2180 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
2181 
2182 	return;
2183  fail:
2184 	/* XXX Failure: Try to fallback to 64-bit only ? */
2185 	if (tce_mem)
2186 		__free_pages(tce_mem, get_order(tce32_segsz * segs));
2187 	if (tbl) {
2188 		pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
2189 		iommu_free_table(tbl, "pnv");
2190 	}
2191 }
2192 
2193 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
2194 		int num, struct iommu_table *tbl)
2195 {
2196 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2197 			table_group);
2198 	struct pnv_phb *phb = pe->phb;
2199 	int64_t rc;
2200 	const unsigned long size = tbl->it_indirect_levels ?
2201 			tbl->it_level_size : tbl->it_size;
2202 	const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
2203 	const __u64 win_size = tbl->it_size << tbl->it_page_shift;
2204 
2205 	pe_info(pe, "Setting up window#%d %llx..%llx pg=%x\n", num,
2206 			start_addr, start_addr + win_size - 1,
2207 			IOMMU_PAGE_SIZE(tbl));
2208 
2209 	/*
2210 	 * Map TCE table through TVT. The TVE index is the PE number
2211 	 * shifted by 1 bit for 32-bits DMA space.
2212 	 */
2213 	rc = opal_pci_map_pe_dma_window(phb->opal_id,
2214 			pe->pe_number,
2215 			(pe->pe_number << 1) + num,
2216 			tbl->it_indirect_levels + 1,
2217 			__pa(tbl->it_base),
2218 			size << 3,
2219 			IOMMU_PAGE_SIZE(tbl));
2220 	if (rc) {
2221 		pe_err(pe, "Failed to configure TCE table, err %ld\n", rc);
2222 		return rc;
2223 	}
2224 
2225 	pnv_pci_link_table_and_group(phb->hose->node, num,
2226 			tbl, &pe->table_group);
2227 	pnv_pci_ioda2_tce_invalidate_pe(pe);
2228 
2229 	return 0;
2230 }
2231 
2232 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
2233 {
2234 	uint16_t window_id = (pe->pe_number << 1 ) + 1;
2235 	int64_t rc;
2236 
2237 	pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
2238 	if (enable) {
2239 		phys_addr_t top = memblock_end_of_DRAM();
2240 
2241 		top = roundup_pow_of_two(top);
2242 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2243 						     pe->pe_number,
2244 						     window_id,
2245 						     pe->tce_bypass_base,
2246 						     top);
2247 	} else {
2248 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2249 						     pe->pe_number,
2250 						     window_id,
2251 						     pe->tce_bypass_base,
2252 						     0);
2253 	}
2254 	if (rc)
2255 		pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
2256 	else
2257 		pe->tce_bypass_enabled = enable;
2258 }
2259 
2260 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset,
2261 		__u32 page_shift, __u64 window_size, __u32 levels,
2262 		struct iommu_table *tbl);
2263 
2264 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group,
2265 		int num, __u32 page_shift, __u64 window_size, __u32 levels,
2266 		struct iommu_table **ptbl)
2267 {
2268 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2269 			table_group);
2270 	int nid = pe->phb->hose->node;
2271 	__u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start;
2272 	long ret;
2273 	struct iommu_table *tbl;
2274 
2275 	tbl = pnv_pci_table_alloc(nid);
2276 	if (!tbl)
2277 		return -ENOMEM;
2278 
2279 	ret = pnv_pci_ioda2_table_alloc_pages(nid,
2280 			bus_offset, page_shift, window_size,
2281 			levels, tbl);
2282 	if (ret) {
2283 		iommu_free_table(tbl, "pnv");
2284 		return ret;
2285 	}
2286 
2287 	tbl->it_ops = &pnv_ioda2_iommu_ops;
2288 
2289 	*ptbl = tbl;
2290 
2291 	return 0;
2292 }
2293 
2294 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe)
2295 {
2296 	struct iommu_table *tbl = NULL;
2297 	long rc;
2298 
2299 	/*
2300 	 * crashkernel= specifies the kdump kernel's maximum memory at
2301 	 * some offset and there is no guaranteed the result is a power
2302 	 * of 2, which will cause errors later.
2303 	 */
2304 	const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max());
2305 
2306 	/*
2307 	 * In memory constrained environments, e.g. kdump kernel, the
2308 	 * DMA window can be larger than available memory, which will
2309 	 * cause errors later.
2310 	 */
2311 	const u64 window_size = min((u64)pe->table_group.tce32_size, max_memory);
2312 
2313 	rc = pnv_pci_ioda2_create_table(&pe->table_group, 0,
2314 			IOMMU_PAGE_SHIFT_4K,
2315 			window_size,
2316 			POWERNV_IOMMU_DEFAULT_LEVELS, &tbl);
2317 	if (rc) {
2318 		pe_err(pe, "Failed to create 32-bit TCE table, err %ld",
2319 				rc);
2320 		return rc;
2321 	}
2322 
2323 	iommu_init_table(tbl, pe->phb->hose->node);
2324 
2325 	rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl);
2326 	if (rc) {
2327 		pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n",
2328 				rc);
2329 		pnv_ioda2_table_free(tbl);
2330 		return rc;
2331 	}
2332 
2333 	if (!pnv_iommu_bypass_disabled)
2334 		pnv_pci_ioda2_set_bypass(pe, true);
2335 
2336 	/*
2337 	 * Setting table base here only for carrying iommu_group
2338 	 * further down to let iommu_add_device() do the job.
2339 	 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2340 	 */
2341 	if (pe->flags & PNV_IODA_PE_DEV)
2342 		set_iommu_table_base(&pe->pdev->dev, tbl);
2343 
2344 	return 0;
2345 }
2346 
2347 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV)
2348 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
2349 		int num)
2350 {
2351 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2352 			table_group);
2353 	struct pnv_phb *phb = pe->phb;
2354 	long ret;
2355 
2356 	pe_info(pe, "Removing DMA window #%d\n", num);
2357 
2358 	ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
2359 			(pe->pe_number << 1) + num,
2360 			0/* levels */, 0/* table address */,
2361 			0/* table size */, 0/* page size */);
2362 	if (ret)
2363 		pe_warn(pe, "Unmapping failed, ret = %ld\n", ret);
2364 	else
2365 		pnv_pci_ioda2_tce_invalidate_pe(pe);
2366 
2367 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
2368 
2369 	return ret;
2370 }
2371 #endif
2372 
2373 #ifdef CONFIG_IOMMU_API
2374 static unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift,
2375 		__u64 window_size, __u32 levels)
2376 {
2377 	unsigned long bytes = 0;
2378 	const unsigned window_shift = ilog2(window_size);
2379 	unsigned entries_shift = window_shift - page_shift;
2380 	unsigned table_shift = entries_shift + 3;
2381 	unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift);
2382 	unsigned long direct_table_size;
2383 
2384 	if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) ||
2385 			(window_size > memory_hotplug_max()) ||
2386 			!is_power_of_2(window_size))
2387 		return 0;
2388 
2389 	/* Calculate a direct table size from window_size and levels */
2390 	entries_shift = (entries_shift + levels - 1) / levels;
2391 	table_shift = entries_shift + 3;
2392 	table_shift = max_t(unsigned, table_shift, PAGE_SHIFT);
2393 	direct_table_size =  1UL << table_shift;
2394 
2395 	for ( ; levels; --levels) {
2396 		bytes += _ALIGN_UP(tce_table_size, direct_table_size);
2397 
2398 		tce_table_size /= direct_table_size;
2399 		tce_table_size <<= 3;
2400 		tce_table_size = _ALIGN_UP(tce_table_size, direct_table_size);
2401 	}
2402 
2403 	return bytes;
2404 }
2405 
2406 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group)
2407 {
2408 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2409 						table_group);
2410 	/* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
2411 	struct iommu_table *tbl = pe->table_group.tables[0];
2412 
2413 	pnv_pci_ioda2_set_bypass(pe, false);
2414 	pnv_pci_ioda2_unset_window(&pe->table_group, 0);
2415 	pnv_ioda2_table_free(tbl);
2416 }
2417 
2418 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group)
2419 {
2420 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2421 						table_group);
2422 
2423 	pnv_pci_ioda2_setup_default_config(pe);
2424 }
2425 
2426 static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
2427 	.get_table_size = pnv_pci_ioda2_get_table_size,
2428 	.create_table = pnv_pci_ioda2_create_table,
2429 	.set_window = pnv_pci_ioda2_set_window,
2430 	.unset_window = pnv_pci_ioda2_unset_window,
2431 	.take_ownership = pnv_ioda2_take_ownership,
2432 	.release_ownership = pnv_ioda2_release_ownership,
2433 };
2434 
2435 static int gpe_table_group_to_npe_cb(struct device *dev, void *opaque)
2436 {
2437 	struct pci_controller *hose;
2438 	struct pnv_phb *phb;
2439 	struct pnv_ioda_pe **ptmppe = opaque;
2440 	struct pci_dev *pdev = container_of(dev, struct pci_dev, dev);
2441 	struct pci_dn *pdn = pci_get_pdn(pdev);
2442 
2443 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
2444 		return 0;
2445 
2446 	hose = pci_bus_to_host(pdev->bus);
2447 	phb = hose->private_data;
2448 	if (phb->type != PNV_PHB_NPU)
2449 		return 0;
2450 
2451 	*ptmppe = &phb->ioda.pe_array[pdn->pe_number];
2452 
2453 	return 1;
2454 }
2455 
2456 /*
2457  * This returns PE of associated NPU.
2458  * This assumes that NPU is in the same IOMMU group with GPU and there is
2459  * no other PEs.
2460  */
2461 static struct pnv_ioda_pe *gpe_table_group_to_npe(
2462 		struct iommu_table_group *table_group)
2463 {
2464 	struct pnv_ioda_pe *npe = NULL;
2465 	int ret = iommu_group_for_each_dev(table_group->group, &npe,
2466 			gpe_table_group_to_npe_cb);
2467 
2468 	BUG_ON(!ret || !npe);
2469 
2470 	return npe;
2471 }
2472 
2473 static long pnv_pci_ioda2_npu_set_window(struct iommu_table_group *table_group,
2474 		int num, struct iommu_table *tbl)
2475 {
2476 	long ret = pnv_pci_ioda2_set_window(table_group, num, tbl);
2477 
2478 	if (ret)
2479 		return ret;
2480 
2481 	ret = pnv_npu_set_window(gpe_table_group_to_npe(table_group), num, tbl);
2482 	if (ret)
2483 		pnv_pci_ioda2_unset_window(table_group, num);
2484 
2485 	return ret;
2486 }
2487 
2488 static long pnv_pci_ioda2_npu_unset_window(
2489 		struct iommu_table_group *table_group,
2490 		int num)
2491 {
2492 	long ret = pnv_pci_ioda2_unset_window(table_group, num);
2493 
2494 	if (ret)
2495 		return ret;
2496 
2497 	return pnv_npu_unset_window(gpe_table_group_to_npe(table_group), num);
2498 }
2499 
2500 static void pnv_ioda2_npu_take_ownership(struct iommu_table_group *table_group)
2501 {
2502 	/*
2503 	 * Detach NPU first as pnv_ioda2_take_ownership() will destroy
2504 	 * the iommu_table if 32bit DMA is enabled.
2505 	 */
2506 	pnv_npu_take_ownership(gpe_table_group_to_npe(table_group));
2507 	pnv_ioda2_take_ownership(table_group);
2508 }
2509 
2510 static struct iommu_table_group_ops pnv_pci_ioda2_npu_ops = {
2511 	.get_table_size = pnv_pci_ioda2_get_table_size,
2512 	.create_table = pnv_pci_ioda2_create_table,
2513 	.set_window = pnv_pci_ioda2_npu_set_window,
2514 	.unset_window = pnv_pci_ioda2_npu_unset_window,
2515 	.take_ownership = pnv_ioda2_npu_take_ownership,
2516 	.release_ownership = pnv_ioda2_release_ownership,
2517 };
2518 
2519 static void pnv_pci_ioda_setup_iommu_api(void)
2520 {
2521 	struct pci_controller *hose, *tmp;
2522 	struct pnv_phb *phb;
2523 	struct pnv_ioda_pe *pe, *gpe;
2524 
2525 	/*
2526 	 * Now we have all PHBs discovered, time to add NPU devices to
2527 	 * the corresponding IOMMU groups.
2528 	 */
2529 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2530 		phb = hose->private_data;
2531 
2532 		if (phb->type != PNV_PHB_NPU)
2533 			continue;
2534 
2535 		list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2536 			gpe = pnv_pci_npu_setup_iommu(pe);
2537 			if (gpe)
2538 				gpe->table_group.ops = &pnv_pci_ioda2_npu_ops;
2539 		}
2540 	}
2541 }
2542 #else /* !CONFIG_IOMMU_API */
2543 static void pnv_pci_ioda_setup_iommu_api(void) { };
2544 #endif
2545 
2546 static __be64 *pnv_pci_ioda2_table_do_alloc_pages(int nid, unsigned shift,
2547 		unsigned levels, unsigned long limit,
2548 		unsigned long *current_offset, unsigned long *total_allocated)
2549 {
2550 	struct page *tce_mem = NULL;
2551 	__be64 *addr, *tmp;
2552 	unsigned order = max_t(unsigned, shift, PAGE_SHIFT) - PAGE_SHIFT;
2553 	unsigned long allocated = 1UL << (order + PAGE_SHIFT);
2554 	unsigned entries = 1UL << (shift - 3);
2555 	long i;
2556 
2557 	tce_mem = alloc_pages_node(nid, GFP_KERNEL, order);
2558 	if (!tce_mem) {
2559 		pr_err("Failed to allocate a TCE memory, order=%d\n", order);
2560 		return NULL;
2561 	}
2562 	addr = page_address(tce_mem);
2563 	memset(addr, 0, allocated);
2564 	*total_allocated += allocated;
2565 
2566 	--levels;
2567 	if (!levels) {
2568 		*current_offset += allocated;
2569 		return addr;
2570 	}
2571 
2572 	for (i = 0; i < entries; ++i) {
2573 		tmp = pnv_pci_ioda2_table_do_alloc_pages(nid, shift,
2574 				levels, limit, current_offset, total_allocated);
2575 		if (!tmp)
2576 			break;
2577 
2578 		addr[i] = cpu_to_be64(__pa(tmp) |
2579 				TCE_PCI_READ | TCE_PCI_WRITE);
2580 
2581 		if (*current_offset >= limit)
2582 			break;
2583 	}
2584 
2585 	return addr;
2586 }
2587 
2588 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr,
2589 		unsigned long size, unsigned level);
2590 
2591 static long pnv_pci_ioda2_table_alloc_pages(int nid, __u64 bus_offset,
2592 		__u32 page_shift, __u64 window_size, __u32 levels,
2593 		struct iommu_table *tbl)
2594 {
2595 	void *addr;
2596 	unsigned long offset = 0, level_shift, total_allocated = 0;
2597 	const unsigned window_shift = ilog2(window_size);
2598 	unsigned entries_shift = window_shift - page_shift;
2599 	unsigned table_shift = max_t(unsigned, entries_shift + 3, PAGE_SHIFT);
2600 	const unsigned long tce_table_size = 1UL << table_shift;
2601 
2602 	if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS))
2603 		return -EINVAL;
2604 
2605 	if ((window_size > memory_hotplug_max()) || !is_power_of_2(window_size))
2606 		return -EINVAL;
2607 
2608 	/* Adjust direct table size from window_size and levels */
2609 	entries_shift = (entries_shift + levels - 1) / levels;
2610 	level_shift = entries_shift + 3;
2611 	level_shift = max_t(unsigned, level_shift, PAGE_SHIFT);
2612 
2613 	/* Allocate TCE table */
2614 	addr = pnv_pci_ioda2_table_do_alloc_pages(nid, level_shift,
2615 			levels, tce_table_size, &offset, &total_allocated);
2616 
2617 	/* addr==NULL means that the first level allocation failed */
2618 	if (!addr)
2619 		return -ENOMEM;
2620 
2621 	/*
2622 	 * First level was allocated but some lower level failed as
2623 	 * we did not allocate as much as we wanted,
2624 	 * release partially allocated table.
2625 	 */
2626 	if (offset < tce_table_size) {
2627 		pnv_pci_ioda2_table_do_free_pages(addr,
2628 				1ULL << (level_shift - 3), levels - 1);
2629 		return -ENOMEM;
2630 	}
2631 
2632 	/* Setup linux iommu table */
2633 	pnv_pci_setup_iommu_table(tbl, addr, tce_table_size, bus_offset,
2634 			page_shift);
2635 	tbl->it_level_size = 1ULL << (level_shift - 3);
2636 	tbl->it_indirect_levels = levels - 1;
2637 	tbl->it_allocated_size = total_allocated;
2638 
2639 	pr_devel("Created TCE table: ws=%08llx ts=%lx @%08llx\n",
2640 			window_size, tce_table_size, bus_offset);
2641 
2642 	return 0;
2643 }
2644 
2645 static void pnv_pci_ioda2_table_do_free_pages(__be64 *addr,
2646 		unsigned long size, unsigned level)
2647 {
2648 	const unsigned long addr_ul = (unsigned long) addr &
2649 			~(TCE_PCI_READ | TCE_PCI_WRITE);
2650 
2651 	if (level) {
2652 		long i;
2653 		u64 *tmp = (u64 *) addr_ul;
2654 
2655 		for (i = 0; i < size; ++i) {
2656 			unsigned long hpa = be64_to_cpu(tmp[i]);
2657 
2658 			if (!(hpa & (TCE_PCI_READ | TCE_PCI_WRITE)))
2659 				continue;
2660 
2661 			pnv_pci_ioda2_table_do_free_pages(__va(hpa), size,
2662 					level - 1);
2663 		}
2664 	}
2665 
2666 	free_pages(addr_ul, get_order(size << 3));
2667 }
2668 
2669 static void pnv_pci_ioda2_table_free_pages(struct iommu_table *tbl)
2670 {
2671 	const unsigned long size = tbl->it_indirect_levels ?
2672 			tbl->it_level_size : tbl->it_size;
2673 
2674 	if (!tbl->it_size)
2675 		return;
2676 
2677 	pnv_pci_ioda2_table_do_free_pages((__be64 *)tbl->it_base, size,
2678 			tbl->it_indirect_levels);
2679 }
2680 
2681 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
2682 				       struct pnv_ioda_pe *pe)
2683 {
2684 	int64_t rc;
2685 
2686 	if (!pnv_pci_ioda_pe_dma_weight(pe))
2687 		return;
2688 
2689 	/* TVE #1 is selected by PCI address bit 59 */
2690 	pe->tce_bypass_base = 1ull << 59;
2691 
2692 	iommu_register_group(&pe->table_group, phb->hose->global_number,
2693 			pe->pe_number);
2694 
2695 	/* The PE will reserve all possible 32-bits space */
2696 	pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
2697 		phb->ioda.m32_pci_base);
2698 
2699 	/* Setup linux iommu table */
2700 	pe->table_group.tce32_start = 0;
2701 	pe->table_group.tce32_size = phb->ioda.m32_pci_base;
2702 	pe->table_group.max_dynamic_windows_supported =
2703 			IOMMU_TABLE_GROUP_MAX_TABLES;
2704 	pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS;
2705 	pe->table_group.pgsizes = SZ_4K | SZ_64K | SZ_16M;
2706 #ifdef CONFIG_IOMMU_API
2707 	pe->table_group.ops = &pnv_pci_ioda2_ops;
2708 #endif
2709 
2710 	rc = pnv_pci_ioda2_setup_default_config(pe);
2711 	if (rc)
2712 		return;
2713 
2714 	if (pe->flags & PNV_IODA_PE_DEV)
2715 		iommu_add_device(&pe->pdev->dev);
2716 	else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2717 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
2718 }
2719 
2720 #ifdef CONFIG_PCI_MSI
2721 int64_t pnv_opal_pci_msi_eoi(struct irq_chip *chip, unsigned int hw_irq)
2722 {
2723 	struct pnv_phb *phb = container_of(chip, struct pnv_phb,
2724 					   ioda.irq_chip);
2725 
2726 	return opal_pci_msi_eoi(phb->opal_id, hw_irq);
2727 }
2728 
2729 static void pnv_ioda2_msi_eoi(struct irq_data *d)
2730 {
2731 	int64_t rc;
2732 	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
2733 	struct irq_chip *chip = irq_data_get_irq_chip(d);
2734 
2735 	rc = pnv_opal_pci_msi_eoi(chip, hw_irq);
2736 	WARN_ON_ONCE(rc);
2737 
2738 	icp_native_eoi(d);
2739 }
2740 
2741 
2742 void pnv_set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2743 {
2744 	struct irq_data *idata;
2745 	struct irq_chip *ichip;
2746 
2747 	/* The MSI EOI OPAL call is only needed on PHB3 */
2748 	if (phb->model != PNV_PHB_MODEL_PHB3)
2749 		return;
2750 
2751 	if (!phb->ioda.irq_chip_init) {
2752 		/*
2753 		 * First time we setup an MSI IRQ, we need to setup the
2754 		 * corresponding IRQ chip to route correctly.
2755 		 */
2756 		idata = irq_get_irq_data(virq);
2757 		ichip = irq_data_get_irq_chip(idata);
2758 		phb->ioda.irq_chip_init = 1;
2759 		phb->ioda.irq_chip = *ichip;
2760 		phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2761 	}
2762 	irq_set_chip(virq, &phb->ioda.irq_chip);
2763 }
2764 
2765 /*
2766  * Returns true iff chip is something that we could call
2767  * pnv_opal_pci_msi_eoi for.
2768  */
2769 bool is_pnv_opal_msi(struct irq_chip *chip)
2770 {
2771 	return chip->irq_eoi == pnv_ioda2_msi_eoi;
2772 }
2773 EXPORT_SYMBOL_GPL(is_pnv_opal_msi);
2774 
2775 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2776 				  unsigned int hwirq, unsigned int virq,
2777 				  unsigned int is_64, struct msi_msg *msg)
2778 {
2779 	struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2780 	unsigned int xive_num = hwirq - phb->msi_base;
2781 	__be32 data;
2782 	int rc;
2783 
2784 	/* No PE assigned ? bail out ... no MSI for you ! */
2785 	if (pe == NULL)
2786 		return -ENXIO;
2787 
2788 	/* Check if we have an MVE */
2789 	if (pe->mve_number < 0)
2790 		return -ENXIO;
2791 
2792 	/* Force 32-bit MSI on some broken devices */
2793 	if (dev->no_64bit_msi)
2794 		is_64 = 0;
2795 
2796 	/* Assign XIVE to PE */
2797 	rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2798 	if (rc) {
2799 		pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2800 			pci_name(dev), rc, xive_num);
2801 		return -EIO;
2802 	}
2803 
2804 	if (is_64) {
2805 		__be64 addr64;
2806 
2807 		rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
2808 				     &addr64, &data);
2809 		if (rc) {
2810 			pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2811 				pci_name(dev), rc);
2812 			return -EIO;
2813 		}
2814 		msg->address_hi = be64_to_cpu(addr64) >> 32;
2815 		msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
2816 	} else {
2817 		__be32 addr32;
2818 
2819 		rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
2820 				     &addr32, &data);
2821 		if (rc) {
2822 			pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2823 				pci_name(dev), rc);
2824 			return -EIO;
2825 		}
2826 		msg->address_hi = 0;
2827 		msg->address_lo = be32_to_cpu(addr32);
2828 	}
2829 	msg->data = be32_to_cpu(data);
2830 
2831 	pnv_set_msi_irq_chip(phb, virq);
2832 
2833 	pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
2834 		 " address=%x_%08x data=%x PE# %d\n",
2835 		 pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
2836 		 msg->address_hi, msg->address_lo, data, pe->pe_number);
2837 
2838 	return 0;
2839 }
2840 
2841 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
2842 {
2843 	unsigned int count;
2844 	const __be32 *prop = of_get_property(phb->hose->dn,
2845 					     "ibm,opal-msi-ranges", NULL);
2846 	if (!prop) {
2847 		/* BML Fallback */
2848 		prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
2849 	}
2850 	if (!prop)
2851 		return;
2852 
2853 	phb->msi_base = be32_to_cpup(prop);
2854 	count = be32_to_cpup(prop + 1);
2855 	if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
2856 		pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2857 		       phb->hose->global_number);
2858 		return;
2859 	}
2860 
2861 	phb->msi_setup = pnv_pci_ioda_msi_setup;
2862 	phb->msi32_support = 1;
2863 	pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2864 		count, phb->msi_base);
2865 }
2866 #else
2867 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb) { }
2868 #endif /* CONFIG_PCI_MSI */
2869 
2870 #ifdef CONFIG_PCI_IOV
2871 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev)
2872 {
2873 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
2874 	struct pnv_phb *phb = hose->private_data;
2875 	const resource_size_t gate = phb->ioda.m64_segsize >> 2;
2876 	struct resource *res;
2877 	int i;
2878 	resource_size_t size, total_vf_bar_sz;
2879 	struct pci_dn *pdn;
2880 	int mul, total_vfs;
2881 
2882 	if (!pdev->is_physfn || pdev->is_added)
2883 		return;
2884 
2885 	pdn = pci_get_pdn(pdev);
2886 	pdn->vfs_expanded = 0;
2887 	pdn->m64_single_mode = false;
2888 
2889 	total_vfs = pci_sriov_get_totalvfs(pdev);
2890 	mul = phb->ioda.total_pe_num;
2891 	total_vf_bar_sz = 0;
2892 
2893 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2894 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
2895 		if (!res->flags || res->parent)
2896 			continue;
2897 		if (!pnv_pci_is_m64_flags(res->flags)) {
2898 			dev_warn(&pdev->dev, "Don't support SR-IOV with"
2899 					" non M64 VF BAR%d: %pR. \n",
2900 				 i, res);
2901 			goto truncate_iov;
2902 		}
2903 
2904 		total_vf_bar_sz += pci_iov_resource_size(pdev,
2905 				i + PCI_IOV_RESOURCES);
2906 
2907 		/*
2908 		 * If bigger than quarter of M64 segment size, just round up
2909 		 * power of two.
2910 		 *
2911 		 * Generally, one M64 BAR maps one IOV BAR. To avoid conflict
2912 		 * with other devices, IOV BAR size is expanded to be
2913 		 * (total_pe * VF_BAR_size).  When VF_BAR_size is half of M64
2914 		 * segment size , the expanded size would equal to half of the
2915 		 * whole M64 space size, which will exhaust the M64 Space and
2916 		 * limit the system flexibility.  This is a design decision to
2917 		 * set the boundary to quarter of the M64 segment size.
2918 		 */
2919 		if (total_vf_bar_sz > gate) {
2920 			mul = roundup_pow_of_two(total_vfs);
2921 			dev_info(&pdev->dev,
2922 				"VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n",
2923 				total_vf_bar_sz, gate, mul);
2924 			pdn->m64_single_mode = true;
2925 			break;
2926 		}
2927 	}
2928 
2929 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2930 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
2931 		if (!res->flags || res->parent)
2932 			continue;
2933 
2934 		size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES);
2935 		/*
2936 		 * On PHB3, the minimum size alignment of M64 BAR in single
2937 		 * mode is 32MB.
2938 		 */
2939 		if (pdn->m64_single_mode && (size < SZ_32M))
2940 			goto truncate_iov;
2941 		dev_dbg(&pdev->dev, " Fixing VF BAR%d: %pR to\n", i, res);
2942 		res->end = res->start + size * mul - 1;
2943 		dev_dbg(&pdev->dev, "                       %pR\n", res);
2944 		dev_info(&pdev->dev, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
2945 			 i, res, mul);
2946 	}
2947 	pdn->vfs_expanded = mul;
2948 
2949 	return;
2950 
2951 truncate_iov:
2952 	/* To save MMIO space, IOV BAR is truncated. */
2953 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2954 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
2955 		res->flags = 0;
2956 		res->end = res->start - 1;
2957 	}
2958 }
2959 #endif /* CONFIG_PCI_IOV */
2960 
2961 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe,
2962 				  struct resource *res)
2963 {
2964 	struct pnv_phb *phb = pe->phb;
2965 	struct pci_bus_region region;
2966 	int index;
2967 	int64_t rc;
2968 
2969 	if (!res || !res->flags || res->start > res->end)
2970 		return;
2971 
2972 	if (res->flags & IORESOURCE_IO) {
2973 		region.start = res->start - phb->ioda.io_pci_base;
2974 		region.end   = res->end - phb->ioda.io_pci_base;
2975 		index = region.start / phb->ioda.io_segsize;
2976 
2977 		while (index < phb->ioda.total_pe_num &&
2978 		       region.start <= region.end) {
2979 			phb->ioda.io_segmap[index] = pe->pe_number;
2980 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2981 				pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
2982 			if (rc != OPAL_SUCCESS) {
2983 				pr_err("%s: Error %lld mapping IO segment#%d to PE#%d\n",
2984 				       __func__, rc, index, pe->pe_number);
2985 				break;
2986 			}
2987 
2988 			region.start += phb->ioda.io_segsize;
2989 			index++;
2990 		}
2991 	} else if ((res->flags & IORESOURCE_MEM) &&
2992 		   !pnv_pci_is_m64(phb, res)) {
2993 		region.start = res->start -
2994 			       phb->hose->mem_offset[0] -
2995 			       phb->ioda.m32_pci_base;
2996 		region.end   = res->end -
2997 			       phb->hose->mem_offset[0] -
2998 			       phb->ioda.m32_pci_base;
2999 		index = region.start / phb->ioda.m32_segsize;
3000 
3001 		while (index < phb->ioda.total_pe_num &&
3002 		       region.start <= region.end) {
3003 			phb->ioda.m32_segmap[index] = pe->pe_number;
3004 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3005 				pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
3006 			if (rc != OPAL_SUCCESS) {
3007 				pr_err("%s: Error %lld mapping M32 segment#%d to PE#%d",
3008 				       __func__, rc, index, pe->pe_number);
3009 				break;
3010 			}
3011 
3012 			region.start += phb->ioda.m32_segsize;
3013 			index++;
3014 		}
3015 	}
3016 }
3017 
3018 /*
3019  * This function is supposed to be called on basis of PE from top
3020  * to bottom style. So the the I/O or MMIO segment assigned to
3021  * parent PE could be overrided by its child PEs if necessary.
3022  */
3023 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe)
3024 {
3025 	struct pci_dev *pdev;
3026 	int i;
3027 
3028 	/*
3029 	 * NOTE: We only care PCI bus based PE for now. For PCI
3030 	 * device based PE, for example SRIOV sensitive VF should
3031 	 * be figured out later.
3032 	 */
3033 	BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
3034 
3035 	list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
3036 		for (i = 0; i <= PCI_ROM_RESOURCE; i++)
3037 			pnv_ioda_setup_pe_res(pe, &pdev->resource[i]);
3038 
3039 		/*
3040 		 * If the PE contains all subordinate PCI buses, the
3041 		 * windows of the child bridges should be mapped to
3042 		 * the PE as well.
3043 		 */
3044 		if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev))
3045 			continue;
3046 		for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++)
3047 			pnv_ioda_setup_pe_res(pe,
3048 				&pdev->resource[PCI_BRIDGE_RESOURCES + i]);
3049 	}
3050 }
3051 
3052 static void pnv_pci_ioda_create_dbgfs(void)
3053 {
3054 #ifdef CONFIG_DEBUG_FS
3055 	struct pci_controller *hose, *tmp;
3056 	struct pnv_phb *phb;
3057 	char name[16];
3058 
3059 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
3060 		phb = hose->private_data;
3061 
3062 		/* Notify initialization of PHB done */
3063 		phb->initialized = 1;
3064 
3065 		sprintf(name, "PCI%04x", hose->global_number);
3066 		phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
3067 		if (!phb->dbgfs)
3068 			pr_warning("%s: Error on creating debugfs on PHB#%x\n",
3069 				__func__, hose->global_number);
3070 	}
3071 #endif /* CONFIG_DEBUG_FS */
3072 }
3073 
3074 static void pnv_pci_ioda_fixup(void)
3075 {
3076 	pnv_pci_ioda_setup_PEs();
3077 	pnv_pci_ioda_setup_iommu_api();
3078 	pnv_pci_ioda_create_dbgfs();
3079 
3080 #ifdef CONFIG_EEH
3081 	eeh_init();
3082 	eeh_addr_cache_build();
3083 #endif
3084 }
3085 
3086 /*
3087  * Returns the alignment for I/O or memory windows for P2P
3088  * bridges. That actually depends on how PEs are segmented.
3089  * For now, we return I/O or M32 segment size for PE sensitive
3090  * P2P bridges. Otherwise, the default values (4KiB for I/O,
3091  * 1MiB for memory) will be returned.
3092  *
3093  * The current PCI bus might be put into one PE, which was
3094  * create against the parent PCI bridge. For that case, we
3095  * needn't enlarge the alignment so that we can save some
3096  * resources.
3097  */
3098 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
3099 						unsigned long type)
3100 {
3101 	struct pci_dev *bridge;
3102 	struct pci_controller *hose = pci_bus_to_host(bus);
3103 	struct pnv_phb *phb = hose->private_data;
3104 	int num_pci_bridges = 0;
3105 
3106 	bridge = bus->self;
3107 	while (bridge) {
3108 		if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
3109 			num_pci_bridges++;
3110 			if (num_pci_bridges >= 2)
3111 				return 1;
3112 		}
3113 
3114 		bridge = bridge->bus->self;
3115 	}
3116 
3117 	/*
3118 	 * We fall back to M32 if M64 isn't supported. We enforce the M64
3119 	 * alignment for any 64-bit resource, PCIe doesn't care and
3120 	 * bridges only do 64-bit prefetchable anyway.
3121 	 */
3122 	if (phb->ioda.m64_segsize && pnv_pci_is_m64_flags(type))
3123 		return phb->ioda.m64_segsize;
3124 	if (type & IORESOURCE_MEM)
3125 		return phb->ioda.m32_segsize;
3126 
3127 	return phb->ioda.io_segsize;
3128 }
3129 
3130 /*
3131  * We are updating root port or the upstream port of the
3132  * bridge behind the root port with PHB's windows in order
3133  * to accommodate the changes on required resources during
3134  * PCI (slot) hotplug, which is connected to either root
3135  * port or the downstream ports of PCIe switch behind the
3136  * root port.
3137  */
3138 static void pnv_pci_fixup_bridge_resources(struct pci_bus *bus,
3139 					   unsigned long type)
3140 {
3141 	struct pci_controller *hose = pci_bus_to_host(bus);
3142 	struct pnv_phb *phb = hose->private_data;
3143 	struct pci_dev *bridge = bus->self;
3144 	struct resource *r, *w;
3145 	bool msi_region = false;
3146 	int i;
3147 
3148 	/* Check if we need apply fixup to the bridge's windows */
3149 	if (!pci_is_root_bus(bridge->bus) &&
3150 	    !pci_is_root_bus(bridge->bus->self->bus))
3151 		return;
3152 
3153 	/* Fixup the resources */
3154 	for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
3155 		r = &bridge->resource[PCI_BRIDGE_RESOURCES + i];
3156 		if (!r->flags || !r->parent)
3157 			continue;
3158 
3159 		w = NULL;
3160 		if (r->flags & type & IORESOURCE_IO)
3161 			w = &hose->io_resource;
3162 		else if (pnv_pci_is_m64(phb, r) &&
3163 			 (type & IORESOURCE_PREFETCH) &&
3164 			 phb->ioda.m64_segsize)
3165 			w = &hose->mem_resources[1];
3166 		else if (r->flags & type & IORESOURCE_MEM) {
3167 			w = &hose->mem_resources[0];
3168 			msi_region = true;
3169 		}
3170 
3171 		r->start = w->start;
3172 		r->end = w->end;
3173 
3174 		/* The 64KB 32-bits MSI region shouldn't be included in
3175 		 * the 32-bits bridge window. Otherwise, we can see strange
3176 		 * issues. One of them is EEH error observed on Garrison.
3177 		 *
3178 		 * Exclude top 1MB region which is the minimal alignment of
3179 		 * 32-bits bridge window.
3180 		 */
3181 		if (msi_region) {
3182 			r->end += 0x10000;
3183 			r->end -= 0x100000;
3184 		}
3185 	}
3186 }
3187 
3188 static void pnv_pci_setup_bridge(struct pci_bus *bus, unsigned long type)
3189 {
3190 	struct pci_controller *hose = pci_bus_to_host(bus);
3191 	struct pnv_phb *phb = hose->private_data;
3192 	struct pci_dev *bridge = bus->self;
3193 	struct pnv_ioda_pe *pe;
3194 	bool all = (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE);
3195 
3196 	/* Extend bridge's windows if necessary */
3197 	pnv_pci_fixup_bridge_resources(bus, type);
3198 
3199 	/* The PE for root bus should be realized before any one else */
3200 	if (!phb->ioda.root_pe_populated) {
3201 		pe = pnv_ioda_setup_bus_PE(phb->hose->bus, false);
3202 		if (pe) {
3203 			phb->ioda.root_pe_idx = pe->pe_number;
3204 			phb->ioda.root_pe_populated = true;
3205 		}
3206 	}
3207 
3208 	/* Don't assign PE to PCI bus, which doesn't have subordinate devices */
3209 	if (list_empty(&bus->devices))
3210 		return;
3211 
3212 	/* Reserve PEs according to used M64 resources */
3213 	if (phb->reserve_m64_pe)
3214 		phb->reserve_m64_pe(bus, NULL, all);
3215 
3216 	/*
3217 	 * Assign PE. We might run here because of partial hotplug.
3218 	 * For the case, we just pick up the existing PE and should
3219 	 * not allocate resources again.
3220 	 */
3221 	pe = pnv_ioda_setup_bus_PE(bus, all);
3222 	if (!pe)
3223 		return;
3224 
3225 	pnv_ioda_setup_pe_seg(pe);
3226 	switch (phb->type) {
3227 	case PNV_PHB_IODA1:
3228 		pnv_pci_ioda1_setup_dma_pe(phb, pe);
3229 		break;
3230 	case PNV_PHB_IODA2:
3231 		pnv_pci_ioda2_setup_dma_pe(phb, pe);
3232 		break;
3233 	default:
3234 		pr_warn("%s: No DMA for PHB#%d (type %d)\n",
3235 			__func__, phb->hose->global_number, phb->type);
3236 	}
3237 }
3238 
3239 #ifdef CONFIG_PCI_IOV
3240 static resource_size_t pnv_pci_iov_resource_alignment(struct pci_dev *pdev,
3241 						      int resno)
3242 {
3243 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3244 	struct pnv_phb *phb = hose->private_data;
3245 	struct pci_dn *pdn = pci_get_pdn(pdev);
3246 	resource_size_t align;
3247 
3248 	/*
3249 	 * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the
3250 	 * SR-IOV. While from hardware perspective, the range mapped by M64
3251 	 * BAR should be size aligned.
3252 	 *
3253 	 * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra
3254 	 * powernv-specific hardware restriction is gone. But if just use the
3255 	 * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with
3256 	 * in one segment of M64 #15, which introduces the PE conflict between
3257 	 * PF and VF. Based on this, the minimum alignment of an IOV BAR is
3258 	 * m64_segsize.
3259 	 *
3260 	 * This function returns the total IOV BAR size if M64 BAR is in
3261 	 * Shared PE mode or just VF BAR size if not.
3262 	 * If the M64 BAR is in Single PE mode, return the VF BAR size or
3263 	 * M64 segment size if IOV BAR size is less.
3264 	 */
3265 	align = pci_iov_resource_size(pdev, resno);
3266 	if (!pdn->vfs_expanded)
3267 		return align;
3268 	if (pdn->m64_single_mode)
3269 		return max(align, (resource_size_t)phb->ioda.m64_segsize);
3270 
3271 	return pdn->vfs_expanded * align;
3272 }
3273 #endif /* CONFIG_PCI_IOV */
3274 
3275 /* Prevent enabling devices for which we couldn't properly
3276  * assign a PE
3277  */
3278 bool pnv_pci_enable_device_hook(struct pci_dev *dev)
3279 {
3280 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
3281 	struct pnv_phb *phb = hose->private_data;
3282 	struct pci_dn *pdn;
3283 
3284 	/* The function is probably called while the PEs have
3285 	 * not be created yet. For example, resource reassignment
3286 	 * during PCI probe period. We just skip the check if
3287 	 * PEs isn't ready.
3288 	 */
3289 	if (!phb->initialized)
3290 		return true;
3291 
3292 	pdn = pci_get_pdn(dev);
3293 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3294 		return false;
3295 
3296 	return true;
3297 }
3298 
3299 static long pnv_pci_ioda1_unset_window(struct iommu_table_group *table_group,
3300 				       int num)
3301 {
3302 	struct pnv_ioda_pe *pe = container_of(table_group,
3303 					      struct pnv_ioda_pe, table_group);
3304 	struct pnv_phb *phb = pe->phb;
3305 	unsigned int idx;
3306 	long rc;
3307 
3308 	pe_info(pe, "Removing DMA window #%d\n", num);
3309 	for (idx = 0; idx < phb->ioda.dma32_count; idx++) {
3310 		if (phb->ioda.dma32_segmap[idx] != pe->pe_number)
3311 			continue;
3312 
3313 		rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
3314 						idx, 0, 0ul, 0ul, 0ul);
3315 		if (rc != OPAL_SUCCESS) {
3316 			pe_warn(pe, "Failure %ld unmapping DMA32 segment#%d\n",
3317 				rc, idx);
3318 			return rc;
3319 		}
3320 
3321 		phb->ioda.dma32_segmap[idx] = IODA_INVALID_PE;
3322 	}
3323 
3324 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
3325 	return OPAL_SUCCESS;
3326 }
3327 
3328 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe *pe)
3329 {
3330 	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3331 	struct iommu_table *tbl = pe->table_group.tables[0];
3332 	int64_t rc;
3333 
3334 	if (!weight)
3335 		return;
3336 
3337 	rc = pnv_pci_ioda1_unset_window(&pe->table_group, 0);
3338 	if (rc != OPAL_SUCCESS)
3339 		return;
3340 
3341 	pnv_pci_p7ioc_tce_invalidate(tbl, tbl->it_offset, tbl->it_size, false);
3342 	if (pe->table_group.group) {
3343 		iommu_group_put(pe->table_group.group);
3344 		WARN_ON(pe->table_group.group);
3345 	}
3346 
3347 	free_pages(tbl->it_base, get_order(tbl->it_size << 3));
3348 	iommu_free_table(tbl, "pnv");
3349 }
3350 
3351 static void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe *pe)
3352 {
3353 	struct iommu_table *tbl = pe->table_group.tables[0];
3354 	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3355 #ifdef CONFIG_IOMMU_API
3356 	int64_t rc;
3357 #endif
3358 
3359 	if (!weight)
3360 		return;
3361 
3362 #ifdef CONFIG_IOMMU_API
3363 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
3364 	if (rc)
3365 		pe_warn(pe, "OPAL error %ld release DMA window\n", rc);
3366 #endif
3367 
3368 	pnv_pci_ioda2_set_bypass(pe, false);
3369 	if (pe->table_group.group) {
3370 		iommu_group_put(pe->table_group.group);
3371 		WARN_ON(pe->table_group.group);
3372 	}
3373 
3374 	pnv_pci_ioda2_table_free_pages(tbl);
3375 	iommu_free_table(tbl, "pnv");
3376 }
3377 
3378 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe *pe,
3379 				 unsigned short win,
3380 				 unsigned int *map)
3381 {
3382 	struct pnv_phb *phb = pe->phb;
3383 	int idx;
3384 	int64_t rc;
3385 
3386 	for (idx = 0; idx < phb->ioda.total_pe_num; idx++) {
3387 		if (map[idx] != pe->pe_number)
3388 			continue;
3389 
3390 		if (win == OPAL_M64_WINDOW_TYPE)
3391 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3392 					phb->ioda.reserved_pe_idx, win,
3393 					idx / PNV_IODA1_M64_SEGS,
3394 					idx % PNV_IODA1_M64_SEGS);
3395 		else
3396 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3397 					phb->ioda.reserved_pe_idx, win, 0, idx);
3398 
3399 		if (rc != OPAL_SUCCESS)
3400 			pe_warn(pe, "Error %ld unmapping (%d) segment#%d\n",
3401 				rc, win, idx);
3402 
3403 		map[idx] = IODA_INVALID_PE;
3404 	}
3405 }
3406 
3407 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
3408 {
3409 	struct pnv_phb *phb = pe->phb;
3410 
3411 	if (phb->type == PNV_PHB_IODA1) {
3412 		pnv_ioda_free_pe_seg(pe, OPAL_IO_WINDOW_TYPE,
3413 				     phb->ioda.io_segmap);
3414 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3415 				     phb->ioda.m32_segmap);
3416 		pnv_ioda_free_pe_seg(pe, OPAL_M64_WINDOW_TYPE,
3417 				     phb->ioda.m64_segmap);
3418 	} else if (phb->type == PNV_PHB_IODA2) {
3419 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3420 				     phb->ioda.m32_segmap);
3421 	}
3422 }
3423 
3424 static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe)
3425 {
3426 	struct pnv_phb *phb = pe->phb;
3427 	struct pnv_ioda_pe *slave, *tmp;
3428 
3429 	list_del(&pe->list);
3430 	switch (phb->type) {
3431 	case PNV_PHB_IODA1:
3432 		pnv_pci_ioda1_release_pe_dma(pe);
3433 		break;
3434 	case PNV_PHB_IODA2:
3435 		pnv_pci_ioda2_release_pe_dma(pe);
3436 		break;
3437 	default:
3438 		WARN_ON(1);
3439 	}
3440 
3441 	pnv_ioda_release_pe_seg(pe);
3442 	pnv_ioda_deconfigure_pe(pe->phb, pe);
3443 
3444 	/* Release slave PEs in the compound PE */
3445 	if (pe->flags & PNV_IODA_PE_MASTER) {
3446 		list_for_each_entry_safe(slave, tmp, &pe->slaves, list) {
3447 			list_del(&slave->list);
3448 			pnv_ioda_free_pe(slave);
3449 		}
3450 	}
3451 
3452 	/*
3453 	 * The PE for root bus can be removed because of hotplug in EEH
3454 	 * recovery for fenced PHB error. We need to mark the PE dead so
3455 	 * that it can be populated again in PCI hot add path. The PE
3456 	 * shouldn't be destroyed as it's the global reserved resource.
3457 	 */
3458 	if (phb->ioda.root_pe_populated &&
3459 	    phb->ioda.root_pe_idx == pe->pe_number)
3460 		phb->ioda.root_pe_populated = false;
3461 	else
3462 		pnv_ioda_free_pe(pe);
3463 }
3464 
3465 static void pnv_pci_release_device(struct pci_dev *pdev)
3466 {
3467 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3468 	struct pnv_phb *phb = hose->private_data;
3469 	struct pci_dn *pdn = pci_get_pdn(pdev);
3470 	struct pnv_ioda_pe *pe;
3471 
3472 	if (pdev->is_virtfn)
3473 		return;
3474 
3475 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3476 		return;
3477 
3478 	/*
3479 	 * PCI hotplug can happen as part of EEH error recovery. The @pdn
3480 	 * isn't removed and added afterwards in this scenario. We should
3481 	 * set the PE number in @pdn to an invalid one. Otherwise, the PE's
3482 	 * device count is decreased on removing devices while failing to
3483 	 * be increased on adding devices. It leads to unbalanced PE's device
3484 	 * count and eventually make normal PCI hotplug path broken.
3485 	 */
3486 	pe = &phb->ioda.pe_array[pdn->pe_number];
3487 	pdn->pe_number = IODA_INVALID_PE;
3488 
3489 	WARN_ON(--pe->device_count < 0);
3490 	if (pe->device_count == 0)
3491 		pnv_ioda_release_pe(pe);
3492 }
3493 
3494 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
3495 {
3496 	struct pnv_phb *phb = hose->private_data;
3497 
3498 	opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
3499 		       OPAL_ASSERT_RESET);
3500 }
3501 
3502 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
3503 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3504 	.dma_bus_setup		= pnv_pci_dma_bus_setup,
3505 #ifdef CONFIG_PCI_MSI
3506 	.setup_msi_irqs		= pnv_setup_msi_irqs,
3507 	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
3508 #endif
3509 	.enable_device_hook	= pnv_pci_enable_device_hook,
3510 	.release_device		= pnv_pci_release_device,
3511 	.window_alignment	= pnv_pci_window_alignment,
3512 	.setup_bridge		= pnv_pci_setup_bridge,
3513 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3514 	.dma_set_mask		= pnv_pci_ioda_dma_set_mask,
3515 	.dma_get_required_mask	= pnv_pci_ioda_dma_get_required_mask,
3516 	.shutdown		= pnv_pci_ioda_shutdown,
3517 };
3518 
3519 static int pnv_npu_dma_set_mask(struct pci_dev *npdev, u64 dma_mask)
3520 {
3521 	dev_err_once(&npdev->dev,
3522 			"%s operation unsupported for NVLink devices\n",
3523 			__func__);
3524 	return -EPERM;
3525 }
3526 
3527 static const struct pci_controller_ops pnv_npu_ioda_controller_ops = {
3528 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3529 #ifdef CONFIG_PCI_MSI
3530 	.setup_msi_irqs		= pnv_setup_msi_irqs,
3531 	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
3532 #endif
3533 	.enable_device_hook	= pnv_pci_enable_device_hook,
3534 	.window_alignment	= pnv_pci_window_alignment,
3535 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3536 	.dma_set_mask		= pnv_npu_dma_set_mask,
3537 	.shutdown		= pnv_pci_ioda_shutdown,
3538 };
3539 
3540 #ifdef CONFIG_CXL_BASE
3541 const struct pci_controller_ops pnv_cxl_cx4_ioda_controller_ops = {
3542 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3543 	.dma_bus_setup		= pnv_pci_dma_bus_setup,
3544 #ifdef CONFIG_PCI_MSI
3545 	.setup_msi_irqs		= pnv_cxl_cx4_setup_msi_irqs,
3546 	.teardown_msi_irqs	= pnv_cxl_cx4_teardown_msi_irqs,
3547 #endif
3548 	.enable_device_hook	= pnv_cxl_enable_device_hook,
3549 	.disable_device		= pnv_cxl_disable_device,
3550 	.release_device		= pnv_pci_release_device,
3551 	.window_alignment	= pnv_pci_window_alignment,
3552 	.setup_bridge		= pnv_pci_setup_bridge,
3553 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3554 	.dma_set_mask		= pnv_pci_ioda_dma_set_mask,
3555 	.dma_get_required_mask	= pnv_pci_ioda_dma_get_required_mask,
3556 	.shutdown		= pnv_pci_ioda_shutdown,
3557 };
3558 #endif
3559 
3560 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
3561 					 u64 hub_id, int ioda_type)
3562 {
3563 	struct pci_controller *hose;
3564 	struct pnv_phb *phb;
3565 	unsigned long size, m64map_off, m32map_off, pemap_off;
3566 	unsigned long iomap_off = 0, dma32map_off = 0;
3567 	struct resource r;
3568 	const __be64 *prop64;
3569 	const __be32 *prop32;
3570 	int len;
3571 	unsigned int segno;
3572 	u64 phb_id;
3573 	void *aux;
3574 	long rc;
3575 
3576 	if (!of_device_is_available(np))
3577 		return;
3578 
3579 	pr_info("Initializing %s PHB (%s)\n",
3580 		pnv_phb_names[ioda_type], of_node_full_name(np));
3581 
3582 	prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
3583 	if (!prop64) {
3584 		pr_err("  Missing \"ibm,opal-phbid\" property !\n");
3585 		return;
3586 	}
3587 	phb_id = be64_to_cpup(prop64);
3588 	pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
3589 
3590 	phb = memblock_virt_alloc(sizeof(struct pnv_phb), 0);
3591 
3592 	/* Allocate PCI controller */
3593 	phb->hose = hose = pcibios_alloc_controller(np);
3594 	if (!phb->hose) {
3595 		pr_err("  Can't allocate PCI controller for %s\n",
3596 		       np->full_name);
3597 		memblock_free(__pa(phb), sizeof(struct pnv_phb));
3598 		return;
3599 	}
3600 
3601 	spin_lock_init(&phb->lock);
3602 	prop32 = of_get_property(np, "bus-range", &len);
3603 	if (prop32 && len == 8) {
3604 		hose->first_busno = be32_to_cpu(prop32[0]);
3605 		hose->last_busno = be32_to_cpu(prop32[1]);
3606 	} else {
3607 		pr_warn("  Broken <bus-range> on %s\n", np->full_name);
3608 		hose->first_busno = 0;
3609 		hose->last_busno = 0xff;
3610 	}
3611 	hose->private_data = phb;
3612 	phb->hub_id = hub_id;
3613 	phb->opal_id = phb_id;
3614 	phb->type = ioda_type;
3615 	mutex_init(&phb->ioda.pe_alloc_mutex);
3616 
3617 	/* Detect specific models for error handling */
3618 	if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
3619 		phb->model = PNV_PHB_MODEL_P7IOC;
3620 	else if (of_device_is_compatible(np, "ibm,power8-pciex"))
3621 		phb->model = PNV_PHB_MODEL_PHB3;
3622 	else if (of_device_is_compatible(np, "ibm,power8-npu-pciex"))
3623 		phb->model = PNV_PHB_MODEL_NPU;
3624 	else
3625 		phb->model = PNV_PHB_MODEL_UNKNOWN;
3626 
3627 	/* Parse 32-bit and IO ranges (if any) */
3628 	pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
3629 
3630 	/* Get registers */
3631 	if (!of_address_to_resource(np, 0, &r)) {
3632 		phb->regs_phys = r.start;
3633 		phb->regs = ioremap(r.start, resource_size(&r));
3634 		if (phb->regs == NULL)
3635 			pr_err("  Failed to map registers !\n");
3636 	}
3637 
3638 	/* Initialize more IODA stuff */
3639 	phb->ioda.total_pe_num = 1;
3640 	prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
3641 	if (prop32)
3642 		phb->ioda.total_pe_num = be32_to_cpup(prop32);
3643 	prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
3644 	if (prop32)
3645 		phb->ioda.reserved_pe_idx = be32_to_cpup(prop32);
3646 
3647 	/* Invalidate RID to PE# mapping */
3648 	for (segno = 0; segno < ARRAY_SIZE(phb->ioda.pe_rmap); segno++)
3649 		phb->ioda.pe_rmap[segno] = IODA_INVALID_PE;
3650 
3651 	/* Parse 64-bit MMIO range */
3652 	pnv_ioda_parse_m64_window(phb);
3653 
3654 	phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
3655 	/* FW Has already off top 64k of M32 space (MSI space) */
3656 	phb->ioda.m32_size += 0x10000;
3657 
3658 	phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num;
3659 	phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
3660 	phb->ioda.io_size = hose->pci_io_size;
3661 	phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num;
3662 	phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
3663 
3664 	/* Calculate how many 32-bit TCE segments we have */
3665 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3666 				PNV_IODA1_DMA32_SEGSIZE;
3667 
3668 	/* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3669 	size = _ALIGN_UP(max_t(unsigned, phb->ioda.total_pe_num, 8) / 8,
3670 			sizeof(unsigned long));
3671 	m64map_off = size;
3672 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]);
3673 	m32map_off = size;
3674 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]);
3675 	if (phb->type == PNV_PHB_IODA1) {
3676 		iomap_off = size;
3677 		size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]);
3678 		dma32map_off = size;
3679 		size += phb->ioda.dma32_count *
3680 			sizeof(phb->ioda.dma32_segmap[0]);
3681 	}
3682 	pemap_off = size;
3683 	size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe);
3684 	aux = memblock_virt_alloc(size, 0);
3685 	phb->ioda.pe_alloc = aux;
3686 	phb->ioda.m64_segmap = aux + m64map_off;
3687 	phb->ioda.m32_segmap = aux + m32map_off;
3688 	for (segno = 0; segno < phb->ioda.total_pe_num; segno++) {
3689 		phb->ioda.m64_segmap[segno] = IODA_INVALID_PE;
3690 		phb->ioda.m32_segmap[segno] = IODA_INVALID_PE;
3691 	}
3692 	if (phb->type == PNV_PHB_IODA1) {
3693 		phb->ioda.io_segmap = aux + iomap_off;
3694 		for (segno = 0; segno < phb->ioda.total_pe_num; segno++)
3695 			phb->ioda.io_segmap[segno] = IODA_INVALID_PE;
3696 
3697 		phb->ioda.dma32_segmap = aux + dma32map_off;
3698 		for (segno = 0; segno < phb->ioda.dma32_count; segno++)
3699 			phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE;
3700 	}
3701 	phb->ioda.pe_array = aux + pemap_off;
3702 
3703 	/*
3704 	 * Choose PE number for root bus, which shouldn't have
3705 	 * M64 resources consumed by its child devices. To pick
3706 	 * the PE number adjacent to the reserved one if possible.
3707 	 */
3708 	pnv_ioda_reserve_pe(phb, phb->ioda.reserved_pe_idx);
3709 	if (phb->ioda.reserved_pe_idx == 0) {
3710 		phb->ioda.root_pe_idx = 1;
3711 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3712 	} else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) {
3713 		phb->ioda.root_pe_idx = phb->ioda.reserved_pe_idx - 1;
3714 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3715 	} else {
3716 		phb->ioda.root_pe_idx = IODA_INVALID_PE;
3717 	}
3718 
3719 	INIT_LIST_HEAD(&phb->ioda.pe_list);
3720 	mutex_init(&phb->ioda.pe_list_mutex);
3721 
3722 	/* Calculate how many 32-bit TCE segments we have */
3723 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3724 				PNV_IODA1_DMA32_SEGSIZE;
3725 
3726 #if 0 /* We should really do that ... */
3727 	rc = opal_pci_set_phb_mem_window(opal->phb_id,
3728 					 window_type,
3729 					 window_num,
3730 					 starting_real_address,
3731 					 starting_pci_address,
3732 					 segment_size);
3733 #endif
3734 
3735 	pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3736 		phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx,
3737 		phb->ioda.m32_size, phb->ioda.m32_segsize);
3738 	if (phb->ioda.m64_size)
3739 		pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
3740 			phb->ioda.m64_size, phb->ioda.m64_segsize);
3741 	if (phb->ioda.io_size)
3742 		pr_info("                  IO: 0x%x [segment=0x%x]\n",
3743 			phb->ioda.io_size, phb->ioda.io_segsize);
3744 
3745 
3746 	phb->hose->ops = &pnv_pci_ops;
3747 	phb->get_pe_state = pnv_ioda_get_pe_state;
3748 	phb->freeze_pe = pnv_ioda_freeze_pe;
3749 	phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
3750 
3751 	/* Setup MSI support */
3752 	pnv_pci_init_ioda_msis(phb);
3753 
3754 	/*
3755 	 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3756 	 * to let the PCI core do resource assignment. It's supposed
3757 	 * that the PCI core will do correct I/O and MMIO alignment
3758 	 * for the P2P bridge bars so that each PCI bus (excluding
3759 	 * the child P2P bridges) can form individual PE.
3760 	 */
3761 	ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
3762 
3763 	if (phb->type == PNV_PHB_NPU) {
3764 		hose->controller_ops = pnv_npu_ioda_controller_ops;
3765 	} else {
3766 		phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
3767 		hose->controller_ops = pnv_pci_ioda_controller_ops;
3768 	}
3769 
3770 #ifdef CONFIG_PCI_IOV
3771 	ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov_resources;
3772 	ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
3773 #endif
3774 
3775 	pci_add_flags(PCI_REASSIGN_ALL_RSRC);
3776 
3777 	/* Reset IODA tables to a clean state */
3778 	rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
3779 	if (rc)
3780 		pr_warning("  OPAL Error %ld performing IODA table reset !\n", rc);
3781 
3782 	/* If we're running in kdump kerenl, the previous kerenl never
3783 	 * shutdown PCI devices correctly. We already got IODA table
3784 	 * cleaned out. So we have to issue PHB reset to stop all PCI
3785 	 * transactions from previous kerenl.
3786 	 */
3787 	if (is_kdump_kernel()) {
3788 		pr_info("  Issue PHB reset ...\n");
3789 		pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
3790 		pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
3791 	}
3792 
3793 	/* Remove M64 resource if we can't configure it successfully */
3794 	if (!phb->init_m64 || phb->init_m64(phb))
3795 		hose->mem_resources[1].flags = 0;
3796 }
3797 
3798 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
3799 {
3800 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
3801 }
3802 
3803 void __init pnv_pci_init_npu_phb(struct device_node *np)
3804 {
3805 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU);
3806 }
3807 
3808 void __init pnv_pci_init_ioda_hub(struct device_node *np)
3809 {
3810 	struct device_node *phbn;
3811 	const __be64 *prop64;
3812 	u64 hub_id;
3813 
3814 	pr_info("Probing IODA IO-Hub %s\n", np->full_name);
3815 
3816 	prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
3817 	if (!prop64) {
3818 		pr_err(" Missing \"ibm,opal-hubid\" property !\n");
3819 		return;
3820 	}
3821 	hub_id = be64_to_cpup(prop64);
3822 	pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
3823 
3824 	/* Count child PHBs */
3825 	for_each_child_of_node(np, phbn) {
3826 		/* Look for IODA1 PHBs */
3827 		if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
3828 			pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
3829 	}
3830 }
3831