xref: /openbmc/linux/arch/powerpc/platforms/powernv/pci-ioda.c (revision 4f727ecefefbd180de10e25b3e74c03dce3f1e75)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Support PCI/PCIe on PowerNV platforms
4  *
5  * Copyright 2011 Benjamin Herrenschmidt, IBM Corp.
6  */
7 
8 #undef DEBUG
9 
10 #include <linux/kernel.h>
11 #include <linux/pci.h>
12 #include <linux/crash_dump.h>
13 #include <linux/delay.h>
14 #include <linux/string.h>
15 #include <linux/init.h>
16 #include <linux/memblock.h>
17 #include <linux/irq.h>
18 #include <linux/io.h>
19 #include <linux/msi.h>
20 #include <linux/iommu.h>
21 #include <linux/rculist.h>
22 #include <linux/sizes.h>
23 
24 #include <asm/sections.h>
25 #include <asm/io.h>
26 #include <asm/prom.h>
27 #include <asm/pci-bridge.h>
28 #include <asm/machdep.h>
29 #include <asm/msi_bitmap.h>
30 #include <asm/ppc-pci.h>
31 #include <asm/opal.h>
32 #include <asm/iommu.h>
33 #include <asm/tce.h>
34 #include <asm/xics.h>
35 #include <asm/debugfs.h>
36 #include <asm/firmware.h>
37 #include <asm/pnv-pci.h>
38 #include <asm/mmzone.h>
39 
40 #include <misc/cxl-base.h>
41 
42 #include "powernv.h"
43 #include "pci.h"
44 #include "../../../../drivers/pci/pci.h"
45 
46 #define PNV_IODA1_M64_NUM	16	/* Number of M64 BARs	*/
47 #define PNV_IODA1_M64_SEGS	8	/* Segments per M64 BAR	*/
48 #define PNV_IODA1_DMA32_SEGSIZE	0x10000000
49 
50 static const char * const pnv_phb_names[] = { "IODA1", "IODA2", "NPU_NVLINK",
51 					      "NPU_OCAPI" };
52 
53 void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
54 			    const char *fmt, ...)
55 {
56 	struct va_format vaf;
57 	va_list args;
58 	char pfix[32];
59 
60 	va_start(args, fmt);
61 
62 	vaf.fmt = fmt;
63 	vaf.va = &args;
64 
65 	if (pe->flags & PNV_IODA_PE_DEV)
66 		strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix));
67 	else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
68 		sprintf(pfix, "%04x:%02x     ",
69 			pci_domain_nr(pe->pbus), pe->pbus->number);
70 #ifdef CONFIG_PCI_IOV
71 	else if (pe->flags & PNV_IODA_PE_VF)
72 		sprintf(pfix, "%04x:%02x:%2x.%d",
73 			pci_domain_nr(pe->parent_dev->bus),
74 			(pe->rid & 0xff00) >> 8,
75 			PCI_SLOT(pe->rid), PCI_FUNC(pe->rid));
76 #endif /* CONFIG_PCI_IOV*/
77 
78 	printk("%spci %s: [PE# %.2x] %pV",
79 	       level, pfix, pe->pe_number, &vaf);
80 
81 	va_end(args);
82 }
83 
84 static bool pnv_iommu_bypass_disabled __read_mostly;
85 static bool pci_reset_phbs __read_mostly;
86 
87 static int __init iommu_setup(char *str)
88 {
89 	if (!str)
90 		return -EINVAL;
91 
92 	while (*str) {
93 		if (!strncmp(str, "nobypass", 8)) {
94 			pnv_iommu_bypass_disabled = true;
95 			pr_info("PowerNV: IOMMU bypass window disabled.\n");
96 			break;
97 		}
98 		str += strcspn(str, ",");
99 		if (*str == ',')
100 			str++;
101 	}
102 
103 	return 0;
104 }
105 early_param("iommu", iommu_setup);
106 
107 static int __init pci_reset_phbs_setup(char *str)
108 {
109 	pci_reset_phbs = true;
110 	return 0;
111 }
112 
113 early_param("ppc_pci_reset_phbs", pci_reset_phbs_setup);
114 
115 static inline bool pnv_pci_is_m64(struct pnv_phb *phb, struct resource *r)
116 {
117 	/*
118 	 * WARNING: We cannot rely on the resource flags. The Linux PCI
119 	 * allocation code sometimes decides to put a 64-bit prefetchable
120 	 * BAR in the 32-bit window, so we have to compare the addresses.
121 	 *
122 	 * For simplicity we only test resource start.
123 	 */
124 	return (r->start >= phb->ioda.m64_base &&
125 		r->start < (phb->ioda.m64_base + phb->ioda.m64_size));
126 }
127 
128 static inline bool pnv_pci_is_m64_flags(unsigned long resource_flags)
129 {
130 	unsigned long flags = (IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
131 
132 	return (resource_flags & flags) == flags;
133 }
134 
135 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
136 {
137 	s64 rc;
138 
139 	phb->ioda.pe_array[pe_no].phb = phb;
140 	phb->ioda.pe_array[pe_no].pe_number = pe_no;
141 
142 	/*
143 	 * Clear the PE frozen state as it might be put into frozen state
144 	 * in the last PCI remove path. It's not harmful to do so when the
145 	 * PE is already in unfrozen state.
146 	 */
147 	rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no,
148 				       OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
149 	if (rc != OPAL_SUCCESS && rc != OPAL_UNSUPPORTED)
150 		pr_warn("%s: Error %lld unfreezing PHB#%x-PE#%x\n",
151 			__func__, rc, phb->hose->global_number, pe_no);
152 
153 	return &phb->ioda.pe_array[pe_no];
154 }
155 
156 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
157 {
158 	if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) {
159 		pr_warn("%s: Invalid PE %x on PHB#%x\n",
160 			__func__, pe_no, phb->hose->global_number);
161 		return;
162 	}
163 
164 	if (test_and_set_bit(pe_no, phb->ioda.pe_alloc))
165 		pr_debug("%s: PE %x was reserved on PHB#%x\n",
166 			 __func__, pe_no, phb->hose->global_number);
167 
168 	pnv_ioda_init_pe(phb, pe_no);
169 }
170 
171 static struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb)
172 {
173 	long pe;
174 
175 	for (pe = phb->ioda.total_pe_num - 1; pe >= 0; pe--) {
176 		if (!test_and_set_bit(pe, phb->ioda.pe_alloc))
177 			return pnv_ioda_init_pe(phb, pe);
178 	}
179 
180 	return NULL;
181 }
182 
183 static void pnv_ioda_free_pe(struct pnv_ioda_pe *pe)
184 {
185 	struct pnv_phb *phb = pe->phb;
186 	unsigned int pe_num = pe->pe_number;
187 
188 	WARN_ON(pe->pdev);
189 	WARN_ON(pe->npucomp); /* NPUs are not supposed to be freed */
190 	kfree(pe->npucomp);
191 	memset(pe, 0, sizeof(struct pnv_ioda_pe));
192 	clear_bit(pe_num, phb->ioda.pe_alloc);
193 }
194 
195 /* The default M64 BAR is shared by all PEs */
196 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
197 {
198 	const char *desc;
199 	struct resource *r;
200 	s64 rc;
201 
202 	/* Configure the default M64 BAR */
203 	rc = opal_pci_set_phb_mem_window(phb->opal_id,
204 					 OPAL_M64_WINDOW_TYPE,
205 					 phb->ioda.m64_bar_idx,
206 					 phb->ioda.m64_base,
207 					 0, /* unused */
208 					 phb->ioda.m64_size);
209 	if (rc != OPAL_SUCCESS) {
210 		desc = "configuring";
211 		goto fail;
212 	}
213 
214 	/* Enable the default M64 BAR */
215 	rc = opal_pci_phb_mmio_enable(phb->opal_id,
216 				      OPAL_M64_WINDOW_TYPE,
217 				      phb->ioda.m64_bar_idx,
218 				      OPAL_ENABLE_M64_SPLIT);
219 	if (rc != OPAL_SUCCESS) {
220 		desc = "enabling";
221 		goto fail;
222 	}
223 
224 	/*
225 	 * Exclude the segments for reserved and root bus PE, which
226 	 * are first or last two PEs.
227 	 */
228 	r = &phb->hose->mem_resources[1];
229 	if (phb->ioda.reserved_pe_idx == 0)
230 		r->start += (2 * phb->ioda.m64_segsize);
231 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
232 		r->end -= (2 * phb->ioda.m64_segsize);
233 	else
234 		pr_warn("  Cannot strip M64 segment for reserved PE#%x\n",
235 			phb->ioda.reserved_pe_idx);
236 
237 	return 0;
238 
239 fail:
240 	pr_warn("  Failure %lld %s M64 BAR#%d\n",
241 		rc, desc, phb->ioda.m64_bar_idx);
242 	opal_pci_phb_mmio_enable(phb->opal_id,
243 				 OPAL_M64_WINDOW_TYPE,
244 				 phb->ioda.m64_bar_idx,
245 				 OPAL_DISABLE_M64);
246 	return -EIO;
247 }
248 
249 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev,
250 					 unsigned long *pe_bitmap)
251 {
252 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
253 	struct pnv_phb *phb = hose->private_data;
254 	struct resource *r;
255 	resource_size_t base, sgsz, start, end;
256 	int segno, i;
257 
258 	base = phb->ioda.m64_base;
259 	sgsz = phb->ioda.m64_segsize;
260 	for (i = 0; i <= PCI_ROM_RESOURCE; i++) {
261 		r = &pdev->resource[i];
262 		if (!r->parent || !pnv_pci_is_m64(phb, r))
263 			continue;
264 
265 		start = _ALIGN_DOWN(r->start - base, sgsz);
266 		end = _ALIGN_UP(r->end - base, sgsz);
267 		for (segno = start / sgsz; segno < end / sgsz; segno++) {
268 			if (pe_bitmap)
269 				set_bit(segno, pe_bitmap);
270 			else
271 				pnv_ioda_reserve_pe(phb, segno);
272 		}
273 	}
274 }
275 
276 static int pnv_ioda1_init_m64(struct pnv_phb *phb)
277 {
278 	struct resource *r;
279 	int index;
280 
281 	/*
282 	 * There are 16 M64 BARs, each of which has 8 segments. So
283 	 * there are as many M64 segments as the maximum number of
284 	 * PEs, which is 128.
285 	 */
286 	for (index = 0; index < PNV_IODA1_M64_NUM; index++) {
287 		unsigned long base, segsz = phb->ioda.m64_segsize;
288 		int64_t rc;
289 
290 		base = phb->ioda.m64_base +
291 		       index * PNV_IODA1_M64_SEGS * segsz;
292 		rc = opal_pci_set_phb_mem_window(phb->opal_id,
293 				OPAL_M64_WINDOW_TYPE, index, base, 0,
294 				PNV_IODA1_M64_SEGS * segsz);
295 		if (rc != OPAL_SUCCESS) {
296 			pr_warn("  Error %lld setting M64 PHB#%x-BAR#%d\n",
297 				rc, phb->hose->global_number, index);
298 			goto fail;
299 		}
300 
301 		rc = opal_pci_phb_mmio_enable(phb->opal_id,
302 				OPAL_M64_WINDOW_TYPE, index,
303 				OPAL_ENABLE_M64_SPLIT);
304 		if (rc != OPAL_SUCCESS) {
305 			pr_warn("  Error %lld enabling M64 PHB#%x-BAR#%d\n",
306 				rc, phb->hose->global_number, index);
307 			goto fail;
308 		}
309 	}
310 
311 	/*
312 	 * Exclude the segments for reserved and root bus PE, which
313 	 * are first or last two PEs.
314 	 */
315 	r = &phb->hose->mem_resources[1];
316 	if (phb->ioda.reserved_pe_idx == 0)
317 		r->start += (2 * phb->ioda.m64_segsize);
318 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
319 		r->end -= (2 * phb->ioda.m64_segsize);
320 	else
321 		WARN(1, "Wrong reserved PE#%x on PHB#%x\n",
322 		     phb->ioda.reserved_pe_idx, phb->hose->global_number);
323 
324 	return 0;
325 
326 fail:
327 	for ( ; index >= 0; index--)
328 		opal_pci_phb_mmio_enable(phb->opal_id,
329 			OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64);
330 
331 	return -EIO;
332 }
333 
334 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus,
335 				    unsigned long *pe_bitmap,
336 				    bool all)
337 {
338 	struct pci_dev *pdev;
339 
340 	list_for_each_entry(pdev, &bus->devices, bus_list) {
341 		pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap);
342 
343 		if (all && pdev->subordinate)
344 			pnv_ioda_reserve_m64_pe(pdev->subordinate,
345 						pe_bitmap, all);
346 	}
347 }
348 
349 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all)
350 {
351 	struct pci_controller *hose = pci_bus_to_host(bus);
352 	struct pnv_phb *phb = hose->private_data;
353 	struct pnv_ioda_pe *master_pe, *pe;
354 	unsigned long size, *pe_alloc;
355 	int i;
356 
357 	/* Root bus shouldn't use M64 */
358 	if (pci_is_root_bus(bus))
359 		return NULL;
360 
361 	/* Allocate bitmap */
362 	size = _ALIGN_UP(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
363 	pe_alloc = kzalloc(size, GFP_KERNEL);
364 	if (!pe_alloc) {
365 		pr_warn("%s: Out of memory !\n",
366 			__func__);
367 		return NULL;
368 	}
369 
370 	/* Figure out reserved PE numbers by the PE */
371 	pnv_ioda_reserve_m64_pe(bus, pe_alloc, all);
372 
373 	/*
374 	 * the current bus might not own M64 window and that's all
375 	 * contributed by its child buses. For the case, we needn't
376 	 * pick M64 dependent PE#.
377 	 */
378 	if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) {
379 		kfree(pe_alloc);
380 		return NULL;
381 	}
382 
383 	/*
384 	 * Figure out the master PE and put all slave PEs to master
385 	 * PE's list to form compound PE.
386 	 */
387 	master_pe = NULL;
388 	i = -1;
389 	while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) <
390 		phb->ioda.total_pe_num) {
391 		pe = &phb->ioda.pe_array[i];
392 
393 		phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number;
394 		if (!master_pe) {
395 			pe->flags |= PNV_IODA_PE_MASTER;
396 			INIT_LIST_HEAD(&pe->slaves);
397 			master_pe = pe;
398 		} else {
399 			pe->flags |= PNV_IODA_PE_SLAVE;
400 			pe->master = master_pe;
401 			list_add_tail(&pe->list, &master_pe->slaves);
402 		}
403 
404 		/*
405 		 * P7IOC supports M64DT, which helps mapping M64 segment
406 		 * to one particular PE#. However, PHB3 has fixed mapping
407 		 * between M64 segment and PE#. In order to have same logic
408 		 * for P7IOC and PHB3, we enforce fixed mapping between M64
409 		 * segment and PE# on P7IOC.
410 		 */
411 		if (phb->type == PNV_PHB_IODA1) {
412 			int64_t rc;
413 
414 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
415 					pe->pe_number, OPAL_M64_WINDOW_TYPE,
416 					pe->pe_number / PNV_IODA1_M64_SEGS,
417 					pe->pe_number % PNV_IODA1_M64_SEGS);
418 			if (rc != OPAL_SUCCESS)
419 				pr_warn("%s: Error %lld mapping M64 for PHB#%x-PE#%x\n",
420 					__func__, rc, phb->hose->global_number,
421 					pe->pe_number);
422 		}
423 	}
424 
425 	kfree(pe_alloc);
426 	return master_pe;
427 }
428 
429 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
430 {
431 	struct pci_controller *hose = phb->hose;
432 	struct device_node *dn = hose->dn;
433 	struct resource *res;
434 	u32 m64_range[2], i;
435 	const __be32 *r;
436 	u64 pci_addr;
437 
438 	if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) {
439 		pr_info("  Not support M64 window\n");
440 		return;
441 	}
442 
443 	if (!firmware_has_feature(FW_FEATURE_OPAL)) {
444 		pr_info("  Firmware too old to support M64 window\n");
445 		return;
446 	}
447 
448 	r = of_get_property(dn, "ibm,opal-m64-window", NULL);
449 	if (!r) {
450 		pr_info("  No <ibm,opal-m64-window> on %pOF\n",
451 			dn);
452 		return;
453 	}
454 
455 	/*
456 	 * Find the available M64 BAR range and pickup the last one for
457 	 * covering the whole 64-bits space. We support only one range.
458 	 */
459 	if (of_property_read_u32_array(dn, "ibm,opal-available-m64-ranges",
460 				       m64_range, 2)) {
461 		/* In absence of the property, assume 0..15 */
462 		m64_range[0] = 0;
463 		m64_range[1] = 16;
464 	}
465 	/* We only support 64 bits in our allocator */
466 	if (m64_range[1] > 63) {
467 		pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n",
468 			__func__, m64_range[1], phb->hose->global_number);
469 		m64_range[1] = 63;
470 	}
471 	/* Empty range, no m64 */
472 	if (m64_range[1] <= m64_range[0]) {
473 		pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n",
474 			__func__, phb->hose->global_number);
475 		return;
476 	}
477 
478 	/* Configure M64 informations */
479 	res = &hose->mem_resources[1];
480 	res->name = dn->full_name;
481 	res->start = of_translate_address(dn, r + 2);
482 	res->end = res->start + of_read_number(r + 4, 2) - 1;
483 	res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
484 	pci_addr = of_read_number(r, 2);
485 	hose->mem_offset[1] = res->start - pci_addr;
486 
487 	phb->ioda.m64_size = resource_size(res);
488 	phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num;
489 	phb->ioda.m64_base = pci_addr;
490 
491 	/* This lines up nicely with the display from processing OF ranges */
492 	pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n",
493 		res->start, res->end, pci_addr, m64_range[0],
494 		m64_range[0] + m64_range[1] - 1);
495 
496 	/* Mark all M64 used up by default */
497 	phb->ioda.m64_bar_alloc = (unsigned long)-1;
498 
499 	/* Use last M64 BAR to cover M64 window */
500 	m64_range[1]--;
501 	phb->ioda.m64_bar_idx = m64_range[0] + m64_range[1];
502 
503 	pr_info(" Using M64 #%d as default window\n", phb->ioda.m64_bar_idx);
504 
505 	/* Mark remaining ones free */
506 	for (i = m64_range[0]; i < m64_range[1]; i++)
507 		clear_bit(i, &phb->ioda.m64_bar_alloc);
508 
509 	/*
510 	 * Setup init functions for M64 based on IODA version, IODA3 uses
511 	 * the IODA2 code.
512 	 */
513 	if (phb->type == PNV_PHB_IODA1)
514 		phb->init_m64 = pnv_ioda1_init_m64;
515 	else
516 		phb->init_m64 = pnv_ioda2_init_m64;
517 }
518 
519 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
520 {
521 	struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
522 	struct pnv_ioda_pe *slave;
523 	s64 rc;
524 
525 	/* Fetch master PE */
526 	if (pe->flags & PNV_IODA_PE_SLAVE) {
527 		pe = pe->master;
528 		if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
529 			return;
530 
531 		pe_no = pe->pe_number;
532 	}
533 
534 	/* Freeze master PE */
535 	rc = opal_pci_eeh_freeze_set(phb->opal_id,
536 				     pe_no,
537 				     OPAL_EEH_ACTION_SET_FREEZE_ALL);
538 	if (rc != OPAL_SUCCESS) {
539 		pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
540 			__func__, rc, phb->hose->global_number, pe_no);
541 		return;
542 	}
543 
544 	/* Freeze slave PEs */
545 	if (!(pe->flags & PNV_IODA_PE_MASTER))
546 		return;
547 
548 	list_for_each_entry(slave, &pe->slaves, list) {
549 		rc = opal_pci_eeh_freeze_set(phb->opal_id,
550 					     slave->pe_number,
551 					     OPAL_EEH_ACTION_SET_FREEZE_ALL);
552 		if (rc != OPAL_SUCCESS)
553 			pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
554 				__func__, rc, phb->hose->global_number,
555 				slave->pe_number);
556 	}
557 }
558 
559 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
560 {
561 	struct pnv_ioda_pe *pe, *slave;
562 	s64 rc;
563 
564 	/* Find master PE */
565 	pe = &phb->ioda.pe_array[pe_no];
566 	if (pe->flags & PNV_IODA_PE_SLAVE) {
567 		pe = pe->master;
568 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
569 		pe_no = pe->pe_number;
570 	}
571 
572 	/* Clear frozen state for master PE */
573 	rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
574 	if (rc != OPAL_SUCCESS) {
575 		pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
576 			__func__, rc, opt, phb->hose->global_number, pe_no);
577 		return -EIO;
578 	}
579 
580 	if (!(pe->flags & PNV_IODA_PE_MASTER))
581 		return 0;
582 
583 	/* Clear frozen state for slave PEs */
584 	list_for_each_entry(slave, &pe->slaves, list) {
585 		rc = opal_pci_eeh_freeze_clear(phb->opal_id,
586 					     slave->pe_number,
587 					     opt);
588 		if (rc != OPAL_SUCCESS) {
589 			pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
590 				__func__, rc, opt, phb->hose->global_number,
591 				slave->pe_number);
592 			return -EIO;
593 		}
594 	}
595 
596 	return 0;
597 }
598 
599 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
600 {
601 	struct pnv_ioda_pe *slave, *pe;
602 	u8 fstate = 0, state;
603 	__be16 pcierr = 0;
604 	s64 rc;
605 
606 	/* Sanity check on PE number */
607 	if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num)
608 		return OPAL_EEH_STOPPED_PERM_UNAVAIL;
609 
610 	/*
611 	 * Fetch the master PE and the PE instance might be
612 	 * not initialized yet.
613 	 */
614 	pe = &phb->ioda.pe_array[pe_no];
615 	if (pe->flags & PNV_IODA_PE_SLAVE) {
616 		pe = pe->master;
617 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
618 		pe_no = pe->pe_number;
619 	}
620 
621 	/* Check the master PE */
622 	rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
623 					&state, &pcierr, NULL);
624 	if (rc != OPAL_SUCCESS) {
625 		pr_warn("%s: Failure %lld getting "
626 			"PHB#%x-PE#%x state\n",
627 			__func__, rc,
628 			phb->hose->global_number, pe_no);
629 		return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
630 	}
631 
632 	/* Check the slave PE */
633 	if (!(pe->flags & PNV_IODA_PE_MASTER))
634 		return state;
635 
636 	list_for_each_entry(slave, &pe->slaves, list) {
637 		rc = opal_pci_eeh_freeze_status(phb->opal_id,
638 						slave->pe_number,
639 						&fstate,
640 						&pcierr,
641 						NULL);
642 		if (rc != OPAL_SUCCESS) {
643 			pr_warn("%s: Failure %lld getting "
644 				"PHB#%x-PE#%x state\n",
645 				__func__, rc,
646 				phb->hose->global_number, slave->pe_number);
647 			return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
648 		}
649 
650 		/*
651 		 * Override the result based on the ascending
652 		 * priority.
653 		 */
654 		if (fstate > state)
655 			state = fstate;
656 	}
657 
658 	return state;
659 }
660 
661 struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
662 {
663 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
664 	struct pnv_phb *phb = hose->private_data;
665 	struct pci_dn *pdn = pci_get_pdn(dev);
666 
667 	if (!pdn)
668 		return NULL;
669 	if (pdn->pe_number == IODA_INVALID_PE)
670 		return NULL;
671 	return &phb->ioda.pe_array[pdn->pe_number];
672 }
673 
674 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
675 				  struct pnv_ioda_pe *parent,
676 				  struct pnv_ioda_pe *child,
677 				  bool is_add)
678 {
679 	const char *desc = is_add ? "adding" : "removing";
680 	uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
681 			      OPAL_REMOVE_PE_FROM_DOMAIN;
682 	struct pnv_ioda_pe *slave;
683 	long rc;
684 
685 	/* Parent PE affects child PE */
686 	rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
687 				child->pe_number, op);
688 	if (rc != OPAL_SUCCESS) {
689 		pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
690 			rc, desc);
691 		return -ENXIO;
692 	}
693 
694 	if (!(child->flags & PNV_IODA_PE_MASTER))
695 		return 0;
696 
697 	/* Compound case: parent PE affects slave PEs */
698 	list_for_each_entry(slave, &child->slaves, list) {
699 		rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
700 					slave->pe_number, op);
701 		if (rc != OPAL_SUCCESS) {
702 			pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
703 				rc, desc);
704 			return -ENXIO;
705 		}
706 	}
707 
708 	return 0;
709 }
710 
711 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
712 			      struct pnv_ioda_pe *pe,
713 			      bool is_add)
714 {
715 	struct pnv_ioda_pe *slave;
716 	struct pci_dev *pdev = NULL;
717 	int ret;
718 
719 	/*
720 	 * Clear PE frozen state. If it's master PE, we need
721 	 * clear slave PE frozen state as well.
722 	 */
723 	if (is_add) {
724 		opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
725 					  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
726 		if (pe->flags & PNV_IODA_PE_MASTER) {
727 			list_for_each_entry(slave, &pe->slaves, list)
728 				opal_pci_eeh_freeze_clear(phb->opal_id,
729 							  slave->pe_number,
730 							  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
731 		}
732 	}
733 
734 	/*
735 	 * Associate PE in PELT. We need add the PE into the
736 	 * corresponding PELT-V as well. Otherwise, the error
737 	 * originated from the PE might contribute to other
738 	 * PEs.
739 	 */
740 	ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
741 	if (ret)
742 		return ret;
743 
744 	/* For compound PEs, any one affects all of them */
745 	if (pe->flags & PNV_IODA_PE_MASTER) {
746 		list_for_each_entry(slave, &pe->slaves, list) {
747 			ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
748 			if (ret)
749 				return ret;
750 		}
751 	}
752 
753 	if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
754 		pdev = pe->pbus->self;
755 	else if (pe->flags & PNV_IODA_PE_DEV)
756 		pdev = pe->pdev->bus->self;
757 #ifdef CONFIG_PCI_IOV
758 	else if (pe->flags & PNV_IODA_PE_VF)
759 		pdev = pe->parent_dev;
760 #endif /* CONFIG_PCI_IOV */
761 	while (pdev) {
762 		struct pci_dn *pdn = pci_get_pdn(pdev);
763 		struct pnv_ioda_pe *parent;
764 
765 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
766 			parent = &phb->ioda.pe_array[pdn->pe_number];
767 			ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
768 			if (ret)
769 				return ret;
770 		}
771 
772 		pdev = pdev->bus->self;
773 	}
774 
775 	return 0;
776 }
777 
778 static int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
779 {
780 	struct pci_dev *parent;
781 	uint8_t bcomp, dcomp, fcomp;
782 	int64_t rc;
783 	long rid_end, rid;
784 
785 	/* Currently, we just deconfigure VF PE. Bus PE will always there.*/
786 	if (pe->pbus) {
787 		int count;
788 
789 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
790 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
791 		parent = pe->pbus->self;
792 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
793 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
794 		else
795 			count = 1;
796 
797 		switch(count) {
798 		case  1: bcomp = OpalPciBusAll;         break;
799 		case  2: bcomp = OpalPciBus7Bits;       break;
800 		case  4: bcomp = OpalPciBus6Bits;       break;
801 		case  8: bcomp = OpalPciBus5Bits;       break;
802 		case 16: bcomp = OpalPciBus4Bits;       break;
803 		case 32: bcomp = OpalPciBus3Bits;       break;
804 		default:
805 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
806 			        count);
807 			/* Do an exact match only */
808 			bcomp = OpalPciBusAll;
809 		}
810 		rid_end = pe->rid + (count << 8);
811 	} else {
812 #ifdef CONFIG_PCI_IOV
813 		if (pe->flags & PNV_IODA_PE_VF)
814 			parent = pe->parent_dev;
815 		else
816 #endif
817 			parent = pe->pdev->bus->self;
818 		bcomp = OpalPciBusAll;
819 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
820 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
821 		rid_end = pe->rid + 1;
822 	}
823 
824 	/* Clear the reverse map */
825 	for (rid = pe->rid; rid < rid_end; rid++)
826 		phb->ioda.pe_rmap[rid] = IODA_INVALID_PE;
827 
828 	/* Release from all parents PELT-V */
829 	while (parent) {
830 		struct pci_dn *pdn = pci_get_pdn(parent);
831 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
832 			rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
833 						pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
834 			/* XXX What to do in case of error ? */
835 		}
836 		parent = parent->bus->self;
837 	}
838 
839 	opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
840 				  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
841 
842 	/* Disassociate PE in PELT */
843 	rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
844 				pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
845 	if (rc)
846 		pe_warn(pe, "OPAL error %lld remove self from PELTV\n", rc);
847 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
848 			     bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
849 	if (rc)
850 		pe_err(pe, "OPAL error %lld trying to setup PELT table\n", rc);
851 
852 	pe->pbus = NULL;
853 	pe->pdev = NULL;
854 #ifdef CONFIG_PCI_IOV
855 	pe->parent_dev = NULL;
856 #endif
857 
858 	return 0;
859 }
860 
861 static int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
862 {
863 	struct pci_dev *parent;
864 	uint8_t bcomp, dcomp, fcomp;
865 	long rc, rid_end, rid;
866 
867 	/* Bus validation ? */
868 	if (pe->pbus) {
869 		int count;
870 
871 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
872 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
873 		parent = pe->pbus->self;
874 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
875 			count = pe->pbus->busn_res.end - pe->pbus->busn_res.start + 1;
876 		else
877 			count = 1;
878 
879 		switch(count) {
880 		case  1: bcomp = OpalPciBusAll;		break;
881 		case  2: bcomp = OpalPciBus7Bits;	break;
882 		case  4: bcomp = OpalPciBus6Bits;	break;
883 		case  8: bcomp = OpalPciBus5Bits;	break;
884 		case 16: bcomp = OpalPciBus4Bits;	break;
885 		case 32: bcomp = OpalPciBus3Bits;	break;
886 		default:
887 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
888 			        count);
889 			/* Do an exact match only */
890 			bcomp = OpalPciBusAll;
891 		}
892 		rid_end = pe->rid + (count << 8);
893 	} else {
894 #ifdef CONFIG_PCI_IOV
895 		if (pe->flags & PNV_IODA_PE_VF)
896 			parent = pe->parent_dev;
897 		else
898 #endif /* CONFIG_PCI_IOV */
899 			parent = pe->pdev->bus->self;
900 		bcomp = OpalPciBusAll;
901 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
902 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
903 		rid_end = pe->rid + 1;
904 	}
905 
906 	/*
907 	 * Associate PE in PELT. We need add the PE into the
908 	 * corresponding PELT-V as well. Otherwise, the error
909 	 * originated from the PE might contribute to other
910 	 * PEs.
911 	 */
912 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
913 			     bcomp, dcomp, fcomp, OPAL_MAP_PE);
914 	if (rc) {
915 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
916 		return -ENXIO;
917 	}
918 
919 	/*
920 	 * Configure PELTV. NPUs don't have a PELTV table so skip
921 	 * configuration on them.
922 	 */
923 	if (phb->type != PNV_PHB_NPU_NVLINK && phb->type != PNV_PHB_NPU_OCAPI)
924 		pnv_ioda_set_peltv(phb, pe, true);
925 
926 	/* Setup reverse map */
927 	for (rid = pe->rid; rid < rid_end; rid++)
928 		phb->ioda.pe_rmap[rid] = pe->pe_number;
929 
930 	/* Setup one MVTs on IODA1 */
931 	if (phb->type != PNV_PHB_IODA1) {
932 		pe->mve_number = 0;
933 		goto out;
934 	}
935 
936 	pe->mve_number = pe->pe_number;
937 	rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
938 	if (rc != OPAL_SUCCESS) {
939 		pe_err(pe, "OPAL error %ld setting up MVE %x\n",
940 		       rc, pe->mve_number);
941 		pe->mve_number = -1;
942 	} else {
943 		rc = opal_pci_set_mve_enable(phb->opal_id,
944 					     pe->mve_number, OPAL_ENABLE_MVE);
945 		if (rc) {
946 			pe_err(pe, "OPAL error %ld enabling MVE %x\n",
947 			       rc, pe->mve_number);
948 			pe->mve_number = -1;
949 		}
950 	}
951 
952 out:
953 	return 0;
954 }
955 
956 #ifdef CONFIG_PCI_IOV
957 static int pnv_pci_vf_resource_shift(struct pci_dev *dev, int offset)
958 {
959 	struct pci_dn *pdn = pci_get_pdn(dev);
960 	int i;
961 	struct resource *res, res2;
962 	resource_size_t size;
963 	u16 num_vfs;
964 
965 	if (!dev->is_physfn)
966 		return -EINVAL;
967 
968 	/*
969 	 * "offset" is in VFs.  The M64 windows are sized so that when they
970 	 * are segmented, each segment is the same size as the IOV BAR.
971 	 * Each segment is in a separate PE, and the high order bits of the
972 	 * address are the PE number.  Therefore, each VF's BAR is in a
973 	 * separate PE, and changing the IOV BAR start address changes the
974 	 * range of PEs the VFs are in.
975 	 */
976 	num_vfs = pdn->num_vfs;
977 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
978 		res = &dev->resource[i + PCI_IOV_RESOURCES];
979 		if (!res->flags || !res->parent)
980 			continue;
981 
982 		/*
983 		 * The actual IOV BAR range is determined by the start address
984 		 * and the actual size for num_vfs VFs BAR.  This check is to
985 		 * make sure that after shifting, the range will not overlap
986 		 * with another device.
987 		 */
988 		size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
989 		res2.flags = res->flags;
990 		res2.start = res->start + (size * offset);
991 		res2.end = res2.start + (size * num_vfs) - 1;
992 
993 		if (res2.end > res->end) {
994 			dev_err(&dev->dev, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
995 				i, &res2, res, num_vfs, offset);
996 			return -EBUSY;
997 		}
998 	}
999 
1000 	/*
1001 	 * Since M64 BAR shares segments among all possible 256 PEs,
1002 	 * we have to shift the beginning of PF IOV BAR to make it start from
1003 	 * the segment which belongs to the PE number assigned to the first VF.
1004 	 * This creates a "hole" in the /proc/iomem which could be used for
1005 	 * allocating other resources so we reserve this area below and
1006 	 * release when IOV is released.
1007 	 */
1008 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1009 		res = &dev->resource[i + PCI_IOV_RESOURCES];
1010 		if (!res->flags || !res->parent)
1011 			continue;
1012 
1013 		size = pci_iov_resource_size(dev, i + PCI_IOV_RESOURCES);
1014 		res2 = *res;
1015 		res->start += size * offset;
1016 
1017 		dev_info(&dev->dev, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n",
1018 			 i, &res2, res, (offset > 0) ? "En" : "Dis",
1019 			 num_vfs, offset);
1020 
1021 		if (offset < 0) {
1022 			devm_release_resource(&dev->dev, &pdn->holes[i]);
1023 			memset(&pdn->holes[i], 0, sizeof(pdn->holes[i]));
1024 		}
1025 
1026 		pci_update_resource(dev, i + PCI_IOV_RESOURCES);
1027 
1028 		if (offset > 0) {
1029 			pdn->holes[i].start = res2.start;
1030 			pdn->holes[i].end = res2.start + size * offset - 1;
1031 			pdn->holes[i].flags = IORESOURCE_BUS;
1032 			pdn->holes[i].name = "pnv_iov_reserved";
1033 			devm_request_resource(&dev->dev, res->parent,
1034 					&pdn->holes[i]);
1035 		}
1036 	}
1037 	return 0;
1038 }
1039 #endif /* CONFIG_PCI_IOV */
1040 
1041 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
1042 {
1043 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
1044 	struct pnv_phb *phb = hose->private_data;
1045 	struct pci_dn *pdn = pci_get_pdn(dev);
1046 	struct pnv_ioda_pe *pe;
1047 
1048 	if (!pdn) {
1049 		pr_err("%s: Device tree node not associated properly\n",
1050 			   pci_name(dev));
1051 		return NULL;
1052 	}
1053 	if (pdn->pe_number != IODA_INVALID_PE)
1054 		return NULL;
1055 
1056 	pe = pnv_ioda_alloc_pe(phb);
1057 	if (!pe) {
1058 		pr_warn("%s: Not enough PE# available, disabling device\n",
1059 			pci_name(dev));
1060 		return NULL;
1061 	}
1062 
1063 	/* NOTE: We get only one ref to the pci_dev for the pdn, not for the
1064 	 * pointer in the PE data structure, both should be destroyed at the
1065 	 * same time. However, this needs to be looked at more closely again
1066 	 * once we actually start removing things (Hotplug, SR-IOV, ...)
1067 	 *
1068 	 * At some point we want to remove the PDN completely anyways
1069 	 */
1070 	pci_dev_get(dev);
1071 	pdn->pe_number = pe->pe_number;
1072 	pe->flags = PNV_IODA_PE_DEV;
1073 	pe->pdev = dev;
1074 	pe->pbus = NULL;
1075 	pe->mve_number = -1;
1076 	pe->rid = dev->bus->number << 8 | pdn->devfn;
1077 
1078 	pe_info(pe, "Associated device to PE\n");
1079 
1080 	if (pnv_ioda_configure_pe(phb, pe)) {
1081 		/* XXX What do we do here ? */
1082 		pnv_ioda_free_pe(pe);
1083 		pdn->pe_number = IODA_INVALID_PE;
1084 		pe->pdev = NULL;
1085 		pci_dev_put(dev);
1086 		return NULL;
1087 	}
1088 
1089 	/* Put PE to the list */
1090 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1091 
1092 	return pe;
1093 }
1094 
1095 static void pnv_ioda_setup_same_PE(struct pci_bus *bus, struct pnv_ioda_pe *pe)
1096 {
1097 	struct pci_dev *dev;
1098 
1099 	list_for_each_entry(dev, &bus->devices, bus_list) {
1100 		struct pci_dn *pdn = pci_get_pdn(dev);
1101 
1102 		if (pdn == NULL) {
1103 			pr_warn("%s: No device node associated with device !\n",
1104 				pci_name(dev));
1105 			continue;
1106 		}
1107 
1108 		/*
1109 		 * In partial hotplug case, the PCI device might be still
1110 		 * associated with the PE and needn't attach it to the PE
1111 		 * again.
1112 		 */
1113 		if (pdn->pe_number != IODA_INVALID_PE)
1114 			continue;
1115 
1116 		pe->device_count++;
1117 		pdn->pe_number = pe->pe_number;
1118 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1119 			pnv_ioda_setup_same_PE(dev->subordinate, pe);
1120 	}
1121 }
1122 
1123 /*
1124  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1125  * single PCI bus. Another one that contains the primary PCI bus and its
1126  * subordinate PCI devices and buses. The second type of PE is normally
1127  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1128  */
1129 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
1130 {
1131 	struct pci_controller *hose = pci_bus_to_host(bus);
1132 	struct pnv_phb *phb = hose->private_data;
1133 	struct pnv_ioda_pe *pe = NULL;
1134 	unsigned int pe_num;
1135 
1136 	/*
1137 	 * In partial hotplug case, the PE instance might be still alive.
1138 	 * We should reuse it instead of allocating a new one.
1139 	 */
1140 	pe_num = phb->ioda.pe_rmap[bus->number << 8];
1141 	if (pe_num != IODA_INVALID_PE) {
1142 		pe = &phb->ioda.pe_array[pe_num];
1143 		pnv_ioda_setup_same_PE(bus, pe);
1144 		return NULL;
1145 	}
1146 
1147 	/* PE number for root bus should have been reserved */
1148 	if (pci_is_root_bus(bus) &&
1149 	    phb->ioda.root_pe_idx != IODA_INVALID_PE)
1150 		pe = &phb->ioda.pe_array[phb->ioda.root_pe_idx];
1151 
1152 	/* Check if PE is determined by M64 */
1153 	if (!pe)
1154 		pe = pnv_ioda_pick_m64_pe(bus, all);
1155 
1156 	/* The PE number isn't pinned by M64 */
1157 	if (!pe)
1158 		pe = pnv_ioda_alloc_pe(phb);
1159 
1160 	if (!pe) {
1161 		pr_warn("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1162 			__func__, pci_domain_nr(bus), bus->number);
1163 		return NULL;
1164 	}
1165 
1166 	pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1167 	pe->pbus = bus;
1168 	pe->pdev = NULL;
1169 	pe->mve_number = -1;
1170 	pe->rid = bus->busn_res.start << 8;
1171 
1172 	if (all)
1173 		pe_info(pe, "Secondary bus %pad..%pad associated with PE#%x\n",
1174 			&bus->busn_res.start, &bus->busn_res.end,
1175 			pe->pe_number);
1176 	else
1177 		pe_info(pe, "Secondary bus %pad associated with PE#%x\n",
1178 			&bus->busn_res.start, pe->pe_number);
1179 
1180 	if (pnv_ioda_configure_pe(phb, pe)) {
1181 		/* XXX What do we do here ? */
1182 		pnv_ioda_free_pe(pe);
1183 		pe->pbus = NULL;
1184 		return NULL;
1185 	}
1186 
1187 	/* Associate it with all child devices */
1188 	pnv_ioda_setup_same_PE(bus, pe);
1189 
1190 	/* Put PE to the list */
1191 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1192 
1193 	return pe;
1194 }
1195 
1196 static struct pnv_ioda_pe *pnv_ioda_setup_npu_PE(struct pci_dev *npu_pdev)
1197 {
1198 	int pe_num, found_pe = false, rc;
1199 	long rid;
1200 	struct pnv_ioda_pe *pe;
1201 	struct pci_dev *gpu_pdev;
1202 	struct pci_dn *npu_pdn;
1203 	struct pci_controller *hose = pci_bus_to_host(npu_pdev->bus);
1204 	struct pnv_phb *phb = hose->private_data;
1205 
1206 	/*
1207 	 * Due to a hardware errata PE#0 on the NPU is reserved for
1208 	 * error handling. This means we only have three PEs remaining
1209 	 * which need to be assigned to four links, implying some
1210 	 * links must share PEs.
1211 	 *
1212 	 * To achieve this we assign PEs such that NPUs linking the
1213 	 * same GPU get assigned the same PE.
1214 	 */
1215 	gpu_pdev = pnv_pci_get_gpu_dev(npu_pdev);
1216 	for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) {
1217 		pe = &phb->ioda.pe_array[pe_num];
1218 		if (!pe->pdev)
1219 			continue;
1220 
1221 		if (pnv_pci_get_gpu_dev(pe->pdev) == gpu_pdev) {
1222 			/*
1223 			 * This device has the same peer GPU so should
1224 			 * be assigned the same PE as the existing
1225 			 * peer NPU.
1226 			 */
1227 			dev_info(&npu_pdev->dev,
1228 				"Associating to existing PE %x\n", pe_num);
1229 			pci_dev_get(npu_pdev);
1230 			npu_pdn = pci_get_pdn(npu_pdev);
1231 			rid = npu_pdev->bus->number << 8 | npu_pdn->devfn;
1232 			npu_pdn->pe_number = pe_num;
1233 			phb->ioda.pe_rmap[rid] = pe->pe_number;
1234 
1235 			/* Map the PE to this link */
1236 			rc = opal_pci_set_pe(phb->opal_id, pe_num, rid,
1237 					OpalPciBusAll,
1238 					OPAL_COMPARE_RID_DEVICE_NUMBER,
1239 					OPAL_COMPARE_RID_FUNCTION_NUMBER,
1240 					OPAL_MAP_PE);
1241 			WARN_ON(rc != OPAL_SUCCESS);
1242 			found_pe = true;
1243 			break;
1244 		}
1245 	}
1246 
1247 	if (!found_pe)
1248 		/*
1249 		 * Could not find an existing PE so allocate a new
1250 		 * one.
1251 		 */
1252 		return pnv_ioda_setup_dev_PE(npu_pdev);
1253 	else
1254 		return pe;
1255 }
1256 
1257 static void pnv_ioda_setup_npu_PEs(struct pci_bus *bus)
1258 {
1259 	struct pci_dev *pdev;
1260 
1261 	list_for_each_entry(pdev, &bus->devices, bus_list)
1262 		pnv_ioda_setup_npu_PE(pdev);
1263 }
1264 
1265 static void pnv_pci_ioda_setup_PEs(void)
1266 {
1267 	struct pci_controller *hose;
1268 	struct pnv_phb *phb;
1269 	struct pci_bus *bus;
1270 	struct pci_dev *pdev;
1271 	struct pnv_ioda_pe *pe;
1272 
1273 	list_for_each_entry(hose, &hose_list, list_node) {
1274 		phb = hose->private_data;
1275 		if (phb->type == PNV_PHB_NPU_NVLINK) {
1276 			/* PE#0 is needed for error reporting */
1277 			pnv_ioda_reserve_pe(phb, 0);
1278 			pnv_ioda_setup_npu_PEs(hose->bus);
1279 			if (phb->model == PNV_PHB_MODEL_NPU2)
1280 				WARN_ON_ONCE(pnv_npu2_init(hose));
1281 		}
1282 		if (phb->type == PNV_PHB_NPU_OCAPI) {
1283 			bus = hose->bus;
1284 			list_for_each_entry(pdev, &bus->devices, bus_list)
1285 				pnv_ioda_setup_dev_PE(pdev);
1286 		}
1287 	}
1288 	list_for_each_entry(hose, &hose_list, list_node) {
1289 		phb = hose->private_data;
1290 		if (phb->type != PNV_PHB_IODA2)
1291 			continue;
1292 
1293 		list_for_each_entry(pe, &phb->ioda.pe_list, list)
1294 			pnv_npu2_map_lpar(pe, MSR_DR | MSR_PR | MSR_HV);
1295 	}
1296 }
1297 
1298 #ifdef CONFIG_PCI_IOV
1299 static int pnv_pci_vf_release_m64(struct pci_dev *pdev, u16 num_vfs)
1300 {
1301 	struct pci_bus        *bus;
1302 	struct pci_controller *hose;
1303 	struct pnv_phb        *phb;
1304 	struct pci_dn         *pdn;
1305 	int                    i, j;
1306 	int                    m64_bars;
1307 
1308 	bus = pdev->bus;
1309 	hose = pci_bus_to_host(bus);
1310 	phb = hose->private_data;
1311 	pdn = pci_get_pdn(pdev);
1312 
1313 	if (pdn->m64_single_mode)
1314 		m64_bars = num_vfs;
1315 	else
1316 		m64_bars = 1;
1317 
1318 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++)
1319 		for (j = 0; j < m64_bars; j++) {
1320 			if (pdn->m64_map[j][i] == IODA_INVALID_M64)
1321 				continue;
1322 			opal_pci_phb_mmio_enable(phb->opal_id,
1323 				OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 0);
1324 			clear_bit(pdn->m64_map[j][i], &phb->ioda.m64_bar_alloc);
1325 			pdn->m64_map[j][i] = IODA_INVALID_M64;
1326 		}
1327 
1328 	kfree(pdn->m64_map);
1329 	return 0;
1330 }
1331 
1332 static int pnv_pci_vf_assign_m64(struct pci_dev *pdev, u16 num_vfs)
1333 {
1334 	struct pci_bus        *bus;
1335 	struct pci_controller *hose;
1336 	struct pnv_phb        *phb;
1337 	struct pci_dn         *pdn;
1338 	unsigned int           win;
1339 	struct resource       *res;
1340 	int                    i, j;
1341 	int64_t                rc;
1342 	int                    total_vfs;
1343 	resource_size_t        size, start;
1344 	int                    pe_num;
1345 	int                    m64_bars;
1346 
1347 	bus = pdev->bus;
1348 	hose = pci_bus_to_host(bus);
1349 	phb = hose->private_data;
1350 	pdn = pci_get_pdn(pdev);
1351 	total_vfs = pci_sriov_get_totalvfs(pdev);
1352 
1353 	if (pdn->m64_single_mode)
1354 		m64_bars = num_vfs;
1355 	else
1356 		m64_bars = 1;
1357 
1358 	pdn->m64_map = kmalloc_array(m64_bars,
1359 				     sizeof(*pdn->m64_map),
1360 				     GFP_KERNEL);
1361 	if (!pdn->m64_map)
1362 		return -ENOMEM;
1363 	/* Initialize the m64_map to IODA_INVALID_M64 */
1364 	for (i = 0; i < m64_bars ; i++)
1365 		for (j = 0; j < PCI_SRIOV_NUM_BARS; j++)
1366 			pdn->m64_map[i][j] = IODA_INVALID_M64;
1367 
1368 
1369 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
1370 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
1371 		if (!res->flags || !res->parent)
1372 			continue;
1373 
1374 		for (j = 0; j < m64_bars; j++) {
1375 			do {
1376 				win = find_next_zero_bit(&phb->ioda.m64_bar_alloc,
1377 						phb->ioda.m64_bar_idx + 1, 0);
1378 
1379 				if (win >= phb->ioda.m64_bar_idx + 1)
1380 					goto m64_failed;
1381 			} while (test_and_set_bit(win, &phb->ioda.m64_bar_alloc));
1382 
1383 			pdn->m64_map[j][i] = win;
1384 
1385 			if (pdn->m64_single_mode) {
1386 				size = pci_iov_resource_size(pdev,
1387 							PCI_IOV_RESOURCES + i);
1388 				start = res->start + size * j;
1389 			} else {
1390 				size = resource_size(res);
1391 				start = res->start;
1392 			}
1393 
1394 			/* Map the M64 here */
1395 			if (pdn->m64_single_mode) {
1396 				pe_num = pdn->pe_num_map[j];
1397 				rc = opal_pci_map_pe_mmio_window(phb->opal_id,
1398 						pe_num, OPAL_M64_WINDOW_TYPE,
1399 						pdn->m64_map[j][i], 0);
1400 			}
1401 
1402 			rc = opal_pci_set_phb_mem_window(phb->opal_id,
1403 						 OPAL_M64_WINDOW_TYPE,
1404 						 pdn->m64_map[j][i],
1405 						 start,
1406 						 0, /* unused */
1407 						 size);
1408 
1409 
1410 			if (rc != OPAL_SUCCESS) {
1411 				dev_err(&pdev->dev, "Failed to map M64 window #%d: %lld\n",
1412 					win, rc);
1413 				goto m64_failed;
1414 			}
1415 
1416 			if (pdn->m64_single_mode)
1417 				rc = opal_pci_phb_mmio_enable(phb->opal_id,
1418 				     OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 2);
1419 			else
1420 				rc = opal_pci_phb_mmio_enable(phb->opal_id,
1421 				     OPAL_M64_WINDOW_TYPE, pdn->m64_map[j][i], 1);
1422 
1423 			if (rc != OPAL_SUCCESS) {
1424 				dev_err(&pdev->dev, "Failed to enable M64 window #%d: %llx\n",
1425 					win, rc);
1426 				goto m64_failed;
1427 			}
1428 		}
1429 	}
1430 	return 0;
1431 
1432 m64_failed:
1433 	pnv_pci_vf_release_m64(pdev, num_vfs);
1434 	return -EBUSY;
1435 }
1436 
1437 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
1438 		int num);
1439 
1440 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev *dev, struct pnv_ioda_pe *pe)
1441 {
1442 	struct iommu_table    *tbl;
1443 	int64_t               rc;
1444 
1445 	tbl = pe->table_group.tables[0];
1446 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
1447 	if (rc)
1448 		pe_warn(pe, "OPAL error %lld release DMA window\n", rc);
1449 
1450 	pnv_pci_ioda2_set_bypass(pe, false);
1451 	if (pe->table_group.group) {
1452 		iommu_group_put(pe->table_group.group);
1453 		BUG_ON(pe->table_group.group);
1454 	}
1455 	iommu_tce_table_put(tbl);
1456 }
1457 
1458 static void pnv_ioda_release_vf_PE(struct pci_dev *pdev)
1459 {
1460 	struct pci_bus        *bus;
1461 	struct pci_controller *hose;
1462 	struct pnv_phb        *phb;
1463 	struct pnv_ioda_pe    *pe, *pe_n;
1464 	struct pci_dn         *pdn;
1465 
1466 	bus = pdev->bus;
1467 	hose = pci_bus_to_host(bus);
1468 	phb = hose->private_data;
1469 	pdn = pci_get_pdn(pdev);
1470 
1471 	if (!pdev->is_physfn)
1472 		return;
1473 
1474 	list_for_each_entry_safe(pe, pe_n, &phb->ioda.pe_list, list) {
1475 		if (pe->parent_dev != pdev)
1476 			continue;
1477 
1478 		pnv_pci_ioda2_release_dma_pe(pdev, pe);
1479 
1480 		/* Remove from list */
1481 		mutex_lock(&phb->ioda.pe_list_mutex);
1482 		list_del(&pe->list);
1483 		mutex_unlock(&phb->ioda.pe_list_mutex);
1484 
1485 		pnv_ioda_deconfigure_pe(phb, pe);
1486 
1487 		pnv_ioda_free_pe(pe);
1488 	}
1489 }
1490 
1491 void pnv_pci_sriov_disable(struct pci_dev *pdev)
1492 {
1493 	struct pci_bus        *bus;
1494 	struct pci_controller *hose;
1495 	struct pnv_phb        *phb;
1496 	struct pnv_ioda_pe    *pe;
1497 	struct pci_dn         *pdn;
1498 	u16                    num_vfs, i;
1499 
1500 	bus = pdev->bus;
1501 	hose = pci_bus_to_host(bus);
1502 	phb = hose->private_data;
1503 	pdn = pci_get_pdn(pdev);
1504 	num_vfs = pdn->num_vfs;
1505 
1506 	/* Release VF PEs */
1507 	pnv_ioda_release_vf_PE(pdev);
1508 
1509 	if (phb->type == PNV_PHB_IODA2) {
1510 		if (!pdn->m64_single_mode)
1511 			pnv_pci_vf_resource_shift(pdev, -*pdn->pe_num_map);
1512 
1513 		/* Release M64 windows */
1514 		pnv_pci_vf_release_m64(pdev, num_vfs);
1515 
1516 		/* Release PE numbers */
1517 		if (pdn->m64_single_mode) {
1518 			for (i = 0; i < num_vfs; i++) {
1519 				if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1520 					continue;
1521 
1522 				pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1523 				pnv_ioda_free_pe(pe);
1524 			}
1525 		} else
1526 			bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1527 		/* Releasing pe_num_map */
1528 		kfree(pdn->pe_num_map);
1529 	}
1530 }
1531 
1532 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1533 				       struct pnv_ioda_pe *pe);
1534 #ifdef CONFIG_IOMMU_API
1535 static void pnv_ioda_setup_bus_iommu_group(struct pnv_ioda_pe *pe,
1536 		struct iommu_table_group *table_group, struct pci_bus *bus);
1537 
1538 #endif
1539 static void pnv_ioda_setup_vf_PE(struct pci_dev *pdev, u16 num_vfs)
1540 {
1541 	struct pci_bus        *bus;
1542 	struct pci_controller *hose;
1543 	struct pnv_phb        *phb;
1544 	struct pnv_ioda_pe    *pe;
1545 	int                    pe_num;
1546 	u16                    vf_index;
1547 	struct pci_dn         *pdn;
1548 
1549 	bus = pdev->bus;
1550 	hose = pci_bus_to_host(bus);
1551 	phb = hose->private_data;
1552 	pdn = pci_get_pdn(pdev);
1553 
1554 	if (!pdev->is_physfn)
1555 		return;
1556 
1557 	/* Reserve PE for each VF */
1558 	for (vf_index = 0; vf_index < num_vfs; vf_index++) {
1559 		if (pdn->m64_single_mode)
1560 			pe_num = pdn->pe_num_map[vf_index];
1561 		else
1562 			pe_num = *pdn->pe_num_map + vf_index;
1563 
1564 		pe = &phb->ioda.pe_array[pe_num];
1565 		pe->pe_number = pe_num;
1566 		pe->phb = phb;
1567 		pe->flags = PNV_IODA_PE_VF;
1568 		pe->pbus = NULL;
1569 		pe->parent_dev = pdev;
1570 		pe->mve_number = -1;
1571 		pe->rid = (pci_iov_virtfn_bus(pdev, vf_index) << 8) |
1572 			   pci_iov_virtfn_devfn(pdev, vf_index);
1573 
1574 		pe_info(pe, "VF %04d:%02d:%02d.%d associated with PE#%x\n",
1575 			hose->global_number, pdev->bus->number,
1576 			PCI_SLOT(pci_iov_virtfn_devfn(pdev, vf_index)),
1577 			PCI_FUNC(pci_iov_virtfn_devfn(pdev, vf_index)), pe_num);
1578 
1579 		if (pnv_ioda_configure_pe(phb, pe)) {
1580 			/* XXX What do we do here ? */
1581 			pnv_ioda_free_pe(pe);
1582 			pe->pdev = NULL;
1583 			continue;
1584 		}
1585 
1586 		/* Put PE to the list */
1587 		mutex_lock(&phb->ioda.pe_list_mutex);
1588 		list_add_tail(&pe->list, &phb->ioda.pe_list);
1589 		mutex_unlock(&phb->ioda.pe_list_mutex);
1590 
1591 		pnv_pci_ioda2_setup_dma_pe(phb, pe);
1592 #ifdef CONFIG_IOMMU_API
1593 		iommu_register_group(&pe->table_group,
1594 				pe->phb->hose->global_number, pe->pe_number);
1595 		pnv_ioda_setup_bus_iommu_group(pe, &pe->table_group, NULL);
1596 #endif
1597 	}
1598 }
1599 
1600 int pnv_pci_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1601 {
1602 	struct pci_bus        *bus;
1603 	struct pci_controller *hose;
1604 	struct pnv_phb        *phb;
1605 	struct pnv_ioda_pe    *pe;
1606 	struct pci_dn         *pdn;
1607 	int                    ret;
1608 	u16                    i;
1609 
1610 	bus = pdev->bus;
1611 	hose = pci_bus_to_host(bus);
1612 	phb = hose->private_data;
1613 	pdn = pci_get_pdn(pdev);
1614 
1615 	if (phb->type == PNV_PHB_IODA2) {
1616 		if (!pdn->vfs_expanded) {
1617 			dev_info(&pdev->dev, "don't support this SRIOV device"
1618 				" with non 64bit-prefetchable IOV BAR\n");
1619 			return -ENOSPC;
1620 		}
1621 
1622 		/*
1623 		 * When M64 BARs functions in Single PE mode, the number of VFs
1624 		 * could be enabled must be less than the number of M64 BARs.
1625 		 */
1626 		if (pdn->m64_single_mode && num_vfs > phb->ioda.m64_bar_idx) {
1627 			dev_info(&pdev->dev, "Not enough M64 BAR for VFs\n");
1628 			return -EBUSY;
1629 		}
1630 
1631 		/* Allocating pe_num_map */
1632 		if (pdn->m64_single_mode)
1633 			pdn->pe_num_map = kmalloc_array(num_vfs,
1634 							sizeof(*pdn->pe_num_map),
1635 							GFP_KERNEL);
1636 		else
1637 			pdn->pe_num_map = kmalloc(sizeof(*pdn->pe_num_map), GFP_KERNEL);
1638 
1639 		if (!pdn->pe_num_map)
1640 			return -ENOMEM;
1641 
1642 		if (pdn->m64_single_mode)
1643 			for (i = 0; i < num_vfs; i++)
1644 				pdn->pe_num_map[i] = IODA_INVALID_PE;
1645 
1646 		/* Calculate available PE for required VFs */
1647 		if (pdn->m64_single_mode) {
1648 			for (i = 0; i < num_vfs; i++) {
1649 				pe = pnv_ioda_alloc_pe(phb);
1650 				if (!pe) {
1651 					ret = -EBUSY;
1652 					goto m64_failed;
1653 				}
1654 
1655 				pdn->pe_num_map[i] = pe->pe_number;
1656 			}
1657 		} else {
1658 			mutex_lock(&phb->ioda.pe_alloc_mutex);
1659 			*pdn->pe_num_map = bitmap_find_next_zero_area(
1660 				phb->ioda.pe_alloc, phb->ioda.total_pe_num,
1661 				0, num_vfs, 0);
1662 			if (*pdn->pe_num_map >= phb->ioda.total_pe_num) {
1663 				mutex_unlock(&phb->ioda.pe_alloc_mutex);
1664 				dev_info(&pdev->dev, "Failed to enable VF%d\n", num_vfs);
1665 				kfree(pdn->pe_num_map);
1666 				return -EBUSY;
1667 			}
1668 			bitmap_set(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1669 			mutex_unlock(&phb->ioda.pe_alloc_mutex);
1670 		}
1671 		pdn->num_vfs = num_vfs;
1672 
1673 		/* Assign M64 window accordingly */
1674 		ret = pnv_pci_vf_assign_m64(pdev, num_vfs);
1675 		if (ret) {
1676 			dev_info(&pdev->dev, "Not enough M64 window resources\n");
1677 			goto m64_failed;
1678 		}
1679 
1680 		/*
1681 		 * When using one M64 BAR to map one IOV BAR, we need to shift
1682 		 * the IOV BAR according to the PE# allocated to the VFs.
1683 		 * Otherwise, the PE# for the VF will conflict with others.
1684 		 */
1685 		if (!pdn->m64_single_mode) {
1686 			ret = pnv_pci_vf_resource_shift(pdev, *pdn->pe_num_map);
1687 			if (ret)
1688 				goto m64_failed;
1689 		}
1690 	}
1691 
1692 	/* Setup VF PEs */
1693 	pnv_ioda_setup_vf_PE(pdev, num_vfs);
1694 
1695 	return 0;
1696 
1697 m64_failed:
1698 	if (pdn->m64_single_mode) {
1699 		for (i = 0; i < num_vfs; i++) {
1700 			if (pdn->pe_num_map[i] == IODA_INVALID_PE)
1701 				continue;
1702 
1703 			pe = &phb->ioda.pe_array[pdn->pe_num_map[i]];
1704 			pnv_ioda_free_pe(pe);
1705 		}
1706 	} else
1707 		bitmap_clear(phb->ioda.pe_alloc, *pdn->pe_num_map, num_vfs);
1708 
1709 	/* Releasing pe_num_map */
1710 	kfree(pdn->pe_num_map);
1711 
1712 	return ret;
1713 }
1714 
1715 int pnv_pcibios_sriov_disable(struct pci_dev *pdev)
1716 {
1717 	pnv_pci_sriov_disable(pdev);
1718 
1719 	/* Release PCI data */
1720 	remove_dev_pci_data(pdev);
1721 	return 0;
1722 }
1723 
1724 int pnv_pcibios_sriov_enable(struct pci_dev *pdev, u16 num_vfs)
1725 {
1726 	/* Allocate PCI data */
1727 	add_dev_pci_data(pdev);
1728 
1729 	return pnv_pci_sriov_enable(pdev, num_vfs);
1730 }
1731 #endif /* CONFIG_PCI_IOV */
1732 
1733 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb *phb, struct pci_dev *pdev)
1734 {
1735 	struct pci_dn *pdn = pci_get_pdn(pdev);
1736 	struct pnv_ioda_pe *pe;
1737 
1738 	/*
1739 	 * The function can be called while the PE#
1740 	 * hasn't been assigned. Do nothing for the
1741 	 * case.
1742 	 */
1743 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
1744 		return;
1745 
1746 	pe = &phb->ioda.pe_array[pdn->pe_number];
1747 	WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1748 	pdev->dev.archdata.dma_offset = pe->tce_bypass_base;
1749 	set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]);
1750 	/*
1751 	 * Note: iommu_add_device() will fail here as
1752 	 * for physical PE: the device is already added by now;
1753 	 * for virtual PE: sysfs entries are not ready yet and
1754 	 * tce_iommu_bus_notifier will add the device to a group later.
1755 	 */
1756 }
1757 
1758 /*
1759  * Reconfigure TVE#0 to be usable as 64-bit DMA space.
1760  *
1761  * The first 4GB of virtual memory for a PE is reserved for 32-bit accesses.
1762  * Devices can only access more than that if bit 59 of the PCI address is set
1763  * by hardware, which indicates TVE#1 should be used instead of TVE#0.
1764  * Many PCI devices are not capable of addressing that many bits, and as a
1765  * result are limited to the 4GB of virtual memory made available to 32-bit
1766  * devices in TVE#0.
1767  *
1768  * In order to work around this, reconfigure TVE#0 to be suitable for 64-bit
1769  * devices by configuring the virtual memory past the first 4GB inaccessible
1770  * by 64-bit DMAs.  This should only be used by devices that want more than
1771  * 4GB, and only on PEs that have no 32-bit devices.
1772  *
1773  * Currently this will only work on PHB3 (POWER8).
1774  */
1775 static int pnv_pci_ioda_dma_64bit_bypass(struct pnv_ioda_pe *pe)
1776 {
1777 	u64 window_size, table_size, tce_count, addr;
1778 	struct page *table_pages;
1779 	u64 tce_order = 28; /* 256MB TCEs */
1780 	__be64 *tces;
1781 	s64 rc;
1782 
1783 	/*
1784 	 * Window size needs to be a power of two, but needs to account for
1785 	 * shifting memory by the 4GB offset required to skip 32bit space.
1786 	 */
1787 	window_size = roundup_pow_of_two(memory_hotplug_max() + (1ULL << 32));
1788 	tce_count = window_size >> tce_order;
1789 	table_size = tce_count << 3;
1790 
1791 	if (table_size < PAGE_SIZE)
1792 		table_size = PAGE_SIZE;
1793 
1794 	table_pages = alloc_pages_node(pe->phb->hose->node, GFP_KERNEL,
1795 				       get_order(table_size));
1796 	if (!table_pages)
1797 		goto err;
1798 
1799 	tces = page_address(table_pages);
1800 	if (!tces)
1801 		goto err;
1802 
1803 	memset(tces, 0, table_size);
1804 
1805 	for (addr = 0; addr < memory_hotplug_max(); addr += (1 << tce_order)) {
1806 		tces[(addr + (1ULL << 32)) >> tce_order] =
1807 			cpu_to_be64(addr | TCE_PCI_READ | TCE_PCI_WRITE);
1808 	}
1809 
1810 	rc = opal_pci_map_pe_dma_window(pe->phb->opal_id,
1811 					pe->pe_number,
1812 					/* reconfigure window 0 */
1813 					(pe->pe_number << 1) + 0,
1814 					1,
1815 					__pa(tces),
1816 					table_size,
1817 					1 << tce_order);
1818 	if (rc == OPAL_SUCCESS) {
1819 		pe_info(pe, "Using 64-bit DMA iommu bypass (through TVE#0)\n");
1820 		return 0;
1821 	}
1822 err:
1823 	pe_err(pe, "Error configuring 64-bit DMA bypass\n");
1824 	return -EIO;
1825 }
1826 
1827 static bool pnv_pci_ioda_iommu_bypass_supported(struct pci_dev *pdev,
1828 		u64 dma_mask)
1829 {
1830 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
1831 	struct pnv_phb *phb = hose->private_data;
1832 	struct pci_dn *pdn = pci_get_pdn(pdev);
1833 	struct pnv_ioda_pe *pe;
1834 
1835 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1836 		return false;
1837 
1838 	pe = &phb->ioda.pe_array[pdn->pe_number];
1839 	if (pe->tce_bypass_enabled) {
1840 		u64 top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1841 		if (dma_mask >= top)
1842 			return true;
1843 	}
1844 
1845 	/*
1846 	 * If the device can't set the TCE bypass bit but still wants
1847 	 * to access 4GB or more, on PHB3 we can reconfigure TVE#0 to
1848 	 * bypass the 32-bit region and be usable for 64-bit DMAs.
1849 	 * The device needs to be able to address all of this space.
1850 	 */
1851 	if (dma_mask >> 32 &&
1852 	    dma_mask > (memory_hotplug_max() + (1ULL << 32)) &&
1853 	    /* pe->pdev should be set if it's a single device, pe->pbus if not */
1854 	    (pe->device_count == 1 || !pe->pbus) &&
1855 	    phb->model == PNV_PHB_MODEL_PHB3) {
1856 		/* Configure the bypass mode */
1857 		s64 rc = pnv_pci_ioda_dma_64bit_bypass(pe);
1858 		if (rc)
1859 			return false;
1860 		/* 4GB offset bypasses 32-bit space */
1861 		pdev->dev.archdata.dma_offset = (1ULL << 32);
1862 		return true;
1863 	}
1864 
1865 	return false;
1866 }
1867 
1868 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe, struct pci_bus *bus)
1869 {
1870 	struct pci_dev *dev;
1871 
1872 	list_for_each_entry(dev, &bus->devices, bus_list) {
1873 		set_iommu_table_base(&dev->dev, pe->table_group.tables[0]);
1874 		dev->dev.archdata.dma_offset = pe->tce_bypass_base;
1875 
1876 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1877 			pnv_ioda_setup_bus_dma(pe, dev->subordinate);
1878 	}
1879 }
1880 
1881 static inline __be64 __iomem *pnv_ioda_get_inval_reg(struct pnv_phb *phb,
1882 						     bool real_mode)
1883 {
1884 	return real_mode ? (__be64 __iomem *)(phb->regs_phys + 0x210) :
1885 		(phb->regs + 0x210);
1886 }
1887 
1888 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table *tbl,
1889 		unsigned long index, unsigned long npages, bool rm)
1890 {
1891 	struct iommu_table_group_link *tgl = list_first_entry_or_null(
1892 			&tbl->it_group_list, struct iommu_table_group_link,
1893 			next);
1894 	struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1895 			struct pnv_ioda_pe, table_group);
1896 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
1897 	unsigned long start, end, inc;
1898 
1899 	start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset);
1900 	end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset +
1901 			npages - 1);
1902 
1903 	/* p7ioc-style invalidation, 2 TCEs per write */
1904 	start |= (1ull << 63);
1905 	end |= (1ull << 63);
1906 	inc = 16;
1907         end |= inc - 1;	/* round up end to be different than start */
1908 
1909         mb(); /* Ensure above stores are visible */
1910         while (start <= end) {
1911 		if (rm)
1912 			__raw_rm_writeq_be(start, invalidate);
1913 		else
1914 			__raw_writeq_be(start, invalidate);
1915 
1916                 start += inc;
1917         }
1918 
1919 	/*
1920 	 * The iommu layer will do another mb() for us on build()
1921 	 * and we don't care on free()
1922 	 */
1923 }
1924 
1925 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index,
1926 		long npages, unsigned long uaddr,
1927 		enum dma_data_direction direction,
1928 		unsigned long attrs)
1929 {
1930 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1931 			attrs);
1932 
1933 	if (!ret)
1934 		pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1935 
1936 	return ret;
1937 }
1938 
1939 #ifdef CONFIG_IOMMU_API
1940 static int pnv_ioda1_tce_xchg(struct iommu_table *tbl, long index,
1941 		unsigned long *hpa, enum dma_data_direction *direction)
1942 {
1943 	long ret = pnv_tce_xchg(tbl, index, hpa, direction, true);
1944 
1945 	if (!ret)
1946 		pnv_pci_p7ioc_tce_invalidate(tbl, index, 1, false);
1947 
1948 	return ret;
1949 }
1950 
1951 static int pnv_ioda1_tce_xchg_rm(struct iommu_table *tbl, long index,
1952 		unsigned long *hpa, enum dma_data_direction *direction)
1953 {
1954 	long ret = pnv_tce_xchg(tbl, index, hpa, direction, false);
1955 
1956 	if (!ret)
1957 		pnv_pci_p7ioc_tce_invalidate(tbl, index, 1, true);
1958 
1959 	return ret;
1960 }
1961 #endif
1962 
1963 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index,
1964 		long npages)
1965 {
1966 	pnv_tce_free(tbl, index, npages);
1967 
1968 	pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1969 }
1970 
1971 static struct iommu_table_ops pnv_ioda1_iommu_ops = {
1972 	.set = pnv_ioda1_tce_build,
1973 #ifdef CONFIG_IOMMU_API
1974 	.exchange = pnv_ioda1_tce_xchg,
1975 	.exchange_rm = pnv_ioda1_tce_xchg_rm,
1976 	.useraddrptr = pnv_tce_useraddrptr,
1977 #endif
1978 	.clear = pnv_ioda1_tce_free,
1979 	.get = pnv_tce_get,
1980 };
1981 
1982 #define PHB3_TCE_KILL_INVAL_ALL		PPC_BIT(0)
1983 #define PHB3_TCE_KILL_INVAL_PE		PPC_BIT(1)
1984 #define PHB3_TCE_KILL_INVAL_ONE		PPC_BIT(2)
1985 
1986 static void pnv_pci_phb3_tce_invalidate_entire(struct pnv_phb *phb, bool rm)
1987 {
1988 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(phb, rm);
1989 	const unsigned long val = PHB3_TCE_KILL_INVAL_ALL;
1990 
1991 	mb(); /* Ensure previous TCE table stores are visible */
1992 	if (rm)
1993 		__raw_rm_writeq_be(val, invalidate);
1994 	else
1995 		__raw_writeq_be(val, invalidate);
1996 }
1997 
1998 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1999 {
2000 	/* 01xb - invalidate TCEs that match the specified PE# */
2001 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, false);
2002 	unsigned long val = PHB3_TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
2003 
2004 	mb(); /* Ensure above stores are visible */
2005 	__raw_writeq_be(val, invalidate);
2006 }
2007 
2008 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe *pe, bool rm,
2009 					unsigned shift, unsigned long index,
2010 					unsigned long npages)
2011 {
2012 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
2013 	unsigned long start, end, inc;
2014 
2015 	/* We'll invalidate DMA address in PE scope */
2016 	start = PHB3_TCE_KILL_INVAL_ONE;
2017 	start |= (pe->pe_number & 0xFF);
2018 	end = start;
2019 
2020 	/* Figure out the start, end and step */
2021 	start |= (index << shift);
2022 	end |= ((index + npages - 1) << shift);
2023 	inc = (0x1ull << shift);
2024 	mb();
2025 
2026 	while (start <= end) {
2027 		if (rm)
2028 			__raw_rm_writeq_be(start, invalidate);
2029 		else
2030 			__raw_writeq_be(start, invalidate);
2031 		start += inc;
2032 	}
2033 }
2034 
2035 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
2036 {
2037 	struct pnv_phb *phb = pe->phb;
2038 
2039 	if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
2040 		pnv_pci_phb3_tce_invalidate_pe(pe);
2041 	else
2042 		opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL_PE,
2043 				  pe->pe_number, 0, 0, 0);
2044 }
2045 
2046 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
2047 		unsigned long index, unsigned long npages, bool rm)
2048 {
2049 	struct iommu_table_group_link *tgl;
2050 
2051 	list_for_each_entry_lockless(tgl, &tbl->it_group_list, next) {
2052 		struct pnv_ioda_pe *pe = container_of(tgl->table_group,
2053 				struct pnv_ioda_pe, table_group);
2054 		struct pnv_phb *phb = pe->phb;
2055 		unsigned int shift = tbl->it_page_shift;
2056 
2057 		/*
2058 		 * NVLink1 can use the TCE kill register directly as
2059 		 * it's the same as PHB3. NVLink2 is different and
2060 		 * should go via the OPAL call.
2061 		 */
2062 		if (phb->model == PNV_PHB_MODEL_NPU) {
2063 			/*
2064 			 * The NVLink hardware does not support TCE kill
2065 			 * per TCE entry so we have to invalidate
2066 			 * the entire cache for it.
2067 			 */
2068 			pnv_pci_phb3_tce_invalidate_entire(phb, rm);
2069 			continue;
2070 		}
2071 		if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
2072 			pnv_pci_phb3_tce_invalidate(pe, rm, shift,
2073 						    index, npages);
2074 		else
2075 			opal_pci_tce_kill(phb->opal_id,
2076 					  OPAL_PCI_TCE_KILL_PAGES,
2077 					  pe->pe_number, 1u << shift,
2078 					  index << shift, npages);
2079 	}
2080 }
2081 
2082 void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_phb *phb, bool rm)
2083 {
2084 	if (phb->model == PNV_PHB_MODEL_NPU || phb->model == PNV_PHB_MODEL_PHB3)
2085 		pnv_pci_phb3_tce_invalidate_entire(phb, rm);
2086 	else
2087 		opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL, 0, 0, 0, 0);
2088 }
2089 
2090 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index,
2091 		long npages, unsigned long uaddr,
2092 		enum dma_data_direction direction,
2093 		unsigned long attrs)
2094 {
2095 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
2096 			attrs);
2097 
2098 	if (!ret)
2099 		pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
2100 
2101 	return ret;
2102 }
2103 
2104 #ifdef CONFIG_IOMMU_API
2105 static int pnv_ioda2_tce_xchg(struct iommu_table *tbl, long index,
2106 		unsigned long *hpa, enum dma_data_direction *direction)
2107 {
2108 	long ret = pnv_tce_xchg(tbl, index, hpa, direction, true);
2109 
2110 	if (!ret)
2111 		pnv_pci_ioda2_tce_invalidate(tbl, index, 1, false);
2112 
2113 	return ret;
2114 }
2115 
2116 static int pnv_ioda2_tce_xchg_rm(struct iommu_table *tbl, long index,
2117 		unsigned long *hpa, enum dma_data_direction *direction)
2118 {
2119 	long ret = pnv_tce_xchg(tbl, index, hpa, direction, false);
2120 
2121 	if (!ret)
2122 		pnv_pci_ioda2_tce_invalidate(tbl, index, 1, true);
2123 
2124 	return ret;
2125 }
2126 #endif
2127 
2128 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index,
2129 		long npages)
2130 {
2131 	pnv_tce_free(tbl, index, npages);
2132 
2133 	pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
2134 }
2135 
2136 static struct iommu_table_ops pnv_ioda2_iommu_ops = {
2137 	.set = pnv_ioda2_tce_build,
2138 #ifdef CONFIG_IOMMU_API
2139 	.exchange = pnv_ioda2_tce_xchg,
2140 	.exchange_rm = pnv_ioda2_tce_xchg_rm,
2141 	.useraddrptr = pnv_tce_useraddrptr,
2142 #endif
2143 	.clear = pnv_ioda2_tce_free,
2144 	.get = pnv_tce_get,
2145 	.free = pnv_pci_ioda2_table_free_pages,
2146 };
2147 
2148 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data)
2149 {
2150 	unsigned int *weight = (unsigned int *)data;
2151 
2152 	/* This is quite simplistic. The "base" weight of a device
2153 	 * is 10. 0 means no DMA is to be accounted for it.
2154 	 */
2155 	if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
2156 		return 0;
2157 
2158 	if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
2159 	    dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
2160 	    dev->class == PCI_CLASS_SERIAL_USB_EHCI)
2161 		*weight += 3;
2162 	else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
2163 		*weight += 15;
2164 	else
2165 		*weight += 10;
2166 
2167 	return 0;
2168 }
2169 
2170 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe)
2171 {
2172 	unsigned int weight = 0;
2173 
2174 	/* SRIOV VF has same DMA32 weight as its PF */
2175 #ifdef CONFIG_PCI_IOV
2176 	if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) {
2177 		pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight);
2178 		return weight;
2179 	}
2180 #endif
2181 
2182 	if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) {
2183 		pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight);
2184 	} else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) {
2185 		struct pci_dev *pdev;
2186 
2187 		list_for_each_entry(pdev, &pe->pbus->devices, bus_list)
2188 			pnv_pci_ioda_dev_dma_weight(pdev, &weight);
2189 	} else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) {
2190 		pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight);
2191 	}
2192 
2193 	return weight;
2194 }
2195 
2196 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
2197 				       struct pnv_ioda_pe *pe)
2198 {
2199 
2200 	struct page *tce_mem = NULL;
2201 	struct iommu_table *tbl;
2202 	unsigned int weight, total_weight = 0;
2203 	unsigned int tce32_segsz, base, segs, avail, i;
2204 	int64_t rc;
2205 	void *addr;
2206 
2207 	/* XXX FIXME: Handle 64-bit only DMA devices */
2208 	/* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
2209 	/* XXX FIXME: Allocate multi-level tables on PHB3 */
2210 	weight = pnv_pci_ioda_pe_dma_weight(pe);
2211 	if (!weight)
2212 		return;
2213 
2214 	pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight,
2215 		     &total_weight);
2216 	segs = (weight * phb->ioda.dma32_count) / total_weight;
2217 	if (!segs)
2218 		segs = 1;
2219 
2220 	/*
2221 	 * Allocate contiguous DMA32 segments. We begin with the expected
2222 	 * number of segments. With one more attempt, the number of DMA32
2223 	 * segments to be allocated is decreased by one until one segment
2224 	 * is allocated successfully.
2225 	 */
2226 	do {
2227 		for (base = 0; base <= phb->ioda.dma32_count - segs; base++) {
2228 			for (avail = 0, i = base; i < base + segs; i++) {
2229 				if (phb->ioda.dma32_segmap[i] ==
2230 				    IODA_INVALID_PE)
2231 					avail++;
2232 			}
2233 
2234 			if (avail == segs)
2235 				goto found;
2236 		}
2237 	} while (--segs);
2238 
2239 	if (!segs) {
2240 		pe_warn(pe, "No available DMA32 segments\n");
2241 		return;
2242 	}
2243 
2244 found:
2245 	tbl = pnv_pci_table_alloc(phb->hose->node);
2246 	if (WARN_ON(!tbl))
2247 		return;
2248 
2249 	iommu_register_group(&pe->table_group, phb->hose->global_number,
2250 			pe->pe_number);
2251 	pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group);
2252 
2253 	/* Grab a 32-bit TCE table */
2254 	pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
2255 		weight, total_weight, base, segs);
2256 	pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
2257 		base * PNV_IODA1_DMA32_SEGSIZE,
2258 		(base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1);
2259 
2260 	/* XXX Currently, we allocate one big contiguous table for the
2261 	 * TCEs. We only really need one chunk per 256M of TCE space
2262 	 * (ie per segment) but that's an optimization for later, it
2263 	 * requires some added smarts with our get/put_tce implementation
2264 	 *
2265 	 * Each TCE page is 4KB in size and each TCE entry occupies 8
2266 	 * bytes
2267 	 */
2268 	tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3);
2269 	tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
2270 				   get_order(tce32_segsz * segs));
2271 	if (!tce_mem) {
2272 		pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
2273 		goto fail;
2274 	}
2275 	addr = page_address(tce_mem);
2276 	memset(addr, 0, tce32_segsz * segs);
2277 
2278 	/* Configure HW */
2279 	for (i = 0; i < segs; i++) {
2280 		rc = opal_pci_map_pe_dma_window(phb->opal_id,
2281 					      pe->pe_number,
2282 					      base + i, 1,
2283 					      __pa(addr) + tce32_segsz * i,
2284 					      tce32_segsz, IOMMU_PAGE_SIZE_4K);
2285 		if (rc) {
2286 			pe_err(pe, " Failed to configure 32-bit TCE table, err %lld\n",
2287 			       rc);
2288 			goto fail;
2289 		}
2290 	}
2291 
2292 	/* Setup DMA32 segment mapping */
2293 	for (i = base; i < base + segs; i++)
2294 		phb->ioda.dma32_segmap[i] = pe->pe_number;
2295 
2296 	/* Setup linux iommu table */
2297 	pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs,
2298 				  base * PNV_IODA1_DMA32_SEGSIZE,
2299 				  IOMMU_PAGE_SHIFT_4K);
2300 
2301 	tbl->it_ops = &pnv_ioda1_iommu_ops;
2302 	pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
2303 	pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
2304 	iommu_init_table(tbl, phb->hose->node);
2305 
2306 	if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2307 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
2308 
2309 	return;
2310  fail:
2311 	/* XXX Failure: Try to fallback to 64-bit only ? */
2312 	if (tce_mem)
2313 		__free_pages(tce_mem, get_order(tce32_segsz * segs));
2314 	if (tbl) {
2315 		pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
2316 		iommu_tce_table_put(tbl);
2317 	}
2318 }
2319 
2320 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
2321 		int num, struct iommu_table *tbl)
2322 {
2323 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2324 			table_group);
2325 	struct pnv_phb *phb = pe->phb;
2326 	int64_t rc;
2327 	const unsigned long size = tbl->it_indirect_levels ?
2328 			tbl->it_level_size : tbl->it_size;
2329 	const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
2330 	const __u64 win_size = tbl->it_size << tbl->it_page_shift;
2331 
2332 	pe_info(pe, "Setting up window#%d %llx..%llx pg=%lx\n",
2333 		num, start_addr, start_addr + win_size - 1,
2334 		IOMMU_PAGE_SIZE(tbl));
2335 
2336 	/*
2337 	 * Map TCE table through TVT. The TVE index is the PE number
2338 	 * shifted by 1 bit for 32-bits DMA space.
2339 	 */
2340 	rc = opal_pci_map_pe_dma_window(phb->opal_id,
2341 			pe->pe_number,
2342 			(pe->pe_number << 1) + num,
2343 			tbl->it_indirect_levels + 1,
2344 			__pa(tbl->it_base),
2345 			size << 3,
2346 			IOMMU_PAGE_SIZE(tbl));
2347 	if (rc) {
2348 		pe_err(pe, "Failed to configure TCE table, err %lld\n", rc);
2349 		return rc;
2350 	}
2351 
2352 	pnv_pci_link_table_and_group(phb->hose->node, num,
2353 			tbl, &pe->table_group);
2354 	pnv_pci_ioda2_tce_invalidate_pe(pe);
2355 
2356 	return 0;
2357 }
2358 
2359 void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
2360 {
2361 	uint16_t window_id = (pe->pe_number << 1 ) + 1;
2362 	int64_t rc;
2363 
2364 	pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
2365 	if (enable) {
2366 		phys_addr_t top = memblock_end_of_DRAM();
2367 
2368 		top = roundup_pow_of_two(top);
2369 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2370 						     pe->pe_number,
2371 						     window_id,
2372 						     pe->tce_bypass_base,
2373 						     top);
2374 	} else {
2375 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
2376 						     pe->pe_number,
2377 						     window_id,
2378 						     pe->tce_bypass_base,
2379 						     0);
2380 	}
2381 	if (rc)
2382 		pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
2383 	else
2384 		pe->tce_bypass_enabled = enable;
2385 }
2386 
2387 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group,
2388 		int num, __u32 page_shift, __u64 window_size, __u32 levels,
2389 		bool alloc_userspace_copy, struct iommu_table **ptbl)
2390 {
2391 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2392 			table_group);
2393 	int nid = pe->phb->hose->node;
2394 	__u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start;
2395 	long ret;
2396 	struct iommu_table *tbl;
2397 
2398 	tbl = pnv_pci_table_alloc(nid);
2399 	if (!tbl)
2400 		return -ENOMEM;
2401 
2402 	tbl->it_ops = &pnv_ioda2_iommu_ops;
2403 
2404 	ret = pnv_pci_ioda2_table_alloc_pages(nid,
2405 			bus_offset, page_shift, window_size,
2406 			levels, alloc_userspace_copy, tbl);
2407 	if (ret) {
2408 		iommu_tce_table_put(tbl);
2409 		return ret;
2410 	}
2411 
2412 	*ptbl = tbl;
2413 
2414 	return 0;
2415 }
2416 
2417 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe)
2418 {
2419 	struct iommu_table *tbl = NULL;
2420 	long rc;
2421 
2422 	/*
2423 	 * crashkernel= specifies the kdump kernel's maximum memory at
2424 	 * some offset and there is no guaranteed the result is a power
2425 	 * of 2, which will cause errors later.
2426 	 */
2427 	const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max());
2428 
2429 	/*
2430 	 * In memory constrained environments, e.g. kdump kernel, the
2431 	 * DMA window can be larger than available memory, which will
2432 	 * cause errors later.
2433 	 */
2434 	const u64 window_size = min((u64)pe->table_group.tce32_size, max_memory);
2435 
2436 	rc = pnv_pci_ioda2_create_table(&pe->table_group, 0,
2437 			IOMMU_PAGE_SHIFT_4K,
2438 			window_size,
2439 			POWERNV_IOMMU_DEFAULT_LEVELS, false, &tbl);
2440 	if (rc) {
2441 		pe_err(pe, "Failed to create 32-bit TCE table, err %ld",
2442 				rc);
2443 		return rc;
2444 	}
2445 
2446 	iommu_init_table(tbl, pe->phb->hose->node);
2447 
2448 	rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl);
2449 	if (rc) {
2450 		pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n",
2451 				rc);
2452 		iommu_tce_table_put(tbl);
2453 		return rc;
2454 	}
2455 
2456 	if (!pnv_iommu_bypass_disabled)
2457 		pnv_pci_ioda2_set_bypass(pe, true);
2458 
2459 	return 0;
2460 }
2461 
2462 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV)
2463 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
2464 		int num)
2465 {
2466 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2467 			table_group);
2468 	struct pnv_phb *phb = pe->phb;
2469 	long ret;
2470 
2471 	pe_info(pe, "Removing DMA window #%d\n", num);
2472 
2473 	ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
2474 			(pe->pe_number << 1) + num,
2475 			0/* levels */, 0/* table address */,
2476 			0/* table size */, 0/* page size */);
2477 	if (ret)
2478 		pe_warn(pe, "Unmapping failed, ret = %ld\n", ret);
2479 	else
2480 		pnv_pci_ioda2_tce_invalidate_pe(pe);
2481 
2482 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
2483 
2484 	return ret;
2485 }
2486 #endif
2487 
2488 #ifdef CONFIG_IOMMU_API
2489 unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift,
2490 		__u64 window_size, __u32 levels)
2491 {
2492 	unsigned long bytes = 0;
2493 	const unsigned window_shift = ilog2(window_size);
2494 	unsigned entries_shift = window_shift - page_shift;
2495 	unsigned table_shift = entries_shift + 3;
2496 	unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift);
2497 	unsigned long direct_table_size;
2498 
2499 	if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) ||
2500 			!is_power_of_2(window_size))
2501 		return 0;
2502 
2503 	/* Calculate a direct table size from window_size and levels */
2504 	entries_shift = (entries_shift + levels - 1) / levels;
2505 	table_shift = entries_shift + 3;
2506 	table_shift = max_t(unsigned, table_shift, PAGE_SHIFT);
2507 	direct_table_size =  1UL << table_shift;
2508 
2509 	for ( ; levels; --levels) {
2510 		bytes += _ALIGN_UP(tce_table_size, direct_table_size);
2511 
2512 		tce_table_size /= direct_table_size;
2513 		tce_table_size <<= 3;
2514 		tce_table_size = max_t(unsigned long,
2515 				tce_table_size, direct_table_size);
2516 	}
2517 
2518 	return bytes + bytes; /* one for HW table, one for userspace copy */
2519 }
2520 
2521 static long pnv_pci_ioda2_create_table_userspace(
2522 		struct iommu_table_group *table_group,
2523 		int num, __u32 page_shift, __u64 window_size, __u32 levels,
2524 		struct iommu_table **ptbl)
2525 {
2526 	long ret = pnv_pci_ioda2_create_table(table_group,
2527 			num, page_shift, window_size, levels, true, ptbl);
2528 
2529 	if (!ret)
2530 		(*ptbl)->it_allocated_size = pnv_pci_ioda2_get_table_size(
2531 				page_shift, window_size, levels);
2532 	return ret;
2533 }
2534 
2535 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group)
2536 {
2537 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2538 						table_group);
2539 	/* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
2540 	struct iommu_table *tbl = pe->table_group.tables[0];
2541 
2542 	pnv_pci_ioda2_set_bypass(pe, false);
2543 	pnv_pci_ioda2_unset_window(&pe->table_group, 0);
2544 	if (pe->pbus)
2545 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
2546 	iommu_tce_table_put(tbl);
2547 }
2548 
2549 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group)
2550 {
2551 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
2552 						table_group);
2553 
2554 	pnv_pci_ioda2_setup_default_config(pe);
2555 	if (pe->pbus)
2556 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
2557 }
2558 
2559 static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
2560 	.get_table_size = pnv_pci_ioda2_get_table_size,
2561 	.create_table = pnv_pci_ioda2_create_table_userspace,
2562 	.set_window = pnv_pci_ioda2_set_window,
2563 	.unset_window = pnv_pci_ioda2_unset_window,
2564 	.take_ownership = pnv_ioda2_take_ownership,
2565 	.release_ownership = pnv_ioda2_release_ownership,
2566 };
2567 
2568 static void pnv_ioda_setup_bus_iommu_group_add_devices(struct pnv_ioda_pe *pe,
2569 		struct iommu_table_group *table_group,
2570 		struct pci_bus *bus)
2571 {
2572 	struct pci_dev *dev;
2573 
2574 	list_for_each_entry(dev, &bus->devices, bus_list) {
2575 		iommu_add_device(table_group, &dev->dev);
2576 
2577 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
2578 			pnv_ioda_setup_bus_iommu_group_add_devices(pe,
2579 					table_group, dev->subordinate);
2580 	}
2581 }
2582 
2583 static void pnv_ioda_setup_bus_iommu_group(struct pnv_ioda_pe *pe,
2584 		struct iommu_table_group *table_group, struct pci_bus *bus)
2585 {
2586 
2587 	if (pe->flags & PNV_IODA_PE_DEV)
2588 		iommu_add_device(table_group, &pe->pdev->dev);
2589 
2590 	if ((pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)) || bus)
2591 		pnv_ioda_setup_bus_iommu_group_add_devices(pe, table_group,
2592 				bus);
2593 }
2594 
2595 static unsigned long pnv_ioda_parse_tce_sizes(struct pnv_phb *phb);
2596 
2597 static void pnv_pci_ioda_setup_iommu_api(void)
2598 {
2599 	struct pci_controller *hose;
2600 	struct pnv_phb *phb;
2601 	struct pnv_ioda_pe *pe;
2602 
2603 	/*
2604 	 * There are 4 types of PEs:
2605 	 * - PNV_IODA_PE_BUS: a downstream port with an adapter,
2606 	 *   created from pnv_pci_setup_bridge();
2607 	 * - PNV_IODA_PE_BUS_ALL: a PCI-PCIX bridge with devices behind it,
2608 	 *   created from pnv_pci_setup_bridge();
2609 	 * - PNV_IODA_PE_VF: a SRIOV virtual function,
2610 	 *   created from pnv_pcibios_sriov_enable();
2611 	 * - PNV_IODA_PE_DEV: an NPU or OCAPI device,
2612 	 *   created from pnv_pci_ioda_fixup().
2613 	 *
2614 	 * Normally a PE is represented by an IOMMU group, however for
2615 	 * devices with side channels the groups need to be more strict.
2616 	 */
2617 	list_for_each_entry(hose, &hose_list, list_node) {
2618 		phb = hose->private_data;
2619 
2620 		if (phb->type == PNV_PHB_NPU_NVLINK ||
2621 		    phb->type == PNV_PHB_NPU_OCAPI)
2622 			continue;
2623 
2624 		list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2625 			struct iommu_table_group *table_group;
2626 
2627 			table_group = pnv_try_setup_npu_table_group(pe);
2628 			if (!table_group) {
2629 				if (!pnv_pci_ioda_pe_dma_weight(pe))
2630 					continue;
2631 
2632 				table_group = &pe->table_group;
2633 				iommu_register_group(&pe->table_group,
2634 						pe->phb->hose->global_number,
2635 						pe->pe_number);
2636 			}
2637 			pnv_ioda_setup_bus_iommu_group(pe, table_group,
2638 					pe->pbus);
2639 		}
2640 	}
2641 
2642 	/*
2643 	 * Now we have all PHBs discovered, time to add NPU devices to
2644 	 * the corresponding IOMMU groups.
2645 	 */
2646 	list_for_each_entry(hose, &hose_list, list_node) {
2647 		unsigned long  pgsizes;
2648 
2649 		phb = hose->private_data;
2650 
2651 		if (phb->type != PNV_PHB_NPU_NVLINK)
2652 			continue;
2653 
2654 		pgsizes = pnv_ioda_parse_tce_sizes(phb);
2655 		list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2656 			/*
2657 			 * IODA2 bridges get this set up from
2658 			 * pci_controller_ops::setup_bridge but NPU bridges
2659 			 * do not have this hook defined so we do it here.
2660 			 */
2661 			pe->table_group.pgsizes = pgsizes;
2662 			pnv_npu_compound_attach(pe);
2663 		}
2664 	}
2665 }
2666 #else /* !CONFIG_IOMMU_API */
2667 static void pnv_pci_ioda_setup_iommu_api(void) { };
2668 #endif
2669 
2670 static unsigned long pnv_ioda_parse_tce_sizes(struct pnv_phb *phb)
2671 {
2672 	struct pci_controller *hose = phb->hose;
2673 	struct device_node *dn = hose->dn;
2674 	unsigned long mask = 0;
2675 	int i, rc, count;
2676 	u32 val;
2677 
2678 	count = of_property_count_u32_elems(dn, "ibm,supported-tce-sizes");
2679 	if (count <= 0) {
2680 		mask = SZ_4K | SZ_64K;
2681 		/* Add 16M for POWER8 by default */
2682 		if (cpu_has_feature(CPU_FTR_ARCH_207S) &&
2683 				!cpu_has_feature(CPU_FTR_ARCH_300))
2684 			mask |= SZ_16M | SZ_256M;
2685 		return mask;
2686 	}
2687 
2688 	for (i = 0; i < count; i++) {
2689 		rc = of_property_read_u32_index(dn, "ibm,supported-tce-sizes",
2690 						i, &val);
2691 		if (rc == 0)
2692 			mask |= 1ULL << val;
2693 	}
2694 
2695 	return mask;
2696 }
2697 
2698 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
2699 				       struct pnv_ioda_pe *pe)
2700 {
2701 	int64_t rc;
2702 
2703 	if (!pnv_pci_ioda_pe_dma_weight(pe))
2704 		return;
2705 
2706 	/* TVE #1 is selected by PCI address bit 59 */
2707 	pe->tce_bypass_base = 1ull << 59;
2708 
2709 	/* The PE will reserve all possible 32-bits space */
2710 	pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
2711 		phb->ioda.m32_pci_base);
2712 
2713 	/* Setup linux iommu table */
2714 	pe->table_group.tce32_start = 0;
2715 	pe->table_group.tce32_size = phb->ioda.m32_pci_base;
2716 	pe->table_group.max_dynamic_windows_supported =
2717 			IOMMU_TABLE_GROUP_MAX_TABLES;
2718 	pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS;
2719 	pe->table_group.pgsizes = pnv_ioda_parse_tce_sizes(phb);
2720 #ifdef CONFIG_IOMMU_API
2721 	pe->table_group.ops = &pnv_pci_ioda2_ops;
2722 #endif
2723 
2724 	rc = pnv_pci_ioda2_setup_default_config(pe);
2725 	if (rc)
2726 		return;
2727 
2728 	if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
2729 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
2730 }
2731 
2732 int64_t pnv_opal_pci_msi_eoi(struct irq_chip *chip, unsigned int hw_irq)
2733 {
2734 	struct pnv_phb *phb = container_of(chip, struct pnv_phb,
2735 					   ioda.irq_chip);
2736 
2737 	return opal_pci_msi_eoi(phb->opal_id, hw_irq);
2738 }
2739 
2740 static void pnv_ioda2_msi_eoi(struct irq_data *d)
2741 {
2742 	int64_t rc;
2743 	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
2744 	struct irq_chip *chip = irq_data_get_irq_chip(d);
2745 
2746 	rc = pnv_opal_pci_msi_eoi(chip, hw_irq);
2747 	WARN_ON_ONCE(rc);
2748 
2749 	icp_native_eoi(d);
2750 }
2751 
2752 
2753 void pnv_set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2754 {
2755 	struct irq_data *idata;
2756 	struct irq_chip *ichip;
2757 
2758 	/* The MSI EOI OPAL call is only needed on PHB3 */
2759 	if (phb->model != PNV_PHB_MODEL_PHB3)
2760 		return;
2761 
2762 	if (!phb->ioda.irq_chip_init) {
2763 		/*
2764 		 * First time we setup an MSI IRQ, we need to setup the
2765 		 * corresponding IRQ chip to route correctly.
2766 		 */
2767 		idata = irq_get_irq_data(virq);
2768 		ichip = irq_data_get_irq_chip(idata);
2769 		phb->ioda.irq_chip_init = 1;
2770 		phb->ioda.irq_chip = *ichip;
2771 		phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2772 	}
2773 	irq_set_chip(virq, &phb->ioda.irq_chip);
2774 }
2775 
2776 /*
2777  * Returns true iff chip is something that we could call
2778  * pnv_opal_pci_msi_eoi for.
2779  */
2780 bool is_pnv_opal_msi(struct irq_chip *chip)
2781 {
2782 	return chip->irq_eoi == pnv_ioda2_msi_eoi;
2783 }
2784 EXPORT_SYMBOL_GPL(is_pnv_opal_msi);
2785 
2786 static int pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2787 				  unsigned int hwirq, unsigned int virq,
2788 				  unsigned int is_64, struct msi_msg *msg)
2789 {
2790 	struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2791 	unsigned int xive_num = hwirq - phb->msi_base;
2792 	__be32 data;
2793 	int rc;
2794 
2795 	/* No PE assigned ? bail out ... no MSI for you ! */
2796 	if (pe == NULL)
2797 		return -ENXIO;
2798 
2799 	/* Check if we have an MVE */
2800 	if (pe->mve_number < 0)
2801 		return -ENXIO;
2802 
2803 	/* Force 32-bit MSI on some broken devices */
2804 	if (dev->no_64bit_msi)
2805 		is_64 = 0;
2806 
2807 	/* Assign XIVE to PE */
2808 	rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2809 	if (rc) {
2810 		pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2811 			pci_name(dev), rc, xive_num);
2812 		return -EIO;
2813 	}
2814 
2815 	if (is_64) {
2816 		__be64 addr64;
2817 
2818 		rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
2819 				     &addr64, &data);
2820 		if (rc) {
2821 			pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2822 				pci_name(dev), rc);
2823 			return -EIO;
2824 		}
2825 		msg->address_hi = be64_to_cpu(addr64) >> 32;
2826 		msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
2827 	} else {
2828 		__be32 addr32;
2829 
2830 		rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
2831 				     &addr32, &data);
2832 		if (rc) {
2833 			pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2834 				pci_name(dev), rc);
2835 			return -EIO;
2836 		}
2837 		msg->address_hi = 0;
2838 		msg->address_lo = be32_to_cpu(addr32);
2839 	}
2840 	msg->data = be32_to_cpu(data);
2841 
2842 	pnv_set_msi_irq_chip(phb, virq);
2843 
2844 	pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
2845 		 " address=%x_%08x data=%x PE# %x\n",
2846 		 pci_name(dev), is_64 ? "64" : "32", hwirq, xive_num,
2847 		 msg->address_hi, msg->address_lo, data, pe->pe_number);
2848 
2849 	return 0;
2850 }
2851 
2852 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
2853 {
2854 	unsigned int count;
2855 	const __be32 *prop = of_get_property(phb->hose->dn,
2856 					     "ibm,opal-msi-ranges", NULL);
2857 	if (!prop) {
2858 		/* BML Fallback */
2859 		prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
2860 	}
2861 	if (!prop)
2862 		return;
2863 
2864 	phb->msi_base = be32_to_cpup(prop);
2865 	count = be32_to_cpup(prop + 1);
2866 	if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
2867 		pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2868 		       phb->hose->global_number);
2869 		return;
2870 	}
2871 
2872 	phb->msi_setup = pnv_pci_ioda_msi_setup;
2873 	phb->msi32_support = 1;
2874 	pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2875 		count, phb->msi_base);
2876 }
2877 
2878 #ifdef CONFIG_PCI_IOV
2879 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev *pdev)
2880 {
2881 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
2882 	struct pnv_phb *phb = hose->private_data;
2883 	const resource_size_t gate = phb->ioda.m64_segsize >> 2;
2884 	struct resource *res;
2885 	int i;
2886 	resource_size_t size, total_vf_bar_sz;
2887 	struct pci_dn *pdn;
2888 	int mul, total_vfs;
2889 
2890 	if (!pdev->is_physfn || pci_dev_is_added(pdev))
2891 		return;
2892 
2893 	pdn = pci_get_pdn(pdev);
2894 	pdn->vfs_expanded = 0;
2895 	pdn->m64_single_mode = false;
2896 
2897 	total_vfs = pci_sriov_get_totalvfs(pdev);
2898 	mul = phb->ioda.total_pe_num;
2899 	total_vf_bar_sz = 0;
2900 
2901 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2902 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
2903 		if (!res->flags || res->parent)
2904 			continue;
2905 		if (!pnv_pci_is_m64_flags(res->flags)) {
2906 			dev_warn(&pdev->dev, "Don't support SR-IOV with"
2907 					" non M64 VF BAR%d: %pR. \n",
2908 				 i, res);
2909 			goto truncate_iov;
2910 		}
2911 
2912 		total_vf_bar_sz += pci_iov_resource_size(pdev,
2913 				i + PCI_IOV_RESOURCES);
2914 
2915 		/*
2916 		 * If bigger than quarter of M64 segment size, just round up
2917 		 * power of two.
2918 		 *
2919 		 * Generally, one M64 BAR maps one IOV BAR. To avoid conflict
2920 		 * with other devices, IOV BAR size is expanded to be
2921 		 * (total_pe * VF_BAR_size).  When VF_BAR_size is half of M64
2922 		 * segment size , the expanded size would equal to half of the
2923 		 * whole M64 space size, which will exhaust the M64 Space and
2924 		 * limit the system flexibility.  This is a design decision to
2925 		 * set the boundary to quarter of the M64 segment size.
2926 		 */
2927 		if (total_vf_bar_sz > gate) {
2928 			mul = roundup_pow_of_two(total_vfs);
2929 			dev_info(&pdev->dev,
2930 				"VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n",
2931 				total_vf_bar_sz, gate, mul);
2932 			pdn->m64_single_mode = true;
2933 			break;
2934 		}
2935 	}
2936 
2937 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2938 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
2939 		if (!res->flags || res->parent)
2940 			continue;
2941 
2942 		size = pci_iov_resource_size(pdev, i + PCI_IOV_RESOURCES);
2943 		/*
2944 		 * On PHB3, the minimum size alignment of M64 BAR in single
2945 		 * mode is 32MB.
2946 		 */
2947 		if (pdn->m64_single_mode && (size < SZ_32M))
2948 			goto truncate_iov;
2949 		dev_dbg(&pdev->dev, " Fixing VF BAR%d: %pR to\n", i, res);
2950 		res->end = res->start + size * mul - 1;
2951 		dev_dbg(&pdev->dev, "                       %pR\n", res);
2952 		dev_info(&pdev->dev, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
2953 			 i, res, mul);
2954 	}
2955 	pdn->vfs_expanded = mul;
2956 
2957 	return;
2958 
2959 truncate_iov:
2960 	/* To save MMIO space, IOV BAR is truncated. */
2961 	for (i = 0; i < PCI_SRIOV_NUM_BARS; i++) {
2962 		res = &pdev->resource[i + PCI_IOV_RESOURCES];
2963 		res->flags = 0;
2964 		res->end = res->start - 1;
2965 	}
2966 }
2967 #endif /* CONFIG_PCI_IOV */
2968 
2969 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe,
2970 				  struct resource *res)
2971 {
2972 	struct pnv_phb *phb = pe->phb;
2973 	struct pci_bus_region region;
2974 	int index;
2975 	int64_t rc;
2976 
2977 	if (!res || !res->flags || res->start > res->end)
2978 		return;
2979 
2980 	if (res->flags & IORESOURCE_IO) {
2981 		region.start = res->start - phb->ioda.io_pci_base;
2982 		region.end   = res->end - phb->ioda.io_pci_base;
2983 		index = region.start / phb->ioda.io_segsize;
2984 
2985 		while (index < phb->ioda.total_pe_num &&
2986 		       region.start <= region.end) {
2987 			phb->ioda.io_segmap[index] = pe->pe_number;
2988 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2989 				pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
2990 			if (rc != OPAL_SUCCESS) {
2991 				pr_err("%s: Error %lld mapping IO segment#%d to PE#%x\n",
2992 				       __func__, rc, index, pe->pe_number);
2993 				break;
2994 			}
2995 
2996 			region.start += phb->ioda.io_segsize;
2997 			index++;
2998 		}
2999 	} else if ((res->flags & IORESOURCE_MEM) &&
3000 		   !pnv_pci_is_m64(phb, res)) {
3001 		region.start = res->start -
3002 			       phb->hose->mem_offset[0] -
3003 			       phb->ioda.m32_pci_base;
3004 		region.end   = res->end -
3005 			       phb->hose->mem_offset[0] -
3006 			       phb->ioda.m32_pci_base;
3007 		index = region.start / phb->ioda.m32_segsize;
3008 
3009 		while (index < phb->ioda.total_pe_num &&
3010 		       region.start <= region.end) {
3011 			phb->ioda.m32_segmap[index] = pe->pe_number;
3012 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3013 				pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
3014 			if (rc != OPAL_SUCCESS) {
3015 				pr_err("%s: Error %lld mapping M32 segment#%d to PE#%x",
3016 				       __func__, rc, index, pe->pe_number);
3017 				break;
3018 			}
3019 
3020 			region.start += phb->ioda.m32_segsize;
3021 			index++;
3022 		}
3023 	}
3024 }
3025 
3026 /*
3027  * This function is supposed to be called on basis of PE from top
3028  * to bottom style. So the the I/O or MMIO segment assigned to
3029  * parent PE could be overridden by its child PEs if necessary.
3030  */
3031 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe)
3032 {
3033 	struct pci_dev *pdev;
3034 	int i;
3035 
3036 	/*
3037 	 * NOTE: We only care PCI bus based PE for now. For PCI
3038 	 * device based PE, for example SRIOV sensitive VF should
3039 	 * be figured out later.
3040 	 */
3041 	BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
3042 
3043 	list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
3044 		for (i = 0; i <= PCI_ROM_RESOURCE; i++)
3045 			pnv_ioda_setup_pe_res(pe, &pdev->resource[i]);
3046 
3047 		/*
3048 		 * If the PE contains all subordinate PCI buses, the
3049 		 * windows of the child bridges should be mapped to
3050 		 * the PE as well.
3051 		 */
3052 		if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev))
3053 			continue;
3054 		for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++)
3055 			pnv_ioda_setup_pe_res(pe,
3056 				&pdev->resource[PCI_BRIDGE_RESOURCES + i]);
3057 	}
3058 }
3059 
3060 #ifdef CONFIG_DEBUG_FS
3061 static int pnv_pci_diag_data_set(void *data, u64 val)
3062 {
3063 	struct pci_controller *hose;
3064 	struct pnv_phb *phb;
3065 	s64 ret;
3066 
3067 	if (val != 1ULL)
3068 		return -EINVAL;
3069 
3070 	hose = (struct pci_controller *)data;
3071 	if (!hose || !hose->private_data)
3072 		return -ENODEV;
3073 
3074 	phb = hose->private_data;
3075 
3076 	/* Retrieve the diag data from firmware */
3077 	ret = opal_pci_get_phb_diag_data2(phb->opal_id, phb->diag_data,
3078 					  phb->diag_data_size);
3079 	if (ret != OPAL_SUCCESS)
3080 		return -EIO;
3081 
3082 	/* Print the diag data to the kernel log */
3083 	pnv_pci_dump_phb_diag_data(phb->hose, phb->diag_data);
3084 	return 0;
3085 }
3086 
3087 DEFINE_SIMPLE_ATTRIBUTE(pnv_pci_diag_data_fops, NULL,
3088 			pnv_pci_diag_data_set, "%llu\n");
3089 
3090 #endif /* CONFIG_DEBUG_FS */
3091 
3092 static void pnv_pci_ioda_create_dbgfs(void)
3093 {
3094 #ifdef CONFIG_DEBUG_FS
3095 	struct pci_controller *hose, *tmp;
3096 	struct pnv_phb *phb;
3097 	char name[16];
3098 
3099 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
3100 		phb = hose->private_data;
3101 
3102 		/* Notify initialization of PHB done */
3103 		phb->initialized = 1;
3104 
3105 		sprintf(name, "PCI%04x", hose->global_number);
3106 		phb->dbgfs = debugfs_create_dir(name, powerpc_debugfs_root);
3107 		if (!phb->dbgfs) {
3108 			pr_warn("%s: Error on creating debugfs on PHB#%x\n",
3109 				__func__, hose->global_number);
3110 			continue;
3111 		}
3112 
3113 		debugfs_create_file("dump_diag_regs", 0200, phb->dbgfs, hose,
3114 				    &pnv_pci_diag_data_fops);
3115 	}
3116 #endif /* CONFIG_DEBUG_FS */
3117 }
3118 
3119 static void pnv_pci_enable_bridge(struct pci_bus *bus)
3120 {
3121 	struct pci_dev *dev = bus->self;
3122 	struct pci_bus *child;
3123 
3124 	/* Empty bus ? bail */
3125 	if (list_empty(&bus->devices))
3126 		return;
3127 
3128 	/*
3129 	 * If there's a bridge associated with that bus enable it. This works
3130 	 * around races in the generic code if the enabling is done during
3131 	 * parallel probing. This can be removed once those races have been
3132 	 * fixed.
3133 	 */
3134 	if (dev) {
3135 		int rc = pci_enable_device(dev);
3136 		if (rc)
3137 			pci_err(dev, "Error enabling bridge (%d)\n", rc);
3138 		pci_set_master(dev);
3139 	}
3140 
3141 	/* Perform the same to child busses */
3142 	list_for_each_entry(child, &bus->children, node)
3143 		pnv_pci_enable_bridge(child);
3144 }
3145 
3146 static void pnv_pci_enable_bridges(void)
3147 {
3148 	struct pci_controller *hose;
3149 
3150 	list_for_each_entry(hose, &hose_list, list_node)
3151 		pnv_pci_enable_bridge(hose->bus);
3152 }
3153 
3154 static void pnv_pci_ioda_fixup(void)
3155 {
3156 	pnv_pci_ioda_setup_PEs();
3157 	pnv_pci_ioda_setup_iommu_api();
3158 	pnv_pci_ioda_create_dbgfs();
3159 
3160 	pnv_pci_enable_bridges();
3161 
3162 #ifdef CONFIG_EEH
3163 	pnv_eeh_post_init();
3164 #endif
3165 }
3166 
3167 /*
3168  * Returns the alignment for I/O or memory windows for P2P
3169  * bridges. That actually depends on how PEs are segmented.
3170  * For now, we return I/O or M32 segment size for PE sensitive
3171  * P2P bridges. Otherwise, the default values (4KiB for I/O,
3172  * 1MiB for memory) will be returned.
3173  *
3174  * The current PCI bus might be put into one PE, which was
3175  * create against the parent PCI bridge. For that case, we
3176  * needn't enlarge the alignment so that we can save some
3177  * resources.
3178  */
3179 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
3180 						unsigned long type)
3181 {
3182 	struct pci_dev *bridge;
3183 	struct pci_controller *hose = pci_bus_to_host(bus);
3184 	struct pnv_phb *phb = hose->private_data;
3185 	int num_pci_bridges = 0;
3186 
3187 	bridge = bus->self;
3188 	while (bridge) {
3189 		if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
3190 			num_pci_bridges++;
3191 			if (num_pci_bridges >= 2)
3192 				return 1;
3193 		}
3194 
3195 		bridge = bridge->bus->self;
3196 	}
3197 
3198 	/*
3199 	 * We fall back to M32 if M64 isn't supported. We enforce the M64
3200 	 * alignment for any 64-bit resource, PCIe doesn't care and
3201 	 * bridges only do 64-bit prefetchable anyway.
3202 	 */
3203 	if (phb->ioda.m64_segsize && pnv_pci_is_m64_flags(type))
3204 		return phb->ioda.m64_segsize;
3205 	if (type & IORESOURCE_MEM)
3206 		return phb->ioda.m32_segsize;
3207 
3208 	return phb->ioda.io_segsize;
3209 }
3210 
3211 /*
3212  * We are updating root port or the upstream port of the
3213  * bridge behind the root port with PHB's windows in order
3214  * to accommodate the changes on required resources during
3215  * PCI (slot) hotplug, which is connected to either root
3216  * port or the downstream ports of PCIe switch behind the
3217  * root port.
3218  */
3219 static void pnv_pci_fixup_bridge_resources(struct pci_bus *bus,
3220 					   unsigned long type)
3221 {
3222 	struct pci_controller *hose = pci_bus_to_host(bus);
3223 	struct pnv_phb *phb = hose->private_data;
3224 	struct pci_dev *bridge = bus->self;
3225 	struct resource *r, *w;
3226 	bool msi_region = false;
3227 	int i;
3228 
3229 	/* Check if we need apply fixup to the bridge's windows */
3230 	if (!pci_is_root_bus(bridge->bus) &&
3231 	    !pci_is_root_bus(bridge->bus->self->bus))
3232 		return;
3233 
3234 	/* Fixup the resources */
3235 	for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
3236 		r = &bridge->resource[PCI_BRIDGE_RESOURCES + i];
3237 		if (!r->flags || !r->parent)
3238 			continue;
3239 
3240 		w = NULL;
3241 		if (r->flags & type & IORESOURCE_IO)
3242 			w = &hose->io_resource;
3243 		else if (pnv_pci_is_m64(phb, r) &&
3244 			 (type & IORESOURCE_PREFETCH) &&
3245 			 phb->ioda.m64_segsize)
3246 			w = &hose->mem_resources[1];
3247 		else if (r->flags & type & IORESOURCE_MEM) {
3248 			w = &hose->mem_resources[0];
3249 			msi_region = true;
3250 		}
3251 
3252 		r->start = w->start;
3253 		r->end = w->end;
3254 
3255 		/* The 64KB 32-bits MSI region shouldn't be included in
3256 		 * the 32-bits bridge window. Otherwise, we can see strange
3257 		 * issues. One of them is EEH error observed on Garrison.
3258 		 *
3259 		 * Exclude top 1MB region which is the minimal alignment of
3260 		 * 32-bits bridge window.
3261 		 */
3262 		if (msi_region) {
3263 			r->end += 0x10000;
3264 			r->end -= 0x100000;
3265 		}
3266 	}
3267 }
3268 
3269 static void pnv_pci_setup_bridge(struct pci_bus *bus, unsigned long type)
3270 {
3271 	struct pci_controller *hose = pci_bus_to_host(bus);
3272 	struct pnv_phb *phb = hose->private_data;
3273 	struct pci_dev *bridge = bus->self;
3274 	struct pnv_ioda_pe *pe;
3275 	bool all = (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE);
3276 
3277 	/* Extend bridge's windows if necessary */
3278 	pnv_pci_fixup_bridge_resources(bus, type);
3279 
3280 	/* The PE for root bus should be realized before any one else */
3281 	if (!phb->ioda.root_pe_populated) {
3282 		pe = pnv_ioda_setup_bus_PE(phb->hose->bus, false);
3283 		if (pe) {
3284 			phb->ioda.root_pe_idx = pe->pe_number;
3285 			phb->ioda.root_pe_populated = true;
3286 		}
3287 	}
3288 
3289 	/* Don't assign PE to PCI bus, which doesn't have subordinate devices */
3290 	if (list_empty(&bus->devices))
3291 		return;
3292 
3293 	/* Reserve PEs according to used M64 resources */
3294 	pnv_ioda_reserve_m64_pe(bus, NULL, all);
3295 
3296 	/*
3297 	 * Assign PE. We might run here because of partial hotplug.
3298 	 * For the case, we just pick up the existing PE and should
3299 	 * not allocate resources again.
3300 	 */
3301 	pe = pnv_ioda_setup_bus_PE(bus, all);
3302 	if (!pe)
3303 		return;
3304 
3305 	pnv_ioda_setup_pe_seg(pe);
3306 	switch (phb->type) {
3307 	case PNV_PHB_IODA1:
3308 		pnv_pci_ioda1_setup_dma_pe(phb, pe);
3309 		break;
3310 	case PNV_PHB_IODA2:
3311 		pnv_pci_ioda2_setup_dma_pe(phb, pe);
3312 		break;
3313 	default:
3314 		pr_warn("%s: No DMA for PHB#%x (type %d)\n",
3315 			__func__, phb->hose->global_number, phb->type);
3316 	}
3317 }
3318 
3319 static resource_size_t pnv_pci_default_alignment(void)
3320 {
3321 	return PAGE_SIZE;
3322 }
3323 
3324 #ifdef CONFIG_PCI_IOV
3325 static resource_size_t pnv_pci_iov_resource_alignment(struct pci_dev *pdev,
3326 						      int resno)
3327 {
3328 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3329 	struct pnv_phb *phb = hose->private_data;
3330 	struct pci_dn *pdn = pci_get_pdn(pdev);
3331 	resource_size_t align;
3332 
3333 	/*
3334 	 * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the
3335 	 * SR-IOV. While from hardware perspective, the range mapped by M64
3336 	 * BAR should be size aligned.
3337 	 *
3338 	 * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra
3339 	 * powernv-specific hardware restriction is gone. But if just use the
3340 	 * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with
3341 	 * in one segment of M64 #15, which introduces the PE conflict between
3342 	 * PF and VF. Based on this, the minimum alignment of an IOV BAR is
3343 	 * m64_segsize.
3344 	 *
3345 	 * This function returns the total IOV BAR size if M64 BAR is in
3346 	 * Shared PE mode or just VF BAR size if not.
3347 	 * If the M64 BAR is in Single PE mode, return the VF BAR size or
3348 	 * M64 segment size if IOV BAR size is less.
3349 	 */
3350 	align = pci_iov_resource_size(pdev, resno);
3351 	if (!pdn->vfs_expanded)
3352 		return align;
3353 	if (pdn->m64_single_mode)
3354 		return max(align, (resource_size_t)phb->ioda.m64_segsize);
3355 
3356 	return pdn->vfs_expanded * align;
3357 }
3358 #endif /* CONFIG_PCI_IOV */
3359 
3360 /* Prevent enabling devices for which we couldn't properly
3361  * assign a PE
3362  */
3363 static bool pnv_pci_enable_device_hook(struct pci_dev *dev)
3364 {
3365 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
3366 	struct pnv_phb *phb = hose->private_data;
3367 	struct pci_dn *pdn;
3368 
3369 	/* The function is probably called while the PEs have
3370 	 * not be created yet. For example, resource reassignment
3371 	 * during PCI probe period. We just skip the check if
3372 	 * PEs isn't ready.
3373 	 */
3374 	if (!phb->initialized)
3375 		return true;
3376 
3377 	pdn = pci_get_pdn(dev);
3378 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3379 		return false;
3380 
3381 	return true;
3382 }
3383 
3384 static long pnv_pci_ioda1_unset_window(struct iommu_table_group *table_group,
3385 				       int num)
3386 {
3387 	struct pnv_ioda_pe *pe = container_of(table_group,
3388 					      struct pnv_ioda_pe, table_group);
3389 	struct pnv_phb *phb = pe->phb;
3390 	unsigned int idx;
3391 	long rc;
3392 
3393 	pe_info(pe, "Removing DMA window #%d\n", num);
3394 	for (idx = 0; idx < phb->ioda.dma32_count; idx++) {
3395 		if (phb->ioda.dma32_segmap[idx] != pe->pe_number)
3396 			continue;
3397 
3398 		rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
3399 						idx, 0, 0ul, 0ul, 0ul);
3400 		if (rc != OPAL_SUCCESS) {
3401 			pe_warn(pe, "Failure %ld unmapping DMA32 segment#%d\n",
3402 				rc, idx);
3403 			return rc;
3404 		}
3405 
3406 		phb->ioda.dma32_segmap[idx] = IODA_INVALID_PE;
3407 	}
3408 
3409 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
3410 	return OPAL_SUCCESS;
3411 }
3412 
3413 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe *pe)
3414 {
3415 	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3416 	struct iommu_table *tbl = pe->table_group.tables[0];
3417 	int64_t rc;
3418 
3419 	if (!weight)
3420 		return;
3421 
3422 	rc = pnv_pci_ioda1_unset_window(&pe->table_group, 0);
3423 	if (rc != OPAL_SUCCESS)
3424 		return;
3425 
3426 	pnv_pci_p7ioc_tce_invalidate(tbl, tbl->it_offset, tbl->it_size, false);
3427 	if (pe->table_group.group) {
3428 		iommu_group_put(pe->table_group.group);
3429 		WARN_ON(pe->table_group.group);
3430 	}
3431 
3432 	free_pages(tbl->it_base, get_order(tbl->it_size << 3));
3433 	iommu_tce_table_put(tbl);
3434 }
3435 
3436 static void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe *pe)
3437 {
3438 	struct iommu_table *tbl = pe->table_group.tables[0];
3439 	unsigned int weight = pnv_pci_ioda_pe_dma_weight(pe);
3440 #ifdef CONFIG_IOMMU_API
3441 	int64_t rc;
3442 #endif
3443 
3444 	if (!weight)
3445 		return;
3446 
3447 #ifdef CONFIG_IOMMU_API
3448 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
3449 	if (rc)
3450 		pe_warn(pe, "OPAL error %lld release DMA window\n", rc);
3451 #endif
3452 
3453 	pnv_pci_ioda2_set_bypass(pe, false);
3454 	if (pe->table_group.group) {
3455 		iommu_group_put(pe->table_group.group);
3456 		WARN_ON(pe->table_group.group);
3457 	}
3458 
3459 	iommu_tce_table_put(tbl);
3460 }
3461 
3462 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe *pe,
3463 				 unsigned short win,
3464 				 unsigned int *map)
3465 {
3466 	struct pnv_phb *phb = pe->phb;
3467 	int idx;
3468 	int64_t rc;
3469 
3470 	for (idx = 0; idx < phb->ioda.total_pe_num; idx++) {
3471 		if (map[idx] != pe->pe_number)
3472 			continue;
3473 
3474 		if (win == OPAL_M64_WINDOW_TYPE)
3475 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3476 					phb->ioda.reserved_pe_idx, win,
3477 					idx / PNV_IODA1_M64_SEGS,
3478 					idx % PNV_IODA1_M64_SEGS);
3479 		else
3480 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
3481 					phb->ioda.reserved_pe_idx, win, 0, idx);
3482 
3483 		if (rc != OPAL_SUCCESS)
3484 			pe_warn(pe, "Error %lld unmapping (%d) segment#%d\n",
3485 				rc, win, idx);
3486 
3487 		map[idx] = IODA_INVALID_PE;
3488 	}
3489 }
3490 
3491 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
3492 {
3493 	struct pnv_phb *phb = pe->phb;
3494 
3495 	if (phb->type == PNV_PHB_IODA1) {
3496 		pnv_ioda_free_pe_seg(pe, OPAL_IO_WINDOW_TYPE,
3497 				     phb->ioda.io_segmap);
3498 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3499 				     phb->ioda.m32_segmap);
3500 		pnv_ioda_free_pe_seg(pe, OPAL_M64_WINDOW_TYPE,
3501 				     phb->ioda.m64_segmap);
3502 	} else if (phb->type == PNV_PHB_IODA2) {
3503 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
3504 				     phb->ioda.m32_segmap);
3505 	}
3506 }
3507 
3508 static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe)
3509 {
3510 	struct pnv_phb *phb = pe->phb;
3511 	struct pnv_ioda_pe *slave, *tmp;
3512 
3513 	list_del(&pe->list);
3514 	switch (phb->type) {
3515 	case PNV_PHB_IODA1:
3516 		pnv_pci_ioda1_release_pe_dma(pe);
3517 		break;
3518 	case PNV_PHB_IODA2:
3519 		pnv_pci_ioda2_release_pe_dma(pe);
3520 		break;
3521 	default:
3522 		WARN_ON(1);
3523 	}
3524 
3525 	pnv_ioda_release_pe_seg(pe);
3526 	pnv_ioda_deconfigure_pe(pe->phb, pe);
3527 
3528 	/* Release slave PEs in the compound PE */
3529 	if (pe->flags & PNV_IODA_PE_MASTER) {
3530 		list_for_each_entry_safe(slave, tmp, &pe->slaves, list) {
3531 			list_del(&slave->list);
3532 			pnv_ioda_free_pe(slave);
3533 		}
3534 	}
3535 
3536 	/*
3537 	 * The PE for root bus can be removed because of hotplug in EEH
3538 	 * recovery for fenced PHB error. We need to mark the PE dead so
3539 	 * that it can be populated again in PCI hot add path. The PE
3540 	 * shouldn't be destroyed as it's the global reserved resource.
3541 	 */
3542 	if (phb->ioda.root_pe_populated &&
3543 	    phb->ioda.root_pe_idx == pe->pe_number)
3544 		phb->ioda.root_pe_populated = false;
3545 	else
3546 		pnv_ioda_free_pe(pe);
3547 }
3548 
3549 static void pnv_pci_release_device(struct pci_dev *pdev)
3550 {
3551 	struct pci_controller *hose = pci_bus_to_host(pdev->bus);
3552 	struct pnv_phb *phb = hose->private_data;
3553 	struct pci_dn *pdn = pci_get_pdn(pdev);
3554 	struct pnv_ioda_pe *pe;
3555 
3556 	if (pdev->is_virtfn)
3557 		return;
3558 
3559 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
3560 		return;
3561 
3562 	/*
3563 	 * PCI hotplug can happen as part of EEH error recovery. The @pdn
3564 	 * isn't removed and added afterwards in this scenario. We should
3565 	 * set the PE number in @pdn to an invalid one. Otherwise, the PE's
3566 	 * device count is decreased on removing devices while failing to
3567 	 * be increased on adding devices. It leads to unbalanced PE's device
3568 	 * count and eventually make normal PCI hotplug path broken.
3569 	 */
3570 	pe = &phb->ioda.pe_array[pdn->pe_number];
3571 	pdn->pe_number = IODA_INVALID_PE;
3572 
3573 	WARN_ON(--pe->device_count < 0);
3574 	if (pe->device_count == 0)
3575 		pnv_ioda_release_pe(pe);
3576 }
3577 
3578 static void pnv_npu_disable_device(struct pci_dev *pdev)
3579 {
3580 	struct eeh_dev *edev = pci_dev_to_eeh_dev(pdev);
3581 	struct eeh_pe *eehpe = edev ? edev->pe : NULL;
3582 
3583 	if (eehpe && eeh_ops && eeh_ops->reset)
3584 		eeh_ops->reset(eehpe, EEH_RESET_HOT);
3585 }
3586 
3587 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
3588 {
3589 	struct pnv_phb *phb = hose->private_data;
3590 
3591 	opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
3592 		       OPAL_ASSERT_RESET);
3593 }
3594 
3595 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
3596 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3597 	.dma_bus_setup		= pnv_pci_dma_bus_setup,
3598 	.iommu_bypass_supported	= pnv_pci_ioda_iommu_bypass_supported,
3599 	.setup_msi_irqs		= pnv_setup_msi_irqs,
3600 	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
3601 	.enable_device_hook	= pnv_pci_enable_device_hook,
3602 	.release_device		= pnv_pci_release_device,
3603 	.window_alignment	= pnv_pci_window_alignment,
3604 	.setup_bridge		= pnv_pci_setup_bridge,
3605 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3606 	.shutdown		= pnv_pci_ioda_shutdown,
3607 };
3608 
3609 static const struct pci_controller_ops pnv_npu_ioda_controller_ops = {
3610 	.dma_dev_setup		= pnv_pci_dma_dev_setup,
3611 	.setup_msi_irqs		= pnv_setup_msi_irqs,
3612 	.teardown_msi_irqs	= pnv_teardown_msi_irqs,
3613 	.enable_device_hook	= pnv_pci_enable_device_hook,
3614 	.window_alignment	= pnv_pci_window_alignment,
3615 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3616 	.shutdown		= pnv_pci_ioda_shutdown,
3617 	.disable_device		= pnv_npu_disable_device,
3618 };
3619 
3620 static const struct pci_controller_ops pnv_npu_ocapi_ioda_controller_ops = {
3621 	.enable_device_hook	= pnv_pci_enable_device_hook,
3622 	.window_alignment	= pnv_pci_window_alignment,
3623 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
3624 	.shutdown		= pnv_pci_ioda_shutdown,
3625 };
3626 
3627 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
3628 					 u64 hub_id, int ioda_type)
3629 {
3630 	struct pci_controller *hose;
3631 	struct pnv_phb *phb;
3632 	unsigned long size, m64map_off, m32map_off, pemap_off;
3633 	unsigned long iomap_off = 0, dma32map_off = 0;
3634 	struct resource r;
3635 	const __be64 *prop64;
3636 	const __be32 *prop32;
3637 	int len;
3638 	unsigned int segno;
3639 	u64 phb_id;
3640 	void *aux;
3641 	long rc;
3642 
3643 	if (!of_device_is_available(np))
3644 		return;
3645 
3646 	pr_info("Initializing %s PHB (%pOF)\n",	pnv_phb_names[ioda_type], np);
3647 
3648 	prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
3649 	if (!prop64) {
3650 		pr_err("  Missing \"ibm,opal-phbid\" property !\n");
3651 		return;
3652 	}
3653 	phb_id = be64_to_cpup(prop64);
3654 	pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
3655 
3656 	phb = memblock_alloc(sizeof(*phb), SMP_CACHE_BYTES);
3657 	if (!phb)
3658 		panic("%s: Failed to allocate %zu bytes\n", __func__,
3659 		      sizeof(*phb));
3660 
3661 	/* Allocate PCI controller */
3662 	phb->hose = hose = pcibios_alloc_controller(np);
3663 	if (!phb->hose) {
3664 		pr_err("  Can't allocate PCI controller for %pOF\n",
3665 		       np);
3666 		memblock_free(__pa(phb), sizeof(struct pnv_phb));
3667 		return;
3668 	}
3669 
3670 	spin_lock_init(&phb->lock);
3671 	prop32 = of_get_property(np, "bus-range", &len);
3672 	if (prop32 && len == 8) {
3673 		hose->first_busno = be32_to_cpu(prop32[0]);
3674 		hose->last_busno = be32_to_cpu(prop32[1]);
3675 	} else {
3676 		pr_warn("  Broken <bus-range> on %pOF\n", np);
3677 		hose->first_busno = 0;
3678 		hose->last_busno = 0xff;
3679 	}
3680 	hose->private_data = phb;
3681 	phb->hub_id = hub_id;
3682 	phb->opal_id = phb_id;
3683 	phb->type = ioda_type;
3684 	mutex_init(&phb->ioda.pe_alloc_mutex);
3685 
3686 	/* Detect specific models for error handling */
3687 	if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
3688 		phb->model = PNV_PHB_MODEL_P7IOC;
3689 	else if (of_device_is_compatible(np, "ibm,power8-pciex"))
3690 		phb->model = PNV_PHB_MODEL_PHB3;
3691 	else if (of_device_is_compatible(np, "ibm,power8-npu-pciex"))
3692 		phb->model = PNV_PHB_MODEL_NPU;
3693 	else if (of_device_is_compatible(np, "ibm,power9-npu-pciex"))
3694 		phb->model = PNV_PHB_MODEL_NPU2;
3695 	else
3696 		phb->model = PNV_PHB_MODEL_UNKNOWN;
3697 
3698 	/* Initialize diagnostic data buffer */
3699 	prop32 = of_get_property(np, "ibm,phb-diag-data-size", NULL);
3700 	if (prop32)
3701 		phb->diag_data_size = be32_to_cpup(prop32);
3702 	else
3703 		phb->diag_data_size = PNV_PCI_DIAG_BUF_SIZE;
3704 
3705 	phb->diag_data = memblock_alloc(phb->diag_data_size, SMP_CACHE_BYTES);
3706 	if (!phb->diag_data)
3707 		panic("%s: Failed to allocate %u bytes\n", __func__,
3708 		      phb->diag_data_size);
3709 
3710 	/* Parse 32-bit and IO ranges (if any) */
3711 	pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
3712 
3713 	/* Get registers */
3714 	if (!of_address_to_resource(np, 0, &r)) {
3715 		phb->regs_phys = r.start;
3716 		phb->regs = ioremap(r.start, resource_size(&r));
3717 		if (phb->regs == NULL)
3718 			pr_err("  Failed to map registers !\n");
3719 	}
3720 
3721 	/* Initialize more IODA stuff */
3722 	phb->ioda.total_pe_num = 1;
3723 	prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
3724 	if (prop32)
3725 		phb->ioda.total_pe_num = be32_to_cpup(prop32);
3726 	prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
3727 	if (prop32)
3728 		phb->ioda.reserved_pe_idx = be32_to_cpup(prop32);
3729 
3730 	/* Invalidate RID to PE# mapping */
3731 	for (segno = 0; segno < ARRAY_SIZE(phb->ioda.pe_rmap); segno++)
3732 		phb->ioda.pe_rmap[segno] = IODA_INVALID_PE;
3733 
3734 	/* Parse 64-bit MMIO range */
3735 	pnv_ioda_parse_m64_window(phb);
3736 
3737 	phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
3738 	/* FW Has already off top 64k of M32 space (MSI space) */
3739 	phb->ioda.m32_size += 0x10000;
3740 
3741 	phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num;
3742 	phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
3743 	phb->ioda.io_size = hose->pci_io_size;
3744 	phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num;
3745 	phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
3746 
3747 	/* Calculate how many 32-bit TCE segments we have */
3748 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3749 				PNV_IODA1_DMA32_SEGSIZE;
3750 
3751 	/* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3752 	size = _ALIGN_UP(max_t(unsigned, phb->ioda.total_pe_num, 8) / 8,
3753 			sizeof(unsigned long));
3754 	m64map_off = size;
3755 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]);
3756 	m32map_off = size;
3757 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]);
3758 	if (phb->type == PNV_PHB_IODA1) {
3759 		iomap_off = size;
3760 		size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]);
3761 		dma32map_off = size;
3762 		size += phb->ioda.dma32_count *
3763 			sizeof(phb->ioda.dma32_segmap[0]);
3764 	}
3765 	pemap_off = size;
3766 	size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe);
3767 	aux = memblock_alloc(size, SMP_CACHE_BYTES);
3768 	if (!aux)
3769 		panic("%s: Failed to allocate %lu bytes\n", __func__, size);
3770 	phb->ioda.pe_alloc = aux;
3771 	phb->ioda.m64_segmap = aux + m64map_off;
3772 	phb->ioda.m32_segmap = aux + m32map_off;
3773 	for (segno = 0; segno < phb->ioda.total_pe_num; segno++) {
3774 		phb->ioda.m64_segmap[segno] = IODA_INVALID_PE;
3775 		phb->ioda.m32_segmap[segno] = IODA_INVALID_PE;
3776 	}
3777 	if (phb->type == PNV_PHB_IODA1) {
3778 		phb->ioda.io_segmap = aux + iomap_off;
3779 		for (segno = 0; segno < phb->ioda.total_pe_num; segno++)
3780 			phb->ioda.io_segmap[segno] = IODA_INVALID_PE;
3781 
3782 		phb->ioda.dma32_segmap = aux + dma32map_off;
3783 		for (segno = 0; segno < phb->ioda.dma32_count; segno++)
3784 			phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE;
3785 	}
3786 	phb->ioda.pe_array = aux + pemap_off;
3787 
3788 	/*
3789 	 * Choose PE number for root bus, which shouldn't have
3790 	 * M64 resources consumed by its child devices. To pick
3791 	 * the PE number adjacent to the reserved one if possible.
3792 	 */
3793 	pnv_ioda_reserve_pe(phb, phb->ioda.reserved_pe_idx);
3794 	if (phb->ioda.reserved_pe_idx == 0) {
3795 		phb->ioda.root_pe_idx = 1;
3796 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3797 	} else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) {
3798 		phb->ioda.root_pe_idx = phb->ioda.reserved_pe_idx - 1;
3799 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3800 	} else {
3801 		phb->ioda.root_pe_idx = IODA_INVALID_PE;
3802 	}
3803 
3804 	INIT_LIST_HEAD(&phb->ioda.pe_list);
3805 	mutex_init(&phb->ioda.pe_list_mutex);
3806 
3807 	/* Calculate how many 32-bit TCE segments we have */
3808 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3809 				PNV_IODA1_DMA32_SEGSIZE;
3810 
3811 #if 0 /* We should really do that ... */
3812 	rc = opal_pci_set_phb_mem_window(opal->phb_id,
3813 					 window_type,
3814 					 window_num,
3815 					 starting_real_address,
3816 					 starting_pci_address,
3817 					 segment_size);
3818 #endif
3819 
3820 	pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3821 		phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx,
3822 		phb->ioda.m32_size, phb->ioda.m32_segsize);
3823 	if (phb->ioda.m64_size)
3824 		pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
3825 			phb->ioda.m64_size, phb->ioda.m64_segsize);
3826 	if (phb->ioda.io_size)
3827 		pr_info("                  IO: 0x%x [segment=0x%x]\n",
3828 			phb->ioda.io_size, phb->ioda.io_segsize);
3829 
3830 
3831 	phb->hose->ops = &pnv_pci_ops;
3832 	phb->get_pe_state = pnv_ioda_get_pe_state;
3833 	phb->freeze_pe = pnv_ioda_freeze_pe;
3834 	phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
3835 
3836 	/* Setup MSI support */
3837 	pnv_pci_init_ioda_msis(phb);
3838 
3839 	/*
3840 	 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3841 	 * to let the PCI core do resource assignment. It's supposed
3842 	 * that the PCI core will do correct I/O and MMIO alignment
3843 	 * for the P2P bridge bars so that each PCI bus (excluding
3844 	 * the child P2P bridges) can form individual PE.
3845 	 */
3846 	ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
3847 
3848 	switch (phb->type) {
3849 	case PNV_PHB_NPU_NVLINK:
3850 		hose->controller_ops = pnv_npu_ioda_controller_ops;
3851 		break;
3852 	case PNV_PHB_NPU_OCAPI:
3853 		hose->controller_ops = pnv_npu_ocapi_ioda_controller_ops;
3854 		break;
3855 	default:
3856 		phb->dma_dev_setup = pnv_pci_ioda_dma_dev_setup;
3857 		hose->controller_ops = pnv_pci_ioda_controller_ops;
3858 	}
3859 
3860 	ppc_md.pcibios_default_alignment = pnv_pci_default_alignment;
3861 
3862 #ifdef CONFIG_PCI_IOV
3863 	ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov_resources;
3864 	ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
3865 	ppc_md.pcibios_sriov_enable = pnv_pcibios_sriov_enable;
3866 	ppc_md.pcibios_sriov_disable = pnv_pcibios_sriov_disable;
3867 #endif
3868 
3869 	pci_add_flags(PCI_REASSIGN_ALL_RSRC);
3870 
3871 	/* Reset IODA tables to a clean state */
3872 	rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
3873 	if (rc)
3874 		pr_warn("  OPAL Error %ld performing IODA table reset !\n", rc);
3875 
3876 	/*
3877 	 * If we're running in kdump kernel, the previous kernel never
3878 	 * shutdown PCI devices correctly. We already got IODA table
3879 	 * cleaned out. So we have to issue PHB reset to stop all PCI
3880 	 * transactions from previous kernel. The ppc_pci_reset_phbs
3881 	 * kernel parameter will force this reset too. Additionally,
3882 	 * if the IODA reset above failed then use a bigger hammer.
3883 	 * This can happen if we get a PHB fatal error in very early
3884 	 * boot.
3885 	 */
3886 	if (is_kdump_kernel() || pci_reset_phbs || rc) {
3887 		pr_info("  Issue PHB reset ...\n");
3888 		pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
3889 		pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
3890 	}
3891 
3892 	/* Remove M64 resource if we can't configure it successfully */
3893 	if (!phb->init_m64 || phb->init_m64(phb))
3894 		hose->mem_resources[1].flags = 0;
3895 }
3896 
3897 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
3898 {
3899 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
3900 }
3901 
3902 void __init pnv_pci_init_npu_phb(struct device_node *np)
3903 {
3904 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_NVLINK);
3905 }
3906 
3907 void __init pnv_pci_init_npu2_opencapi_phb(struct device_node *np)
3908 {
3909 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_OCAPI);
3910 }
3911 
3912 static void pnv_npu2_opencapi_cfg_size_fixup(struct pci_dev *dev)
3913 {
3914 	struct pci_controller *hose = pci_bus_to_host(dev->bus);
3915 	struct pnv_phb *phb = hose->private_data;
3916 
3917 	if (!machine_is(powernv))
3918 		return;
3919 
3920 	if (phb->type == PNV_PHB_NPU_OCAPI)
3921 		dev->cfg_size = PCI_CFG_SPACE_EXP_SIZE;
3922 }
3923 DECLARE_PCI_FIXUP_EARLY(PCI_ANY_ID, PCI_ANY_ID, pnv_npu2_opencapi_cfg_size_fixup);
3924 
3925 void __init pnv_pci_init_ioda_hub(struct device_node *np)
3926 {
3927 	struct device_node *phbn;
3928 	const __be64 *prop64;
3929 	u64 hub_id;
3930 
3931 	pr_info("Probing IODA IO-Hub %pOF\n", np);
3932 
3933 	prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
3934 	if (!prop64) {
3935 		pr_err(" Missing \"ibm,opal-hubid\" property !\n");
3936 		return;
3937 	}
3938 	hub_id = be64_to_cpup(prop64);
3939 	pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
3940 
3941 	/* Count child PHBs */
3942 	for_each_child_of_node(np, phbn) {
3943 		/* Look for IODA1 PHBs */
3944 		if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
3945 			pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
3946 	}
3947 }
3948