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 #include <linux/debugfs.h>
24 
25 #include <asm/sections.h>
26 #include <asm/io.h>
27 #include <asm/prom.h>
28 #include <asm/pci-bridge.h>
29 #include <asm/machdep.h>
30 #include <asm/msi_bitmap.h>
31 #include <asm/ppc-pci.h>
32 #include <asm/opal.h>
33 #include <asm/iommu.h>
34 #include <asm/tce.h>
35 #include <asm/xics.h>
36 #include <asm/firmware.h>
37 #include <asm/pnv-pci.h>
38 #include <asm/mmzone.h>
39 #include <asm/xive.h>
40 
41 #include <misc/cxl-base.h>
42 
43 #include "powernv.h"
44 #include "pci.h"
45 #include "../../../../drivers/pci/pci.h"
46 
47 #define PNV_IODA1_M64_NUM	16	/* Number of M64 BARs	*/
48 #define PNV_IODA1_M64_SEGS	8	/* Segments per M64 BAR	*/
49 #define PNV_IODA1_DMA32_SEGSIZE	0x10000000
50 
51 static const char * const pnv_phb_names[] = { "IODA1", "IODA2", "NPU_OCAPI" };
52 
53 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable);
54 static void pnv_pci_configure_bus(struct pci_bus *bus);
55 
56 void pe_level_printk(const struct pnv_ioda_pe *pe, const char *level,
57 			    const char *fmt, ...)
58 {
59 	struct va_format vaf;
60 	va_list args;
61 	char pfix[32];
62 
63 	va_start(args, fmt);
64 
65 	vaf.fmt = fmt;
66 	vaf.va = &args;
67 
68 	if (pe->flags & PNV_IODA_PE_DEV)
69 		strlcpy(pfix, dev_name(&pe->pdev->dev), sizeof(pfix));
70 	else if (pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL))
71 		sprintf(pfix, "%04x:%02x     ",
72 			pci_domain_nr(pe->pbus), pe->pbus->number);
73 #ifdef CONFIG_PCI_IOV
74 	else if (pe->flags & PNV_IODA_PE_VF)
75 		sprintf(pfix, "%04x:%02x:%2x.%d",
76 			pci_domain_nr(pe->parent_dev->bus),
77 			(pe->rid & 0xff00) >> 8,
78 			PCI_SLOT(pe->rid), PCI_FUNC(pe->rid));
79 #endif /* CONFIG_PCI_IOV*/
80 
81 	printk("%spci %s: [PE# %.2x] %pV",
82 	       level, pfix, pe->pe_number, &vaf);
83 
84 	va_end(args);
85 }
86 
87 static bool pnv_iommu_bypass_disabled __read_mostly;
88 static bool pci_reset_phbs __read_mostly;
89 
90 static int __init iommu_setup(char *str)
91 {
92 	if (!str)
93 		return -EINVAL;
94 
95 	while (*str) {
96 		if (!strncmp(str, "nobypass", 8)) {
97 			pnv_iommu_bypass_disabled = true;
98 			pr_info("PowerNV: IOMMU bypass window disabled.\n");
99 			break;
100 		}
101 		str += strcspn(str, ",");
102 		if (*str == ',')
103 			str++;
104 	}
105 
106 	return 0;
107 }
108 early_param("iommu", iommu_setup);
109 
110 static int __init pci_reset_phbs_setup(char *str)
111 {
112 	pci_reset_phbs = true;
113 	return 0;
114 }
115 
116 early_param("ppc_pci_reset_phbs", pci_reset_phbs_setup);
117 
118 static struct pnv_ioda_pe *pnv_ioda_init_pe(struct pnv_phb *phb, int pe_no)
119 {
120 	s64 rc;
121 
122 	phb->ioda.pe_array[pe_no].phb = phb;
123 	phb->ioda.pe_array[pe_no].pe_number = pe_no;
124 	phb->ioda.pe_array[pe_no].dma_setup_done = false;
125 
126 	/*
127 	 * Clear the PE frozen state as it might be put into frozen state
128 	 * in the last PCI remove path. It's not harmful to do so when the
129 	 * PE is already in unfrozen state.
130 	 */
131 	rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no,
132 				       OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
133 	if (rc != OPAL_SUCCESS && rc != OPAL_UNSUPPORTED)
134 		pr_warn("%s: Error %lld unfreezing PHB#%x-PE#%x\n",
135 			__func__, rc, phb->hose->global_number, pe_no);
136 
137 	return &phb->ioda.pe_array[pe_no];
138 }
139 
140 static void pnv_ioda_reserve_pe(struct pnv_phb *phb, int pe_no)
141 {
142 	if (!(pe_no >= 0 && pe_no < phb->ioda.total_pe_num)) {
143 		pr_warn("%s: Invalid PE %x on PHB#%x\n",
144 			__func__, pe_no, phb->hose->global_number);
145 		return;
146 	}
147 
148 	mutex_lock(&phb->ioda.pe_alloc_mutex);
149 	if (test_and_set_bit(pe_no, phb->ioda.pe_alloc))
150 		pr_debug("%s: PE %x was reserved on PHB#%x\n",
151 			 __func__, pe_no, phb->hose->global_number);
152 	mutex_unlock(&phb->ioda.pe_alloc_mutex);
153 
154 	pnv_ioda_init_pe(phb, pe_no);
155 }
156 
157 struct pnv_ioda_pe *pnv_ioda_alloc_pe(struct pnv_phb *phb, int count)
158 {
159 	struct pnv_ioda_pe *ret = NULL;
160 	int run = 0, pe, i;
161 
162 	mutex_lock(&phb->ioda.pe_alloc_mutex);
163 
164 	/* scan backwards for a run of @count cleared bits */
165 	for (pe = phb->ioda.total_pe_num - 1; pe >= 0; pe--) {
166 		if (test_bit(pe, phb->ioda.pe_alloc)) {
167 			run = 0;
168 			continue;
169 		}
170 
171 		run++;
172 		if (run == count)
173 			break;
174 	}
175 	if (run != count)
176 		goto out;
177 
178 	for (i = pe; i < pe + count; i++) {
179 		set_bit(i, phb->ioda.pe_alloc);
180 		pnv_ioda_init_pe(phb, i);
181 	}
182 	ret = &phb->ioda.pe_array[pe];
183 
184 out:
185 	mutex_unlock(&phb->ioda.pe_alloc_mutex);
186 	return ret;
187 }
188 
189 void pnv_ioda_free_pe(struct pnv_ioda_pe *pe)
190 {
191 	struct pnv_phb *phb = pe->phb;
192 	unsigned int pe_num = pe->pe_number;
193 
194 	WARN_ON(pe->pdev);
195 	memset(pe, 0, sizeof(struct pnv_ioda_pe));
196 
197 	mutex_lock(&phb->ioda.pe_alloc_mutex);
198 	clear_bit(pe_num, phb->ioda.pe_alloc);
199 	mutex_unlock(&phb->ioda.pe_alloc_mutex);
200 }
201 
202 /* The default M64 BAR is shared by all PEs */
203 static int pnv_ioda2_init_m64(struct pnv_phb *phb)
204 {
205 	const char *desc;
206 	struct resource *r;
207 	s64 rc;
208 
209 	/* Configure the default M64 BAR */
210 	rc = opal_pci_set_phb_mem_window(phb->opal_id,
211 					 OPAL_M64_WINDOW_TYPE,
212 					 phb->ioda.m64_bar_idx,
213 					 phb->ioda.m64_base,
214 					 0, /* unused */
215 					 phb->ioda.m64_size);
216 	if (rc != OPAL_SUCCESS) {
217 		desc = "configuring";
218 		goto fail;
219 	}
220 
221 	/* Enable the default M64 BAR */
222 	rc = opal_pci_phb_mmio_enable(phb->opal_id,
223 				      OPAL_M64_WINDOW_TYPE,
224 				      phb->ioda.m64_bar_idx,
225 				      OPAL_ENABLE_M64_SPLIT);
226 	if (rc != OPAL_SUCCESS) {
227 		desc = "enabling";
228 		goto fail;
229 	}
230 
231 	/*
232 	 * Exclude the segments for reserved and root bus PE, which
233 	 * are first or last two PEs.
234 	 */
235 	r = &phb->hose->mem_resources[1];
236 	if (phb->ioda.reserved_pe_idx == 0)
237 		r->start += (2 * phb->ioda.m64_segsize);
238 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
239 		r->end -= (2 * phb->ioda.m64_segsize);
240 	else
241 		pr_warn("  Cannot strip M64 segment for reserved PE#%x\n",
242 			phb->ioda.reserved_pe_idx);
243 
244 	return 0;
245 
246 fail:
247 	pr_warn("  Failure %lld %s M64 BAR#%d\n",
248 		rc, desc, phb->ioda.m64_bar_idx);
249 	opal_pci_phb_mmio_enable(phb->opal_id,
250 				 OPAL_M64_WINDOW_TYPE,
251 				 phb->ioda.m64_bar_idx,
252 				 OPAL_DISABLE_M64);
253 	return -EIO;
254 }
255 
256 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev *pdev,
257 					 unsigned long *pe_bitmap)
258 {
259 	struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
260 	struct resource *r;
261 	resource_size_t base, sgsz, start, end;
262 	int segno, i;
263 
264 	base = phb->ioda.m64_base;
265 	sgsz = phb->ioda.m64_segsize;
266 	for (i = 0; i <= PCI_ROM_RESOURCE; i++) {
267 		r = &pdev->resource[i];
268 		if (!r->parent || !pnv_pci_is_m64(phb, r))
269 			continue;
270 
271 		start = ALIGN_DOWN(r->start - base, sgsz);
272 		end = ALIGN(r->end - base, sgsz);
273 		for (segno = start / sgsz; segno < end / sgsz; segno++) {
274 			if (pe_bitmap)
275 				set_bit(segno, pe_bitmap);
276 			else
277 				pnv_ioda_reserve_pe(phb, segno);
278 		}
279 	}
280 }
281 
282 static int pnv_ioda1_init_m64(struct pnv_phb *phb)
283 {
284 	struct resource *r;
285 	int index;
286 
287 	/*
288 	 * There are 16 M64 BARs, each of which has 8 segments. So
289 	 * there are as many M64 segments as the maximum number of
290 	 * PEs, which is 128.
291 	 */
292 	for (index = 0; index < PNV_IODA1_M64_NUM; index++) {
293 		unsigned long base, segsz = phb->ioda.m64_segsize;
294 		int64_t rc;
295 
296 		base = phb->ioda.m64_base +
297 		       index * PNV_IODA1_M64_SEGS * segsz;
298 		rc = opal_pci_set_phb_mem_window(phb->opal_id,
299 				OPAL_M64_WINDOW_TYPE, index, base, 0,
300 				PNV_IODA1_M64_SEGS * segsz);
301 		if (rc != OPAL_SUCCESS) {
302 			pr_warn("  Error %lld setting M64 PHB#%x-BAR#%d\n",
303 				rc, phb->hose->global_number, index);
304 			goto fail;
305 		}
306 
307 		rc = opal_pci_phb_mmio_enable(phb->opal_id,
308 				OPAL_M64_WINDOW_TYPE, index,
309 				OPAL_ENABLE_M64_SPLIT);
310 		if (rc != OPAL_SUCCESS) {
311 			pr_warn("  Error %lld enabling M64 PHB#%x-BAR#%d\n",
312 				rc, phb->hose->global_number, index);
313 			goto fail;
314 		}
315 	}
316 
317 	for (index = 0; index < phb->ioda.total_pe_num; index++) {
318 		int64_t rc;
319 
320 		/*
321 		 * P7IOC supports M64DT, which helps mapping M64 segment
322 		 * to one particular PE#. However, PHB3 has fixed mapping
323 		 * between M64 segment and PE#. In order to have same logic
324 		 * for P7IOC and PHB3, we enforce fixed mapping between M64
325 		 * segment and PE# on P7IOC.
326 		 */
327 		rc = opal_pci_map_pe_mmio_window(phb->opal_id,
328 				index, OPAL_M64_WINDOW_TYPE,
329 				index / PNV_IODA1_M64_SEGS,
330 				index % PNV_IODA1_M64_SEGS);
331 		if (rc != OPAL_SUCCESS) {
332 			pr_warn("%s: Error %lld mapping M64 for PHB#%x-PE#%x\n",
333 				__func__, rc, phb->hose->global_number,
334 				index);
335 			goto fail;
336 		}
337 	}
338 
339 	/*
340 	 * Exclude the segments for reserved and root bus PE, which
341 	 * are first or last two PEs.
342 	 */
343 	r = &phb->hose->mem_resources[1];
344 	if (phb->ioda.reserved_pe_idx == 0)
345 		r->start += (2 * phb->ioda.m64_segsize);
346 	else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1))
347 		r->end -= (2 * phb->ioda.m64_segsize);
348 	else
349 		WARN(1, "Wrong reserved PE#%x on PHB#%x\n",
350 		     phb->ioda.reserved_pe_idx, phb->hose->global_number);
351 
352 	return 0;
353 
354 fail:
355 	for ( ; index >= 0; index--)
356 		opal_pci_phb_mmio_enable(phb->opal_id,
357 			OPAL_M64_WINDOW_TYPE, index, OPAL_DISABLE_M64);
358 
359 	return -EIO;
360 }
361 
362 static void pnv_ioda_reserve_m64_pe(struct pci_bus *bus,
363 				    unsigned long *pe_bitmap,
364 				    bool all)
365 {
366 	struct pci_dev *pdev;
367 
368 	list_for_each_entry(pdev, &bus->devices, bus_list) {
369 		pnv_ioda_reserve_dev_m64_pe(pdev, pe_bitmap);
370 
371 		if (all && pdev->subordinate)
372 			pnv_ioda_reserve_m64_pe(pdev->subordinate,
373 						pe_bitmap, all);
374 	}
375 }
376 
377 static struct pnv_ioda_pe *pnv_ioda_pick_m64_pe(struct pci_bus *bus, bool all)
378 {
379 	struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
380 	struct pnv_ioda_pe *master_pe, *pe;
381 	unsigned long size, *pe_alloc;
382 	int i;
383 
384 	/* Root bus shouldn't use M64 */
385 	if (pci_is_root_bus(bus))
386 		return NULL;
387 
388 	/* Allocate bitmap */
389 	size = ALIGN(phb->ioda.total_pe_num / 8, sizeof(unsigned long));
390 	pe_alloc = kzalloc(size, GFP_KERNEL);
391 	if (!pe_alloc) {
392 		pr_warn("%s: Out of memory !\n",
393 			__func__);
394 		return NULL;
395 	}
396 
397 	/* Figure out reserved PE numbers by the PE */
398 	pnv_ioda_reserve_m64_pe(bus, pe_alloc, all);
399 
400 	/*
401 	 * the current bus might not own M64 window and that's all
402 	 * contributed by its child buses. For the case, we needn't
403 	 * pick M64 dependent PE#.
404 	 */
405 	if (bitmap_empty(pe_alloc, phb->ioda.total_pe_num)) {
406 		kfree(pe_alloc);
407 		return NULL;
408 	}
409 
410 	/*
411 	 * Figure out the master PE and put all slave PEs to master
412 	 * PE's list to form compound PE.
413 	 */
414 	master_pe = NULL;
415 	i = -1;
416 	while ((i = find_next_bit(pe_alloc, phb->ioda.total_pe_num, i + 1)) <
417 		phb->ioda.total_pe_num) {
418 		pe = &phb->ioda.pe_array[i];
419 
420 		phb->ioda.m64_segmap[pe->pe_number] = pe->pe_number;
421 		if (!master_pe) {
422 			pe->flags |= PNV_IODA_PE_MASTER;
423 			INIT_LIST_HEAD(&pe->slaves);
424 			master_pe = pe;
425 		} else {
426 			pe->flags |= PNV_IODA_PE_SLAVE;
427 			pe->master = master_pe;
428 			list_add_tail(&pe->list, &master_pe->slaves);
429 		}
430 	}
431 
432 	kfree(pe_alloc);
433 	return master_pe;
434 }
435 
436 static void __init pnv_ioda_parse_m64_window(struct pnv_phb *phb)
437 {
438 	struct pci_controller *hose = phb->hose;
439 	struct device_node *dn = hose->dn;
440 	struct resource *res;
441 	u32 m64_range[2], i;
442 	const __be32 *r;
443 	u64 pci_addr;
444 
445 	if (phb->type != PNV_PHB_IODA1 && phb->type != PNV_PHB_IODA2) {
446 		pr_info("  Not support M64 window\n");
447 		return;
448 	}
449 
450 	if (!firmware_has_feature(FW_FEATURE_OPAL)) {
451 		pr_info("  Firmware too old to support M64 window\n");
452 		return;
453 	}
454 
455 	r = of_get_property(dn, "ibm,opal-m64-window", NULL);
456 	if (!r) {
457 		pr_info("  No <ibm,opal-m64-window> on %pOF\n",
458 			dn);
459 		return;
460 	}
461 
462 	/*
463 	 * Find the available M64 BAR range and pickup the last one for
464 	 * covering the whole 64-bits space. We support only one range.
465 	 */
466 	if (of_property_read_u32_array(dn, "ibm,opal-available-m64-ranges",
467 				       m64_range, 2)) {
468 		/* In absence of the property, assume 0..15 */
469 		m64_range[0] = 0;
470 		m64_range[1] = 16;
471 	}
472 	/* We only support 64 bits in our allocator */
473 	if (m64_range[1] > 63) {
474 		pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n",
475 			__func__, m64_range[1], phb->hose->global_number);
476 		m64_range[1] = 63;
477 	}
478 	/* Empty range, no m64 */
479 	if (m64_range[1] <= m64_range[0]) {
480 		pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n",
481 			__func__, phb->hose->global_number);
482 		return;
483 	}
484 
485 	/* Configure M64 informations */
486 	res = &hose->mem_resources[1];
487 	res->name = dn->full_name;
488 	res->start = of_translate_address(dn, r + 2);
489 	res->end = res->start + of_read_number(r + 4, 2) - 1;
490 	res->flags = (IORESOURCE_MEM | IORESOURCE_MEM_64 | IORESOURCE_PREFETCH);
491 	pci_addr = of_read_number(r, 2);
492 	hose->mem_offset[1] = res->start - pci_addr;
493 
494 	phb->ioda.m64_size = resource_size(res);
495 	phb->ioda.m64_segsize = phb->ioda.m64_size / phb->ioda.total_pe_num;
496 	phb->ioda.m64_base = pci_addr;
497 
498 	/* This lines up nicely with the display from processing OF ranges */
499 	pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n",
500 		res->start, res->end, pci_addr, m64_range[0],
501 		m64_range[0] + m64_range[1] - 1);
502 
503 	/* Mark all M64 used up by default */
504 	phb->ioda.m64_bar_alloc = (unsigned long)-1;
505 
506 	/* Use last M64 BAR to cover M64 window */
507 	m64_range[1]--;
508 	phb->ioda.m64_bar_idx = m64_range[0] + m64_range[1];
509 
510 	pr_info(" Using M64 #%d as default window\n", phb->ioda.m64_bar_idx);
511 
512 	/* Mark remaining ones free */
513 	for (i = m64_range[0]; i < m64_range[1]; i++)
514 		clear_bit(i, &phb->ioda.m64_bar_alloc);
515 
516 	/*
517 	 * Setup init functions for M64 based on IODA version, IODA3 uses
518 	 * the IODA2 code.
519 	 */
520 	if (phb->type == PNV_PHB_IODA1)
521 		phb->init_m64 = pnv_ioda1_init_m64;
522 	else
523 		phb->init_m64 = pnv_ioda2_init_m64;
524 }
525 
526 static void pnv_ioda_freeze_pe(struct pnv_phb *phb, int pe_no)
527 {
528 	struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_no];
529 	struct pnv_ioda_pe *slave;
530 	s64 rc;
531 
532 	/* Fetch master PE */
533 	if (pe->flags & PNV_IODA_PE_SLAVE) {
534 		pe = pe->master;
535 		if (WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER)))
536 			return;
537 
538 		pe_no = pe->pe_number;
539 	}
540 
541 	/* Freeze master PE */
542 	rc = opal_pci_eeh_freeze_set(phb->opal_id,
543 				     pe_no,
544 				     OPAL_EEH_ACTION_SET_FREEZE_ALL);
545 	if (rc != OPAL_SUCCESS) {
546 		pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
547 			__func__, rc, phb->hose->global_number, pe_no);
548 		return;
549 	}
550 
551 	/* Freeze slave PEs */
552 	if (!(pe->flags & PNV_IODA_PE_MASTER))
553 		return;
554 
555 	list_for_each_entry(slave, &pe->slaves, list) {
556 		rc = opal_pci_eeh_freeze_set(phb->opal_id,
557 					     slave->pe_number,
558 					     OPAL_EEH_ACTION_SET_FREEZE_ALL);
559 		if (rc != OPAL_SUCCESS)
560 			pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
561 				__func__, rc, phb->hose->global_number,
562 				slave->pe_number);
563 	}
564 }
565 
566 static int pnv_ioda_unfreeze_pe(struct pnv_phb *phb, int pe_no, int opt)
567 {
568 	struct pnv_ioda_pe *pe, *slave;
569 	s64 rc;
570 
571 	/* Find master PE */
572 	pe = &phb->ioda.pe_array[pe_no];
573 	if (pe->flags & PNV_IODA_PE_SLAVE) {
574 		pe = pe->master;
575 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
576 		pe_no = pe->pe_number;
577 	}
578 
579 	/* Clear frozen state for master PE */
580 	rc = opal_pci_eeh_freeze_clear(phb->opal_id, pe_no, opt);
581 	if (rc != OPAL_SUCCESS) {
582 		pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
583 			__func__, rc, opt, phb->hose->global_number, pe_no);
584 		return -EIO;
585 	}
586 
587 	if (!(pe->flags & PNV_IODA_PE_MASTER))
588 		return 0;
589 
590 	/* Clear frozen state for slave PEs */
591 	list_for_each_entry(slave, &pe->slaves, list) {
592 		rc = opal_pci_eeh_freeze_clear(phb->opal_id,
593 					     slave->pe_number,
594 					     opt);
595 		if (rc != OPAL_SUCCESS) {
596 			pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
597 				__func__, rc, opt, phb->hose->global_number,
598 				slave->pe_number);
599 			return -EIO;
600 		}
601 	}
602 
603 	return 0;
604 }
605 
606 static int pnv_ioda_get_pe_state(struct pnv_phb *phb, int pe_no)
607 {
608 	struct pnv_ioda_pe *slave, *pe;
609 	u8 fstate = 0, state;
610 	__be16 pcierr = 0;
611 	s64 rc;
612 
613 	/* Sanity check on PE number */
614 	if (pe_no < 0 || pe_no >= phb->ioda.total_pe_num)
615 		return OPAL_EEH_STOPPED_PERM_UNAVAIL;
616 
617 	/*
618 	 * Fetch the master PE and the PE instance might be
619 	 * not initialized yet.
620 	 */
621 	pe = &phb->ioda.pe_array[pe_no];
622 	if (pe->flags & PNV_IODA_PE_SLAVE) {
623 		pe = pe->master;
624 		WARN_ON(!pe || !(pe->flags & PNV_IODA_PE_MASTER));
625 		pe_no = pe->pe_number;
626 	}
627 
628 	/* Check the master PE */
629 	rc = opal_pci_eeh_freeze_status(phb->opal_id, pe_no,
630 					&state, &pcierr, NULL);
631 	if (rc != OPAL_SUCCESS) {
632 		pr_warn("%s: Failure %lld getting "
633 			"PHB#%x-PE#%x state\n",
634 			__func__, rc,
635 			phb->hose->global_number, pe_no);
636 		return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
637 	}
638 
639 	/* Check the slave PE */
640 	if (!(pe->flags & PNV_IODA_PE_MASTER))
641 		return state;
642 
643 	list_for_each_entry(slave, &pe->slaves, list) {
644 		rc = opal_pci_eeh_freeze_status(phb->opal_id,
645 						slave->pe_number,
646 						&fstate,
647 						&pcierr,
648 						NULL);
649 		if (rc != OPAL_SUCCESS) {
650 			pr_warn("%s: Failure %lld getting "
651 				"PHB#%x-PE#%x state\n",
652 				__func__, rc,
653 				phb->hose->global_number, slave->pe_number);
654 			return OPAL_EEH_STOPPED_TEMP_UNAVAIL;
655 		}
656 
657 		/*
658 		 * Override the result based on the ascending
659 		 * priority.
660 		 */
661 		if (fstate > state)
662 			state = fstate;
663 	}
664 
665 	return state;
666 }
667 
668 struct pnv_ioda_pe *pnv_pci_bdfn_to_pe(struct pnv_phb *phb, u16 bdfn)
669 {
670 	int pe_number = phb->ioda.pe_rmap[bdfn];
671 
672 	if (pe_number == IODA_INVALID_PE)
673 		return NULL;
674 
675 	return &phb->ioda.pe_array[pe_number];
676 }
677 
678 struct pnv_ioda_pe *pnv_ioda_get_pe(struct pci_dev *dev)
679 {
680 	struct pnv_phb *phb = pci_bus_to_pnvhb(dev->bus);
681 	struct pci_dn *pdn = pci_get_pdn(dev);
682 
683 	if (!pdn)
684 		return NULL;
685 	if (pdn->pe_number == IODA_INVALID_PE)
686 		return NULL;
687 	return &phb->ioda.pe_array[pdn->pe_number];
688 }
689 
690 static int pnv_ioda_set_one_peltv(struct pnv_phb *phb,
691 				  struct pnv_ioda_pe *parent,
692 				  struct pnv_ioda_pe *child,
693 				  bool is_add)
694 {
695 	const char *desc = is_add ? "adding" : "removing";
696 	uint8_t op = is_add ? OPAL_ADD_PE_TO_DOMAIN :
697 			      OPAL_REMOVE_PE_FROM_DOMAIN;
698 	struct pnv_ioda_pe *slave;
699 	long rc;
700 
701 	/* Parent PE affects child PE */
702 	rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
703 				child->pe_number, op);
704 	if (rc != OPAL_SUCCESS) {
705 		pe_warn(child, "OPAL error %ld %s to parent PELTV\n",
706 			rc, desc);
707 		return -ENXIO;
708 	}
709 
710 	if (!(child->flags & PNV_IODA_PE_MASTER))
711 		return 0;
712 
713 	/* Compound case: parent PE affects slave PEs */
714 	list_for_each_entry(slave, &child->slaves, list) {
715 		rc = opal_pci_set_peltv(phb->opal_id, parent->pe_number,
716 					slave->pe_number, op);
717 		if (rc != OPAL_SUCCESS) {
718 			pe_warn(slave, "OPAL error %ld %s to parent PELTV\n",
719 				rc, desc);
720 			return -ENXIO;
721 		}
722 	}
723 
724 	return 0;
725 }
726 
727 static int pnv_ioda_set_peltv(struct pnv_phb *phb,
728 			      struct pnv_ioda_pe *pe,
729 			      bool is_add)
730 {
731 	struct pnv_ioda_pe *slave;
732 	struct pci_dev *pdev = NULL;
733 	int ret;
734 
735 	/*
736 	 * Clear PE frozen state. If it's master PE, we need
737 	 * clear slave PE frozen state as well.
738 	 */
739 	if (is_add) {
740 		opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
741 					  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
742 		if (pe->flags & PNV_IODA_PE_MASTER) {
743 			list_for_each_entry(slave, &pe->slaves, list)
744 				opal_pci_eeh_freeze_clear(phb->opal_id,
745 							  slave->pe_number,
746 							  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
747 		}
748 	}
749 
750 	/*
751 	 * Associate PE in PELT. We need add the PE into the
752 	 * corresponding PELT-V as well. Otherwise, the error
753 	 * originated from the PE might contribute to other
754 	 * PEs.
755 	 */
756 	ret = pnv_ioda_set_one_peltv(phb, pe, pe, is_add);
757 	if (ret)
758 		return ret;
759 
760 	/* For compound PEs, any one affects all of them */
761 	if (pe->flags & PNV_IODA_PE_MASTER) {
762 		list_for_each_entry(slave, &pe->slaves, list) {
763 			ret = pnv_ioda_set_one_peltv(phb, slave, pe, is_add);
764 			if (ret)
765 				return ret;
766 		}
767 	}
768 
769 	if (pe->flags & (PNV_IODA_PE_BUS_ALL | PNV_IODA_PE_BUS))
770 		pdev = pe->pbus->self;
771 	else if (pe->flags & PNV_IODA_PE_DEV)
772 		pdev = pe->pdev->bus->self;
773 #ifdef CONFIG_PCI_IOV
774 	else if (pe->flags & PNV_IODA_PE_VF)
775 		pdev = pe->parent_dev;
776 #endif /* CONFIG_PCI_IOV */
777 	while (pdev) {
778 		struct pci_dn *pdn = pci_get_pdn(pdev);
779 		struct pnv_ioda_pe *parent;
780 
781 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
782 			parent = &phb->ioda.pe_array[pdn->pe_number];
783 			ret = pnv_ioda_set_one_peltv(phb, parent, pe, is_add);
784 			if (ret)
785 				return ret;
786 		}
787 
788 		pdev = pdev->bus->self;
789 	}
790 
791 	return 0;
792 }
793 
794 static void pnv_ioda_unset_peltv(struct pnv_phb *phb,
795 				 struct pnv_ioda_pe *pe,
796 				 struct pci_dev *parent)
797 {
798 	int64_t rc;
799 
800 	while (parent) {
801 		struct pci_dn *pdn = pci_get_pdn(parent);
802 
803 		if (pdn && pdn->pe_number != IODA_INVALID_PE) {
804 			rc = opal_pci_set_peltv(phb->opal_id, pdn->pe_number,
805 						pe->pe_number,
806 						OPAL_REMOVE_PE_FROM_DOMAIN);
807 			/* XXX What to do in case of error ? */
808 		}
809 		parent = parent->bus->self;
810 	}
811 
812 	opal_pci_eeh_freeze_clear(phb->opal_id, pe->pe_number,
813 				  OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
814 
815 	/* Disassociate PE in PELT */
816 	rc = opal_pci_set_peltv(phb->opal_id, pe->pe_number,
817 				pe->pe_number, OPAL_REMOVE_PE_FROM_DOMAIN);
818 	if (rc)
819 		pe_warn(pe, "OPAL error %lld remove self from PELTV\n", rc);
820 }
821 
822 int pnv_ioda_deconfigure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
823 {
824 	struct pci_dev *parent;
825 	uint8_t bcomp, dcomp, fcomp;
826 	int64_t rc;
827 	long rid_end, rid;
828 
829 	/* Currently, we just deconfigure VF PE. Bus PE will always there.*/
830 	if (pe->pbus) {
831 		int count;
832 
833 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
834 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
835 		parent = pe->pbus->self;
836 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
837 			count = resource_size(&pe->pbus->busn_res);
838 		else
839 			count = 1;
840 
841 		switch(count) {
842 		case  1: bcomp = OpalPciBusAll;         break;
843 		case  2: bcomp = OpalPciBus7Bits;       break;
844 		case  4: bcomp = OpalPciBus6Bits;       break;
845 		case  8: bcomp = OpalPciBus5Bits;       break;
846 		case 16: bcomp = OpalPciBus4Bits;       break;
847 		case 32: bcomp = OpalPciBus3Bits;       break;
848 		default:
849 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
850 			        count);
851 			/* Do an exact match only */
852 			bcomp = OpalPciBusAll;
853 		}
854 		rid_end = pe->rid + (count << 8);
855 	} else {
856 #ifdef CONFIG_PCI_IOV
857 		if (pe->flags & PNV_IODA_PE_VF)
858 			parent = pe->parent_dev;
859 		else
860 #endif
861 			parent = pe->pdev->bus->self;
862 		bcomp = OpalPciBusAll;
863 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
864 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
865 		rid_end = pe->rid + 1;
866 	}
867 
868 	/* Clear the reverse map */
869 	for (rid = pe->rid; rid < rid_end; rid++)
870 		phb->ioda.pe_rmap[rid] = IODA_INVALID_PE;
871 
872 	/*
873 	 * Release from all parents PELT-V. NPUs don't have a PELTV
874 	 * table
875 	 */
876 	if (phb->type != PNV_PHB_NPU_OCAPI)
877 		pnv_ioda_unset_peltv(phb, pe, parent);
878 
879 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
880 			     bcomp, dcomp, fcomp, OPAL_UNMAP_PE);
881 	if (rc)
882 		pe_err(pe, "OPAL error %lld trying to setup PELT table\n", rc);
883 
884 	pe->pbus = NULL;
885 	pe->pdev = NULL;
886 #ifdef CONFIG_PCI_IOV
887 	pe->parent_dev = NULL;
888 #endif
889 
890 	return 0;
891 }
892 
893 int pnv_ioda_configure_pe(struct pnv_phb *phb, struct pnv_ioda_pe *pe)
894 {
895 	uint8_t bcomp, dcomp, fcomp;
896 	long rc, rid_end, rid;
897 
898 	/* Bus validation ? */
899 	if (pe->pbus) {
900 		int count;
901 
902 		dcomp = OPAL_IGNORE_RID_DEVICE_NUMBER;
903 		fcomp = OPAL_IGNORE_RID_FUNCTION_NUMBER;
904 		if (pe->flags & PNV_IODA_PE_BUS_ALL)
905 			count = resource_size(&pe->pbus->busn_res);
906 		else
907 			count = 1;
908 
909 		switch(count) {
910 		case  1: bcomp = OpalPciBusAll;		break;
911 		case  2: bcomp = OpalPciBus7Bits;	break;
912 		case  4: bcomp = OpalPciBus6Bits;	break;
913 		case  8: bcomp = OpalPciBus5Bits;	break;
914 		case 16: bcomp = OpalPciBus4Bits;	break;
915 		case 32: bcomp = OpalPciBus3Bits;	break;
916 		default:
917 			dev_err(&pe->pbus->dev, "Number of subordinate buses %d unsupported\n",
918 			        count);
919 			/* Do an exact match only */
920 			bcomp = OpalPciBusAll;
921 		}
922 		rid_end = pe->rid + (count << 8);
923 	} else {
924 		bcomp = OpalPciBusAll;
925 		dcomp = OPAL_COMPARE_RID_DEVICE_NUMBER;
926 		fcomp = OPAL_COMPARE_RID_FUNCTION_NUMBER;
927 		rid_end = pe->rid + 1;
928 	}
929 
930 	/*
931 	 * Associate PE in PELT. We need add the PE into the
932 	 * corresponding PELT-V as well. Otherwise, the error
933 	 * originated from the PE might contribute to other
934 	 * PEs.
935 	 */
936 	rc = opal_pci_set_pe(phb->opal_id, pe->pe_number, pe->rid,
937 			     bcomp, dcomp, fcomp, OPAL_MAP_PE);
938 	if (rc) {
939 		pe_err(pe, "OPAL error %ld trying to setup PELT table\n", rc);
940 		return -ENXIO;
941 	}
942 
943 	/*
944 	 * Configure PELTV. NPUs don't have a PELTV table so skip
945 	 * configuration on them.
946 	 */
947 	if (phb->type != PNV_PHB_NPU_OCAPI)
948 		pnv_ioda_set_peltv(phb, pe, true);
949 
950 	/* Setup reverse map */
951 	for (rid = pe->rid; rid < rid_end; rid++)
952 		phb->ioda.pe_rmap[rid] = pe->pe_number;
953 
954 	/* Setup one MVTs on IODA1 */
955 	if (phb->type != PNV_PHB_IODA1) {
956 		pe->mve_number = 0;
957 		goto out;
958 	}
959 
960 	pe->mve_number = pe->pe_number;
961 	rc = opal_pci_set_mve(phb->opal_id, pe->mve_number, pe->pe_number);
962 	if (rc != OPAL_SUCCESS) {
963 		pe_err(pe, "OPAL error %ld setting up MVE %x\n",
964 		       rc, pe->mve_number);
965 		pe->mve_number = -1;
966 	} else {
967 		rc = opal_pci_set_mve_enable(phb->opal_id,
968 					     pe->mve_number, OPAL_ENABLE_MVE);
969 		if (rc) {
970 			pe_err(pe, "OPAL error %ld enabling MVE %x\n",
971 			       rc, pe->mve_number);
972 			pe->mve_number = -1;
973 		}
974 	}
975 
976 out:
977 	return 0;
978 }
979 
980 static struct pnv_ioda_pe *pnv_ioda_setup_dev_PE(struct pci_dev *dev)
981 {
982 	struct pnv_phb *phb = pci_bus_to_pnvhb(dev->bus);
983 	struct pci_dn *pdn = pci_get_pdn(dev);
984 	struct pnv_ioda_pe *pe;
985 
986 	if (!pdn) {
987 		pr_err("%s: Device tree node not associated properly\n",
988 			   pci_name(dev));
989 		return NULL;
990 	}
991 	if (pdn->pe_number != IODA_INVALID_PE)
992 		return NULL;
993 
994 	pe = pnv_ioda_alloc_pe(phb, 1);
995 	if (!pe) {
996 		pr_warn("%s: Not enough PE# available, disabling device\n",
997 			pci_name(dev));
998 		return NULL;
999 	}
1000 
1001 	/* NOTE: We don't get a reference for the pointer in the PE
1002 	 * data structure, both the device and PE structures should be
1003 	 * destroyed at the same time.
1004 	 *
1005 	 * At some point we want to remove the PDN completely anyways
1006 	 */
1007 	pdn->pe_number = pe->pe_number;
1008 	pe->flags = PNV_IODA_PE_DEV;
1009 	pe->pdev = dev;
1010 	pe->pbus = NULL;
1011 	pe->mve_number = -1;
1012 	pe->rid = dev->bus->number << 8 | pdn->devfn;
1013 	pe->device_count++;
1014 
1015 	pe_info(pe, "Associated device to PE\n");
1016 
1017 	if (pnv_ioda_configure_pe(phb, pe)) {
1018 		/* XXX What do we do here ? */
1019 		pnv_ioda_free_pe(pe);
1020 		pdn->pe_number = IODA_INVALID_PE;
1021 		pe->pdev = NULL;
1022 		return NULL;
1023 	}
1024 
1025 	/* Put PE to the list */
1026 	mutex_lock(&phb->ioda.pe_list_mutex);
1027 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1028 	mutex_unlock(&phb->ioda.pe_list_mutex);
1029 	return pe;
1030 }
1031 
1032 /*
1033  * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1034  * single PCI bus. Another one that contains the primary PCI bus and its
1035  * subordinate PCI devices and buses. The second type of PE is normally
1036  * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1037  */
1038 static struct pnv_ioda_pe *pnv_ioda_setup_bus_PE(struct pci_bus *bus, bool all)
1039 {
1040 	struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
1041 	struct pnv_ioda_pe *pe = NULL;
1042 	unsigned int pe_num;
1043 
1044 	/*
1045 	 * In partial hotplug case, the PE instance might be still alive.
1046 	 * We should reuse it instead of allocating a new one.
1047 	 */
1048 	pe_num = phb->ioda.pe_rmap[bus->number << 8];
1049 	if (WARN_ON(pe_num != IODA_INVALID_PE)) {
1050 		pe = &phb->ioda.pe_array[pe_num];
1051 		return NULL;
1052 	}
1053 
1054 	/* PE number for root bus should have been reserved */
1055 	if (pci_is_root_bus(bus))
1056 		pe = &phb->ioda.pe_array[phb->ioda.root_pe_idx];
1057 
1058 	/* Check if PE is determined by M64 */
1059 	if (!pe)
1060 		pe = pnv_ioda_pick_m64_pe(bus, all);
1061 
1062 	/* The PE number isn't pinned by M64 */
1063 	if (!pe)
1064 		pe = pnv_ioda_alloc_pe(phb, 1);
1065 
1066 	if (!pe) {
1067 		pr_warn("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1068 			__func__, pci_domain_nr(bus), bus->number);
1069 		return NULL;
1070 	}
1071 
1072 	pe->flags |= (all ? PNV_IODA_PE_BUS_ALL : PNV_IODA_PE_BUS);
1073 	pe->pbus = bus;
1074 	pe->pdev = NULL;
1075 	pe->mve_number = -1;
1076 	pe->rid = bus->busn_res.start << 8;
1077 
1078 	if (all)
1079 		pe_info(pe, "Secondary bus %pad..%pad associated with PE#%x\n",
1080 			&bus->busn_res.start, &bus->busn_res.end,
1081 			pe->pe_number);
1082 	else
1083 		pe_info(pe, "Secondary bus %pad associated with PE#%x\n",
1084 			&bus->busn_res.start, pe->pe_number);
1085 
1086 	if (pnv_ioda_configure_pe(phb, pe)) {
1087 		/* XXX What do we do here ? */
1088 		pnv_ioda_free_pe(pe);
1089 		pe->pbus = NULL;
1090 		return NULL;
1091 	}
1092 
1093 	/* Put PE to the list */
1094 	list_add_tail(&pe->list, &phb->ioda.pe_list);
1095 
1096 	return pe;
1097 }
1098 
1099 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
1100 				       struct pnv_ioda_pe *pe);
1101 
1102 static void pnv_pci_ioda_dma_dev_setup(struct pci_dev *pdev)
1103 {
1104 	struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
1105 	struct pci_dn *pdn = pci_get_pdn(pdev);
1106 	struct pnv_ioda_pe *pe;
1107 
1108 	/* Check if the BDFN for this device is associated with a PE yet */
1109 	pe = pnv_pci_bdfn_to_pe(phb, pdev->devfn | (pdev->bus->number << 8));
1110 	if (!pe) {
1111 		/* VF PEs should be pre-configured in pnv_pci_sriov_enable() */
1112 		if (WARN_ON(pdev->is_virtfn))
1113 			return;
1114 
1115 		pnv_pci_configure_bus(pdev->bus);
1116 		pe = pnv_pci_bdfn_to_pe(phb, pdev->devfn | (pdev->bus->number << 8));
1117 		pci_info(pdev, "Configured PE#%x\n", pe ? pe->pe_number : 0xfffff);
1118 
1119 
1120 		/*
1121 		 * If we can't setup the IODA PE something has gone horribly
1122 		 * wrong and we can't enable DMA for the device.
1123 		 */
1124 		if (WARN_ON(!pe))
1125 			return;
1126 	} else {
1127 		pci_info(pdev, "Added to existing PE#%x\n", pe->pe_number);
1128 	}
1129 
1130 	/*
1131 	 * We assume that bridges *probably* don't need to do any DMA so we can
1132 	 * skip allocating a TCE table, etc unless we get a non-bridge device.
1133 	 */
1134 	if (!pe->dma_setup_done && !pci_is_bridge(pdev)) {
1135 		switch (phb->type) {
1136 		case PNV_PHB_IODA1:
1137 			pnv_pci_ioda1_setup_dma_pe(phb, pe);
1138 			break;
1139 		case PNV_PHB_IODA2:
1140 			pnv_pci_ioda2_setup_dma_pe(phb, pe);
1141 			break;
1142 		default:
1143 			pr_warn("%s: No DMA for PHB#%x (type %d)\n",
1144 				__func__, phb->hose->global_number, phb->type);
1145 		}
1146 	}
1147 
1148 	if (pdn)
1149 		pdn->pe_number = pe->pe_number;
1150 	pe->device_count++;
1151 
1152 	WARN_ON(get_dma_ops(&pdev->dev) != &dma_iommu_ops);
1153 	pdev->dev.archdata.dma_offset = pe->tce_bypass_base;
1154 	set_iommu_table_base(&pdev->dev, pe->table_group.tables[0]);
1155 
1156 	/* PEs with a DMA weight of zero won't have a group */
1157 	if (pe->table_group.group)
1158 		iommu_add_device(&pe->table_group, &pdev->dev);
1159 }
1160 
1161 /*
1162  * Reconfigure TVE#0 to be usable as 64-bit DMA space.
1163  *
1164  * The first 4GB of virtual memory for a PE is reserved for 32-bit accesses.
1165  * Devices can only access more than that if bit 59 of the PCI address is set
1166  * by hardware, which indicates TVE#1 should be used instead of TVE#0.
1167  * Many PCI devices are not capable of addressing that many bits, and as a
1168  * result are limited to the 4GB of virtual memory made available to 32-bit
1169  * devices in TVE#0.
1170  *
1171  * In order to work around this, reconfigure TVE#0 to be suitable for 64-bit
1172  * devices by configuring the virtual memory past the first 4GB inaccessible
1173  * by 64-bit DMAs.  This should only be used by devices that want more than
1174  * 4GB, and only on PEs that have no 32-bit devices.
1175  *
1176  * Currently this will only work on PHB3 (POWER8).
1177  */
1178 static int pnv_pci_ioda_dma_64bit_bypass(struct pnv_ioda_pe *pe)
1179 {
1180 	u64 window_size, table_size, tce_count, addr;
1181 	struct page *table_pages;
1182 	u64 tce_order = 28; /* 256MB TCEs */
1183 	__be64 *tces;
1184 	s64 rc;
1185 
1186 	/*
1187 	 * Window size needs to be a power of two, but needs to account for
1188 	 * shifting memory by the 4GB offset required to skip 32bit space.
1189 	 */
1190 	window_size = roundup_pow_of_two(memory_hotplug_max() + (1ULL << 32));
1191 	tce_count = window_size >> tce_order;
1192 	table_size = tce_count << 3;
1193 
1194 	if (table_size < PAGE_SIZE)
1195 		table_size = PAGE_SIZE;
1196 
1197 	table_pages = alloc_pages_node(pe->phb->hose->node, GFP_KERNEL,
1198 				       get_order(table_size));
1199 	if (!table_pages)
1200 		goto err;
1201 
1202 	tces = page_address(table_pages);
1203 	if (!tces)
1204 		goto err;
1205 
1206 	memset(tces, 0, table_size);
1207 
1208 	for (addr = 0; addr < memory_hotplug_max(); addr += (1 << tce_order)) {
1209 		tces[(addr + (1ULL << 32)) >> tce_order] =
1210 			cpu_to_be64(addr | TCE_PCI_READ | TCE_PCI_WRITE);
1211 	}
1212 
1213 	rc = opal_pci_map_pe_dma_window(pe->phb->opal_id,
1214 					pe->pe_number,
1215 					/* reconfigure window 0 */
1216 					(pe->pe_number << 1) + 0,
1217 					1,
1218 					__pa(tces),
1219 					table_size,
1220 					1 << tce_order);
1221 	if (rc == OPAL_SUCCESS) {
1222 		pe_info(pe, "Using 64-bit DMA iommu bypass (through TVE#0)\n");
1223 		return 0;
1224 	}
1225 err:
1226 	pe_err(pe, "Error configuring 64-bit DMA bypass\n");
1227 	return -EIO;
1228 }
1229 
1230 static bool pnv_pci_ioda_iommu_bypass_supported(struct pci_dev *pdev,
1231 		u64 dma_mask)
1232 {
1233 	struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
1234 	struct pci_dn *pdn = pci_get_pdn(pdev);
1235 	struct pnv_ioda_pe *pe;
1236 
1237 	if (WARN_ON(!pdn || pdn->pe_number == IODA_INVALID_PE))
1238 		return false;
1239 
1240 	pe = &phb->ioda.pe_array[pdn->pe_number];
1241 	if (pe->tce_bypass_enabled) {
1242 		u64 top = pe->tce_bypass_base + memblock_end_of_DRAM() - 1;
1243 		if (dma_mask >= top)
1244 			return true;
1245 	}
1246 
1247 	/*
1248 	 * If the device can't set the TCE bypass bit but still wants
1249 	 * to access 4GB or more, on PHB3 we can reconfigure TVE#0 to
1250 	 * bypass the 32-bit region and be usable for 64-bit DMAs.
1251 	 * The device needs to be able to address all of this space.
1252 	 */
1253 	if (dma_mask >> 32 &&
1254 	    dma_mask > (memory_hotplug_max() + (1ULL << 32)) &&
1255 	    /* pe->pdev should be set if it's a single device, pe->pbus if not */
1256 	    (pe->device_count == 1 || !pe->pbus) &&
1257 	    phb->model == PNV_PHB_MODEL_PHB3) {
1258 		/* Configure the bypass mode */
1259 		s64 rc = pnv_pci_ioda_dma_64bit_bypass(pe);
1260 		if (rc)
1261 			return false;
1262 		/* 4GB offset bypasses 32-bit space */
1263 		pdev->dev.archdata.dma_offset = (1ULL << 32);
1264 		return true;
1265 	}
1266 
1267 	return false;
1268 }
1269 
1270 static inline __be64 __iomem *pnv_ioda_get_inval_reg(struct pnv_phb *phb,
1271 						     bool real_mode)
1272 {
1273 	return real_mode ? (__be64 __iomem *)(phb->regs_phys + 0x210) :
1274 		(phb->regs + 0x210);
1275 }
1276 
1277 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table *tbl,
1278 		unsigned long index, unsigned long npages, bool rm)
1279 {
1280 	struct iommu_table_group_link *tgl = list_first_entry_or_null(
1281 			&tbl->it_group_list, struct iommu_table_group_link,
1282 			next);
1283 	struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1284 			struct pnv_ioda_pe, table_group);
1285 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
1286 	unsigned long start, end, inc;
1287 
1288 	start = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset);
1289 	end = __pa(((__be64 *)tbl->it_base) + index - tbl->it_offset +
1290 			npages - 1);
1291 
1292 	/* p7ioc-style invalidation, 2 TCEs per write */
1293 	start |= (1ull << 63);
1294 	end |= (1ull << 63);
1295 	inc = 16;
1296         end |= inc - 1;	/* round up end to be different than start */
1297 
1298         mb(); /* Ensure above stores are visible */
1299         while (start <= end) {
1300 		if (rm)
1301 			__raw_rm_writeq_be(start, invalidate);
1302 		else
1303 			__raw_writeq_be(start, invalidate);
1304 
1305                 start += inc;
1306         }
1307 
1308 	/*
1309 	 * The iommu layer will do another mb() for us on build()
1310 	 * and we don't care on free()
1311 	 */
1312 }
1313 
1314 static int pnv_ioda1_tce_build(struct iommu_table *tbl, long index,
1315 		long npages, unsigned long uaddr,
1316 		enum dma_data_direction direction,
1317 		unsigned long attrs)
1318 {
1319 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1320 			attrs);
1321 
1322 	if (!ret)
1323 		pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1324 
1325 	return ret;
1326 }
1327 
1328 #ifdef CONFIG_IOMMU_API
1329 /* Common for IODA1 and IODA2 */
1330 static int pnv_ioda_tce_xchg_no_kill(struct iommu_table *tbl, long index,
1331 		unsigned long *hpa, enum dma_data_direction *direction,
1332 		bool realmode)
1333 {
1334 	return pnv_tce_xchg(tbl, index, hpa, direction, !realmode);
1335 }
1336 #endif
1337 
1338 static void pnv_ioda1_tce_free(struct iommu_table *tbl, long index,
1339 		long npages)
1340 {
1341 	pnv_tce_free(tbl, index, npages);
1342 
1343 	pnv_pci_p7ioc_tce_invalidate(tbl, index, npages, false);
1344 }
1345 
1346 static struct iommu_table_ops pnv_ioda1_iommu_ops = {
1347 	.set = pnv_ioda1_tce_build,
1348 #ifdef CONFIG_IOMMU_API
1349 	.xchg_no_kill = pnv_ioda_tce_xchg_no_kill,
1350 	.tce_kill = pnv_pci_p7ioc_tce_invalidate,
1351 	.useraddrptr = pnv_tce_useraddrptr,
1352 #endif
1353 	.clear = pnv_ioda1_tce_free,
1354 	.get = pnv_tce_get,
1355 };
1356 
1357 #define PHB3_TCE_KILL_INVAL_ALL		PPC_BIT(0)
1358 #define PHB3_TCE_KILL_INVAL_PE		PPC_BIT(1)
1359 #define PHB3_TCE_KILL_INVAL_ONE		PPC_BIT(2)
1360 
1361 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1362 {
1363 	/* 01xb - invalidate TCEs that match the specified PE# */
1364 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, false);
1365 	unsigned long val = PHB3_TCE_KILL_INVAL_PE | (pe->pe_number & 0xFF);
1366 
1367 	mb(); /* Ensure above stores are visible */
1368 	__raw_writeq_be(val, invalidate);
1369 }
1370 
1371 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe *pe, bool rm,
1372 					unsigned shift, unsigned long index,
1373 					unsigned long npages)
1374 {
1375 	__be64 __iomem *invalidate = pnv_ioda_get_inval_reg(pe->phb, rm);
1376 	unsigned long start, end, inc;
1377 
1378 	/* We'll invalidate DMA address in PE scope */
1379 	start = PHB3_TCE_KILL_INVAL_ONE;
1380 	start |= (pe->pe_number & 0xFF);
1381 	end = start;
1382 
1383 	/* Figure out the start, end and step */
1384 	start |= (index << shift);
1385 	end |= ((index + npages - 1) << shift);
1386 	inc = (0x1ull << shift);
1387 	mb();
1388 
1389 	while (start <= end) {
1390 		if (rm)
1391 			__raw_rm_writeq_be(start, invalidate);
1392 		else
1393 			__raw_writeq_be(start, invalidate);
1394 		start += inc;
1395 	}
1396 }
1397 
1398 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe *pe)
1399 {
1400 	struct pnv_phb *phb = pe->phb;
1401 
1402 	if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
1403 		pnv_pci_phb3_tce_invalidate_pe(pe);
1404 	else
1405 		opal_pci_tce_kill(phb->opal_id, OPAL_PCI_TCE_KILL_PE,
1406 				  pe->pe_number, 0, 0, 0);
1407 }
1408 
1409 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table *tbl,
1410 		unsigned long index, unsigned long npages, bool rm)
1411 {
1412 	struct iommu_table_group_link *tgl;
1413 
1414 	list_for_each_entry_lockless(tgl, &tbl->it_group_list, next) {
1415 		struct pnv_ioda_pe *pe = container_of(tgl->table_group,
1416 				struct pnv_ioda_pe, table_group);
1417 		struct pnv_phb *phb = pe->phb;
1418 		unsigned int shift = tbl->it_page_shift;
1419 
1420 		if (phb->model == PNV_PHB_MODEL_PHB3 && phb->regs)
1421 			pnv_pci_phb3_tce_invalidate(pe, rm, shift,
1422 						    index, npages);
1423 		else
1424 			opal_pci_tce_kill(phb->opal_id,
1425 					  OPAL_PCI_TCE_KILL_PAGES,
1426 					  pe->pe_number, 1u << shift,
1427 					  index << shift, npages);
1428 	}
1429 }
1430 
1431 static int pnv_ioda2_tce_build(struct iommu_table *tbl, long index,
1432 		long npages, unsigned long uaddr,
1433 		enum dma_data_direction direction,
1434 		unsigned long attrs)
1435 {
1436 	int ret = pnv_tce_build(tbl, index, npages, uaddr, direction,
1437 			attrs);
1438 
1439 	if (!ret)
1440 		pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1441 
1442 	return ret;
1443 }
1444 
1445 static void pnv_ioda2_tce_free(struct iommu_table *tbl, long index,
1446 		long npages)
1447 {
1448 	pnv_tce_free(tbl, index, npages);
1449 
1450 	pnv_pci_ioda2_tce_invalidate(tbl, index, npages, false);
1451 }
1452 
1453 static struct iommu_table_ops pnv_ioda2_iommu_ops = {
1454 	.set = pnv_ioda2_tce_build,
1455 #ifdef CONFIG_IOMMU_API
1456 	.xchg_no_kill = pnv_ioda_tce_xchg_no_kill,
1457 	.tce_kill = pnv_pci_ioda2_tce_invalidate,
1458 	.useraddrptr = pnv_tce_useraddrptr,
1459 #endif
1460 	.clear = pnv_ioda2_tce_free,
1461 	.get = pnv_tce_get,
1462 	.free = pnv_pci_ioda2_table_free_pages,
1463 };
1464 
1465 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev *dev, void *data)
1466 {
1467 	unsigned int *weight = (unsigned int *)data;
1468 
1469 	/* This is quite simplistic. The "base" weight of a device
1470 	 * is 10. 0 means no DMA is to be accounted for it.
1471 	 */
1472 	if (dev->hdr_type != PCI_HEADER_TYPE_NORMAL)
1473 		return 0;
1474 
1475 	if (dev->class == PCI_CLASS_SERIAL_USB_UHCI ||
1476 	    dev->class == PCI_CLASS_SERIAL_USB_OHCI ||
1477 	    dev->class == PCI_CLASS_SERIAL_USB_EHCI)
1478 		*weight += 3;
1479 	else if ((dev->class >> 8) == PCI_CLASS_STORAGE_RAID)
1480 		*weight += 15;
1481 	else
1482 		*weight += 10;
1483 
1484 	return 0;
1485 }
1486 
1487 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe *pe)
1488 {
1489 	unsigned int weight = 0;
1490 
1491 	/* SRIOV VF has same DMA32 weight as its PF */
1492 #ifdef CONFIG_PCI_IOV
1493 	if ((pe->flags & PNV_IODA_PE_VF) && pe->parent_dev) {
1494 		pnv_pci_ioda_dev_dma_weight(pe->parent_dev, &weight);
1495 		return weight;
1496 	}
1497 #endif
1498 
1499 	if ((pe->flags & PNV_IODA_PE_DEV) && pe->pdev) {
1500 		pnv_pci_ioda_dev_dma_weight(pe->pdev, &weight);
1501 	} else if ((pe->flags & PNV_IODA_PE_BUS) && pe->pbus) {
1502 		struct pci_dev *pdev;
1503 
1504 		list_for_each_entry(pdev, &pe->pbus->devices, bus_list)
1505 			pnv_pci_ioda_dev_dma_weight(pdev, &weight);
1506 	} else if ((pe->flags & PNV_IODA_PE_BUS_ALL) && pe->pbus) {
1507 		pci_walk_bus(pe->pbus, pnv_pci_ioda_dev_dma_weight, &weight);
1508 	}
1509 
1510 	return weight;
1511 }
1512 
1513 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb *phb,
1514 				       struct pnv_ioda_pe *pe)
1515 {
1516 
1517 	struct page *tce_mem = NULL;
1518 	struct iommu_table *tbl;
1519 	unsigned int weight, total_weight = 0;
1520 	unsigned int tce32_segsz, base, segs, avail, i;
1521 	int64_t rc;
1522 	void *addr;
1523 
1524 	/* XXX FIXME: Handle 64-bit only DMA devices */
1525 	/* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
1526 	/* XXX FIXME: Allocate multi-level tables on PHB3 */
1527 	weight = pnv_pci_ioda_pe_dma_weight(pe);
1528 	if (!weight)
1529 		return;
1530 
1531 	pci_walk_bus(phb->hose->bus, pnv_pci_ioda_dev_dma_weight,
1532 		     &total_weight);
1533 	segs = (weight * phb->ioda.dma32_count) / total_weight;
1534 	if (!segs)
1535 		segs = 1;
1536 
1537 	/*
1538 	 * Allocate contiguous DMA32 segments. We begin with the expected
1539 	 * number of segments. With one more attempt, the number of DMA32
1540 	 * segments to be allocated is decreased by one until one segment
1541 	 * is allocated successfully.
1542 	 */
1543 	do {
1544 		for (base = 0; base <= phb->ioda.dma32_count - segs; base++) {
1545 			for (avail = 0, i = base; i < base + segs; i++) {
1546 				if (phb->ioda.dma32_segmap[i] ==
1547 				    IODA_INVALID_PE)
1548 					avail++;
1549 			}
1550 
1551 			if (avail == segs)
1552 				goto found;
1553 		}
1554 	} while (--segs);
1555 
1556 	if (!segs) {
1557 		pe_warn(pe, "No available DMA32 segments\n");
1558 		return;
1559 	}
1560 
1561 found:
1562 	tbl = pnv_pci_table_alloc(phb->hose->node);
1563 	if (WARN_ON(!tbl))
1564 		return;
1565 
1566 	iommu_register_group(&pe->table_group, phb->hose->global_number,
1567 			pe->pe_number);
1568 	pnv_pci_link_table_and_group(phb->hose->node, 0, tbl, &pe->table_group);
1569 
1570 	/* Grab a 32-bit TCE table */
1571 	pe_info(pe, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
1572 		weight, total_weight, base, segs);
1573 	pe_info(pe, " Setting up 32-bit TCE table at %08x..%08x\n",
1574 		base * PNV_IODA1_DMA32_SEGSIZE,
1575 		(base + segs) * PNV_IODA1_DMA32_SEGSIZE - 1);
1576 
1577 	/* XXX Currently, we allocate one big contiguous table for the
1578 	 * TCEs. We only really need one chunk per 256M of TCE space
1579 	 * (ie per segment) but that's an optimization for later, it
1580 	 * requires some added smarts with our get/put_tce implementation
1581 	 *
1582 	 * Each TCE page is 4KB in size and each TCE entry occupies 8
1583 	 * bytes
1584 	 */
1585 	tce32_segsz = PNV_IODA1_DMA32_SEGSIZE >> (IOMMU_PAGE_SHIFT_4K - 3);
1586 	tce_mem = alloc_pages_node(phb->hose->node, GFP_KERNEL,
1587 				   get_order(tce32_segsz * segs));
1588 	if (!tce_mem) {
1589 		pe_err(pe, " Failed to allocate a 32-bit TCE memory\n");
1590 		goto fail;
1591 	}
1592 	addr = page_address(tce_mem);
1593 	memset(addr, 0, tce32_segsz * segs);
1594 
1595 	/* Configure HW */
1596 	for (i = 0; i < segs; i++) {
1597 		rc = opal_pci_map_pe_dma_window(phb->opal_id,
1598 					      pe->pe_number,
1599 					      base + i, 1,
1600 					      __pa(addr) + tce32_segsz * i,
1601 					      tce32_segsz, IOMMU_PAGE_SIZE_4K);
1602 		if (rc) {
1603 			pe_err(pe, " Failed to configure 32-bit TCE table, err %lld\n",
1604 			       rc);
1605 			goto fail;
1606 		}
1607 	}
1608 
1609 	/* Setup DMA32 segment mapping */
1610 	for (i = base; i < base + segs; i++)
1611 		phb->ioda.dma32_segmap[i] = pe->pe_number;
1612 
1613 	/* Setup linux iommu table */
1614 	pnv_pci_setup_iommu_table(tbl, addr, tce32_segsz * segs,
1615 				  base * PNV_IODA1_DMA32_SEGSIZE,
1616 				  IOMMU_PAGE_SHIFT_4K);
1617 
1618 	tbl->it_ops = &pnv_ioda1_iommu_ops;
1619 	pe->table_group.tce32_start = tbl->it_offset << tbl->it_page_shift;
1620 	pe->table_group.tce32_size = tbl->it_size << tbl->it_page_shift;
1621 	if (!iommu_init_table(tbl, phb->hose->node, 0, 0))
1622 		panic("Failed to initialize iommu table");
1623 
1624 	pe->dma_setup_done = true;
1625 	return;
1626  fail:
1627 	/* XXX Failure: Try to fallback to 64-bit only ? */
1628 	if (tce_mem)
1629 		__free_pages(tce_mem, get_order(tce32_segsz * segs));
1630 	if (tbl) {
1631 		pnv_pci_unlink_table_and_group(tbl, &pe->table_group);
1632 		iommu_tce_table_put(tbl);
1633 	}
1634 }
1635 
1636 static long pnv_pci_ioda2_set_window(struct iommu_table_group *table_group,
1637 		int num, struct iommu_table *tbl)
1638 {
1639 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1640 			table_group);
1641 	struct pnv_phb *phb = pe->phb;
1642 	int64_t rc;
1643 	const unsigned long size = tbl->it_indirect_levels ?
1644 			tbl->it_level_size : tbl->it_size;
1645 	const __u64 start_addr = tbl->it_offset << tbl->it_page_shift;
1646 	const __u64 win_size = tbl->it_size << tbl->it_page_shift;
1647 
1648 	pe_info(pe, "Setting up window#%d %llx..%llx pg=%lx\n",
1649 		num, start_addr, start_addr + win_size - 1,
1650 		IOMMU_PAGE_SIZE(tbl));
1651 
1652 	/*
1653 	 * Map TCE table through TVT. The TVE index is the PE number
1654 	 * shifted by 1 bit for 32-bits DMA space.
1655 	 */
1656 	rc = opal_pci_map_pe_dma_window(phb->opal_id,
1657 			pe->pe_number,
1658 			(pe->pe_number << 1) + num,
1659 			tbl->it_indirect_levels + 1,
1660 			__pa(tbl->it_base),
1661 			size << 3,
1662 			IOMMU_PAGE_SIZE(tbl));
1663 	if (rc) {
1664 		pe_err(pe, "Failed to configure TCE table, err %lld\n", rc);
1665 		return rc;
1666 	}
1667 
1668 	pnv_pci_link_table_and_group(phb->hose->node, num,
1669 			tbl, &pe->table_group);
1670 	pnv_pci_ioda2_tce_invalidate_pe(pe);
1671 
1672 	return 0;
1673 }
1674 
1675 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe *pe, bool enable)
1676 {
1677 	uint16_t window_id = (pe->pe_number << 1 ) + 1;
1678 	int64_t rc;
1679 
1680 	pe_info(pe, "%sabling 64-bit DMA bypass\n", enable ? "En" : "Dis");
1681 	if (enable) {
1682 		phys_addr_t top = memblock_end_of_DRAM();
1683 
1684 		top = roundup_pow_of_two(top);
1685 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1686 						     pe->pe_number,
1687 						     window_id,
1688 						     pe->tce_bypass_base,
1689 						     top);
1690 	} else {
1691 		rc = opal_pci_map_pe_dma_window_real(pe->phb->opal_id,
1692 						     pe->pe_number,
1693 						     window_id,
1694 						     pe->tce_bypass_base,
1695 						     0);
1696 	}
1697 	if (rc)
1698 		pe_err(pe, "OPAL error %lld configuring bypass window\n", rc);
1699 	else
1700 		pe->tce_bypass_enabled = enable;
1701 }
1702 
1703 static long pnv_pci_ioda2_create_table(struct iommu_table_group *table_group,
1704 		int num, __u32 page_shift, __u64 window_size, __u32 levels,
1705 		bool alloc_userspace_copy, struct iommu_table **ptbl)
1706 {
1707 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1708 			table_group);
1709 	int nid = pe->phb->hose->node;
1710 	__u64 bus_offset = num ? pe->tce_bypass_base : table_group->tce32_start;
1711 	long ret;
1712 	struct iommu_table *tbl;
1713 
1714 	tbl = pnv_pci_table_alloc(nid);
1715 	if (!tbl)
1716 		return -ENOMEM;
1717 
1718 	tbl->it_ops = &pnv_ioda2_iommu_ops;
1719 
1720 	ret = pnv_pci_ioda2_table_alloc_pages(nid,
1721 			bus_offset, page_shift, window_size,
1722 			levels, alloc_userspace_copy, tbl);
1723 	if (ret) {
1724 		iommu_tce_table_put(tbl);
1725 		return ret;
1726 	}
1727 
1728 	*ptbl = tbl;
1729 
1730 	return 0;
1731 }
1732 
1733 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe *pe)
1734 {
1735 	struct iommu_table *tbl = NULL;
1736 	long rc;
1737 	unsigned long res_start, res_end;
1738 
1739 	/*
1740 	 * crashkernel= specifies the kdump kernel's maximum memory at
1741 	 * some offset and there is no guaranteed the result is a power
1742 	 * of 2, which will cause errors later.
1743 	 */
1744 	const u64 max_memory = __rounddown_pow_of_two(memory_hotplug_max());
1745 
1746 	/*
1747 	 * In memory constrained environments, e.g. kdump kernel, the
1748 	 * DMA window can be larger than available memory, which will
1749 	 * cause errors later.
1750 	 */
1751 	const u64 maxblock = 1UL << (PAGE_SHIFT + MAX_ORDER - 1);
1752 
1753 	/*
1754 	 * We create the default window as big as we can. The constraint is
1755 	 * the max order of allocation possible. The TCE table is likely to
1756 	 * end up being multilevel and with on-demand allocation in place,
1757 	 * the initial use is not going to be huge as the default window aims
1758 	 * to support crippled devices (i.e. not fully 64bit DMAble) only.
1759 	 */
1760 	/* iommu_table::it_map uses 1 bit per IOMMU page, hence 8 */
1761 	const u64 window_size = min((maxblock * 8) << PAGE_SHIFT, max_memory);
1762 	/* Each TCE level cannot exceed maxblock so go multilevel if needed */
1763 	unsigned long tces_order = ilog2(window_size >> PAGE_SHIFT);
1764 	unsigned long tcelevel_order = ilog2(maxblock >> 3);
1765 	unsigned int levels = tces_order / tcelevel_order;
1766 
1767 	if (tces_order % tcelevel_order)
1768 		levels += 1;
1769 	/*
1770 	 * We try to stick to default levels (which is >1 at the moment) in
1771 	 * order to save memory by relying on on-demain TCE level allocation.
1772 	 */
1773 	levels = max_t(unsigned int, levels, POWERNV_IOMMU_DEFAULT_LEVELS);
1774 
1775 	rc = pnv_pci_ioda2_create_table(&pe->table_group, 0, PAGE_SHIFT,
1776 			window_size, levels, false, &tbl);
1777 	if (rc) {
1778 		pe_err(pe, "Failed to create 32-bit TCE table, err %ld",
1779 				rc);
1780 		return rc;
1781 	}
1782 
1783 	/* We use top part of 32bit space for MMIO so exclude it from DMA */
1784 	res_start = 0;
1785 	res_end = 0;
1786 	if (window_size > pe->phb->ioda.m32_pci_base) {
1787 		res_start = pe->phb->ioda.m32_pci_base >> tbl->it_page_shift;
1788 		res_end = min(window_size, SZ_4G) >> tbl->it_page_shift;
1789 	}
1790 
1791 	if (iommu_init_table(tbl, pe->phb->hose->node, res_start, res_end))
1792 		rc = pnv_pci_ioda2_set_window(&pe->table_group, 0, tbl);
1793 	else
1794 		rc = -ENOMEM;
1795 	if (rc) {
1796 		pe_err(pe, "Failed to configure 32-bit TCE table, err %ld\n", rc);
1797 		iommu_tce_table_put(tbl);
1798 		tbl = NULL; /* This clears iommu_table_base below */
1799 	}
1800 	if (!pnv_iommu_bypass_disabled)
1801 		pnv_pci_ioda2_set_bypass(pe, true);
1802 
1803 	/*
1804 	 * Set table base for the case of IOMMU DMA use. Usually this is done
1805 	 * from dma_dev_setup() which is not called when a device is returned
1806 	 * from VFIO so do it here.
1807 	 */
1808 	if (pe->pdev)
1809 		set_iommu_table_base(&pe->pdev->dev, tbl);
1810 
1811 	return 0;
1812 }
1813 
1814 static long pnv_pci_ioda2_unset_window(struct iommu_table_group *table_group,
1815 		int num)
1816 {
1817 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1818 			table_group);
1819 	struct pnv_phb *phb = pe->phb;
1820 	long ret;
1821 
1822 	pe_info(pe, "Removing DMA window #%d\n", num);
1823 
1824 	ret = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
1825 			(pe->pe_number << 1) + num,
1826 			0/* levels */, 0/* table address */,
1827 			0/* table size */, 0/* page size */);
1828 	if (ret)
1829 		pe_warn(pe, "Unmapping failed, ret = %ld\n", ret);
1830 	else
1831 		pnv_pci_ioda2_tce_invalidate_pe(pe);
1832 
1833 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
1834 
1835 	return ret;
1836 }
1837 
1838 #ifdef CONFIG_IOMMU_API
1839 unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift,
1840 		__u64 window_size, __u32 levels)
1841 {
1842 	unsigned long bytes = 0;
1843 	const unsigned window_shift = ilog2(window_size);
1844 	unsigned entries_shift = window_shift - page_shift;
1845 	unsigned table_shift = entries_shift + 3;
1846 	unsigned long tce_table_size = max(0x1000UL, 1UL << table_shift);
1847 	unsigned long direct_table_size;
1848 
1849 	if (!levels || (levels > POWERNV_IOMMU_MAX_LEVELS) ||
1850 			!is_power_of_2(window_size))
1851 		return 0;
1852 
1853 	/* Calculate a direct table size from window_size and levels */
1854 	entries_shift = (entries_shift + levels - 1) / levels;
1855 	table_shift = entries_shift + 3;
1856 	table_shift = max_t(unsigned, table_shift, PAGE_SHIFT);
1857 	direct_table_size =  1UL << table_shift;
1858 
1859 	for ( ; levels; --levels) {
1860 		bytes += ALIGN(tce_table_size, direct_table_size);
1861 
1862 		tce_table_size /= direct_table_size;
1863 		tce_table_size <<= 3;
1864 		tce_table_size = max_t(unsigned long,
1865 				tce_table_size, direct_table_size);
1866 	}
1867 
1868 	return bytes + bytes; /* one for HW table, one for userspace copy */
1869 }
1870 
1871 static long pnv_pci_ioda2_create_table_userspace(
1872 		struct iommu_table_group *table_group,
1873 		int num, __u32 page_shift, __u64 window_size, __u32 levels,
1874 		struct iommu_table **ptbl)
1875 {
1876 	long ret = pnv_pci_ioda2_create_table(table_group,
1877 			num, page_shift, window_size, levels, true, ptbl);
1878 
1879 	if (!ret)
1880 		(*ptbl)->it_allocated_size = pnv_pci_ioda2_get_table_size(
1881 				page_shift, window_size, levels);
1882 	return ret;
1883 }
1884 
1885 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe *pe, struct pci_bus *bus)
1886 {
1887 	struct pci_dev *dev;
1888 
1889 	list_for_each_entry(dev, &bus->devices, bus_list) {
1890 		set_iommu_table_base(&dev->dev, pe->table_group.tables[0]);
1891 		dev->dev.archdata.dma_offset = pe->tce_bypass_base;
1892 
1893 		if ((pe->flags & PNV_IODA_PE_BUS_ALL) && dev->subordinate)
1894 			pnv_ioda_setup_bus_dma(pe, dev->subordinate);
1895 	}
1896 }
1897 
1898 static void pnv_ioda2_take_ownership(struct iommu_table_group *table_group)
1899 {
1900 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1901 						table_group);
1902 	/* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
1903 	struct iommu_table *tbl = pe->table_group.tables[0];
1904 
1905 	pnv_pci_ioda2_set_bypass(pe, false);
1906 	pnv_pci_ioda2_unset_window(&pe->table_group, 0);
1907 	if (pe->pbus)
1908 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
1909 	else if (pe->pdev)
1910 		set_iommu_table_base(&pe->pdev->dev, NULL);
1911 	iommu_tce_table_put(tbl);
1912 }
1913 
1914 static void pnv_ioda2_release_ownership(struct iommu_table_group *table_group)
1915 {
1916 	struct pnv_ioda_pe *pe = container_of(table_group, struct pnv_ioda_pe,
1917 						table_group);
1918 
1919 	pnv_pci_ioda2_setup_default_config(pe);
1920 	if (pe->pbus)
1921 		pnv_ioda_setup_bus_dma(pe, pe->pbus);
1922 }
1923 
1924 static struct iommu_table_group_ops pnv_pci_ioda2_ops = {
1925 	.get_table_size = pnv_pci_ioda2_get_table_size,
1926 	.create_table = pnv_pci_ioda2_create_table_userspace,
1927 	.set_window = pnv_pci_ioda2_set_window,
1928 	.unset_window = pnv_pci_ioda2_unset_window,
1929 	.take_ownership = pnv_ioda2_take_ownership,
1930 	.release_ownership = pnv_ioda2_release_ownership,
1931 };
1932 #endif
1933 
1934 void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb *phb,
1935 				struct pnv_ioda_pe *pe)
1936 {
1937 	int64_t rc;
1938 
1939 	/* TVE #1 is selected by PCI address bit 59 */
1940 	pe->tce_bypass_base = 1ull << 59;
1941 
1942 	/* The PE will reserve all possible 32-bits space */
1943 	pe_info(pe, "Setting up 32-bit TCE table at 0..%08x\n",
1944 		phb->ioda.m32_pci_base);
1945 
1946 	/* Setup linux iommu table */
1947 	pe->table_group.tce32_start = 0;
1948 	pe->table_group.tce32_size = phb->ioda.m32_pci_base;
1949 	pe->table_group.max_dynamic_windows_supported =
1950 			IOMMU_TABLE_GROUP_MAX_TABLES;
1951 	pe->table_group.max_levels = POWERNV_IOMMU_MAX_LEVELS;
1952 	pe->table_group.pgsizes = pnv_ioda_parse_tce_sizes(phb);
1953 
1954 	rc = pnv_pci_ioda2_setup_default_config(pe);
1955 	if (rc)
1956 		return;
1957 
1958 #ifdef CONFIG_IOMMU_API
1959 	pe->table_group.ops = &pnv_pci_ioda2_ops;
1960 	iommu_register_group(&pe->table_group, phb->hose->global_number,
1961 			     pe->pe_number);
1962 #endif
1963 	pe->dma_setup_done = true;
1964 }
1965 
1966 /*
1967  * Called from KVM in real mode to EOI passthru interrupts. The ICP
1968  * EOI is handled directly in KVM in kvmppc_deliver_irq_passthru().
1969  *
1970  * The IRQ data is mapped in the PCI-MSI domain and the EOI OPAL call
1971  * needs an HW IRQ number mapped in the XICS IRQ domain. The HW IRQ
1972  * numbers of the in-the-middle MSI domain are vector numbers and it's
1973  * good enough for OPAL. Use that.
1974  */
1975 int64_t pnv_opal_pci_msi_eoi(struct irq_data *d)
1976 {
1977 	struct pci_controller *hose = irq_data_get_irq_chip_data(d->parent_data);
1978 	struct pnv_phb *phb = hose->private_data;
1979 
1980 	return opal_pci_msi_eoi(phb->opal_id, d->parent_data->hwirq);
1981 }
1982 
1983 /*
1984  * The IRQ data is mapped in the XICS domain, with OPAL HW IRQ numbers
1985  */
1986 static void pnv_ioda2_msi_eoi(struct irq_data *d)
1987 {
1988 	int64_t rc;
1989 	unsigned int hw_irq = (unsigned int)irqd_to_hwirq(d);
1990 	struct pci_controller *hose = irq_data_get_irq_chip_data(d);
1991 	struct pnv_phb *phb = hose->private_data;
1992 
1993 	rc = opal_pci_msi_eoi(phb->opal_id, hw_irq);
1994 	WARN_ON_ONCE(rc);
1995 
1996 	icp_native_eoi(d);
1997 }
1998 
1999 /* P8/CXL only */
2000 void pnv_set_msi_irq_chip(struct pnv_phb *phb, unsigned int virq)
2001 {
2002 	struct irq_data *idata;
2003 	struct irq_chip *ichip;
2004 
2005 	/* The MSI EOI OPAL call is only needed on PHB3 */
2006 	if (phb->model != PNV_PHB_MODEL_PHB3)
2007 		return;
2008 
2009 	if (!phb->ioda.irq_chip_init) {
2010 		/*
2011 		 * First time we setup an MSI IRQ, we need to setup the
2012 		 * corresponding IRQ chip to route correctly.
2013 		 */
2014 		idata = irq_get_irq_data(virq);
2015 		ichip = irq_data_get_irq_chip(idata);
2016 		phb->ioda.irq_chip_init = 1;
2017 		phb->ioda.irq_chip = *ichip;
2018 		phb->ioda.irq_chip.irq_eoi = pnv_ioda2_msi_eoi;
2019 	}
2020 	irq_set_chip(virq, &phb->ioda.irq_chip);
2021 	irq_set_chip_data(virq, phb->hose);
2022 }
2023 
2024 static struct irq_chip pnv_pci_msi_irq_chip;
2025 
2026 /*
2027  * Returns true iff chip is something that we could call
2028  * pnv_opal_pci_msi_eoi for.
2029  */
2030 bool is_pnv_opal_msi(struct irq_chip *chip)
2031 {
2032 	return chip == &pnv_pci_msi_irq_chip;
2033 }
2034 EXPORT_SYMBOL_GPL(is_pnv_opal_msi);
2035 
2036 static int __pnv_pci_ioda_msi_setup(struct pnv_phb *phb, struct pci_dev *dev,
2037 				    unsigned int xive_num,
2038 				    unsigned int is_64, struct msi_msg *msg)
2039 {
2040 	struct pnv_ioda_pe *pe = pnv_ioda_get_pe(dev);
2041 	__be32 data;
2042 	int rc;
2043 
2044 	dev_dbg(&dev->dev, "%s: setup %s-bit MSI for vector #%d\n", __func__,
2045 		is_64 ? "64" : "32", xive_num);
2046 
2047 	/* No PE assigned ? bail out ... no MSI for you ! */
2048 	if (pe == NULL)
2049 		return -ENXIO;
2050 
2051 	/* Check if we have an MVE */
2052 	if (pe->mve_number < 0)
2053 		return -ENXIO;
2054 
2055 	/* Force 32-bit MSI on some broken devices */
2056 	if (dev->no_64bit_msi)
2057 		is_64 = 0;
2058 
2059 	/* Assign XIVE to PE */
2060 	rc = opal_pci_set_xive_pe(phb->opal_id, pe->pe_number, xive_num);
2061 	if (rc) {
2062 		pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2063 			pci_name(dev), rc, xive_num);
2064 		return -EIO;
2065 	}
2066 
2067 	if (is_64) {
2068 		__be64 addr64;
2069 
2070 		rc = opal_get_msi_64(phb->opal_id, pe->mve_number, xive_num, 1,
2071 				     &addr64, &data);
2072 		if (rc) {
2073 			pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2074 				pci_name(dev), rc);
2075 			return -EIO;
2076 		}
2077 		msg->address_hi = be64_to_cpu(addr64) >> 32;
2078 		msg->address_lo = be64_to_cpu(addr64) & 0xfffffffful;
2079 	} else {
2080 		__be32 addr32;
2081 
2082 		rc = opal_get_msi_32(phb->opal_id, pe->mve_number, xive_num, 1,
2083 				     &addr32, &data);
2084 		if (rc) {
2085 			pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2086 				pci_name(dev), rc);
2087 			return -EIO;
2088 		}
2089 		msg->address_hi = 0;
2090 		msg->address_lo = be32_to_cpu(addr32);
2091 	}
2092 	msg->data = be32_to_cpu(data);
2093 
2094 	return 0;
2095 }
2096 
2097 /*
2098  * The msi_free() op is called before irq_domain_free_irqs_top() when
2099  * the handler data is still available. Use that to clear the XIVE
2100  * controller.
2101  */
2102 static void pnv_msi_ops_msi_free(struct irq_domain *domain,
2103 				 struct msi_domain_info *info,
2104 				 unsigned int irq)
2105 {
2106 	if (xive_enabled())
2107 		xive_irq_free_data(irq);
2108 }
2109 
2110 static struct msi_domain_ops pnv_pci_msi_domain_ops = {
2111 	.msi_free	= pnv_msi_ops_msi_free,
2112 };
2113 
2114 static void pnv_msi_shutdown(struct irq_data *d)
2115 {
2116 	d = d->parent_data;
2117 	if (d->chip->irq_shutdown)
2118 		d->chip->irq_shutdown(d);
2119 }
2120 
2121 static void pnv_msi_mask(struct irq_data *d)
2122 {
2123 	pci_msi_mask_irq(d);
2124 	irq_chip_mask_parent(d);
2125 }
2126 
2127 static void pnv_msi_unmask(struct irq_data *d)
2128 {
2129 	pci_msi_unmask_irq(d);
2130 	irq_chip_unmask_parent(d);
2131 }
2132 
2133 static struct irq_chip pnv_pci_msi_irq_chip = {
2134 	.name		= "PNV-PCI-MSI",
2135 	.irq_shutdown	= pnv_msi_shutdown,
2136 	.irq_mask	= pnv_msi_mask,
2137 	.irq_unmask	= pnv_msi_unmask,
2138 	.irq_eoi	= irq_chip_eoi_parent,
2139 };
2140 
2141 static struct msi_domain_info pnv_msi_domain_info = {
2142 	.flags = (MSI_FLAG_USE_DEF_DOM_OPS | MSI_FLAG_USE_DEF_CHIP_OPS |
2143 		  MSI_FLAG_MULTI_PCI_MSI  | MSI_FLAG_PCI_MSIX),
2144 	.ops   = &pnv_pci_msi_domain_ops,
2145 	.chip  = &pnv_pci_msi_irq_chip,
2146 };
2147 
2148 static void pnv_msi_compose_msg(struct irq_data *d, struct msi_msg *msg)
2149 {
2150 	struct msi_desc *entry = irq_data_get_msi_desc(d);
2151 	struct pci_dev *pdev = msi_desc_to_pci_dev(entry);
2152 	struct pci_controller *hose = irq_data_get_irq_chip_data(d);
2153 	struct pnv_phb *phb = hose->private_data;
2154 	int rc;
2155 
2156 	rc = __pnv_pci_ioda_msi_setup(phb, pdev, d->hwirq,
2157 				      entry->msi_attrib.is_64, msg);
2158 	if (rc)
2159 		dev_err(&pdev->dev, "Failed to setup %s-bit MSI #%ld : %d\n",
2160 			entry->msi_attrib.is_64 ? "64" : "32", d->hwirq, rc);
2161 }
2162 
2163 /*
2164  * The IRQ data is mapped in the MSI domain in which HW IRQ numbers
2165  * correspond to vector numbers.
2166  */
2167 static void pnv_msi_eoi(struct irq_data *d)
2168 {
2169 	struct pci_controller *hose = irq_data_get_irq_chip_data(d);
2170 	struct pnv_phb *phb = hose->private_data;
2171 
2172 	if (phb->model == PNV_PHB_MODEL_PHB3) {
2173 		/*
2174 		 * The EOI OPAL call takes an OPAL HW IRQ number but
2175 		 * since it is translated into a vector number in
2176 		 * OPAL, use that directly.
2177 		 */
2178 		WARN_ON_ONCE(opal_pci_msi_eoi(phb->opal_id, d->hwirq));
2179 	}
2180 
2181 	irq_chip_eoi_parent(d);
2182 }
2183 
2184 static struct irq_chip pnv_msi_irq_chip = {
2185 	.name			= "PNV-MSI",
2186 	.irq_shutdown		= pnv_msi_shutdown,
2187 	.irq_mask		= irq_chip_mask_parent,
2188 	.irq_unmask		= irq_chip_unmask_parent,
2189 	.irq_eoi		= pnv_msi_eoi,
2190 	.irq_set_affinity	= irq_chip_set_affinity_parent,
2191 	.irq_compose_msi_msg	= pnv_msi_compose_msg,
2192 };
2193 
2194 static int pnv_irq_parent_domain_alloc(struct irq_domain *domain,
2195 				       unsigned int virq, int hwirq)
2196 {
2197 	struct irq_fwspec parent_fwspec;
2198 	int ret;
2199 
2200 	parent_fwspec.fwnode = domain->parent->fwnode;
2201 	parent_fwspec.param_count = 2;
2202 	parent_fwspec.param[0] = hwirq;
2203 	parent_fwspec.param[1] = IRQ_TYPE_EDGE_RISING;
2204 
2205 	ret = irq_domain_alloc_irqs_parent(domain, virq, 1, &parent_fwspec);
2206 	if (ret)
2207 		return ret;
2208 
2209 	return 0;
2210 }
2211 
2212 static int pnv_irq_domain_alloc(struct irq_domain *domain, unsigned int virq,
2213 				unsigned int nr_irqs, void *arg)
2214 {
2215 	struct pci_controller *hose = domain->host_data;
2216 	struct pnv_phb *phb = hose->private_data;
2217 	msi_alloc_info_t *info = arg;
2218 	struct pci_dev *pdev = msi_desc_to_pci_dev(info->desc);
2219 	int hwirq;
2220 	int i, ret;
2221 
2222 	hwirq = msi_bitmap_alloc_hwirqs(&phb->msi_bmp, nr_irqs);
2223 	if (hwirq < 0) {
2224 		dev_warn(&pdev->dev, "failed to find a free MSI\n");
2225 		return -ENOSPC;
2226 	}
2227 
2228 	dev_dbg(&pdev->dev, "%s bridge %pOF %d/%x #%d\n", __func__,
2229 		hose->dn, virq, hwirq, nr_irqs);
2230 
2231 	for (i = 0; i < nr_irqs; i++) {
2232 		ret = pnv_irq_parent_domain_alloc(domain, virq + i,
2233 						  phb->msi_base + hwirq + i);
2234 		if (ret)
2235 			goto out;
2236 
2237 		irq_domain_set_hwirq_and_chip(domain, virq + i, hwirq + i,
2238 					      &pnv_msi_irq_chip, hose);
2239 	}
2240 
2241 	return 0;
2242 
2243 out:
2244 	irq_domain_free_irqs_parent(domain, virq, i - 1);
2245 	msi_bitmap_free_hwirqs(&phb->msi_bmp, hwirq, nr_irqs);
2246 	return ret;
2247 }
2248 
2249 static void pnv_irq_domain_free(struct irq_domain *domain, unsigned int virq,
2250 				unsigned int nr_irqs)
2251 {
2252 	struct irq_data *d = irq_domain_get_irq_data(domain, virq);
2253 	struct pci_controller *hose = irq_data_get_irq_chip_data(d);
2254 	struct pnv_phb *phb = hose->private_data;
2255 
2256 	pr_debug("%s bridge %pOF %d/%lx #%d\n", __func__, hose->dn,
2257 		 virq, d->hwirq, nr_irqs);
2258 
2259 	msi_bitmap_free_hwirqs(&phb->msi_bmp, d->hwirq, nr_irqs);
2260 	/* XIVE domain is cleared through ->msi_free() */
2261 }
2262 
2263 static const struct irq_domain_ops pnv_irq_domain_ops = {
2264 	.alloc  = pnv_irq_domain_alloc,
2265 	.free   = pnv_irq_domain_free,
2266 };
2267 
2268 static int pnv_msi_allocate_domains(struct pci_controller *hose, unsigned int count)
2269 {
2270 	struct pnv_phb *phb = hose->private_data;
2271 	struct irq_domain *parent = irq_get_default_host();
2272 
2273 	hose->fwnode = irq_domain_alloc_named_id_fwnode("PNV-MSI", phb->opal_id);
2274 	if (!hose->fwnode)
2275 		return -ENOMEM;
2276 
2277 	hose->dev_domain = irq_domain_create_hierarchy(parent, 0, count,
2278 						       hose->fwnode,
2279 						       &pnv_irq_domain_ops, hose);
2280 	if (!hose->dev_domain) {
2281 		pr_err("PCI: failed to create IRQ domain bridge %pOF (domain %d)\n",
2282 		       hose->dn, hose->global_number);
2283 		irq_domain_free_fwnode(hose->fwnode);
2284 		return -ENOMEM;
2285 	}
2286 
2287 	hose->msi_domain = pci_msi_create_irq_domain(of_node_to_fwnode(hose->dn),
2288 						     &pnv_msi_domain_info,
2289 						     hose->dev_domain);
2290 	if (!hose->msi_domain) {
2291 		pr_err("PCI: failed to create MSI IRQ domain bridge %pOF (domain %d)\n",
2292 		       hose->dn, hose->global_number);
2293 		irq_domain_free_fwnode(hose->fwnode);
2294 		irq_domain_remove(hose->dev_domain);
2295 		return -ENOMEM;
2296 	}
2297 
2298 	return 0;
2299 }
2300 
2301 static void pnv_pci_init_ioda_msis(struct pnv_phb *phb)
2302 {
2303 	unsigned int count;
2304 	const __be32 *prop = of_get_property(phb->hose->dn,
2305 					     "ibm,opal-msi-ranges", NULL);
2306 	if (!prop) {
2307 		/* BML Fallback */
2308 		prop = of_get_property(phb->hose->dn, "msi-ranges", NULL);
2309 	}
2310 	if (!prop)
2311 		return;
2312 
2313 	phb->msi_base = be32_to_cpup(prop);
2314 	count = be32_to_cpup(prop + 1);
2315 	if (msi_bitmap_alloc(&phb->msi_bmp, count, phb->hose->dn)) {
2316 		pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2317 		       phb->hose->global_number);
2318 		return;
2319 	}
2320 
2321 	pr_info("  Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2322 		count, phb->msi_base);
2323 
2324 	pnv_msi_allocate_domains(phb->hose, count);
2325 }
2326 
2327 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe *pe,
2328 				  struct resource *res)
2329 {
2330 	struct pnv_phb *phb = pe->phb;
2331 	struct pci_bus_region region;
2332 	int index;
2333 	int64_t rc;
2334 
2335 	if (!res || !res->flags || res->start > res->end)
2336 		return;
2337 
2338 	if (res->flags & IORESOURCE_IO) {
2339 		region.start = res->start - phb->ioda.io_pci_base;
2340 		region.end   = res->end - phb->ioda.io_pci_base;
2341 		index = region.start / phb->ioda.io_segsize;
2342 
2343 		while (index < phb->ioda.total_pe_num &&
2344 		       region.start <= region.end) {
2345 			phb->ioda.io_segmap[index] = pe->pe_number;
2346 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2347 				pe->pe_number, OPAL_IO_WINDOW_TYPE, 0, index);
2348 			if (rc != OPAL_SUCCESS) {
2349 				pr_err("%s: Error %lld mapping IO segment#%d to PE#%x\n",
2350 				       __func__, rc, index, pe->pe_number);
2351 				break;
2352 			}
2353 
2354 			region.start += phb->ioda.io_segsize;
2355 			index++;
2356 		}
2357 	} else if ((res->flags & IORESOURCE_MEM) &&
2358 		   !pnv_pci_is_m64(phb, res)) {
2359 		region.start = res->start -
2360 			       phb->hose->mem_offset[0] -
2361 			       phb->ioda.m32_pci_base;
2362 		region.end   = res->end -
2363 			       phb->hose->mem_offset[0] -
2364 			       phb->ioda.m32_pci_base;
2365 		index = region.start / phb->ioda.m32_segsize;
2366 
2367 		while (index < phb->ioda.total_pe_num &&
2368 		       region.start <= region.end) {
2369 			phb->ioda.m32_segmap[index] = pe->pe_number;
2370 			rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2371 				pe->pe_number, OPAL_M32_WINDOW_TYPE, 0, index);
2372 			if (rc != OPAL_SUCCESS) {
2373 				pr_err("%s: Error %lld mapping M32 segment#%d to PE#%x",
2374 				       __func__, rc, index, pe->pe_number);
2375 				break;
2376 			}
2377 
2378 			region.start += phb->ioda.m32_segsize;
2379 			index++;
2380 		}
2381 	}
2382 }
2383 
2384 /*
2385  * This function is supposed to be called on basis of PE from top
2386  * to bottom style. So the the I/O or MMIO segment assigned to
2387  * parent PE could be overridden by its child PEs if necessary.
2388  */
2389 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe *pe)
2390 {
2391 	struct pci_dev *pdev;
2392 	int i;
2393 
2394 	/*
2395 	 * NOTE: We only care PCI bus based PE for now. For PCI
2396 	 * device based PE, for example SRIOV sensitive VF should
2397 	 * be figured out later.
2398 	 */
2399 	BUG_ON(!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)));
2400 
2401 	list_for_each_entry(pdev, &pe->pbus->devices, bus_list) {
2402 		for (i = 0; i <= PCI_ROM_RESOURCE; i++)
2403 			pnv_ioda_setup_pe_res(pe, &pdev->resource[i]);
2404 
2405 		/*
2406 		 * If the PE contains all subordinate PCI buses, the
2407 		 * windows of the child bridges should be mapped to
2408 		 * the PE as well.
2409 		 */
2410 		if (!(pe->flags & PNV_IODA_PE_BUS_ALL) || !pci_is_bridge(pdev))
2411 			continue;
2412 		for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++)
2413 			pnv_ioda_setup_pe_res(pe,
2414 				&pdev->resource[PCI_BRIDGE_RESOURCES + i]);
2415 	}
2416 }
2417 
2418 #ifdef CONFIG_DEBUG_FS
2419 static int pnv_pci_diag_data_set(void *data, u64 val)
2420 {
2421 	struct pnv_phb *phb = data;
2422 	s64 ret;
2423 
2424 	/* Retrieve the diag data from firmware */
2425 	ret = opal_pci_get_phb_diag_data2(phb->opal_id, phb->diag_data,
2426 					  phb->diag_data_size);
2427 	if (ret != OPAL_SUCCESS)
2428 		return -EIO;
2429 
2430 	/* Print the diag data to the kernel log */
2431 	pnv_pci_dump_phb_diag_data(phb->hose, phb->diag_data);
2432 	return 0;
2433 }
2434 
2435 DEFINE_DEBUGFS_ATTRIBUTE(pnv_pci_diag_data_fops, NULL, pnv_pci_diag_data_set,
2436 			 "%llu\n");
2437 
2438 static int pnv_pci_ioda_pe_dump(void *data, u64 val)
2439 {
2440 	struct pnv_phb *phb = data;
2441 	int pe_num;
2442 
2443 	for (pe_num = 0; pe_num < phb->ioda.total_pe_num; pe_num++) {
2444 		struct pnv_ioda_pe *pe = &phb->ioda.pe_array[pe_num];
2445 
2446 		if (!test_bit(pe_num, phb->ioda.pe_alloc))
2447 			continue;
2448 
2449 		pe_warn(pe, "rid: %04x dev count: %2d flags: %s%s%s%s%s%s\n",
2450 			pe->rid, pe->device_count,
2451 			(pe->flags & PNV_IODA_PE_DEV) ? "dev " : "",
2452 			(pe->flags & PNV_IODA_PE_BUS) ? "bus " : "",
2453 			(pe->flags & PNV_IODA_PE_BUS_ALL) ? "all " : "",
2454 			(pe->flags & PNV_IODA_PE_MASTER) ? "master " : "",
2455 			(pe->flags & PNV_IODA_PE_SLAVE) ? "slave " : "",
2456 			(pe->flags & PNV_IODA_PE_VF) ? "vf " : "");
2457 	}
2458 
2459 	return 0;
2460 }
2461 
2462 DEFINE_DEBUGFS_ATTRIBUTE(pnv_pci_ioda_pe_dump_fops, NULL,
2463 			 pnv_pci_ioda_pe_dump, "%llu\n");
2464 
2465 #endif /* CONFIG_DEBUG_FS */
2466 
2467 static void pnv_pci_ioda_create_dbgfs(void)
2468 {
2469 #ifdef CONFIG_DEBUG_FS
2470 	struct pci_controller *hose, *tmp;
2471 	struct pnv_phb *phb;
2472 	char name[16];
2473 
2474 	list_for_each_entry_safe(hose, tmp, &hose_list, list_node) {
2475 		phb = hose->private_data;
2476 
2477 		sprintf(name, "PCI%04x", hose->global_number);
2478 		phb->dbgfs = debugfs_create_dir(name, arch_debugfs_dir);
2479 
2480 		debugfs_create_file_unsafe("dump_diag_regs", 0200, phb->dbgfs,
2481 					   phb, &pnv_pci_diag_data_fops);
2482 		debugfs_create_file_unsafe("dump_ioda_pe_state", 0200, phb->dbgfs,
2483 					   phb, &pnv_pci_ioda_pe_dump_fops);
2484 	}
2485 #endif /* CONFIG_DEBUG_FS */
2486 }
2487 
2488 static void pnv_pci_enable_bridge(struct pci_bus *bus)
2489 {
2490 	struct pci_dev *dev = bus->self;
2491 	struct pci_bus *child;
2492 
2493 	/* Empty bus ? bail */
2494 	if (list_empty(&bus->devices))
2495 		return;
2496 
2497 	/*
2498 	 * If there's a bridge associated with that bus enable it. This works
2499 	 * around races in the generic code if the enabling is done during
2500 	 * parallel probing. This can be removed once those races have been
2501 	 * fixed.
2502 	 */
2503 	if (dev) {
2504 		int rc = pci_enable_device(dev);
2505 		if (rc)
2506 			pci_err(dev, "Error enabling bridge (%d)\n", rc);
2507 		pci_set_master(dev);
2508 	}
2509 
2510 	/* Perform the same to child busses */
2511 	list_for_each_entry(child, &bus->children, node)
2512 		pnv_pci_enable_bridge(child);
2513 }
2514 
2515 static void pnv_pci_enable_bridges(void)
2516 {
2517 	struct pci_controller *hose;
2518 
2519 	list_for_each_entry(hose, &hose_list, list_node)
2520 		pnv_pci_enable_bridge(hose->bus);
2521 }
2522 
2523 static void pnv_pci_ioda_fixup(void)
2524 {
2525 	pnv_pci_ioda_create_dbgfs();
2526 
2527 	pnv_pci_enable_bridges();
2528 
2529 #ifdef CONFIG_EEH
2530 	pnv_eeh_post_init();
2531 #endif
2532 }
2533 
2534 /*
2535  * Returns the alignment for I/O or memory windows for P2P
2536  * bridges. That actually depends on how PEs are segmented.
2537  * For now, we return I/O or M32 segment size for PE sensitive
2538  * P2P bridges. Otherwise, the default values (4KiB for I/O,
2539  * 1MiB for memory) will be returned.
2540  *
2541  * The current PCI bus might be put into one PE, which was
2542  * create against the parent PCI bridge. For that case, we
2543  * needn't enlarge the alignment so that we can save some
2544  * resources.
2545  */
2546 static resource_size_t pnv_pci_window_alignment(struct pci_bus *bus,
2547 						unsigned long type)
2548 {
2549 	struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
2550 	int num_pci_bridges = 0;
2551 	struct pci_dev *bridge;
2552 
2553 	bridge = bus->self;
2554 	while (bridge) {
2555 		if (pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE) {
2556 			num_pci_bridges++;
2557 			if (num_pci_bridges >= 2)
2558 				return 1;
2559 		}
2560 
2561 		bridge = bridge->bus->self;
2562 	}
2563 
2564 	/*
2565 	 * We fall back to M32 if M64 isn't supported. We enforce the M64
2566 	 * alignment for any 64-bit resource, PCIe doesn't care and
2567 	 * bridges only do 64-bit prefetchable anyway.
2568 	 */
2569 	if (phb->ioda.m64_segsize && pnv_pci_is_m64_flags(type))
2570 		return phb->ioda.m64_segsize;
2571 	if (type & IORESOURCE_MEM)
2572 		return phb->ioda.m32_segsize;
2573 
2574 	return phb->ioda.io_segsize;
2575 }
2576 
2577 /*
2578  * We are updating root port or the upstream port of the
2579  * bridge behind the root port with PHB's windows in order
2580  * to accommodate the changes on required resources during
2581  * PCI (slot) hotplug, which is connected to either root
2582  * port or the downstream ports of PCIe switch behind the
2583  * root port.
2584  */
2585 static void pnv_pci_fixup_bridge_resources(struct pci_bus *bus,
2586 					   unsigned long type)
2587 {
2588 	struct pci_controller *hose = pci_bus_to_host(bus);
2589 	struct pnv_phb *phb = hose->private_data;
2590 	struct pci_dev *bridge = bus->self;
2591 	struct resource *r, *w;
2592 	bool msi_region = false;
2593 	int i;
2594 
2595 	/* Check if we need apply fixup to the bridge's windows */
2596 	if (!pci_is_root_bus(bridge->bus) &&
2597 	    !pci_is_root_bus(bridge->bus->self->bus))
2598 		return;
2599 
2600 	/* Fixup the resources */
2601 	for (i = 0; i < PCI_BRIDGE_RESOURCE_NUM; i++) {
2602 		r = &bridge->resource[PCI_BRIDGE_RESOURCES + i];
2603 		if (!r->flags || !r->parent)
2604 			continue;
2605 
2606 		w = NULL;
2607 		if (r->flags & type & IORESOURCE_IO)
2608 			w = &hose->io_resource;
2609 		else if (pnv_pci_is_m64(phb, r) &&
2610 			 (type & IORESOURCE_PREFETCH) &&
2611 			 phb->ioda.m64_segsize)
2612 			w = &hose->mem_resources[1];
2613 		else if (r->flags & type & IORESOURCE_MEM) {
2614 			w = &hose->mem_resources[0];
2615 			msi_region = true;
2616 		}
2617 
2618 		r->start = w->start;
2619 		r->end = w->end;
2620 
2621 		/* The 64KB 32-bits MSI region shouldn't be included in
2622 		 * the 32-bits bridge window. Otherwise, we can see strange
2623 		 * issues. One of them is EEH error observed on Garrison.
2624 		 *
2625 		 * Exclude top 1MB region which is the minimal alignment of
2626 		 * 32-bits bridge window.
2627 		 */
2628 		if (msi_region) {
2629 			r->end += 0x10000;
2630 			r->end -= 0x100000;
2631 		}
2632 	}
2633 }
2634 
2635 static void pnv_pci_configure_bus(struct pci_bus *bus)
2636 {
2637 	struct pci_dev *bridge = bus->self;
2638 	struct pnv_ioda_pe *pe;
2639 	bool all = (bridge && pci_pcie_type(bridge) == PCI_EXP_TYPE_PCI_BRIDGE);
2640 
2641 	dev_info(&bus->dev, "Configuring PE for bus\n");
2642 
2643 	/* Don't assign PE to PCI bus, which doesn't have subordinate devices */
2644 	if (WARN_ON(list_empty(&bus->devices)))
2645 		return;
2646 
2647 	/* Reserve PEs according to used M64 resources */
2648 	pnv_ioda_reserve_m64_pe(bus, NULL, all);
2649 
2650 	/*
2651 	 * Assign PE. We might run here because of partial hotplug.
2652 	 * For the case, we just pick up the existing PE and should
2653 	 * not allocate resources again.
2654 	 */
2655 	pe = pnv_ioda_setup_bus_PE(bus, all);
2656 	if (!pe)
2657 		return;
2658 
2659 	pnv_ioda_setup_pe_seg(pe);
2660 }
2661 
2662 static resource_size_t pnv_pci_default_alignment(void)
2663 {
2664 	return PAGE_SIZE;
2665 }
2666 
2667 /* Prevent enabling devices for which we couldn't properly
2668  * assign a PE
2669  */
2670 static bool pnv_pci_enable_device_hook(struct pci_dev *dev)
2671 {
2672 	struct pci_dn *pdn;
2673 
2674 	pdn = pci_get_pdn(dev);
2675 	if (!pdn || pdn->pe_number == IODA_INVALID_PE) {
2676 		pci_err(dev, "pci_enable_device() blocked, no PE assigned.\n");
2677 		return false;
2678 	}
2679 
2680 	return true;
2681 }
2682 
2683 static bool pnv_ocapi_enable_device_hook(struct pci_dev *dev)
2684 {
2685 	struct pci_dn *pdn;
2686 	struct pnv_ioda_pe *pe;
2687 
2688 	pdn = pci_get_pdn(dev);
2689 	if (!pdn)
2690 		return false;
2691 
2692 	if (pdn->pe_number == IODA_INVALID_PE) {
2693 		pe = pnv_ioda_setup_dev_PE(dev);
2694 		if (!pe)
2695 			return false;
2696 	}
2697 	return true;
2698 }
2699 
2700 static long pnv_pci_ioda1_unset_window(struct iommu_table_group *table_group,
2701 				       int num)
2702 {
2703 	struct pnv_ioda_pe *pe = container_of(table_group,
2704 					      struct pnv_ioda_pe, table_group);
2705 	struct pnv_phb *phb = pe->phb;
2706 	unsigned int idx;
2707 	long rc;
2708 
2709 	pe_info(pe, "Removing DMA window #%d\n", num);
2710 	for (idx = 0; idx < phb->ioda.dma32_count; idx++) {
2711 		if (phb->ioda.dma32_segmap[idx] != pe->pe_number)
2712 			continue;
2713 
2714 		rc = opal_pci_map_pe_dma_window(phb->opal_id, pe->pe_number,
2715 						idx, 0, 0ul, 0ul, 0ul);
2716 		if (rc != OPAL_SUCCESS) {
2717 			pe_warn(pe, "Failure %ld unmapping DMA32 segment#%d\n",
2718 				rc, idx);
2719 			return rc;
2720 		}
2721 
2722 		phb->ioda.dma32_segmap[idx] = IODA_INVALID_PE;
2723 	}
2724 
2725 	pnv_pci_unlink_table_and_group(table_group->tables[num], table_group);
2726 	return OPAL_SUCCESS;
2727 }
2728 
2729 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe *pe)
2730 {
2731 	struct iommu_table *tbl = pe->table_group.tables[0];
2732 	int64_t rc;
2733 
2734 	if (!pe->dma_setup_done)
2735 		return;
2736 
2737 	rc = pnv_pci_ioda1_unset_window(&pe->table_group, 0);
2738 	if (rc != OPAL_SUCCESS)
2739 		return;
2740 
2741 	pnv_pci_p7ioc_tce_invalidate(tbl, tbl->it_offset, tbl->it_size, false);
2742 	if (pe->table_group.group) {
2743 		iommu_group_put(pe->table_group.group);
2744 		WARN_ON(pe->table_group.group);
2745 	}
2746 
2747 	free_pages(tbl->it_base, get_order(tbl->it_size << 3));
2748 	iommu_tce_table_put(tbl);
2749 }
2750 
2751 void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe *pe)
2752 {
2753 	struct iommu_table *tbl = pe->table_group.tables[0];
2754 	int64_t rc;
2755 
2756 	if (!pe->dma_setup_done)
2757 		return;
2758 
2759 	rc = pnv_pci_ioda2_unset_window(&pe->table_group, 0);
2760 	if (rc)
2761 		pe_warn(pe, "OPAL error %lld release DMA window\n", rc);
2762 
2763 	pnv_pci_ioda2_set_bypass(pe, false);
2764 	if (pe->table_group.group) {
2765 		iommu_group_put(pe->table_group.group);
2766 		WARN_ON(pe->table_group.group);
2767 	}
2768 
2769 	iommu_tce_table_put(tbl);
2770 }
2771 
2772 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe *pe,
2773 				 unsigned short win,
2774 				 unsigned int *map)
2775 {
2776 	struct pnv_phb *phb = pe->phb;
2777 	int idx;
2778 	int64_t rc;
2779 
2780 	for (idx = 0; idx < phb->ioda.total_pe_num; idx++) {
2781 		if (map[idx] != pe->pe_number)
2782 			continue;
2783 
2784 		rc = opal_pci_map_pe_mmio_window(phb->opal_id,
2785 				phb->ioda.reserved_pe_idx, win, 0, idx);
2786 
2787 		if (rc != OPAL_SUCCESS)
2788 			pe_warn(pe, "Error %lld unmapping (%d) segment#%d\n",
2789 				rc, win, idx);
2790 
2791 		map[idx] = IODA_INVALID_PE;
2792 	}
2793 }
2794 
2795 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe *pe)
2796 {
2797 	struct pnv_phb *phb = pe->phb;
2798 
2799 	if (phb->type == PNV_PHB_IODA1) {
2800 		pnv_ioda_free_pe_seg(pe, OPAL_IO_WINDOW_TYPE,
2801 				     phb->ioda.io_segmap);
2802 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
2803 				     phb->ioda.m32_segmap);
2804 		/* M64 is pre-configured by pnv_ioda1_init_m64() */
2805 	} else if (phb->type == PNV_PHB_IODA2) {
2806 		pnv_ioda_free_pe_seg(pe, OPAL_M32_WINDOW_TYPE,
2807 				     phb->ioda.m32_segmap);
2808 	}
2809 }
2810 
2811 static void pnv_ioda_release_pe(struct pnv_ioda_pe *pe)
2812 {
2813 	struct pnv_phb *phb = pe->phb;
2814 	struct pnv_ioda_pe *slave, *tmp;
2815 
2816 	pe_info(pe, "Releasing PE\n");
2817 
2818 	mutex_lock(&phb->ioda.pe_list_mutex);
2819 	list_del(&pe->list);
2820 	mutex_unlock(&phb->ioda.pe_list_mutex);
2821 
2822 	switch (phb->type) {
2823 	case PNV_PHB_IODA1:
2824 		pnv_pci_ioda1_release_pe_dma(pe);
2825 		break;
2826 	case PNV_PHB_IODA2:
2827 		pnv_pci_ioda2_release_pe_dma(pe);
2828 		break;
2829 	case PNV_PHB_NPU_OCAPI:
2830 		break;
2831 	default:
2832 		WARN_ON(1);
2833 	}
2834 
2835 	pnv_ioda_release_pe_seg(pe);
2836 	pnv_ioda_deconfigure_pe(pe->phb, pe);
2837 
2838 	/* Release slave PEs in the compound PE */
2839 	if (pe->flags & PNV_IODA_PE_MASTER) {
2840 		list_for_each_entry_safe(slave, tmp, &pe->slaves, list) {
2841 			list_del(&slave->list);
2842 			pnv_ioda_free_pe(slave);
2843 		}
2844 	}
2845 
2846 	/*
2847 	 * The PE for root bus can be removed because of hotplug in EEH
2848 	 * recovery for fenced PHB error. We need to mark the PE dead so
2849 	 * that it can be populated again in PCI hot add path. The PE
2850 	 * shouldn't be destroyed as it's the global reserved resource.
2851 	 */
2852 	if (phb->ioda.root_pe_idx == pe->pe_number)
2853 		return;
2854 
2855 	pnv_ioda_free_pe(pe);
2856 }
2857 
2858 static void pnv_pci_release_device(struct pci_dev *pdev)
2859 {
2860 	struct pnv_phb *phb = pci_bus_to_pnvhb(pdev->bus);
2861 	struct pci_dn *pdn = pci_get_pdn(pdev);
2862 	struct pnv_ioda_pe *pe;
2863 
2864 	/* The VF PE state is torn down when sriov_disable() is called */
2865 	if (pdev->is_virtfn)
2866 		return;
2867 
2868 	if (!pdn || pdn->pe_number == IODA_INVALID_PE)
2869 		return;
2870 
2871 #ifdef CONFIG_PCI_IOV
2872 	/*
2873 	 * FIXME: Try move this to sriov_disable(). It's here since we allocate
2874 	 * the iov state at probe time since we need to fiddle with the IOV
2875 	 * resources.
2876 	 */
2877 	if (pdev->is_physfn)
2878 		kfree(pdev->dev.archdata.iov_data);
2879 #endif
2880 
2881 	/*
2882 	 * PCI hotplug can happen as part of EEH error recovery. The @pdn
2883 	 * isn't removed and added afterwards in this scenario. We should
2884 	 * set the PE number in @pdn to an invalid one. Otherwise, the PE's
2885 	 * device count is decreased on removing devices while failing to
2886 	 * be increased on adding devices. It leads to unbalanced PE's device
2887 	 * count and eventually make normal PCI hotplug path broken.
2888 	 */
2889 	pe = &phb->ioda.pe_array[pdn->pe_number];
2890 	pdn->pe_number = IODA_INVALID_PE;
2891 
2892 	WARN_ON(--pe->device_count < 0);
2893 	if (pe->device_count == 0)
2894 		pnv_ioda_release_pe(pe);
2895 }
2896 
2897 static void pnv_pci_ioda_shutdown(struct pci_controller *hose)
2898 {
2899 	struct pnv_phb *phb = hose->private_data;
2900 
2901 	opal_pci_reset(phb->opal_id, OPAL_RESET_PCI_IODA_TABLE,
2902 		       OPAL_ASSERT_RESET);
2903 }
2904 
2905 static void pnv_pci_ioda_dma_bus_setup(struct pci_bus *bus)
2906 {
2907 	struct pnv_phb *phb = pci_bus_to_pnvhb(bus);
2908 	struct pnv_ioda_pe *pe;
2909 
2910 	list_for_each_entry(pe, &phb->ioda.pe_list, list) {
2911 		if (!(pe->flags & (PNV_IODA_PE_BUS | PNV_IODA_PE_BUS_ALL)))
2912 			continue;
2913 
2914 		if (!pe->pbus)
2915 			continue;
2916 
2917 		if (bus->number == ((pe->rid >> 8) & 0xFF)) {
2918 			pe->pbus = bus;
2919 			break;
2920 		}
2921 	}
2922 }
2923 
2924 static const struct pci_controller_ops pnv_pci_ioda_controller_ops = {
2925 	.dma_dev_setup		= pnv_pci_ioda_dma_dev_setup,
2926 	.dma_bus_setup		= pnv_pci_ioda_dma_bus_setup,
2927 	.iommu_bypass_supported	= pnv_pci_ioda_iommu_bypass_supported,
2928 	.enable_device_hook	= pnv_pci_enable_device_hook,
2929 	.release_device		= pnv_pci_release_device,
2930 	.window_alignment	= pnv_pci_window_alignment,
2931 	.setup_bridge		= pnv_pci_fixup_bridge_resources,
2932 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
2933 	.shutdown		= pnv_pci_ioda_shutdown,
2934 };
2935 
2936 static const struct pci_controller_ops pnv_npu_ocapi_ioda_controller_ops = {
2937 	.enable_device_hook	= pnv_ocapi_enable_device_hook,
2938 	.release_device		= pnv_pci_release_device,
2939 	.window_alignment	= pnv_pci_window_alignment,
2940 	.reset_secondary_bus	= pnv_pci_reset_secondary_bus,
2941 	.shutdown		= pnv_pci_ioda_shutdown,
2942 };
2943 
2944 static void __init pnv_pci_init_ioda_phb(struct device_node *np,
2945 					 u64 hub_id, int ioda_type)
2946 {
2947 	struct pci_controller *hose;
2948 	struct pnv_phb *phb;
2949 	unsigned long size, m64map_off, m32map_off, pemap_off;
2950 	unsigned long iomap_off = 0, dma32map_off = 0;
2951 	struct pnv_ioda_pe *root_pe;
2952 	struct resource r;
2953 	const __be64 *prop64;
2954 	const __be32 *prop32;
2955 	int len;
2956 	unsigned int segno;
2957 	u64 phb_id;
2958 	void *aux;
2959 	long rc;
2960 
2961 	if (!of_device_is_available(np))
2962 		return;
2963 
2964 	pr_info("Initializing %s PHB (%pOF)\n",	pnv_phb_names[ioda_type], np);
2965 
2966 	prop64 = of_get_property(np, "ibm,opal-phbid", NULL);
2967 	if (!prop64) {
2968 		pr_err("  Missing \"ibm,opal-phbid\" property !\n");
2969 		return;
2970 	}
2971 	phb_id = be64_to_cpup(prop64);
2972 	pr_debug("  PHB-ID  : 0x%016llx\n", phb_id);
2973 
2974 	phb = kzalloc(sizeof(*phb), GFP_KERNEL);
2975 	if (!phb)
2976 		panic("%s: Failed to allocate %zu bytes\n", __func__,
2977 		      sizeof(*phb));
2978 
2979 	/* Allocate PCI controller */
2980 	phb->hose = hose = pcibios_alloc_controller(np);
2981 	if (!phb->hose) {
2982 		pr_err("  Can't allocate PCI controller for %pOF\n",
2983 		       np);
2984 		memblock_free(__pa(phb), sizeof(struct pnv_phb));
2985 		return;
2986 	}
2987 
2988 	spin_lock_init(&phb->lock);
2989 	prop32 = of_get_property(np, "bus-range", &len);
2990 	if (prop32 && len == 8) {
2991 		hose->first_busno = be32_to_cpu(prop32[0]);
2992 		hose->last_busno = be32_to_cpu(prop32[1]);
2993 	} else {
2994 		pr_warn("  Broken <bus-range> on %pOF\n", np);
2995 		hose->first_busno = 0;
2996 		hose->last_busno = 0xff;
2997 	}
2998 	hose->private_data = phb;
2999 	phb->hub_id = hub_id;
3000 	phb->opal_id = phb_id;
3001 	phb->type = ioda_type;
3002 	mutex_init(&phb->ioda.pe_alloc_mutex);
3003 
3004 	/* Detect specific models for error handling */
3005 	if (of_device_is_compatible(np, "ibm,p7ioc-pciex"))
3006 		phb->model = PNV_PHB_MODEL_P7IOC;
3007 	else if (of_device_is_compatible(np, "ibm,power8-pciex"))
3008 		phb->model = PNV_PHB_MODEL_PHB3;
3009 	else
3010 		phb->model = PNV_PHB_MODEL_UNKNOWN;
3011 
3012 	/* Initialize diagnostic data buffer */
3013 	prop32 = of_get_property(np, "ibm,phb-diag-data-size", NULL);
3014 	if (prop32)
3015 		phb->diag_data_size = be32_to_cpup(prop32);
3016 	else
3017 		phb->diag_data_size = PNV_PCI_DIAG_BUF_SIZE;
3018 
3019 	phb->diag_data = kzalloc(phb->diag_data_size, GFP_KERNEL);
3020 	if (!phb->diag_data)
3021 		panic("%s: Failed to allocate %u bytes\n", __func__,
3022 		      phb->diag_data_size);
3023 
3024 	/* Parse 32-bit and IO ranges (if any) */
3025 	pci_process_bridge_OF_ranges(hose, np, !hose->global_number);
3026 
3027 	/* Get registers */
3028 	if (!of_address_to_resource(np, 0, &r)) {
3029 		phb->regs_phys = r.start;
3030 		phb->regs = ioremap(r.start, resource_size(&r));
3031 		if (phb->regs == NULL)
3032 			pr_err("  Failed to map registers !\n");
3033 	}
3034 
3035 	/* Initialize more IODA stuff */
3036 	phb->ioda.total_pe_num = 1;
3037 	prop32 = of_get_property(np, "ibm,opal-num-pes", NULL);
3038 	if (prop32)
3039 		phb->ioda.total_pe_num = be32_to_cpup(prop32);
3040 	prop32 = of_get_property(np, "ibm,opal-reserved-pe", NULL);
3041 	if (prop32)
3042 		phb->ioda.reserved_pe_idx = be32_to_cpup(prop32);
3043 
3044 	/* Invalidate RID to PE# mapping */
3045 	for (segno = 0; segno < ARRAY_SIZE(phb->ioda.pe_rmap); segno++)
3046 		phb->ioda.pe_rmap[segno] = IODA_INVALID_PE;
3047 
3048 	/* Parse 64-bit MMIO range */
3049 	pnv_ioda_parse_m64_window(phb);
3050 
3051 	phb->ioda.m32_size = resource_size(&hose->mem_resources[0]);
3052 	/* FW Has already off top 64k of M32 space (MSI space) */
3053 	phb->ioda.m32_size += 0x10000;
3054 
3055 	phb->ioda.m32_segsize = phb->ioda.m32_size / phb->ioda.total_pe_num;
3056 	phb->ioda.m32_pci_base = hose->mem_resources[0].start - hose->mem_offset[0];
3057 	phb->ioda.io_size = hose->pci_io_size;
3058 	phb->ioda.io_segsize = phb->ioda.io_size / phb->ioda.total_pe_num;
3059 	phb->ioda.io_pci_base = 0; /* XXX calculate this ? */
3060 
3061 	/* Calculate how many 32-bit TCE segments we have */
3062 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3063 				PNV_IODA1_DMA32_SEGSIZE;
3064 
3065 	/* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3066 	size = ALIGN(max_t(unsigned, phb->ioda.total_pe_num, 8) / 8,
3067 			sizeof(unsigned long));
3068 	m64map_off = size;
3069 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m64_segmap[0]);
3070 	m32map_off = size;
3071 	size += phb->ioda.total_pe_num * sizeof(phb->ioda.m32_segmap[0]);
3072 	if (phb->type == PNV_PHB_IODA1) {
3073 		iomap_off = size;
3074 		size += phb->ioda.total_pe_num * sizeof(phb->ioda.io_segmap[0]);
3075 		dma32map_off = size;
3076 		size += phb->ioda.dma32_count *
3077 			sizeof(phb->ioda.dma32_segmap[0]);
3078 	}
3079 	pemap_off = size;
3080 	size += phb->ioda.total_pe_num * sizeof(struct pnv_ioda_pe);
3081 	aux = kzalloc(size, GFP_KERNEL);
3082 	if (!aux)
3083 		panic("%s: Failed to allocate %lu bytes\n", __func__, size);
3084 
3085 	phb->ioda.pe_alloc = aux;
3086 	phb->ioda.m64_segmap = aux + m64map_off;
3087 	phb->ioda.m32_segmap = aux + m32map_off;
3088 	for (segno = 0; segno < phb->ioda.total_pe_num; segno++) {
3089 		phb->ioda.m64_segmap[segno] = IODA_INVALID_PE;
3090 		phb->ioda.m32_segmap[segno] = IODA_INVALID_PE;
3091 	}
3092 	if (phb->type == PNV_PHB_IODA1) {
3093 		phb->ioda.io_segmap = aux + iomap_off;
3094 		for (segno = 0; segno < phb->ioda.total_pe_num; segno++)
3095 			phb->ioda.io_segmap[segno] = IODA_INVALID_PE;
3096 
3097 		phb->ioda.dma32_segmap = aux + dma32map_off;
3098 		for (segno = 0; segno < phb->ioda.dma32_count; segno++)
3099 			phb->ioda.dma32_segmap[segno] = IODA_INVALID_PE;
3100 	}
3101 	phb->ioda.pe_array = aux + pemap_off;
3102 
3103 	/*
3104 	 * Choose PE number for root bus, which shouldn't have
3105 	 * M64 resources consumed by its child devices. To pick
3106 	 * the PE number adjacent to the reserved one if possible.
3107 	 */
3108 	pnv_ioda_reserve_pe(phb, phb->ioda.reserved_pe_idx);
3109 	if (phb->ioda.reserved_pe_idx == 0) {
3110 		phb->ioda.root_pe_idx = 1;
3111 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3112 	} else if (phb->ioda.reserved_pe_idx == (phb->ioda.total_pe_num - 1)) {
3113 		phb->ioda.root_pe_idx = phb->ioda.reserved_pe_idx - 1;
3114 		pnv_ioda_reserve_pe(phb, phb->ioda.root_pe_idx);
3115 	} else {
3116 		/* otherwise just allocate one */
3117 		root_pe = pnv_ioda_alloc_pe(phb, 1);
3118 		phb->ioda.root_pe_idx = root_pe->pe_number;
3119 	}
3120 
3121 	INIT_LIST_HEAD(&phb->ioda.pe_list);
3122 	mutex_init(&phb->ioda.pe_list_mutex);
3123 
3124 	/* Calculate how many 32-bit TCE segments we have */
3125 	phb->ioda.dma32_count = phb->ioda.m32_pci_base /
3126 				PNV_IODA1_DMA32_SEGSIZE;
3127 
3128 #if 0 /* We should really do that ... */
3129 	rc = opal_pci_set_phb_mem_window(opal->phb_id,
3130 					 window_type,
3131 					 window_num,
3132 					 starting_real_address,
3133 					 starting_pci_address,
3134 					 segment_size);
3135 #endif
3136 
3137 	pr_info("  %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3138 		phb->ioda.total_pe_num, phb->ioda.reserved_pe_idx,
3139 		phb->ioda.m32_size, phb->ioda.m32_segsize);
3140 	if (phb->ioda.m64_size)
3141 		pr_info("                 M64: 0x%lx [segment=0x%lx]\n",
3142 			phb->ioda.m64_size, phb->ioda.m64_segsize);
3143 	if (phb->ioda.io_size)
3144 		pr_info("                  IO: 0x%x [segment=0x%x]\n",
3145 			phb->ioda.io_size, phb->ioda.io_segsize);
3146 
3147 
3148 	phb->hose->ops = &pnv_pci_ops;
3149 	phb->get_pe_state = pnv_ioda_get_pe_state;
3150 	phb->freeze_pe = pnv_ioda_freeze_pe;
3151 	phb->unfreeze_pe = pnv_ioda_unfreeze_pe;
3152 
3153 	/* Setup MSI support */
3154 	pnv_pci_init_ioda_msis(phb);
3155 
3156 	/*
3157 	 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3158 	 * to let the PCI core do resource assignment. It's supposed
3159 	 * that the PCI core will do correct I/O and MMIO alignment
3160 	 * for the P2P bridge bars so that each PCI bus (excluding
3161 	 * the child P2P bridges) can form individual PE.
3162 	 */
3163 	ppc_md.pcibios_fixup = pnv_pci_ioda_fixup;
3164 
3165 	switch (phb->type) {
3166 	case PNV_PHB_NPU_OCAPI:
3167 		hose->controller_ops = pnv_npu_ocapi_ioda_controller_ops;
3168 		break;
3169 	default:
3170 		hose->controller_ops = pnv_pci_ioda_controller_ops;
3171 	}
3172 
3173 	ppc_md.pcibios_default_alignment = pnv_pci_default_alignment;
3174 
3175 #ifdef CONFIG_PCI_IOV
3176 	ppc_md.pcibios_fixup_sriov = pnv_pci_ioda_fixup_iov;
3177 	ppc_md.pcibios_iov_resource_alignment = pnv_pci_iov_resource_alignment;
3178 	ppc_md.pcibios_sriov_enable = pnv_pcibios_sriov_enable;
3179 	ppc_md.pcibios_sriov_disable = pnv_pcibios_sriov_disable;
3180 #endif
3181 
3182 	pci_add_flags(PCI_REASSIGN_ALL_RSRC);
3183 
3184 	/* Reset IODA tables to a clean state */
3185 	rc = opal_pci_reset(phb_id, OPAL_RESET_PCI_IODA_TABLE, OPAL_ASSERT_RESET);
3186 	if (rc)
3187 		pr_warn("  OPAL Error %ld performing IODA table reset !\n", rc);
3188 
3189 	/*
3190 	 * If we're running in kdump kernel, the previous kernel never
3191 	 * shutdown PCI devices correctly. We already got IODA table
3192 	 * cleaned out. So we have to issue PHB reset to stop all PCI
3193 	 * transactions from previous kernel. The ppc_pci_reset_phbs
3194 	 * kernel parameter will force this reset too. Additionally,
3195 	 * if the IODA reset above failed then use a bigger hammer.
3196 	 * This can happen if we get a PHB fatal error in very early
3197 	 * boot.
3198 	 */
3199 	if (is_kdump_kernel() || pci_reset_phbs || rc) {
3200 		pr_info("  Issue PHB reset ...\n");
3201 		pnv_eeh_phb_reset(hose, EEH_RESET_FUNDAMENTAL);
3202 		pnv_eeh_phb_reset(hose, EEH_RESET_DEACTIVATE);
3203 	}
3204 
3205 	/* Remove M64 resource if we can't configure it successfully */
3206 	if (!phb->init_m64 || phb->init_m64(phb))
3207 		hose->mem_resources[1].flags = 0;
3208 
3209 	/* create pci_dn's for DT nodes under this PHB */
3210 	pci_devs_phb_init_dynamic(hose);
3211 }
3212 
3213 void __init pnv_pci_init_ioda2_phb(struct device_node *np)
3214 {
3215 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_IODA2);
3216 }
3217 
3218 void __init pnv_pci_init_npu2_opencapi_phb(struct device_node *np)
3219 {
3220 	pnv_pci_init_ioda_phb(np, 0, PNV_PHB_NPU_OCAPI);
3221 }
3222 
3223 static void pnv_npu2_opencapi_cfg_size_fixup(struct pci_dev *dev)
3224 {
3225 	struct pnv_phb *phb = pci_bus_to_pnvhb(dev->bus);
3226 
3227 	if (!machine_is(powernv))
3228 		return;
3229 
3230 	if (phb->type == PNV_PHB_NPU_OCAPI)
3231 		dev->cfg_size = PCI_CFG_SPACE_EXP_SIZE;
3232 }
3233 DECLARE_PCI_FIXUP_EARLY(PCI_ANY_ID, PCI_ANY_ID, pnv_npu2_opencapi_cfg_size_fixup);
3234 
3235 void __init pnv_pci_init_ioda_hub(struct device_node *np)
3236 {
3237 	struct device_node *phbn;
3238 	const __be64 *prop64;
3239 	u64 hub_id;
3240 
3241 	pr_info("Probing IODA IO-Hub %pOF\n", np);
3242 
3243 	prop64 = of_get_property(np, "ibm,opal-hubid", NULL);
3244 	if (!prop64) {
3245 		pr_err(" Missing \"ibm,opal-hubid\" property !\n");
3246 		return;
3247 	}
3248 	hub_id = be64_to_cpup(prop64);
3249 	pr_devel(" HUB-ID : 0x%016llx\n", hub_id);
3250 
3251 	/* Count child PHBs */
3252 	for_each_child_of_node(np, phbn) {
3253 		/* Look for IODA1 PHBs */
3254 		if (of_device_is_compatible(phbn, "ibm,ioda-phb"))
3255 			pnv_pci_init_ioda_phb(phbn, hub_id, PNV_PHB_IODA1);
3256 	}
3257 }
3258