2 * Support PCI/PCIe on PowerNV platforms
4 * Copyright 2011 Benjamin Herrenschmidt, IBM Corp.
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
14 #include <linux/kernel.h>
15 #include <linux/pci.h>
16 #include <linux/crash_dump.h>
17 #include <linux/delay.h>
18 #include <linux/string.h>
19 #include <linux/init.h>
20 #include <linux/bootmem.h>
21 #include <linux/irq.h>
23 #include <linux/msi.h>
24 #include <linux/memblock.h>
25 #include <linux/iommu.h>
26 #include <linux/rculist.h>
27 #include <linux/sizes.h>
29 #include <asm/sections.h>
32 #include <asm/pci-bridge.h>
33 #include <asm/machdep.h>
34 #include <asm/msi_bitmap.h>
35 #include <asm/ppc-pci.h>
37 #include <asm/iommu.h>
40 #include <asm/debugfs.h>
41 #include <asm/firmware.h>
42 #include <asm/pnv-pci.h>
43 #include <asm/mmzone.h>
45 #include <misc/cxl-base.h>
50 #define PNV_IODA1_M64_NUM 16 /* Number of M64 BARs */
51 #define PNV_IODA1_M64_SEGS 8 /* Segments per M64 BAR */
52 #define PNV_IODA1_DMA32_SEGSIZE 0x10000000
54 #define POWERNV_IOMMU_DEFAULT_LEVELS 1
55 #define POWERNV_IOMMU_MAX_LEVELS 5
57 static const char * const pnv_phb_names
[] = { "IODA1", "IODA2", "NPU" };
58 static void pnv_pci_ioda2_table_free_pages(struct iommu_table
*tbl
);
60 void pe_level_printk(const struct pnv_ioda_pe
*pe
, const char *level
,
72 if (pe
->flags
& PNV_IODA_PE_DEV
)
73 strlcpy(pfix
, dev_name(&pe
->pdev
->dev
), sizeof(pfix
));
74 else if (pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
))
75 sprintf(pfix
, "%04x:%02x ",
76 pci_domain_nr(pe
->pbus
), pe
->pbus
->number
);
78 else if (pe
->flags
& PNV_IODA_PE_VF
)
79 sprintf(pfix
, "%04x:%02x:%2x.%d",
80 pci_domain_nr(pe
->parent_dev
->bus
),
81 (pe
->rid
& 0xff00) >> 8,
82 PCI_SLOT(pe
->rid
), PCI_FUNC(pe
->rid
));
83 #endif /* CONFIG_PCI_IOV*/
85 printk("%spci %s: [PE# %.2x] %pV",
86 level
, pfix
, pe
->pe_number
, &vaf
);
91 static bool pnv_iommu_bypass_disabled __read_mostly
;
93 static int __init
iommu_setup(char *str
)
99 if (!strncmp(str
, "nobypass", 8)) {
100 pnv_iommu_bypass_disabled
= true;
101 pr_info("PowerNV: IOMMU bypass window disabled.\n");
104 str
+= strcspn(str
, ",");
111 early_param("iommu", iommu_setup
);
113 static inline bool pnv_pci_is_m64(struct pnv_phb
*phb
, struct resource
*r
)
116 * WARNING: We cannot rely on the resource flags. The Linux PCI
117 * allocation code sometimes decides to put a 64-bit prefetchable
118 * BAR in the 32-bit window, so we have to compare the addresses.
120 * For simplicity we only test resource start.
122 return (r
->start
>= phb
->ioda
.m64_base
&&
123 r
->start
< (phb
->ioda
.m64_base
+ phb
->ioda
.m64_size
));
126 static inline bool pnv_pci_is_m64_flags(unsigned long resource_flags
)
128 unsigned long flags
= (IORESOURCE_MEM_64
| IORESOURCE_PREFETCH
);
130 return (resource_flags
& flags
) == flags
;
133 static struct pnv_ioda_pe
*pnv_ioda_init_pe(struct pnv_phb
*phb
, int pe_no
)
137 phb
->ioda
.pe_array
[pe_no
].phb
= phb
;
138 phb
->ioda
.pe_array
[pe_no
].pe_number
= pe_no
;
141 * Clear the PE frozen state as it might be put into frozen state
142 * in the last PCI remove path. It's not harmful to do so when the
143 * PE is already in unfrozen state.
145 rc
= opal_pci_eeh_freeze_clear(phb
->opal_id
, pe_no
,
146 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
147 if (rc
!= OPAL_SUCCESS
&& rc
!= OPAL_UNSUPPORTED
)
148 pr_warn("%s: Error %lld unfreezing PHB#%x-PE#%x\n",
149 __func__
, rc
, phb
->hose
->global_number
, pe_no
);
151 return &phb
->ioda
.pe_array
[pe_no
];
154 static void pnv_ioda_reserve_pe(struct pnv_phb
*phb
, int pe_no
)
156 if (!(pe_no
>= 0 && pe_no
< phb
->ioda
.total_pe_num
)) {
157 pr_warn("%s: Invalid PE %x on PHB#%x\n",
158 __func__
, pe_no
, phb
->hose
->global_number
);
162 if (test_and_set_bit(pe_no
, phb
->ioda
.pe_alloc
))
163 pr_debug("%s: PE %x was reserved on PHB#%x\n",
164 __func__
, pe_no
, phb
->hose
->global_number
);
166 pnv_ioda_init_pe(phb
, pe_no
);
169 static struct pnv_ioda_pe
*pnv_ioda_alloc_pe(struct pnv_phb
*phb
)
173 for (pe
= phb
->ioda
.total_pe_num
- 1; pe
>= 0; pe
--) {
174 if (!test_and_set_bit(pe
, phb
->ioda
.pe_alloc
))
175 return pnv_ioda_init_pe(phb
, pe
);
181 static void pnv_ioda_free_pe(struct pnv_ioda_pe
*pe
)
183 struct pnv_phb
*phb
= pe
->phb
;
184 unsigned int pe_num
= pe
->pe_number
;
188 memset(pe
, 0, sizeof(struct pnv_ioda_pe
));
189 clear_bit(pe_num
, phb
->ioda
.pe_alloc
);
192 /* The default M64 BAR is shared by all PEs */
193 static int pnv_ioda2_init_m64(struct pnv_phb
*phb
)
199 /* Configure the default M64 BAR */
200 rc
= opal_pci_set_phb_mem_window(phb
->opal_id
,
201 OPAL_M64_WINDOW_TYPE
,
202 phb
->ioda
.m64_bar_idx
,
206 if (rc
!= OPAL_SUCCESS
) {
207 desc
= "configuring";
211 /* Enable the default M64 BAR */
212 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
213 OPAL_M64_WINDOW_TYPE
,
214 phb
->ioda
.m64_bar_idx
,
215 OPAL_ENABLE_M64_SPLIT
);
216 if (rc
!= OPAL_SUCCESS
) {
222 * Exclude the segments for reserved and root bus PE, which
223 * are first or last two PEs.
225 r
= &phb
->hose
->mem_resources
[1];
226 if (phb
->ioda
.reserved_pe_idx
== 0)
227 r
->start
+= (2 * phb
->ioda
.m64_segsize
);
228 else if (phb
->ioda
.reserved_pe_idx
== (phb
->ioda
.total_pe_num
- 1))
229 r
->end
-= (2 * phb
->ioda
.m64_segsize
);
231 pr_warn(" Cannot strip M64 segment for reserved PE#%x\n",
232 phb
->ioda
.reserved_pe_idx
);
237 pr_warn(" Failure %lld %s M64 BAR#%d\n",
238 rc
, desc
, phb
->ioda
.m64_bar_idx
);
239 opal_pci_phb_mmio_enable(phb
->opal_id
,
240 OPAL_M64_WINDOW_TYPE
,
241 phb
->ioda
.m64_bar_idx
,
246 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev
*pdev
,
247 unsigned long *pe_bitmap
)
249 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
250 struct pnv_phb
*phb
= hose
->private_data
;
252 resource_size_t base
, sgsz
, start
, end
;
255 base
= phb
->ioda
.m64_base
;
256 sgsz
= phb
->ioda
.m64_segsize
;
257 for (i
= 0; i
<= PCI_ROM_RESOURCE
; i
++) {
258 r
= &pdev
->resource
[i
];
259 if (!r
->parent
|| !pnv_pci_is_m64(phb
, r
))
262 start
= _ALIGN_DOWN(r
->start
- base
, sgsz
);
263 end
= _ALIGN_UP(r
->end
- base
, sgsz
);
264 for (segno
= start
/ sgsz
; segno
< end
/ sgsz
; segno
++) {
266 set_bit(segno
, pe_bitmap
);
268 pnv_ioda_reserve_pe(phb
, segno
);
273 static int pnv_ioda1_init_m64(struct pnv_phb
*phb
)
279 * There are 16 M64 BARs, each of which has 8 segments. So
280 * there are as many M64 segments as the maximum number of
283 for (index
= 0; index
< PNV_IODA1_M64_NUM
; index
++) {
284 unsigned long base
, segsz
= phb
->ioda
.m64_segsize
;
287 base
= phb
->ioda
.m64_base
+
288 index
* PNV_IODA1_M64_SEGS
* segsz
;
289 rc
= opal_pci_set_phb_mem_window(phb
->opal_id
,
290 OPAL_M64_WINDOW_TYPE
, index
, base
, 0,
291 PNV_IODA1_M64_SEGS
* segsz
);
292 if (rc
!= OPAL_SUCCESS
) {
293 pr_warn(" Error %lld setting M64 PHB#%x-BAR#%d\n",
294 rc
, phb
->hose
->global_number
, index
);
298 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
299 OPAL_M64_WINDOW_TYPE
, index
,
300 OPAL_ENABLE_M64_SPLIT
);
301 if (rc
!= OPAL_SUCCESS
) {
302 pr_warn(" Error %lld enabling M64 PHB#%x-BAR#%d\n",
303 rc
, phb
->hose
->global_number
, index
);
309 * Exclude the segments for reserved and root bus PE, which
310 * are first or last two PEs.
312 r
= &phb
->hose
->mem_resources
[1];
313 if (phb
->ioda
.reserved_pe_idx
== 0)
314 r
->start
+= (2 * phb
->ioda
.m64_segsize
);
315 else if (phb
->ioda
.reserved_pe_idx
== (phb
->ioda
.total_pe_num
- 1))
316 r
->end
-= (2 * phb
->ioda
.m64_segsize
);
318 WARN(1, "Wrong reserved PE#%x on PHB#%x\n",
319 phb
->ioda
.reserved_pe_idx
, phb
->hose
->global_number
);
324 for ( ; index
>= 0; index
--)
325 opal_pci_phb_mmio_enable(phb
->opal_id
,
326 OPAL_M64_WINDOW_TYPE
, index
, OPAL_DISABLE_M64
);
331 static void pnv_ioda_reserve_m64_pe(struct pci_bus
*bus
,
332 unsigned long *pe_bitmap
,
335 struct pci_dev
*pdev
;
337 list_for_each_entry(pdev
, &bus
->devices
, bus_list
) {
338 pnv_ioda_reserve_dev_m64_pe(pdev
, pe_bitmap
);
340 if (all
&& pdev
->subordinate
)
341 pnv_ioda_reserve_m64_pe(pdev
->subordinate
,
346 static struct pnv_ioda_pe
*pnv_ioda_pick_m64_pe(struct pci_bus
*bus
, bool all
)
348 struct pci_controller
*hose
= pci_bus_to_host(bus
);
349 struct pnv_phb
*phb
= hose
->private_data
;
350 struct pnv_ioda_pe
*master_pe
, *pe
;
351 unsigned long size
, *pe_alloc
;
354 /* Root bus shouldn't use M64 */
355 if (pci_is_root_bus(bus
))
358 /* Allocate bitmap */
359 size
= _ALIGN_UP(phb
->ioda
.total_pe_num
/ 8, sizeof(unsigned long));
360 pe_alloc
= kzalloc(size
, GFP_KERNEL
);
362 pr_warn("%s: Out of memory !\n",
367 /* Figure out reserved PE numbers by the PE */
368 pnv_ioda_reserve_m64_pe(bus
, pe_alloc
, all
);
371 * the current bus might not own M64 window and that's all
372 * contributed by its child buses. For the case, we needn't
373 * pick M64 dependent PE#.
375 if (bitmap_empty(pe_alloc
, phb
->ioda
.total_pe_num
)) {
381 * Figure out the master PE and put all slave PEs to master
382 * PE's list to form compound PE.
386 while ((i
= find_next_bit(pe_alloc
, phb
->ioda
.total_pe_num
, i
+ 1)) <
387 phb
->ioda
.total_pe_num
) {
388 pe
= &phb
->ioda
.pe_array
[i
];
390 phb
->ioda
.m64_segmap
[pe
->pe_number
] = pe
->pe_number
;
392 pe
->flags
|= PNV_IODA_PE_MASTER
;
393 INIT_LIST_HEAD(&pe
->slaves
);
396 pe
->flags
|= PNV_IODA_PE_SLAVE
;
397 pe
->master
= master_pe
;
398 list_add_tail(&pe
->list
, &master_pe
->slaves
);
402 * P7IOC supports M64DT, which helps mapping M64 segment
403 * to one particular PE#. However, PHB3 has fixed mapping
404 * between M64 segment and PE#. In order to have same logic
405 * for P7IOC and PHB3, we enforce fixed mapping between M64
406 * segment and PE# on P7IOC.
408 if (phb
->type
== PNV_PHB_IODA1
) {
411 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
412 pe
->pe_number
, OPAL_M64_WINDOW_TYPE
,
413 pe
->pe_number
/ PNV_IODA1_M64_SEGS
,
414 pe
->pe_number
% PNV_IODA1_M64_SEGS
);
415 if (rc
!= OPAL_SUCCESS
)
416 pr_warn("%s: Error %lld mapping M64 for PHB#%x-PE#%x\n",
417 __func__
, rc
, phb
->hose
->global_number
,
426 static void __init
pnv_ioda_parse_m64_window(struct pnv_phb
*phb
)
428 struct pci_controller
*hose
= phb
->hose
;
429 struct device_node
*dn
= hose
->dn
;
430 struct resource
*res
;
435 if (phb
->type
!= PNV_PHB_IODA1
&& phb
->type
!= PNV_PHB_IODA2
) {
436 pr_info(" Not support M64 window\n");
440 if (!firmware_has_feature(FW_FEATURE_OPAL
)) {
441 pr_info(" Firmware too old to support M64 window\n");
445 r
= of_get_property(dn
, "ibm,opal-m64-window", NULL
);
447 pr_info(" No <ibm,opal-m64-window> on %s\n",
453 * Find the available M64 BAR range and pickup the last one for
454 * covering the whole 64-bits space. We support only one range.
456 if (of_property_read_u32_array(dn
, "ibm,opal-available-m64-ranges",
458 /* In absence of the property, assume 0..15 */
462 /* We only support 64 bits in our allocator */
463 if (m64_range
[1] > 63) {
464 pr_warn("%s: Limiting M64 range to 63 (from %d) on PHB#%x\n",
465 __func__
, m64_range
[1], phb
->hose
->global_number
);
468 /* Empty range, no m64 */
469 if (m64_range
[1] <= m64_range
[0]) {
470 pr_warn("%s: M64 empty, disabling M64 usage on PHB#%x\n",
471 __func__
, phb
->hose
->global_number
);
475 /* Configure M64 informations */
476 res
= &hose
->mem_resources
[1];
477 res
->name
= dn
->full_name
;
478 res
->start
= of_translate_address(dn
, r
+ 2);
479 res
->end
= res
->start
+ of_read_number(r
+ 4, 2) - 1;
480 res
->flags
= (IORESOURCE_MEM
| IORESOURCE_MEM_64
| IORESOURCE_PREFETCH
);
481 pci_addr
= of_read_number(r
, 2);
482 hose
->mem_offset
[1] = res
->start
- pci_addr
;
484 phb
->ioda
.m64_size
= resource_size(res
);
485 phb
->ioda
.m64_segsize
= phb
->ioda
.m64_size
/ phb
->ioda
.total_pe_num
;
486 phb
->ioda
.m64_base
= pci_addr
;
488 /* This lines up nicely with the display from processing OF ranges */
489 pr_info(" MEM 0x%016llx..0x%016llx -> 0x%016llx (M64 #%d..%d)\n",
490 res
->start
, res
->end
, pci_addr
, m64_range
[0],
491 m64_range
[0] + m64_range
[1] - 1);
493 /* Mark all M64 used up by default */
494 phb
->ioda
.m64_bar_alloc
= (unsigned long)-1;
496 /* Use last M64 BAR to cover M64 window */
498 phb
->ioda
.m64_bar_idx
= m64_range
[0] + m64_range
[1];
500 pr_info(" Using M64 #%d as default window\n", phb
->ioda
.m64_bar_idx
);
502 /* Mark remaining ones free */
503 for (i
= m64_range
[0]; i
< m64_range
[1]; i
++)
504 clear_bit(i
, &phb
->ioda
.m64_bar_alloc
);
507 * Setup init functions for M64 based on IODA version, IODA3 uses
510 if (phb
->type
== PNV_PHB_IODA1
)
511 phb
->init_m64
= pnv_ioda1_init_m64
;
513 phb
->init_m64
= pnv_ioda2_init_m64
;
514 phb
->reserve_m64_pe
= pnv_ioda_reserve_m64_pe
;
515 phb
->pick_m64_pe
= pnv_ioda_pick_m64_pe
;
518 static void pnv_ioda_freeze_pe(struct pnv_phb
*phb
, int pe_no
)
520 struct pnv_ioda_pe
*pe
= &phb
->ioda
.pe_array
[pe_no
];
521 struct pnv_ioda_pe
*slave
;
524 /* Fetch master PE */
525 if (pe
->flags
& PNV_IODA_PE_SLAVE
) {
527 if (WARN_ON(!pe
|| !(pe
->flags
& PNV_IODA_PE_MASTER
)))
530 pe_no
= pe
->pe_number
;
533 /* Freeze master PE */
534 rc
= opal_pci_eeh_freeze_set(phb
->opal_id
,
536 OPAL_EEH_ACTION_SET_FREEZE_ALL
);
537 if (rc
!= OPAL_SUCCESS
) {
538 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
539 __func__
, rc
, phb
->hose
->global_number
, pe_no
);
543 /* Freeze slave PEs */
544 if (!(pe
->flags
& PNV_IODA_PE_MASTER
))
547 list_for_each_entry(slave
, &pe
->slaves
, list
) {
548 rc
= opal_pci_eeh_freeze_set(phb
->opal_id
,
550 OPAL_EEH_ACTION_SET_FREEZE_ALL
);
551 if (rc
!= OPAL_SUCCESS
)
552 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
553 __func__
, rc
, phb
->hose
->global_number
,
558 static int pnv_ioda_unfreeze_pe(struct pnv_phb
*phb
, int pe_no
, int opt
)
560 struct pnv_ioda_pe
*pe
, *slave
;
564 pe
= &phb
->ioda
.pe_array
[pe_no
];
565 if (pe
->flags
& PNV_IODA_PE_SLAVE
) {
567 WARN_ON(!pe
|| !(pe
->flags
& PNV_IODA_PE_MASTER
));
568 pe_no
= pe
->pe_number
;
571 /* Clear frozen state for master PE */
572 rc
= opal_pci_eeh_freeze_clear(phb
->opal_id
, pe_no
, opt
);
573 if (rc
!= OPAL_SUCCESS
) {
574 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
575 __func__
, rc
, opt
, phb
->hose
->global_number
, pe_no
);
579 if (!(pe
->flags
& PNV_IODA_PE_MASTER
))
582 /* Clear frozen state for slave PEs */
583 list_for_each_entry(slave
, &pe
->slaves
, list
) {
584 rc
= opal_pci_eeh_freeze_clear(phb
->opal_id
,
587 if (rc
!= OPAL_SUCCESS
) {
588 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
589 __func__
, rc
, opt
, phb
->hose
->global_number
,
598 static int pnv_ioda_get_pe_state(struct pnv_phb
*phb
, int pe_no
)
600 struct pnv_ioda_pe
*slave
, *pe
;
605 /* Sanity check on PE number */
606 if (pe_no
< 0 || pe_no
>= phb
->ioda
.total_pe_num
)
607 return OPAL_EEH_STOPPED_PERM_UNAVAIL
;
610 * Fetch the master PE and the PE instance might be
611 * not initialized yet.
613 pe
= &phb
->ioda
.pe_array
[pe_no
];
614 if (pe
->flags
& PNV_IODA_PE_SLAVE
) {
616 WARN_ON(!pe
|| !(pe
->flags
& PNV_IODA_PE_MASTER
));
617 pe_no
= pe
->pe_number
;
620 /* Check the master PE */
621 rc
= opal_pci_eeh_freeze_status(phb
->opal_id
, pe_no
,
622 &state
, &pcierr
, NULL
);
623 if (rc
!= OPAL_SUCCESS
) {
624 pr_warn("%s: Failure %lld getting "
625 "PHB#%x-PE#%x state\n",
627 phb
->hose
->global_number
, pe_no
);
628 return OPAL_EEH_STOPPED_TEMP_UNAVAIL
;
631 /* Check the slave PE */
632 if (!(pe
->flags
& PNV_IODA_PE_MASTER
))
635 list_for_each_entry(slave
, &pe
->slaves
, list
) {
636 rc
= opal_pci_eeh_freeze_status(phb
->opal_id
,
641 if (rc
!= OPAL_SUCCESS
) {
642 pr_warn("%s: Failure %lld getting "
643 "PHB#%x-PE#%x state\n",
645 phb
->hose
->global_number
, slave
->pe_number
);
646 return OPAL_EEH_STOPPED_TEMP_UNAVAIL
;
650 * Override the result based on the ascending
660 /* Currently those 2 are only used when MSIs are enabled, this will change
661 * but in the meantime, we need to protect them to avoid warnings
663 #ifdef CONFIG_PCI_MSI
664 struct pnv_ioda_pe
*pnv_ioda_get_pe(struct pci_dev
*dev
)
666 struct pci_controller
*hose
= pci_bus_to_host(dev
->bus
);
667 struct pnv_phb
*phb
= hose
->private_data
;
668 struct pci_dn
*pdn
= pci_get_pdn(dev
);
672 if (pdn
->pe_number
== IODA_INVALID_PE
)
674 return &phb
->ioda
.pe_array
[pdn
->pe_number
];
676 #endif /* CONFIG_PCI_MSI */
678 static int pnv_ioda_set_one_peltv(struct pnv_phb
*phb
,
679 struct pnv_ioda_pe
*parent
,
680 struct pnv_ioda_pe
*child
,
683 const char *desc
= is_add
? "adding" : "removing";
684 uint8_t op
= is_add
? OPAL_ADD_PE_TO_DOMAIN
:
685 OPAL_REMOVE_PE_FROM_DOMAIN
;
686 struct pnv_ioda_pe
*slave
;
689 /* Parent PE affects child PE */
690 rc
= opal_pci_set_peltv(phb
->opal_id
, parent
->pe_number
,
691 child
->pe_number
, op
);
692 if (rc
!= OPAL_SUCCESS
) {
693 pe_warn(child
, "OPAL error %ld %s to parent PELTV\n",
698 if (!(child
->flags
& PNV_IODA_PE_MASTER
))
701 /* Compound case: parent PE affects slave PEs */
702 list_for_each_entry(slave
, &child
->slaves
, list
) {
703 rc
= opal_pci_set_peltv(phb
->opal_id
, parent
->pe_number
,
704 slave
->pe_number
, op
);
705 if (rc
!= OPAL_SUCCESS
) {
706 pe_warn(slave
, "OPAL error %ld %s to parent PELTV\n",
715 static int pnv_ioda_set_peltv(struct pnv_phb
*phb
,
716 struct pnv_ioda_pe
*pe
,
719 struct pnv_ioda_pe
*slave
;
720 struct pci_dev
*pdev
= NULL
;
724 * Clear PE frozen state. If it's master PE, we need
725 * clear slave PE frozen state as well.
728 opal_pci_eeh_freeze_clear(phb
->opal_id
, pe
->pe_number
,
729 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
730 if (pe
->flags
& PNV_IODA_PE_MASTER
) {
731 list_for_each_entry(slave
, &pe
->slaves
, list
)
732 opal_pci_eeh_freeze_clear(phb
->opal_id
,
734 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
739 * Associate PE in PELT. We need add the PE into the
740 * corresponding PELT-V as well. Otherwise, the error
741 * originated from the PE might contribute to other
744 ret
= pnv_ioda_set_one_peltv(phb
, pe
, pe
, is_add
);
748 /* For compound PEs, any one affects all of them */
749 if (pe
->flags
& PNV_IODA_PE_MASTER
) {
750 list_for_each_entry(slave
, &pe
->slaves
, list
) {
751 ret
= pnv_ioda_set_one_peltv(phb
, slave
, pe
, is_add
);
757 if (pe
->flags
& (PNV_IODA_PE_BUS_ALL
| PNV_IODA_PE_BUS
))
758 pdev
= pe
->pbus
->self
;
759 else if (pe
->flags
& PNV_IODA_PE_DEV
)
760 pdev
= pe
->pdev
->bus
->self
;
761 #ifdef CONFIG_PCI_IOV
762 else if (pe
->flags
& PNV_IODA_PE_VF
)
763 pdev
= pe
->parent_dev
;
764 #endif /* CONFIG_PCI_IOV */
766 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
767 struct pnv_ioda_pe
*parent
;
769 if (pdn
&& pdn
->pe_number
!= IODA_INVALID_PE
) {
770 parent
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
771 ret
= pnv_ioda_set_one_peltv(phb
, parent
, pe
, is_add
);
776 pdev
= pdev
->bus
->self
;
782 static int pnv_ioda_deconfigure_pe(struct pnv_phb
*phb
, struct pnv_ioda_pe
*pe
)
784 struct pci_dev
*parent
;
785 uint8_t bcomp
, dcomp
, fcomp
;
789 /* Currently, we just deconfigure VF PE. Bus PE will always there.*/
793 dcomp
= OPAL_IGNORE_RID_DEVICE_NUMBER
;
794 fcomp
= OPAL_IGNORE_RID_FUNCTION_NUMBER
;
795 parent
= pe
->pbus
->self
;
796 if (pe
->flags
& PNV_IODA_PE_BUS_ALL
)
797 count
= pe
->pbus
->busn_res
.end
- pe
->pbus
->busn_res
.start
+ 1;
802 case 1: bcomp
= OpalPciBusAll
; break;
803 case 2: bcomp
= OpalPciBus7Bits
; break;
804 case 4: bcomp
= OpalPciBus6Bits
; break;
805 case 8: bcomp
= OpalPciBus5Bits
; break;
806 case 16: bcomp
= OpalPciBus4Bits
; break;
807 case 32: bcomp
= OpalPciBus3Bits
; break;
809 dev_err(&pe
->pbus
->dev
, "Number of subordinate buses %d unsupported\n",
811 /* Do an exact match only */
812 bcomp
= OpalPciBusAll
;
814 rid_end
= pe
->rid
+ (count
<< 8);
816 #ifdef CONFIG_PCI_IOV
817 if (pe
->flags
& PNV_IODA_PE_VF
)
818 parent
= pe
->parent_dev
;
821 parent
= pe
->pdev
->bus
->self
;
822 bcomp
= OpalPciBusAll
;
823 dcomp
= OPAL_COMPARE_RID_DEVICE_NUMBER
;
824 fcomp
= OPAL_COMPARE_RID_FUNCTION_NUMBER
;
825 rid_end
= pe
->rid
+ 1;
828 /* Clear the reverse map */
829 for (rid
= pe
->rid
; rid
< rid_end
; rid
++)
830 phb
->ioda
.pe_rmap
[rid
] = IODA_INVALID_PE
;
832 /* Release from all parents PELT-V */
834 struct pci_dn
*pdn
= pci_get_pdn(parent
);
835 if (pdn
&& pdn
->pe_number
!= IODA_INVALID_PE
) {
836 rc
= opal_pci_set_peltv(phb
->opal_id
, pdn
->pe_number
,
837 pe
->pe_number
, OPAL_REMOVE_PE_FROM_DOMAIN
);
838 /* XXX What to do in case of error ? */
840 parent
= parent
->bus
->self
;
843 opal_pci_eeh_freeze_clear(phb
->opal_id
, pe
->pe_number
,
844 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
846 /* Disassociate PE in PELT */
847 rc
= opal_pci_set_peltv(phb
->opal_id
, pe
->pe_number
,
848 pe
->pe_number
, OPAL_REMOVE_PE_FROM_DOMAIN
);
850 pe_warn(pe
, "OPAL error %ld remove self from PELTV\n", rc
);
851 rc
= opal_pci_set_pe(phb
->opal_id
, pe
->pe_number
, pe
->rid
,
852 bcomp
, dcomp
, fcomp
, OPAL_UNMAP_PE
);
854 pe_err(pe
, "OPAL error %ld trying to setup PELT table\n", rc
);
858 #ifdef CONFIG_PCI_IOV
859 pe
->parent_dev
= NULL
;
865 static int pnv_ioda_configure_pe(struct pnv_phb
*phb
, struct pnv_ioda_pe
*pe
)
867 struct pci_dev
*parent
;
868 uint8_t bcomp
, dcomp
, fcomp
;
869 long rc
, rid_end
, rid
;
871 /* Bus validation ? */
875 dcomp
= OPAL_IGNORE_RID_DEVICE_NUMBER
;
876 fcomp
= OPAL_IGNORE_RID_FUNCTION_NUMBER
;
877 parent
= pe
->pbus
->self
;
878 if (pe
->flags
& PNV_IODA_PE_BUS_ALL
)
879 count
= pe
->pbus
->busn_res
.end
- pe
->pbus
->busn_res
.start
+ 1;
884 case 1: bcomp
= OpalPciBusAll
; break;
885 case 2: bcomp
= OpalPciBus7Bits
; break;
886 case 4: bcomp
= OpalPciBus6Bits
; break;
887 case 8: bcomp
= OpalPciBus5Bits
; break;
888 case 16: bcomp
= OpalPciBus4Bits
; break;
889 case 32: bcomp
= OpalPciBus3Bits
; break;
891 dev_err(&pe
->pbus
->dev
, "Number of subordinate buses %d unsupported\n",
893 /* Do an exact match only */
894 bcomp
= OpalPciBusAll
;
896 rid_end
= pe
->rid
+ (count
<< 8);
898 #ifdef CONFIG_PCI_IOV
899 if (pe
->flags
& PNV_IODA_PE_VF
)
900 parent
= pe
->parent_dev
;
902 #endif /* CONFIG_PCI_IOV */
903 parent
= pe
->pdev
->bus
->self
;
904 bcomp
= OpalPciBusAll
;
905 dcomp
= OPAL_COMPARE_RID_DEVICE_NUMBER
;
906 fcomp
= OPAL_COMPARE_RID_FUNCTION_NUMBER
;
907 rid_end
= pe
->rid
+ 1;
911 * Associate PE in PELT. We need add the PE into the
912 * corresponding PELT-V as well. Otherwise, the error
913 * originated from the PE might contribute to other
916 rc
= opal_pci_set_pe(phb
->opal_id
, pe
->pe_number
, pe
->rid
,
917 bcomp
, dcomp
, fcomp
, OPAL_MAP_PE
);
919 pe_err(pe
, "OPAL error %ld trying to setup PELT table\n", rc
);
924 * Configure PELTV. NPUs don't have a PELTV table so skip
925 * configuration on them.
927 if (phb
->type
!= PNV_PHB_NPU
)
928 pnv_ioda_set_peltv(phb
, pe
, true);
930 /* Setup reverse map */
931 for (rid
= pe
->rid
; rid
< rid_end
; rid
++)
932 phb
->ioda
.pe_rmap
[rid
] = pe
->pe_number
;
934 /* Setup one MVTs on IODA1 */
935 if (phb
->type
!= PNV_PHB_IODA1
) {
940 pe
->mve_number
= pe
->pe_number
;
941 rc
= opal_pci_set_mve(phb
->opal_id
, pe
->mve_number
, pe
->pe_number
);
942 if (rc
!= OPAL_SUCCESS
) {
943 pe_err(pe
, "OPAL error %ld setting up MVE %x\n",
947 rc
= opal_pci_set_mve_enable(phb
->opal_id
,
948 pe
->mve_number
, OPAL_ENABLE_MVE
);
950 pe_err(pe
, "OPAL error %ld enabling MVE %x\n",
960 #ifdef CONFIG_PCI_IOV
961 static int pnv_pci_vf_resource_shift(struct pci_dev
*dev
, int offset
)
963 struct pci_dn
*pdn
= pci_get_pdn(dev
);
965 struct resource
*res
, res2
;
966 resource_size_t size
;
973 * "offset" is in VFs. The M64 windows are sized so that when they
974 * are segmented, each segment is the same size as the IOV BAR.
975 * Each segment is in a separate PE, and the high order bits of the
976 * address are the PE number. Therefore, each VF's BAR is in a
977 * separate PE, and changing the IOV BAR start address changes the
978 * range of PEs the VFs are in.
980 num_vfs
= pdn
->num_vfs
;
981 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
982 res
= &dev
->resource
[i
+ PCI_IOV_RESOURCES
];
983 if (!res
->flags
|| !res
->parent
)
987 * The actual IOV BAR range is determined by the start address
988 * and the actual size for num_vfs VFs BAR. This check is to
989 * make sure that after shifting, the range will not overlap
990 * with another device.
992 size
= pci_iov_resource_size(dev
, i
+ PCI_IOV_RESOURCES
);
993 res2
.flags
= res
->flags
;
994 res2
.start
= res
->start
+ (size
* offset
);
995 res2
.end
= res2
.start
+ (size
* num_vfs
) - 1;
997 if (res2
.end
> res
->end
) {
998 dev_err(&dev
->dev
, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
999 i
, &res2
, res
, num_vfs
, offset
);
1005 * After doing so, there would be a "hole" in the /proc/iomem when
1006 * offset is a positive value. It looks like the device return some
1007 * mmio back to the system, which actually no one could use it.
1009 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
1010 res
= &dev
->resource
[i
+ PCI_IOV_RESOURCES
];
1011 if (!res
->flags
|| !res
->parent
)
1014 size
= pci_iov_resource_size(dev
, i
+ PCI_IOV_RESOURCES
);
1016 res
->start
+= size
* offset
;
1018 dev_info(&dev
->dev
, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n",
1019 i
, &res2
, res
, (offset
> 0) ? "En" : "Dis",
1021 pci_update_resource(dev
, i
+ PCI_IOV_RESOURCES
);
1025 #endif /* CONFIG_PCI_IOV */
1027 static struct pnv_ioda_pe
*pnv_ioda_setup_dev_PE(struct pci_dev
*dev
)
1029 struct pci_controller
*hose
= pci_bus_to_host(dev
->bus
);
1030 struct pnv_phb
*phb
= hose
->private_data
;
1031 struct pci_dn
*pdn
= pci_get_pdn(dev
);
1032 struct pnv_ioda_pe
*pe
;
1035 pr_err("%s: Device tree node not associated properly\n",
1039 if (pdn
->pe_number
!= IODA_INVALID_PE
)
1042 pe
= pnv_ioda_alloc_pe(phb
);
1044 pr_warning("%s: Not enough PE# available, disabling device\n",
1049 /* NOTE: We get only one ref to the pci_dev for the pdn, not for the
1050 * pointer in the PE data structure, both should be destroyed at the
1051 * same time. However, this needs to be looked at more closely again
1052 * once we actually start removing things (Hotplug, SR-IOV, ...)
1054 * At some point we want to remove the PDN completely anyways
1058 pdn
->pe_number
= pe
->pe_number
;
1059 pe
->flags
= PNV_IODA_PE_DEV
;
1062 pe
->mve_number
= -1;
1063 pe
->rid
= dev
->bus
->number
<< 8 | pdn
->devfn
;
1065 pe_info(pe
, "Associated device to PE\n");
1067 if (pnv_ioda_configure_pe(phb
, pe
)) {
1068 /* XXX What do we do here ? */
1069 pnv_ioda_free_pe(pe
);
1070 pdn
->pe_number
= IODA_INVALID_PE
;
1076 /* Put PE to the list */
1077 list_add_tail(&pe
->list
, &phb
->ioda
.pe_list
);
1082 static void pnv_ioda_setup_same_PE(struct pci_bus
*bus
, struct pnv_ioda_pe
*pe
)
1084 struct pci_dev
*dev
;
1086 list_for_each_entry(dev
, &bus
->devices
, bus_list
) {
1087 struct pci_dn
*pdn
= pci_get_pdn(dev
);
1090 pr_warn("%s: No device node associated with device !\n",
1096 * In partial hotplug case, the PCI device might be still
1097 * associated with the PE and needn't attach it to the PE
1100 if (pdn
->pe_number
!= IODA_INVALID_PE
)
1105 pdn
->pe_number
= pe
->pe_number
;
1106 if ((pe
->flags
& PNV_IODA_PE_BUS_ALL
) && dev
->subordinate
)
1107 pnv_ioda_setup_same_PE(dev
->subordinate
, pe
);
1112 * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1113 * single PCI bus. Another one that contains the primary PCI bus and its
1114 * subordinate PCI devices and buses. The second type of PE is normally
1115 * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1117 static struct pnv_ioda_pe
*pnv_ioda_setup_bus_PE(struct pci_bus
*bus
, bool all
)
1119 struct pci_controller
*hose
= pci_bus_to_host(bus
);
1120 struct pnv_phb
*phb
= hose
->private_data
;
1121 struct pnv_ioda_pe
*pe
= NULL
;
1122 unsigned int pe_num
;
1125 * In partial hotplug case, the PE instance might be still alive.
1126 * We should reuse it instead of allocating a new one.
1128 pe_num
= phb
->ioda
.pe_rmap
[bus
->number
<< 8];
1129 if (pe_num
!= IODA_INVALID_PE
) {
1130 pe
= &phb
->ioda
.pe_array
[pe_num
];
1131 pnv_ioda_setup_same_PE(bus
, pe
);
1135 /* PE number for root bus should have been reserved */
1136 if (pci_is_root_bus(bus
) &&
1137 phb
->ioda
.root_pe_idx
!= IODA_INVALID_PE
)
1138 pe
= &phb
->ioda
.pe_array
[phb
->ioda
.root_pe_idx
];
1140 /* Check if PE is determined by M64 */
1141 if (!pe
&& phb
->pick_m64_pe
)
1142 pe
= phb
->pick_m64_pe(bus
, all
);
1144 /* The PE number isn't pinned by M64 */
1146 pe
= pnv_ioda_alloc_pe(phb
);
1149 pr_warning("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1150 __func__
, pci_domain_nr(bus
), bus
->number
);
1154 pe
->flags
|= (all
? PNV_IODA_PE_BUS_ALL
: PNV_IODA_PE_BUS
);
1157 pe
->mve_number
= -1;
1158 pe
->rid
= bus
->busn_res
.start
<< 8;
1161 pe_info(pe
, "Secondary bus %d..%d associated with PE#%x\n",
1162 bus
->busn_res
.start
, bus
->busn_res
.end
, pe
->pe_number
);
1164 pe_info(pe
, "Secondary bus %d associated with PE#%x\n",
1165 bus
->busn_res
.start
, pe
->pe_number
);
1167 if (pnv_ioda_configure_pe(phb
, pe
)) {
1168 /* XXX What do we do here ? */
1169 pnv_ioda_free_pe(pe
);
1174 /* Associate it with all child devices */
1175 pnv_ioda_setup_same_PE(bus
, pe
);
1177 /* Put PE to the list */
1178 list_add_tail(&pe
->list
, &phb
->ioda
.pe_list
);
1183 static struct pnv_ioda_pe
*pnv_ioda_setup_npu_PE(struct pci_dev
*npu_pdev
)
1185 int pe_num
, found_pe
= false, rc
;
1187 struct pnv_ioda_pe
*pe
;
1188 struct pci_dev
*gpu_pdev
;
1189 struct pci_dn
*npu_pdn
;
1190 struct pci_controller
*hose
= pci_bus_to_host(npu_pdev
->bus
);
1191 struct pnv_phb
*phb
= hose
->private_data
;
1194 * Due to a hardware errata PE#0 on the NPU is reserved for
1195 * error handling. This means we only have three PEs remaining
1196 * which need to be assigned to four links, implying some
1197 * links must share PEs.
1199 * To achieve this we assign PEs such that NPUs linking the
1200 * same GPU get assigned the same PE.
1202 gpu_pdev
= pnv_pci_get_gpu_dev(npu_pdev
);
1203 for (pe_num
= 0; pe_num
< phb
->ioda
.total_pe_num
; pe_num
++) {
1204 pe
= &phb
->ioda
.pe_array
[pe_num
];
1208 if (pnv_pci_get_gpu_dev(pe
->pdev
) == gpu_pdev
) {
1210 * This device has the same peer GPU so should
1211 * be assigned the same PE as the existing
1214 dev_info(&npu_pdev
->dev
,
1215 "Associating to existing PE %x\n", pe_num
);
1216 pci_dev_get(npu_pdev
);
1217 npu_pdn
= pci_get_pdn(npu_pdev
);
1218 rid
= npu_pdev
->bus
->number
<< 8 | npu_pdn
->devfn
;
1219 npu_pdn
->pcidev
= npu_pdev
;
1220 npu_pdn
->pe_number
= pe_num
;
1221 phb
->ioda
.pe_rmap
[rid
] = pe
->pe_number
;
1223 /* Map the PE to this link */
1224 rc
= opal_pci_set_pe(phb
->opal_id
, pe_num
, rid
,
1226 OPAL_COMPARE_RID_DEVICE_NUMBER
,
1227 OPAL_COMPARE_RID_FUNCTION_NUMBER
,
1229 WARN_ON(rc
!= OPAL_SUCCESS
);
1237 * Could not find an existing PE so allocate a new
1240 return pnv_ioda_setup_dev_PE(npu_pdev
);
1245 static void pnv_ioda_setup_npu_PEs(struct pci_bus
*bus
)
1247 struct pci_dev
*pdev
;
1249 list_for_each_entry(pdev
, &bus
->devices
, bus_list
)
1250 pnv_ioda_setup_npu_PE(pdev
);
1253 static void pnv_pci_ioda_setup_PEs(void)
1255 struct pci_controller
*hose
, *tmp
;
1256 struct pnv_phb
*phb
;
1258 list_for_each_entry_safe(hose
, tmp
, &hose_list
, list_node
) {
1259 phb
= hose
->private_data
;
1260 if (phb
->type
== PNV_PHB_NPU
) {
1261 /* PE#0 is needed for error reporting */
1262 pnv_ioda_reserve_pe(phb
, 0);
1263 pnv_ioda_setup_npu_PEs(hose
->bus
);
1264 if (phb
->model
== PNV_PHB_MODEL_NPU2
)
1270 #ifdef CONFIG_PCI_IOV
1271 static int pnv_pci_vf_release_m64(struct pci_dev
*pdev
, u16 num_vfs
)
1273 struct pci_bus
*bus
;
1274 struct pci_controller
*hose
;
1275 struct pnv_phb
*phb
;
1281 hose
= pci_bus_to_host(bus
);
1282 phb
= hose
->private_data
;
1283 pdn
= pci_get_pdn(pdev
);
1285 if (pdn
->m64_single_mode
)
1290 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++)
1291 for (j
= 0; j
< m64_bars
; j
++) {
1292 if (pdn
->m64_map
[j
][i
] == IODA_INVALID_M64
)
1294 opal_pci_phb_mmio_enable(phb
->opal_id
,
1295 OPAL_M64_WINDOW_TYPE
, pdn
->m64_map
[j
][i
], 0);
1296 clear_bit(pdn
->m64_map
[j
][i
], &phb
->ioda
.m64_bar_alloc
);
1297 pdn
->m64_map
[j
][i
] = IODA_INVALID_M64
;
1300 kfree(pdn
->m64_map
);
1304 static int pnv_pci_vf_assign_m64(struct pci_dev
*pdev
, u16 num_vfs
)
1306 struct pci_bus
*bus
;
1307 struct pci_controller
*hose
;
1308 struct pnv_phb
*phb
;
1311 struct resource
*res
;
1315 resource_size_t size
, start
;
1320 hose
= pci_bus_to_host(bus
);
1321 phb
= hose
->private_data
;
1322 pdn
= pci_get_pdn(pdev
);
1323 total_vfs
= pci_sriov_get_totalvfs(pdev
);
1325 if (pdn
->m64_single_mode
)
1330 pdn
->m64_map
= kmalloc_array(m64_bars
,
1331 sizeof(*pdn
->m64_map
),
1335 /* Initialize the m64_map to IODA_INVALID_M64 */
1336 for (i
= 0; i
< m64_bars
; i
++)
1337 for (j
= 0; j
< PCI_SRIOV_NUM_BARS
; j
++)
1338 pdn
->m64_map
[i
][j
] = IODA_INVALID_M64
;
1341 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
1342 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
1343 if (!res
->flags
|| !res
->parent
)
1346 for (j
= 0; j
< m64_bars
; j
++) {
1348 win
= find_next_zero_bit(&phb
->ioda
.m64_bar_alloc
,
1349 phb
->ioda
.m64_bar_idx
+ 1, 0);
1351 if (win
>= phb
->ioda
.m64_bar_idx
+ 1)
1353 } while (test_and_set_bit(win
, &phb
->ioda
.m64_bar_alloc
));
1355 pdn
->m64_map
[j
][i
] = win
;
1357 if (pdn
->m64_single_mode
) {
1358 size
= pci_iov_resource_size(pdev
,
1359 PCI_IOV_RESOURCES
+ i
);
1360 start
= res
->start
+ size
* j
;
1362 size
= resource_size(res
);
1366 /* Map the M64 here */
1367 if (pdn
->m64_single_mode
) {
1368 pe_num
= pdn
->pe_num_map
[j
];
1369 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
1370 pe_num
, OPAL_M64_WINDOW_TYPE
,
1371 pdn
->m64_map
[j
][i
], 0);
1374 rc
= opal_pci_set_phb_mem_window(phb
->opal_id
,
1375 OPAL_M64_WINDOW_TYPE
,
1382 if (rc
!= OPAL_SUCCESS
) {
1383 dev_err(&pdev
->dev
, "Failed to map M64 window #%d: %lld\n",
1388 if (pdn
->m64_single_mode
)
1389 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
1390 OPAL_M64_WINDOW_TYPE
, pdn
->m64_map
[j
][i
], 2);
1392 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
1393 OPAL_M64_WINDOW_TYPE
, pdn
->m64_map
[j
][i
], 1);
1395 if (rc
!= OPAL_SUCCESS
) {
1396 dev_err(&pdev
->dev
, "Failed to enable M64 window #%d: %llx\n",
1405 pnv_pci_vf_release_m64(pdev
, num_vfs
);
1409 static long pnv_pci_ioda2_unset_window(struct iommu_table_group
*table_group
,
1411 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe
*pe
, bool enable
);
1413 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev
*dev
, struct pnv_ioda_pe
*pe
)
1415 struct iommu_table
*tbl
;
1418 tbl
= pe
->table_group
.tables
[0];
1419 rc
= pnv_pci_ioda2_unset_window(&pe
->table_group
, 0);
1421 pe_warn(pe
, "OPAL error %ld release DMA window\n", rc
);
1423 pnv_pci_ioda2_set_bypass(pe
, false);
1424 if (pe
->table_group
.group
) {
1425 iommu_group_put(pe
->table_group
.group
);
1426 BUG_ON(pe
->table_group
.group
);
1428 iommu_tce_table_put(tbl
);
1431 static void pnv_ioda_release_vf_PE(struct pci_dev
*pdev
)
1433 struct pci_bus
*bus
;
1434 struct pci_controller
*hose
;
1435 struct pnv_phb
*phb
;
1436 struct pnv_ioda_pe
*pe
, *pe_n
;
1440 hose
= pci_bus_to_host(bus
);
1441 phb
= hose
->private_data
;
1442 pdn
= pci_get_pdn(pdev
);
1444 if (!pdev
->is_physfn
)
1447 list_for_each_entry_safe(pe
, pe_n
, &phb
->ioda
.pe_list
, list
) {
1448 if (pe
->parent_dev
!= pdev
)
1451 pnv_pci_ioda2_release_dma_pe(pdev
, pe
);
1453 /* Remove from list */
1454 mutex_lock(&phb
->ioda
.pe_list_mutex
);
1455 list_del(&pe
->list
);
1456 mutex_unlock(&phb
->ioda
.pe_list_mutex
);
1458 pnv_ioda_deconfigure_pe(phb
, pe
);
1460 pnv_ioda_free_pe(pe
);
1464 void pnv_pci_sriov_disable(struct pci_dev
*pdev
)
1466 struct pci_bus
*bus
;
1467 struct pci_controller
*hose
;
1468 struct pnv_phb
*phb
;
1469 struct pnv_ioda_pe
*pe
;
1474 hose
= pci_bus_to_host(bus
);
1475 phb
= hose
->private_data
;
1476 pdn
= pci_get_pdn(pdev
);
1477 num_vfs
= pdn
->num_vfs
;
1479 /* Release VF PEs */
1480 pnv_ioda_release_vf_PE(pdev
);
1482 if (phb
->type
== PNV_PHB_IODA2
) {
1483 if (!pdn
->m64_single_mode
)
1484 pnv_pci_vf_resource_shift(pdev
, -*pdn
->pe_num_map
);
1486 /* Release M64 windows */
1487 pnv_pci_vf_release_m64(pdev
, num_vfs
);
1489 /* Release PE numbers */
1490 if (pdn
->m64_single_mode
) {
1491 for (i
= 0; i
< num_vfs
; i
++) {
1492 if (pdn
->pe_num_map
[i
] == IODA_INVALID_PE
)
1495 pe
= &phb
->ioda
.pe_array
[pdn
->pe_num_map
[i
]];
1496 pnv_ioda_free_pe(pe
);
1499 bitmap_clear(phb
->ioda
.pe_alloc
, *pdn
->pe_num_map
, num_vfs
);
1500 /* Releasing pe_num_map */
1501 kfree(pdn
->pe_num_map
);
1505 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb
*phb
,
1506 struct pnv_ioda_pe
*pe
);
1507 static void pnv_ioda_setup_vf_PE(struct pci_dev
*pdev
, u16 num_vfs
)
1509 struct pci_bus
*bus
;
1510 struct pci_controller
*hose
;
1511 struct pnv_phb
*phb
;
1512 struct pnv_ioda_pe
*pe
;
1518 hose
= pci_bus_to_host(bus
);
1519 phb
= hose
->private_data
;
1520 pdn
= pci_get_pdn(pdev
);
1522 if (!pdev
->is_physfn
)
1525 /* Reserve PE for each VF */
1526 for (vf_index
= 0; vf_index
< num_vfs
; vf_index
++) {
1527 if (pdn
->m64_single_mode
)
1528 pe_num
= pdn
->pe_num_map
[vf_index
];
1530 pe_num
= *pdn
->pe_num_map
+ vf_index
;
1532 pe
= &phb
->ioda
.pe_array
[pe_num
];
1533 pe
->pe_number
= pe_num
;
1535 pe
->flags
= PNV_IODA_PE_VF
;
1537 pe
->parent_dev
= pdev
;
1538 pe
->mve_number
= -1;
1539 pe
->rid
= (pci_iov_virtfn_bus(pdev
, vf_index
) << 8) |
1540 pci_iov_virtfn_devfn(pdev
, vf_index
);
1542 pe_info(pe
, "VF %04d:%02d:%02d.%d associated with PE#%x\n",
1543 hose
->global_number
, pdev
->bus
->number
,
1544 PCI_SLOT(pci_iov_virtfn_devfn(pdev
, vf_index
)),
1545 PCI_FUNC(pci_iov_virtfn_devfn(pdev
, vf_index
)), pe_num
);
1547 if (pnv_ioda_configure_pe(phb
, pe
)) {
1548 /* XXX What do we do here ? */
1549 pnv_ioda_free_pe(pe
);
1554 /* Put PE to the list */
1555 mutex_lock(&phb
->ioda
.pe_list_mutex
);
1556 list_add_tail(&pe
->list
, &phb
->ioda
.pe_list
);
1557 mutex_unlock(&phb
->ioda
.pe_list_mutex
);
1559 pnv_pci_ioda2_setup_dma_pe(phb
, pe
);
1563 int pnv_pci_sriov_enable(struct pci_dev
*pdev
, u16 num_vfs
)
1565 struct pci_bus
*bus
;
1566 struct pci_controller
*hose
;
1567 struct pnv_phb
*phb
;
1568 struct pnv_ioda_pe
*pe
;
1574 hose
= pci_bus_to_host(bus
);
1575 phb
= hose
->private_data
;
1576 pdn
= pci_get_pdn(pdev
);
1578 if (phb
->type
== PNV_PHB_IODA2
) {
1579 if (!pdn
->vfs_expanded
) {
1580 dev_info(&pdev
->dev
, "don't support this SRIOV device"
1581 " with non 64bit-prefetchable IOV BAR\n");
1586 * When M64 BARs functions in Single PE mode, the number of VFs
1587 * could be enabled must be less than the number of M64 BARs.
1589 if (pdn
->m64_single_mode
&& num_vfs
> phb
->ioda
.m64_bar_idx
) {
1590 dev_info(&pdev
->dev
, "Not enough M64 BAR for VFs\n");
1594 /* Allocating pe_num_map */
1595 if (pdn
->m64_single_mode
)
1596 pdn
->pe_num_map
= kmalloc_array(num_vfs
,
1597 sizeof(*pdn
->pe_num_map
),
1600 pdn
->pe_num_map
= kmalloc(sizeof(*pdn
->pe_num_map
), GFP_KERNEL
);
1602 if (!pdn
->pe_num_map
)
1605 if (pdn
->m64_single_mode
)
1606 for (i
= 0; i
< num_vfs
; i
++)
1607 pdn
->pe_num_map
[i
] = IODA_INVALID_PE
;
1609 /* Calculate available PE for required VFs */
1610 if (pdn
->m64_single_mode
) {
1611 for (i
= 0; i
< num_vfs
; i
++) {
1612 pe
= pnv_ioda_alloc_pe(phb
);
1618 pdn
->pe_num_map
[i
] = pe
->pe_number
;
1621 mutex_lock(&phb
->ioda
.pe_alloc_mutex
);
1622 *pdn
->pe_num_map
= bitmap_find_next_zero_area(
1623 phb
->ioda
.pe_alloc
, phb
->ioda
.total_pe_num
,
1625 if (*pdn
->pe_num_map
>= phb
->ioda
.total_pe_num
) {
1626 mutex_unlock(&phb
->ioda
.pe_alloc_mutex
);
1627 dev_info(&pdev
->dev
, "Failed to enable VF%d\n", num_vfs
);
1628 kfree(pdn
->pe_num_map
);
1631 bitmap_set(phb
->ioda
.pe_alloc
, *pdn
->pe_num_map
, num_vfs
);
1632 mutex_unlock(&phb
->ioda
.pe_alloc_mutex
);
1634 pdn
->num_vfs
= num_vfs
;
1636 /* Assign M64 window accordingly */
1637 ret
= pnv_pci_vf_assign_m64(pdev
, num_vfs
);
1639 dev_info(&pdev
->dev
, "Not enough M64 window resources\n");
1644 * When using one M64 BAR to map one IOV BAR, we need to shift
1645 * the IOV BAR according to the PE# allocated to the VFs.
1646 * Otherwise, the PE# for the VF will conflict with others.
1648 if (!pdn
->m64_single_mode
) {
1649 ret
= pnv_pci_vf_resource_shift(pdev
, *pdn
->pe_num_map
);
1656 pnv_ioda_setup_vf_PE(pdev
, num_vfs
);
1661 if (pdn
->m64_single_mode
) {
1662 for (i
= 0; i
< num_vfs
; i
++) {
1663 if (pdn
->pe_num_map
[i
] == IODA_INVALID_PE
)
1666 pe
= &phb
->ioda
.pe_array
[pdn
->pe_num_map
[i
]];
1667 pnv_ioda_free_pe(pe
);
1670 bitmap_clear(phb
->ioda
.pe_alloc
, *pdn
->pe_num_map
, num_vfs
);
1672 /* Releasing pe_num_map */
1673 kfree(pdn
->pe_num_map
);
1678 int pcibios_sriov_disable(struct pci_dev
*pdev
)
1680 pnv_pci_sriov_disable(pdev
);
1682 /* Release PCI data */
1683 remove_dev_pci_data(pdev
);
1687 int pcibios_sriov_enable(struct pci_dev
*pdev
, u16 num_vfs
)
1689 /* Allocate PCI data */
1690 add_dev_pci_data(pdev
);
1692 return pnv_pci_sriov_enable(pdev
, num_vfs
);
1694 #endif /* CONFIG_PCI_IOV */
1696 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb
*phb
, struct pci_dev
*pdev
)
1698 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
1699 struct pnv_ioda_pe
*pe
;
1702 * The function can be called while the PE#
1703 * hasn't been assigned. Do nothing for the
1706 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
1709 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
1710 WARN_ON(get_dma_ops(&pdev
->dev
) != &dma_iommu_ops
);
1711 set_dma_offset(&pdev
->dev
, pe
->tce_bypass_base
);
1712 set_iommu_table_base(&pdev
->dev
, pe
->table_group
.tables
[0]);
1714 * Note: iommu_add_device() will fail here as
1715 * for physical PE: the device is already added by now;
1716 * for virtual PE: sysfs entries are not ready yet and
1717 * tce_iommu_bus_notifier will add the device to a group later.
1721 static bool pnv_pci_ioda_pe_single_vendor(struct pnv_ioda_pe
*pe
)
1723 unsigned short vendor
= 0;
1724 struct pci_dev
*pdev
;
1726 if (pe
->device_count
== 1)
1729 /* pe->pdev should be set if it's a single device, pe->pbus if not */
1733 list_for_each_entry(pdev
, &pe
->pbus
->devices
, bus_list
) {
1735 vendor
= pdev
->vendor
;
1739 if (pdev
->vendor
!= vendor
)
1747 * Reconfigure TVE#0 to be usable as 64-bit DMA space.
1749 * The first 4GB of virtual memory for a PE is reserved for 32-bit accesses.
1750 * Devices can only access more than that if bit 59 of the PCI address is set
1751 * by hardware, which indicates TVE#1 should be used instead of TVE#0.
1752 * Many PCI devices are not capable of addressing that many bits, and as a
1753 * result are limited to the 4GB of virtual memory made available to 32-bit
1756 * In order to work around this, reconfigure TVE#0 to be suitable for 64-bit
1757 * devices by configuring the virtual memory past the first 4GB inaccessible
1758 * by 64-bit DMAs. This should only be used by devices that want more than
1759 * 4GB, and only on PEs that have no 32-bit devices.
1761 * Currently this will only work on PHB3 (POWER8).
1763 static int pnv_pci_ioda_dma_64bit_bypass(struct pnv_ioda_pe
*pe
)
1765 u64 window_size
, table_size
, tce_count
, addr
;
1766 struct page
*table_pages
;
1767 u64 tce_order
= 28; /* 256MB TCEs */
1772 * Window size needs to be a power of two, but needs to account for
1773 * shifting memory by the 4GB offset required to skip 32bit space.
1775 window_size
= roundup_pow_of_two(memory_hotplug_max() + (1ULL << 32));
1776 tce_count
= window_size
>> tce_order
;
1777 table_size
= tce_count
<< 3;
1779 if (table_size
< PAGE_SIZE
)
1780 table_size
= PAGE_SIZE
;
1782 table_pages
= alloc_pages_node(pe
->phb
->hose
->node
, GFP_KERNEL
,
1783 get_order(table_size
));
1787 tces
= page_address(table_pages
);
1791 memset(tces
, 0, table_size
);
1793 for (addr
= 0; addr
< memory_hotplug_max(); addr
+= (1 << tce_order
)) {
1794 tces
[(addr
+ (1ULL << 32)) >> tce_order
] =
1795 cpu_to_be64(addr
| TCE_PCI_READ
| TCE_PCI_WRITE
);
1798 rc
= opal_pci_map_pe_dma_window(pe
->phb
->opal_id
,
1800 /* reconfigure window 0 */
1801 (pe
->pe_number
<< 1) + 0,
1806 if (rc
== OPAL_SUCCESS
) {
1807 pe_info(pe
, "Using 64-bit DMA iommu bypass (through TVE#0)\n");
1811 pe_err(pe
, "Error configuring 64-bit DMA bypass\n");
1815 static int pnv_pci_ioda_dma_set_mask(struct pci_dev
*pdev
, u64 dma_mask
)
1817 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
1818 struct pnv_phb
*phb
= hose
->private_data
;
1819 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
1820 struct pnv_ioda_pe
*pe
;
1822 bool bypass
= false;
1825 if (WARN_ON(!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
))
1828 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
1829 if (pe
->tce_bypass_enabled
) {
1830 top
= pe
->tce_bypass_base
+ memblock_end_of_DRAM() - 1;
1831 bypass
= (dma_mask
>= top
);
1835 dev_info(&pdev
->dev
, "Using 64-bit DMA iommu bypass\n");
1836 set_dma_ops(&pdev
->dev
, &dma_direct_ops
);
1839 * If the device can't set the TCE bypass bit but still wants
1840 * to access 4GB or more, on PHB3 we can reconfigure TVE#0 to
1841 * bypass the 32-bit region and be usable for 64-bit DMAs.
1842 * The device needs to be able to address all of this space.
1844 if (dma_mask
>> 32 &&
1845 dma_mask
> (memory_hotplug_max() + (1ULL << 32)) &&
1846 pnv_pci_ioda_pe_single_vendor(pe
) &&
1847 phb
->model
== PNV_PHB_MODEL_PHB3
) {
1848 /* Configure the bypass mode */
1849 rc
= pnv_pci_ioda_dma_64bit_bypass(pe
);
1852 /* 4GB offset bypasses 32-bit space */
1853 set_dma_offset(&pdev
->dev
, (1ULL << 32));
1854 set_dma_ops(&pdev
->dev
, &dma_direct_ops
);
1855 } else if (dma_mask
>> 32 && dma_mask
!= DMA_BIT_MASK(64)) {
1857 * Fail the request if a DMA mask between 32 and 64 bits
1858 * was requested but couldn't be fulfilled. Ideally we
1859 * would do this for 64-bits but historically we have
1860 * always fallen back to 32-bits.
1864 dev_info(&pdev
->dev
, "Using 32-bit DMA via iommu\n");
1865 set_dma_ops(&pdev
->dev
, &dma_iommu_ops
);
1868 *pdev
->dev
.dma_mask
= dma_mask
;
1870 /* Update peer npu devices */
1871 pnv_npu_try_dma_set_bypass(pdev
, bypass
);
1876 static u64
pnv_pci_ioda_dma_get_required_mask(struct pci_dev
*pdev
)
1878 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
1879 struct pnv_phb
*phb
= hose
->private_data
;
1880 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
1881 struct pnv_ioda_pe
*pe
;
1884 if (WARN_ON(!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
))
1887 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
1888 if (!pe
->tce_bypass_enabled
)
1889 return __dma_get_required_mask(&pdev
->dev
);
1892 end
= pe
->tce_bypass_base
+ memblock_end_of_DRAM();
1893 mask
= 1ULL << (fls64(end
) - 1);
1899 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe
*pe
,
1900 struct pci_bus
*bus
,
1903 struct pci_dev
*dev
;
1905 list_for_each_entry(dev
, &bus
->devices
, bus_list
) {
1906 set_iommu_table_base(&dev
->dev
, pe
->table_group
.tables
[0]);
1907 set_dma_offset(&dev
->dev
, pe
->tce_bypass_base
);
1909 iommu_add_device(&dev
->dev
);
1911 if ((pe
->flags
& PNV_IODA_PE_BUS_ALL
) && dev
->subordinate
)
1912 pnv_ioda_setup_bus_dma(pe
, dev
->subordinate
,
1917 static inline __be64 __iomem
*pnv_ioda_get_inval_reg(struct pnv_phb
*phb
,
1920 return real_mode
? (__be64 __iomem
*)(phb
->regs_phys
+ 0x210) :
1921 (phb
->regs
+ 0x210);
1924 static void pnv_pci_p7ioc_tce_invalidate(struct iommu_table
*tbl
,
1925 unsigned long index
, unsigned long npages
, bool rm
)
1927 struct iommu_table_group_link
*tgl
= list_first_entry_or_null(
1928 &tbl
->it_group_list
, struct iommu_table_group_link
,
1930 struct pnv_ioda_pe
*pe
= container_of(tgl
->table_group
,
1931 struct pnv_ioda_pe
, table_group
);
1932 __be64 __iomem
*invalidate
= pnv_ioda_get_inval_reg(pe
->phb
, rm
);
1933 unsigned long start
, end
, inc
;
1935 start
= __pa(((__be64
*)tbl
->it_base
) + index
- tbl
->it_offset
);
1936 end
= __pa(((__be64
*)tbl
->it_base
) + index
- tbl
->it_offset
+
1939 /* p7ioc-style invalidation, 2 TCEs per write */
1940 start
|= (1ull << 63);
1941 end
|= (1ull << 63);
1943 end
|= inc
- 1; /* round up end to be different than start */
1945 mb(); /* Ensure above stores are visible */
1946 while (start
<= end
) {
1948 __raw_rm_writeq(cpu_to_be64(start
), invalidate
);
1950 __raw_writeq(cpu_to_be64(start
), invalidate
);
1955 * The iommu layer will do another mb() for us on build()
1956 * and we don't care on free()
1960 static int pnv_ioda1_tce_build(struct iommu_table
*tbl
, long index
,
1961 long npages
, unsigned long uaddr
,
1962 enum dma_data_direction direction
,
1963 unsigned long attrs
)
1965 int ret
= pnv_tce_build(tbl
, index
, npages
, uaddr
, direction
,
1969 pnv_pci_p7ioc_tce_invalidate(tbl
, index
, npages
, false);
1974 #ifdef CONFIG_IOMMU_API
1975 static int pnv_ioda1_tce_xchg(struct iommu_table
*tbl
, long index
,
1976 unsigned long *hpa
, enum dma_data_direction
*direction
)
1978 long ret
= pnv_tce_xchg(tbl
, index
, hpa
, direction
);
1981 pnv_pci_p7ioc_tce_invalidate(tbl
, index
, 1, false);
1986 static int pnv_ioda1_tce_xchg_rm(struct iommu_table
*tbl
, long index
,
1987 unsigned long *hpa
, enum dma_data_direction
*direction
)
1989 long ret
= pnv_tce_xchg(tbl
, index
, hpa
, direction
);
1992 pnv_pci_p7ioc_tce_invalidate(tbl
, index
, 1, true);
1998 static void pnv_ioda1_tce_free(struct iommu_table
*tbl
, long index
,
2001 pnv_tce_free(tbl
, index
, npages
);
2003 pnv_pci_p7ioc_tce_invalidate(tbl
, index
, npages
, false);
2006 static struct iommu_table_ops pnv_ioda1_iommu_ops
= {
2007 .set
= pnv_ioda1_tce_build
,
2008 #ifdef CONFIG_IOMMU_API
2009 .exchange
= pnv_ioda1_tce_xchg
,
2010 .exchange_rm
= pnv_ioda1_tce_xchg_rm
,
2012 .clear
= pnv_ioda1_tce_free
,
2016 #define PHB3_TCE_KILL_INVAL_ALL PPC_BIT(0)
2017 #define PHB3_TCE_KILL_INVAL_PE PPC_BIT(1)
2018 #define PHB3_TCE_KILL_INVAL_ONE PPC_BIT(2)
2020 static void pnv_pci_phb3_tce_invalidate_entire(struct pnv_phb
*phb
, bool rm
)
2022 __be64 __iomem
*invalidate
= pnv_ioda_get_inval_reg(phb
, rm
);
2023 const unsigned long val
= PHB3_TCE_KILL_INVAL_ALL
;
2025 mb(); /* Ensure previous TCE table stores are visible */
2027 __raw_rm_writeq(cpu_to_be64(val
), invalidate
);
2029 __raw_writeq(cpu_to_be64(val
), invalidate
);
2032 static inline void pnv_pci_phb3_tce_invalidate_pe(struct pnv_ioda_pe
*pe
)
2034 /* 01xb - invalidate TCEs that match the specified PE# */
2035 __be64 __iomem
*invalidate
= pnv_ioda_get_inval_reg(pe
->phb
, false);
2036 unsigned long val
= PHB3_TCE_KILL_INVAL_PE
| (pe
->pe_number
& 0xFF);
2038 mb(); /* Ensure above stores are visible */
2039 __raw_writeq(cpu_to_be64(val
), invalidate
);
2042 static void pnv_pci_phb3_tce_invalidate(struct pnv_ioda_pe
*pe
, bool rm
,
2043 unsigned shift
, unsigned long index
,
2044 unsigned long npages
)
2046 __be64 __iomem
*invalidate
= pnv_ioda_get_inval_reg(pe
->phb
, rm
);
2047 unsigned long start
, end
, inc
;
2049 /* We'll invalidate DMA address in PE scope */
2050 start
= PHB3_TCE_KILL_INVAL_ONE
;
2051 start
|= (pe
->pe_number
& 0xFF);
2054 /* Figure out the start, end and step */
2055 start
|= (index
<< shift
);
2056 end
|= ((index
+ npages
- 1) << shift
);
2057 inc
= (0x1ull
<< shift
);
2060 while (start
<= end
) {
2062 __raw_rm_writeq(cpu_to_be64(start
), invalidate
);
2064 __raw_writeq(cpu_to_be64(start
), invalidate
);
2069 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe
*pe
)
2071 struct pnv_phb
*phb
= pe
->phb
;
2073 if (phb
->model
== PNV_PHB_MODEL_PHB3
&& phb
->regs
)
2074 pnv_pci_phb3_tce_invalidate_pe(pe
);
2076 opal_pci_tce_kill(phb
->opal_id
, OPAL_PCI_TCE_KILL_PE
,
2077 pe
->pe_number
, 0, 0, 0);
2080 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table
*tbl
,
2081 unsigned long index
, unsigned long npages
, bool rm
)
2083 struct iommu_table_group_link
*tgl
;
2085 list_for_each_entry_lockless(tgl
, &tbl
->it_group_list
, next
) {
2086 struct pnv_ioda_pe
*pe
= container_of(tgl
->table_group
,
2087 struct pnv_ioda_pe
, table_group
);
2088 struct pnv_phb
*phb
= pe
->phb
;
2089 unsigned int shift
= tbl
->it_page_shift
;
2092 * NVLink1 can use the TCE kill register directly as
2093 * it's the same as PHB3. NVLink2 is different and
2094 * should go via the OPAL call.
2096 if (phb
->model
== PNV_PHB_MODEL_NPU
) {
2098 * The NVLink hardware does not support TCE kill
2099 * per TCE entry so we have to invalidate
2100 * the entire cache for it.
2102 pnv_pci_phb3_tce_invalidate_entire(phb
, rm
);
2105 if (phb
->model
== PNV_PHB_MODEL_PHB3
&& phb
->regs
)
2106 pnv_pci_phb3_tce_invalidate(pe
, rm
, shift
,
2109 opal_pci_tce_kill(phb
->opal_id
,
2110 OPAL_PCI_TCE_KILL_PAGES
,
2111 pe
->pe_number
, 1u << shift
,
2112 index
<< shift
, npages
);
2116 void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_phb
*phb
, bool rm
)
2118 if (phb
->model
== PNV_PHB_MODEL_NPU
|| phb
->model
== PNV_PHB_MODEL_PHB3
)
2119 pnv_pci_phb3_tce_invalidate_entire(phb
, rm
);
2121 opal_pci_tce_kill(phb
->opal_id
, OPAL_PCI_TCE_KILL
, 0, 0, 0, 0);
2124 static int pnv_ioda2_tce_build(struct iommu_table
*tbl
, long index
,
2125 long npages
, unsigned long uaddr
,
2126 enum dma_data_direction direction
,
2127 unsigned long attrs
)
2129 int ret
= pnv_tce_build(tbl
, index
, npages
, uaddr
, direction
,
2133 pnv_pci_ioda2_tce_invalidate(tbl
, index
, npages
, false);
2138 #ifdef CONFIG_IOMMU_API
2139 static int pnv_ioda2_tce_xchg(struct iommu_table
*tbl
, long index
,
2140 unsigned long *hpa
, enum dma_data_direction
*direction
)
2142 long ret
= pnv_tce_xchg(tbl
, index
, hpa
, direction
);
2145 pnv_pci_ioda2_tce_invalidate(tbl
, index
, 1, false);
2150 static int pnv_ioda2_tce_xchg_rm(struct iommu_table
*tbl
, long index
,
2151 unsigned long *hpa
, enum dma_data_direction
*direction
)
2153 long ret
= pnv_tce_xchg(tbl
, index
, hpa
, direction
);
2156 pnv_pci_ioda2_tce_invalidate(tbl
, index
, 1, true);
2162 static void pnv_ioda2_tce_free(struct iommu_table
*tbl
, long index
,
2165 pnv_tce_free(tbl
, index
, npages
);
2167 pnv_pci_ioda2_tce_invalidate(tbl
, index
, npages
, false);
2170 static void pnv_ioda2_table_free(struct iommu_table
*tbl
)
2172 pnv_pci_ioda2_table_free_pages(tbl
);
2175 static struct iommu_table_ops pnv_ioda2_iommu_ops
= {
2176 .set
= pnv_ioda2_tce_build
,
2177 #ifdef CONFIG_IOMMU_API
2178 .exchange
= pnv_ioda2_tce_xchg
,
2179 .exchange_rm
= pnv_ioda2_tce_xchg_rm
,
2181 .clear
= pnv_ioda2_tce_free
,
2183 .free
= pnv_ioda2_table_free
,
2186 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev
*dev
, void *data
)
2188 unsigned int *weight
= (unsigned int *)data
;
2190 /* This is quite simplistic. The "base" weight of a device
2191 * is 10. 0 means no DMA is to be accounted for it.
2193 if (dev
->hdr_type
!= PCI_HEADER_TYPE_NORMAL
)
2196 if (dev
->class == PCI_CLASS_SERIAL_USB_UHCI
||
2197 dev
->class == PCI_CLASS_SERIAL_USB_OHCI
||
2198 dev
->class == PCI_CLASS_SERIAL_USB_EHCI
)
2200 else if ((dev
->class >> 8) == PCI_CLASS_STORAGE_RAID
)
2208 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe
*pe
)
2210 unsigned int weight
= 0;
2212 /* SRIOV VF has same DMA32 weight as its PF */
2213 #ifdef CONFIG_PCI_IOV
2214 if ((pe
->flags
& PNV_IODA_PE_VF
) && pe
->parent_dev
) {
2215 pnv_pci_ioda_dev_dma_weight(pe
->parent_dev
, &weight
);
2220 if ((pe
->flags
& PNV_IODA_PE_DEV
) && pe
->pdev
) {
2221 pnv_pci_ioda_dev_dma_weight(pe
->pdev
, &weight
);
2222 } else if ((pe
->flags
& PNV_IODA_PE_BUS
) && pe
->pbus
) {
2223 struct pci_dev
*pdev
;
2225 list_for_each_entry(pdev
, &pe
->pbus
->devices
, bus_list
)
2226 pnv_pci_ioda_dev_dma_weight(pdev
, &weight
);
2227 } else if ((pe
->flags
& PNV_IODA_PE_BUS_ALL
) && pe
->pbus
) {
2228 pci_walk_bus(pe
->pbus
, pnv_pci_ioda_dev_dma_weight
, &weight
);
2234 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb
*phb
,
2235 struct pnv_ioda_pe
*pe
)
2238 struct page
*tce_mem
= NULL
;
2239 struct iommu_table
*tbl
;
2240 unsigned int weight
, total_weight
= 0;
2241 unsigned int tce32_segsz
, base
, segs
, avail
, i
;
2245 /* XXX FIXME: Handle 64-bit only DMA devices */
2246 /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
2247 /* XXX FIXME: Allocate multi-level tables on PHB3 */
2248 weight
= pnv_pci_ioda_pe_dma_weight(pe
);
2252 pci_walk_bus(phb
->hose
->bus
, pnv_pci_ioda_dev_dma_weight
,
2254 segs
= (weight
* phb
->ioda
.dma32_count
) / total_weight
;
2259 * Allocate contiguous DMA32 segments. We begin with the expected
2260 * number of segments. With one more attempt, the number of DMA32
2261 * segments to be allocated is decreased by one until one segment
2262 * is allocated successfully.
2265 for (base
= 0; base
<= phb
->ioda
.dma32_count
- segs
; base
++) {
2266 for (avail
= 0, i
= base
; i
< base
+ segs
; i
++) {
2267 if (phb
->ioda
.dma32_segmap
[i
] ==
2278 pe_warn(pe
, "No available DMA32 segments\n");
2283 tbl
= pnv_pci_table_alloc(phb
->hose
->node
);
2287 iommu_register_group(&pe
->table_group
, phb
->hose
->global_number
,
2289 pnv_pci_link_table_and_group(phb
->hose
->node
, 0, tbl
, &pe
->table_group
);
2291 /* Grab a 32-bit TCE table */
2292 pe_info(pe
, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
2293 weight
, total_weight
, base
, segs
);
2294 pe_info(pe
, " Setting up 32-bit TCE table at %08x..%08x\n",
2295 base
* PNV_IODA1_DMA32_SEGSIZE
,
2296 (base
+ segs
) * PNV_IODA1_DMA32_SEGSIZE
- 1);
2298 /* XXX Currently, we allocate one big contiguous table for the
2299 * TCEs. We only really need one chunk per 256M of TCE space
2300 * (ie per segment) but that's an optimization for later, it
2301 * requires some added smarts with our get/put_tce implementation
2303 * Each TCE page is 4KB in size and each TCE entry occupies 8
2306 tce32_segsz
= PNV_IODA1_DMA32_SEGSIZE
>> (IOMMU_PAGE_SHIFT_4K
- 3);
2307 tce_mem
= alloc_pages_node(phb
->hose
->node
, GFP_KERNEL
,
2308 get_order(tce32_segsz
* segs
));
2310 pe_err(pe
, " Failed to allocate a 32-bit TCE memory\n");
2313 addr
= page_address(tce_mem
);
2314 memset(addr
, 0, tce32_segsz
* segs
);
2317 for (i
= 0; i
< segs
; i
++) {
2318 rc
= opal_pci_map_pe_dma_window(phb
->opal_id
,
2321 __pa(addr
) + tce32_segsz
* i
,
2322 tce32_segsz
, IOMMU_PAGE_SIZE_4K
);
2324 pe_err(pe
, " Failed to configure 32-bit TCE table,"
2330 /* Setup DMA32 segment mapping */
2331 for (i
= base
; i
< base
+ segs
; i
++)
2332 phb
->ioda
.dma32_segmap
[i
] = pe
->pe_number
;
2334 /* Setup linux iommu table */
2335 pnv_pci_setup_iommu_table(tbl
, addr
, tce32_segsz
* segs
,
2336 base
* PNV_IODA1_DMA32_SEGSIZE
,
2337 IOMMU_PAGE_SHIFT_4K
);
2339 tbl
->it_ops
= &pnv_ioda1_iommu_ops
;
2340 pe
->table_group
.tce32_start
= tbl
->it_offset
<< tbl
->it_page_shift
;
2341 pe
->table_group
.tce32_size
= tbl
->it_size
<< tbl
->it_page_shift
;
2342 iommu_init_table(tbl
, phb
->hose
->node
);
2344 if (pe
->flags
& PNV_IODA_PE_DEV
) {
2346 * Setting table base here only for carrying iommu_group
2347 * further down to let iommu_add_device() do the job.
2348 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2350 set_iommu_table_base(&pe
->pdev
->dev
, tbl
);
2351 iommu_add_device(&pe
->pdev
->dev
);
2352 } else if (pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
))
2353 pnv_ioda_setup_bus_dma(pe
, pe
->pbus
, true);
2357 /* XXX Failure: Try to fallback to 64-bit only ? */
2359 __free_pages(tce_mem
, get_order(tce32_segsz
* segs
));
2361 pnv_pci_unlink_table_and_group(tbl
, &pe
->table_group
);
2362 iommu_tce_table_put(tbl
);
2366 static long pnv_pci_ioda2_set_window(struct iommu_table_group
*table_group
,
2367 int num
, struct iommu_table
*tbl
)
2369 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2371 struct pnv_phb
*phb
= pe
->phb
;
2373 const unsigned long size
= tbl
->it_indirect_levels
?
2374 tbl
->it_level_size
: tbl
->it_size
;
2375 const __u64 start_addr
= tbl
->it_offset
<< tbl
->it_page_shift
;
2376 const __u64 win_size
= tbl
->it_size
<< tbl
->it_page_shift
;
2378 pe_info(pe
, "Setting up window#%d %llx..%llx pg=%x\n", num
,
2379 start_addr
, start_addr
+ win_size
- 1,
2380 IOMMU_PAGE_SIZE(tbl
));
2383 * Map TCE table through TVT. The TVE index is the PE number
2384 * shifted by 1 bit for 32-bits DMA space.
2386 rc
= opal_pci_map_pe_dma_window(phb
->opal_id
,
2388 (pe
->pe_number
<< 1) + num
,
2389 tbl
->it_indirect_levels
+ 1,
2392 IOMMU_PAGE_SIZE(tbl
));
2394 pe_err(pe
, "Failed to configure TCE table, err %ld\n", rc
);
2398 pnv_pci_link_table_and_group(phb
->hose
->node
, num
,
2399 tbl
, &pe
->table_group
);
2400 pnv_pci_ioda2_tce_invalidate_pe(pe
);
2405 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe
*pe
, bool enable
)
2407 uint16_t window_id
= (pe
->pe_number
<< 1 ) + 1;
2410 pe_info(pe
, "%sabling 64-bit DMA bypass\n", enable
? "En" : "Dis");
2412 phys_addr_t top
= memblock_end_of_DRAM();
2414 top
= roundup_pow_of_two(top
);
2415 rc
= opal_pci_map_pe_dma_window_real(pe
->phb
->opal_id
,
2418 pe
->tce_bypass_base
,
2421 rc
= opal_pci_map_pe_dma_window_real(pe
->phb
->opal_id
,
2424 pe
->tce_bypass_base
,
2428 pe_err(pe
, "OPAL error %lld configuring bypass window\n", rc
);
2430 pe
->tce_bypass_enabled
= enable
;
2433 static long pnv_pci_ioda2_table_alloc_pages(int nid
, __u64 bus_offset
,
2434 __u32 page_shift
, __u64 window_size
, __u32 levels
,
2435 struct iommu_table
*tbl
);
2437 static long pnv_pci_ioda2_create_table(struct iommu_table_group
*table_group
,
2438 int num
, __u32 page_shift
, __u64 window_size
, __u32 levels
,
2439 struct iommu_table
**ptbl
)
2441 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2443 int nid
= pe
->phb
->hose
->node
;
2444 __u64 bus_offset
= num
? pe
->tce_bypass_base
: table_group
->tce32_start
;
2446 struct iommu_table
*tbl
;
2448 tbl
= pnv_pci_table_alloc(nid
);
2452 tbl
->it_ops
= &pnv_ioda2_iommu_ops
;
2454 ret
= pnv_pci_ioda2_table_alloc_pages(nid
,
2455 bus_offset
, page_shift
, window_size
,
2458 iommu_tce_table_put(tbl
);
2467 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe
*pe
)
2469 struct iommu_table
*tbl
= NULL
;
2473 * crashkernel= specifies the kdump kernel's maximum memory at
2474 * some offset and there is no guaranteed the result is a power
2475 * of 2, which will cause errors later.
2477 const u64 max_memory
= __rounddown_pow_of_two(memory_hotplug_max());
2480 * In memory constrained environments, e.g. kdump kernel, the
2481 * DMA window can be larger than available memory, which will
2482 * cause errors later.
2484 const u64 window_size
= min((u64
)pe
->table_group
.tce32_size
, max_memory
);
2486 rc
= pnv_pci_ioda2_create_table(&pe
->table_group
, 0,
2487 IOMMU_PAGE_SHIFT_4K
,
2489 POWERNV_IOMMU_DEFAULT_LEVELS
, &tbl
);
2491 pe_err(pe
, "Failed to create 32-bit TCE table, err %ld",
2496 iommu_init_table(tbl
, pe
->phb
->hose
->node
);
2498 rc
= pnv_pci_ioda2_set_window(&pe
->table_group
, 0, tbl
);
2500 pe_err(pe
, "Failed to configure 32-bit TCE table, err %ld\n",
2502 iommu_tce_table_put(tbl
);
2506 if (!pnv_iommu_bypass_disabled
)
2507 pnv_pci_ioda2_set_bypass(pe
, true);
2510 * Setting table base here only for carrying iommu_group
2511 * further down to let iommu_add_device() do the job.
2512 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2514 if (pe
->flags
& PNV_IODA_PE_DEV
)
2515 set_iommu_table_base(&pe
->pdev
->dev
, tbl
);
2520 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV)
2521 static long pnv_pci_ioda2_unset_window(struct iommu_table_group
*table_group
,
2524 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2526 struct pnv_phb
*phb
= pe
->phb
;
2529 pe_info(pe
, "Removing DMA window #%d\n", num
);
2531 ret
= opal_pci_map_pe_dma_window(phb
->opal_id
, pe
->pe_number
,
2532 (pe
->pe_number
<< 1) + num
,
2533 0/* levels */, 0/* table address */,
2534 0/* table size */, 0/* page size */);
2536 pe_warn(pe
, "Unmapping failed, ret = %ld\n", ret
);
2538 pnv_pci_ioda2_tce_invalidate_pe(pe
);
2540 pnv_pci_unlink_table_and_group(table_group
->tables
[num
], table_group
);
2546 #ifdef CONFIG_IOMMU_API
2547 static unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift
,
2548 __u64 window_size
, __u32 levels
)
2550 unsigned long bytes
= 0;
2551 const unsigned window_shift
= ilog2(window_size
);
2552 unsigned entries_shift
= window_shift
- page_shift
;
2553 unsigned table_shift
= entries_shift
+ 3;
2554 unsigned long tce_table_size
= max(0x1000UL
, 1UL << table_shift
);
2555 unsigned long direct_table_size
;
2557 if (!levels
|| (levels
> POWERNV_IOMMU_MAX_LEVELS
) ||
2558 (window_size
> memory_hotplug_max()) ||
2559 !is_power_of_2(window_size
))
2562 /* Calculate a direct table size from window_size and levels */
2563 entries_shift
= (entries_shift
+ levels
- 1) / levels
;
2564 table_shift
= entries_shift
+ 3;
2565 table_shift
= max_t(unsigned, table_shift
, PAGE_SHIFT
);
2566 direct_table_size
= 1UL << table_shift
;
2568 for ( ; levels
; --levels
) {
2569 bytes
+= _ALIGN_UP(tce_table_size
, direct_table_size
);
2571 tce_table_size
/= direct_table_size
;
2572 tce_table_size
<<= 3;
2573 tce_table_size
= max_t(unsigned long,
2574 tce_table_size
, direct_table_size
);
2580 static void pnv_ioda2_take_ownership(struct iommu_table_group
*table_group
)
2582 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2584 /* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
2585 struct iommu_table
*tbl
= pe
->table_group
.tables
[0];
2587 pnv_pci_ioda2_set_bypass(pe
, false);
2588 pnv_pci_ioda2_unset_window(&pe
->table_group
, 0);
2590 pnv_ioda_setup_bus_dma(pe
, pe
->pbus
, false);
2591 iommu_tce_table_put(tbl
);
2594 static void pnv_ioda2_release_ownership(struct iommu_table_group
*table_group
)
2596 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2599 pnv_pci_ioda2_setup_default_config(pe
);
2601 pnv_ioda_setup_bus_dma(pe
, pe
->pbus
, false);
2604 static struct iommu_table_group_ops pnv_pci_ioda2_ops
= {
2605 .get_table_size
= pnv_pci_ioda2_get_table_size
,
2606 .create_table
= pnv_pci_ioda2_create_table
,
2607 .set_window
= pnv_pci_ioda2_set_window
,
2608 .unset_window
= pnv_pci_ioda2_unset_window
,
2609 .take_ownership
= pnv_ioda2_take_ownership
,
2610 .release_ownership
= pnv_ioda2_release_ownership
,
2613 static int gpe_table_group_to_npe_cb(struct device
*dev
, void *opaque
)
2615 struct pci_controller
*hose
;
2616 struct pnv_phb
*phb
;
2617 struct pnv_ioda_pe
**ptmppe
= opaque
;
2618 struct pci_dev
*pdev
= container_of(dev
, struct pci_dev
, dev
);
2619 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
2621 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
2624 hose
= pci_bus_to_host(pdev
->bus
);
2625 phb
= hose
->private_data
;
2626 if (phb
->type
!= PNV_PHB_NPU
)
2629 *ptmppe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
2635 * This returns PE of associated NPU.
2636 * This assumes that NPU is in the same IOMMU group with GPU and there is
2639 static struct pnv_ioda_pe
*gpe_table_group_to_npe(
2640 struct iommu_table_group
*table_group
)
2642 struct pnv_ioda_pe
*npe
= NULL
;
2643 int ret
= iommu_group_for_each_dev(table_group
->group
, &npe
,
2644 gpe_table_group_to_npe_cb
);
2646 BUG_ON(!ret
|| !npe
);
2651 static long pnv_pci_ioda2_npu_set_window(struct iommu_table_group
*table_group
,
2652 int num
, struct iommu_table
*tbl
)
2654 long ret
= pnv_pci_ioda2_set_window(table_group
, num
, tbl
);
2659 ret
= pnv_npu_set_window(gpe_table_group_to_npe(table_group
), num
, tbl
);
2661 pnv_pci_ioda2_unset_window(table_group
, num
);
2666 static long pnv_pci_ioda2_npu_unset_window(
2667 struct iommu_table_group
*table_group
,
2670 long ret
= pnv_pci_ioda2_unset_window(table_group
, num
);
2675 return pnv_npu_unset_window(gpe_table_group_to_npe(table_group
), num
);
2678 static void pnv_ioda2_npu_take_ownership(struct iommu_table_group
*table_group
)
2681 * Detach NPU first as pnv_ioda2_take_ownership() will destroy
2682 * the iommu_table if 32bit DMA is enabled.
2684 pnv_npu_take_ownership(gpe_table_group_to_npe(table_group
));
2685 pnv_ioda2_take_ownership(table_group
);
2688 static struct iommu_table_group_ops pnv_pci_ioda2_npu_ops
= {
2689 .get_table_size
= pnv_pci_ioda2_get_table_size
,
2690 .create_table
= pnv_pci_ioda2_create_table
,
2691 .set_window
= pnv_pci_ioda2_npu_set_window
,
2692 .unset_window
= pnv_pci_ioda2_npu_unset_window
,
2693 .take_ownership
= pnv_ioda2_npu_take_ownership
,
2694 .release_ownership
= pnv_ioda2_release_ownership
,
2697 static void pnv_pci_ioda_setup_iommu_api(void)
2699 struct pci_controller
*hose
, *tmp
;
2700 struct pnv_phb
*phb
;
2701 struct pnv_ioda_pe
*pe
, *gpe
;
2704 * Now we have all PHBs discovered, time to add NPU devices to
2705 * the corresponding IOMMU groups.
2707 list_for_each_entry_safe(hose
, tmp
, &hose_list
, list_node
) {
2708 phb
= hose
->private_data
;
2710 if (phb
->type
!= PNV_PHB_NPU
)
2713 list_for_each_entry(pe
, &phb
->ioda
.pe_list
, list
) {
2714 gpe
= pnv_pci_npu_setup_iommu(pe
);
2716 gpe
->table_group
.ops
= &pnv_pci_ioda2_npu_ops
;
2720 #else /* !CONFIG_IOMMU_API */
2721 static void pnv_pci_ioda_setup_iommu_api(void) { };
2724 static __be64
*pnv_pci_ioda2_table_do_alloc_pages(int nid
, unsigned shift
,
2725 unsigned levels
, unsigned long limit
,
2726 unsigned long *current_offset
, unsigned long *total_allocated
)
2728 struct page
*tce_mem
= NULL
;
2730 unsigned order
= max_t(unsigned, shift
, PAGE_SHIFT
) - PAGE_SHIFT
;
2731 unsigned long allocated
= 1UL << (order
+ PAGE_SHIFT
);
2732 unsigned entries
= 1UL << (shift
- 3);
2735 tce_mem
= alloc_pages_node(nid
, GFP_KERNEL
, order
);
2737 pr_err("Failed to allocate a TCE memory, order=%d\n", order
);
2740 addr
= page_address(tce_mem
);
2741 memset(addr
, 0, allocated
);
2742 *total_allocated
+= allocated
;
2746 *current_offset
+= allocated
;
2750 for (i
= 0; i
< entries
; ++i
) {
2751 tmp
= pnv_pci_ioda2_table_do_alloc_pages(nid
, shift
,
2752 levels
, limit
, current_offset
, total_allocated
);
2756 addr
[i
] = cpu_to_be64(__pa(tmp
) |
2757 TCE_PCI_READ
| TCE_PCI_WRITE
);
2759 if (*current_offset
>= limit
)
2766 static void pnv_pci_ioda2_table_do_free_pages(__be64
*addr
,
2767 unsigned long size
, unsigned level
);
2769 static long pnv_pci_ioda2_table_alloc_pages(int nid
, __u64 bus_offset
,
2770 __u32 page_shift
, __u64 window_size
, __u32 levels
,
2771 struct iommu_table
*tbl
)
2774 unsigned long offset
= 0, level_shift
, total_allocated
= 0;
2775 const unsigned window_shift
= ilog2(window_size
);
2776 unsigned entries_shift
= window_shift
- page_shift
;
2777 unsigned table_shift
= max_t(unsigned, entries_shift
+ 3, PAGE_SHIFT
);
2778 const unsigned long tce_table_size
= 1UL << table_shift
;
2780 if (!levels
|| (levels
> POWERNV_IOMMU_MAX_LEVELS
))
2783 if ((window_size
> memory_hotplug_max()) || !is_power_of_2(window_size
))
2786 /* Adjust direct table size from window_size and levels */
2787 entries_shift
= (entries_shift
+ levels
- 1) / levels
;
2788 level_shift
= entries_shift
+ 3;
2789 level_shift
= max_t(unsigned, level_shift
, PAGE_SHIFT
);
2791 if ((level_shift
- 3) * levels
+ page_shift
>= 60)
2794 /* Allocate TCE table */
2795 addr
= pnv_pci_ioda2_table_do_alloc_pages(nid
, level_shift
,
2796 levels
, tce_table_size
, &offset
, &total_allocated
);
2798 /* addr==NULL means that the first level allocation failed */
2803 * First level was allocated but some lower level failed as
2804 * we did not allocate as much as we wanted,
2805 * release partially allocated table.
2807 if (offset
< tce_table_size
) {
2808 pnv_pci_ioda2_table_do_free_pages(addr
,
2809 1ULL << (level_shift
- 3), levels
- 1);
2813 /* Setup linux iommu table */
2814 pnv_pci_setup_iommu_table(tbl
, addr
, tce_table_size
, bus_offset
,
2816 tbl
->it_level_size
= 1ULL << (level_shift
- 3);
2817 tbl
->it_indirect_levels
= levels
- 1;
2818 tbl
->it_allocated_size
= total_allocated
;
2820 pr_devel("Created TCE table: ws=%08llx ts=%lx @%08llx\n",
2821 window_size
, tce_table_size
, bus_offset
);
2826 static void pnv_pci_ioda2_table_do_free_pages(__be64
*addr
,
2827 unsigned long size
, unsigned level
)
2829 const unsigned long addr_ul
= (unsigned long) addr
&
2830 ~(TCE_PCI_READ
| TCE_PCI_WRITE
);
2834 u64
*tmp
= (u64
*) addr_ul
;
2836 for (i
= 0; i
< size
; ++i
) {
2837 unsigned long hpa
= be64_to_cpu(tmp
[i
]);
2839 if (!(hpa
& (TCE_PCI_READ
| TCE_PCI_WRITE
)))
2842 pnv_pci_ioda2_table_do_free_pages(__va(hpa
), size
,
2847 free_pages(addr_ul
, get_order(size
<< 3));
2850 static void pnv_pci_ioda2_table_free_pages(struct iommu_table
*tbl
)
2852 const unsigned long size
= tbl
->it_indirect_levels
?
2853 tbl
->it_level_size
: tbl
->it_size
;
2858 pnv_pci_ioda2_table_do_free_pages((__be64
*)tbl
->it_base
, size
,
2859 tbl
->it_indirect_levels
);
2862 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb
*phb
,
2863 struct pnv_ioda_pe
*pe
)
2867 if (!pnv_pci_ioda_pe_dma_weight(pe
))
2870 /* TVE #1 is selected by PCI address bit 59 */
2871 pe
->tce_bypass_base
= 1ull << 59;
2873 iommu_register_group(&pe
->table_group
, phb
->hose
->global_number
,
2876 /* The PE will reserve all possible 32-bits space */
2877 pe_info(pe
, "Setting up 32-bit TCE table at 0..%08x\n",
2878 phb
->ioda
.m32_pci_base
);
2880 /* Setup linux iommu table */
2881 pe
->table_group
.tce32_start
= 0;
2882 pe
->table_group
.tce32_size
= phb
->ioda
.m32_pci_base
;
2883 pe
->table_group
.max_dynamic_windows_supported
=
2884 IOMMU_TABLE_GROUP_MAX_TABLES
;
2885 pe
->table_group
.max_levels
= POWERNV_IOMMU_MAX_LEVELS
;
2886 pe
->table_group
.pgsizes
= SZ_4K
| SZ_64K
| SZ_16M
;
2887 #ifdef CONFIG_IOMMU_API
2888 pe
->table_group
.ops
= &pnv_pci_ioda2_ops
;
2891 rc
= pnv_pci_ioda2_setup_default_config(pe
);
2895 if (pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
))
2896 pnv_ioda_setup_bus_dma(pe
, pe
->pbus
, true);
2899 #ifdef CONFIG_PCI_MSI
2900 int64_t pnv_opal_pci_msi_eoi(struct irq_chip
*chip
, unsigned int hw_irq
)
2902 struct pnv_phb
*phb
= container_of(chip
, struct pnv_phb
,
2905 return opal_pci_msi_eoi(phb
->opal_id
, hw_irq
);
2908 static void pnv_ioda2_msi_eoi(struct irq_data
*d
)
2911 unsigned int hw_irq
= (unsigned int)irqd_to_hwirq(d
);
2912 struct irq_chip
*chip
= irq_data_get_irq_chip(d
);
2914 rc
= pnv_opal_pci_msi_eoi(chip
, hw_irq
);
2921 void pnv_set_msi_irq_chip(struct pnv_phb
*phb
, unsigned int virq
)
2923 struct irq_data
*idata
;
2924 struct irq_chip
*ichip
;
2926 /* The MSI EOI OPAL call is only needed on PHB3 */
2927 if (phb
->model
!= PNV_PHB_MODEL_PHB3
)
2930 if (!phb
->ioda
.irq_chip_init
) {
2932 * First time we setup an MSI IRQ, we need to setup the
2933 * corresponding IRQ chip to route correctly.
2935 idata
= irq_get_irq_data(virq
);
2936 ichip
= irq_data_get_irq_chip(idata
);
2937 phb
->ioda
.irq_chip_init
= 1;
2938 phb
->ioda
.irq_chip
= *ichip
;
2939 phb
->ioda
.irq_chip
.irq_eoi
= pnv_ioda2_msi_eoi
;
2941 irq_set_chip(virq
, &phb
->ioda
.irq_chip
);
2945 * Returns true iff chip is something that we could call
2946 * pnv_opal_pci_msi_eoi for.
2948 bool is_pnv_opal_msi(struct irq_chip
*chip
)
2950 return chip
->irq_eoi
== pnv_ioda2_msi_eoi
;
2952 EXPORT_SYMBOL_GPL(is_pnv_opal_msi
);
2954 static int pnv_pci_ioda_msi_setup(struct pnv_phb
*phb
, struct pci_dev
*dev
,
2955 unsigned int hwirq
, unsigned int virq
,
2956 unsigned int is_64
, struct msi_msg
*msg
)
2958 struct pnv_ioda_pe
*pe
= pnv_ioda_get_pe(dev
);
2959 unsigned int xive_num
= hwirq
- phb
->msi_base
;
2963 /* No PE assigned ? bail out ... no MSI for you ! */
2967 /* Check if we have an MVE */
2968 if (pe
->mve_number
< 0)
2971 /* Force 32-bit MSI on some broken devices */
2972 if (dev
->no_64bit_msi
)
2975 /* Assign XIVE to PE */
2976 rc
= opal_pci_set_xive_pe(phb
->opal_id
, pe
->pe_number
, xive_num
);
2978 pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2979 pci_name(dev
), rc
, xive_num
);
2986 rc
= opal_get_msi_64(phb
->opal_id
, pe
->mve_number
, xive_num
, 1,
2989 pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2993 msg
->address_hi
= be64_to_cpu(addr64
) >> 32;
2994 msg
->address_lo
= be64_to_cpu(addr64
) & 0xfffffffful
;
2998 rc
= opal_get_msi_32(phb
->opal_id
, pe
->mve_number
, xive_num
, 1,
3001 pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
3005 msg
->address_hi
= 0;
3006 msg
->address_lo
= be32_to_cpu(addr32
);
3008 msg
->data
= be32_to_cpu(data
);
3010 pnv_set_msi_irq_chip(phb
, virq
);
3012 pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
3013 " address=%x_%08x data=%x PE# %x\n",
3014 pci_name(dev
), is_64
? "64" : "32", hwirq
, xive_num
,
3015 msg
->address_hi
, msg
->address_lo
, data
, pe
->pe_number
);
3020 static void pnv_pci_init_ioda_msis(struct pnv_phb
*phb
)
3023 const __be32
*prop
= of_get_property(phb
->hose
->dn
,
3024 "ibm,opal-msi-ranges", NULL
);
3027 prop
= of_get_property(phb
->hose
->dn
, "msi-ranges", NULL
);
3032 phb
->msi_base
= be32_to_cpup(prop
);
3033 count
= be32_to_cpup(prop
+ 1);
3034 if (msi_bitmap_alloc(&phb
->msi_bmp
, count
, phb
->hose
->dn
)) {
3035 pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
3036 phb
->hose
->global_number
);
3040 phb
->msi_setup
= pnv_pci_ioda_msi_setup
;
3041 phb
->msi32_support
= 1;
3042 pr_info(" Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
3043 count
, phb
->msi_base
);
3046 static void pnv_pci_init_ioda_msis(struct pnv_phb
*phb
) { }
3047 #endif /* CONFIG_PCI_MSI */
3049 #ifdef CONFIG_PCI_IOV
3050 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev
*pdev
)
3052 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
3053 struct pnv_phb
*phb
= hose
->private_data
;
3054 const resource_size_t gate
= phb
->ioda
.m64_segsize
>> 2;
3055 struct resource
*res
;
3057 resource_size_t size
, total_vf_bar_sz
;
3061 if (!pdev
->is_physfn
|| pdev
->is_added
)
3064 pdn
= pci_get_pdn(pdev
);
3065 pdn
->vfs_expanded
= 0;
3066 pdn
->m64_single_mode
= false;
3068 total_vfs
= pci_sriov_get_totalvfs(pdev
);
3069 mul
= phb
->ioda
.total_pe_num
;
3070 total_vf_bar_sz
= 0;
3072 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
3073 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
3074 if (!res
->flags
|| res
->parent
)
3076 if (!pnv_pci_is_m64_flags(res
->flags
)) {
3077 dev_warn(&pdev
->dev
, "Don't support SR-IOV with"
3078 " non M64 VF BAR%d: %pR. \n",
3083 total_vf_bar_sz
+= pci_iov_resource_size(pdev
,
3084 i
+ PCI_IOV_RESOURCES
);
3087 * If bigger than quarter of M64 segment size, just round up
3090 * Generally, one M64 BAR maps one IOV BAR. To avoid conflict
3091 * with other devices, IOV BAR size is expanded to be
3092 * (total_pe * VF_BAR_size). When VF_BAR_size is half of M64
3093 * segment size , the expanded size would equal to half of the
3094 * whole M64 space size, which will exhaust the M64 Space and
3095 * limit the system flexibility. This is a design decision to
3096 * set the boundary to quarter of the M64 segment size.
3098 if (total_vf_bar_sz
> gate
) {
3099 mul
= roundup_pow_of_two(total_vfs
);
3100 dev_info(&pdev
->dev
,
3101 "VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n",
3102 total_vf_bar_sz
, gate
, mul
);
3103 pdn
->m64_single_mode
= true;
3108 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
3109 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
3110 if (!res
->flags
|| res
->parent
)
3113 size
= pci_iov_resource_size(pdev
, i
+ PCI_IOV_RESOURCES
);
3115 * On PHB3, the minimum size alignment of M64 BAR in single
3118 if (pdn
->m64_single_mode
&& (size
< SZ_32M
))
3120 dev_dbg(&pdev
->dev
, " Fixing VF BAR%d: %pR to\n", i
, res
);
3121 res
->end
= res
->start
+ size
* mul
- 1;
3122 dev_dbg(&pdev
->dev
, " %pR\n", res
);
3123 dev_info(&pdev
->dev
, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
3126 pdn
->vfs_expanded
= mul
;
3131 /* To save MMIO space, IOV BAR is truncated. */
3132 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
3133 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
3135 res
->end
= res
->start
- 1;
3138 #endif /* CONFIG_PCI_IOV */
3140 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe
*pe
,
3141 struct resource
*res
)
3143 struct pnv_phb
*phb
= pe
->phb
;
3144 struct pci_bus_region region
;
3148 if (!res
|| !res
->flags
|| res
->start
> res
->end
)
3151 if (res
->flags
& IORESOURCE_IO
) {
3152 region
.start
= res
->start
- phb
->ioda
.io_pci_base
;
3153 region
.end
= res
->end
- phb
->ioda
.io_pci_base
;
3154 index
= region
.start
/ phb
->ioda
.io_segsize
;
3156 while (index
< phb
->ioda
.total_pe_num
&&
3157 region
.start
<= region
.end
) {
3158 phb
->ioda
.io_segmap
[index
] = pe
->pe_number
;
3159 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
3160 pe
->pe_number
, OPAL_IO_WINDOW_TYPE
, 0, index
);
3161 if (rc
!= OPAL_SUCCESS
) {
3162 pr_err("%s: Error %lld mapping IO segment#%d to PE#%x\n",
3163 __func__
, rc
, index
, pe
->pe_number
);
3167 region
.start
+= phb
->ioda
.io_segsize
;
3170 } else if ((res
->flags
& IORESOURCE_MEM
) &&
3171 !pnv_pci_is_m64(phb
, res
)) {
3172 region
.start
= res
->start
-
3173 phb
->hose
->mem_offset
[0] -
3174 phb
->ioda
.m32_pci_base
;
3175 region
.end
= res
->end
-
3176 phb
->hose
->mem_offset
[0] -
3177 phb
->ioda
.m32_pci_base
;
3178 index
= region
.start
/ phb
->ioda
.m32_segsize
;
3180 while (index
< phb
->ioda
.total_pe_num
&&
3181 region
.start
<= region
.end
) {
3182 phb
->ioda
.m32_segmap
[index
] = pe
->pe_number
;
3183 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
3184 pe
->pe_number
, OPAL_M32_WINDOW_TYPE
, 0, index
);
3185 if (rc
!= OPAL_SUCCESS
) {
3186 pr_err("%s: Error %lld mapping M32 segment#%d to PE#%x",
3187 __func__
, rc
, index
, pe
->pe_number
);
3191 region
.start
+= phb
->ioda
.m32_segsize
;
3198 * This function is supposed to be called on basis of PE from top
3199 * to bottom style. So the the I/O or MMIO segment assigned to
3200 * parent PE could be overridden by its child PEs if necessary.
3202 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe
*pe
)
3204 struct pci_dev
*pdev
;
3208 * NOTE: We only care PCI bus based PE for now. For PCI
3209 * device based PE, for example SRIOV sensitive VF should
3210 * be figured out later.
3212 BUG_ON(!(pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
)));
3214 list_for_each_entry(pdev
, &pe
->pbus
->devices
, bus_list
) {
3215 for (i
= 0; i
<= PCI_ROM_RESOURCE
; i
++)
3216 pnv_ioda_setup_pe_res(pe
, &pdev
->resource
[i
]);
3219 * If the PE contains all subordinate PCI buses, the
3220 * windows of the child bridges should be mapped to
3223 if (!(pe
->flags
& PNV_IODA_PE_BUS_ALL
) || !pci_is_bridge(pdev
))
3225 for (i
= 0; i
< PCI_BRIDGE_RESOURCE_NUM
; i
++)
3226 pnv_ioda_setup_pe_res(pe
,
3227 &pdev
->resource
[PCI_BRIDGE_RESOURCES
+ i
]);
3231 #ifdef CONFIG_DEBUG_FS
3232 static int pnv_pci_diag_data_set(void *data
, u64 val
)
3234 struct pci_controller
*hose
;
3235 struct pnv_phb
*phb
;
3241 hose
= (struct pci_controller
*)data
;
3242 if (!hose
|| !hose
->private_data
)
3245 phb
= hose
->private_data
;
3247 /* Retrieve the diag data from firmware */
3248 ret
= opal_pci_get_phb_diag_data2(phb
->opal_id
, phb
->diag_data
,
3249 phb
->diag_data_size
);
3250 if (ret
!= OPAL_SUCCESS
)
3253 /* Print the diag data to the kernel log */
3254 pnv_pci_dump_phb_diag_data(phb
->hose
, phb
->diag_data
);
3258 DEFINE_SIMPLE_ATTRIBUTE(pnv_pci_diag_data_fops
, NULL
,
3259 pnv_pci_diag_data_set
, "%llu\n");
3261 #endif /* CONFIG_DEBUG_FS */
3263 static void pnv_pci_ioda_create_dbgfs(void)
3265 #ifdef CONFIG_DEBUG_FS
3266 struct pci_controller
*hose
, *tmp
;
3267 struct pnv_phb
*phb
;
3270 list_for_each_entry_safe(hose
, tmp
, &hose_list
, list_node
) {
3271 phb
= hose
->private_data
;
3273 /* Notify initialization of PHB done */
3274 phb
->initialized
= 1;
3276 sprintf(name
, "PCI%04x", hose
->global_number
);
3277 phb
->dbgfs
= debugfs_create_dir(name
, powerpc_debugfs_root
);
3279 pr_warning("%s: Error on creating debugfs on PHB#%x\n",
3280 __func__
, hose
->global_number
);
3284 debugfs_create_file("dump_diag_regs", 0200, phb
->dbgfs
, hose
,
3285 &pnv_pci_diag_data_fops
);
3287 #endif /* CONFIG_DEBUG_FS */
3290 static void pnv_pci_ioda_fixup(void)
3292 pnv_pci_ioda_setup_PEs();
3293 pnv_pci_ioda_setup_iommu_api();
3294 pnv_pci_ioda_create_dbgfs();
3298 eeh_addr_cache_build();
3303 * Returns the alignment for I/O or memory windows for P2P
3304 * bridges. That actually depends on how PEs are segmented.
3305 * For now, we return I/O or M32 segment size for PE sensitive
3306 * P2P bridges. Otherwise, the default values (4KiB for I/O,
3307 * 1MiB for memory) will be returned.
3309 * The current PCI bus might be put into one PE, which was
3310 * create against the parent PCI bridge. For that case, we
3311 * needn't enlarge the alignment so that we can save some
3314 static resource_size_t
pnv_pci_window_alignment(struct pci_bus
*bus
,
3317 struct pci_dev
*bridge
;
3318 struct pci_controller
*hose
= pci_bus_to_host(bus
);
3319 struct pnv_phb
*phb
= hose
->private_data
;
3320 int num_pci_bridges
= 0;
3324 if (pci_pcie_type(bridge
) == PCI_EXP_TYPE_PCI_BRIDGE
) {
3326 if (num_pci_bridges
>= 2)
3330 bridge
= bridge
->bus
->self
;
3334 * We fall back to M32 if M64 isn't supported. We enforce the M64
3335 * alignment for any 64-bit resource, PCIe doesn't care and
3336 * bridges only do 64-bit prefetchable anyway.
3338 if (phb
->ioda
.m64_segsize
&& pnv_pci_is_m64_flags(type
))
3339 return phb
->ioda
.m64_segsize
;
3340 if (type
& IORESOURCE_MEM
)
3341 return phb
->ioda
.m32_segsize
;
3343 return phb
->ioda
.io_segsize
;
3347 * We are updating root port or the upstream port of the
3348 * bridge behind the root port with PHB's windows in order
3349 * to accommodate the changes on required resources during
3350 * PCI (slot) hotplug, which is connected to either root
3351 * port or the downstream ports of PCIe switch behind the
3354 static void pnv_pci_fixup_bridge_resources(struct pci_bus
*bus
,
3357 struct pci_controller
*hose
= pci_bus_to_host(bus
);
3358 struct pnv_phb
*phb
= hose
->private_data
;
3359 struct pci_dev
*bridge
= bus
->self
;
3360 struct resource
*r
, *w
;
3361 bool msi_region
= false;
3364 /* Check if we need apply fixup to the bridge's windows */
3365 if (!pci_is_root_bus(bridge
->bus
) &&
3366 !pci_is_root_bus(bridge
->bus
->self
->bus
))
3369 /* Fixup the resources */
3370 for (i
= 0; i
< PCI_BRIDGE_RESOURCE_NUM
; i
++) {
3371 r
= &bridge
->resource
[PCI_BRIDGE_RESOURCES
+ i
];
3372 if (!r
->flags
|| !r
->parent
)
3376 if (r
->flags
& type
& IORESOURCE_IO
)
3377 w
= &hose
->io_resource
;
3378 else if (pnv_pci_is_m64(phb
, r
) &&
3379 (type
& IORESOURCE_PREFETCH
) &&
3380 phb
->ioda
.m64_segsize
)
3381 w
= &hose
->mem_resources
[1];
3382 else if (r
->flags
& type
& IORESOURCE_MEM
) {
3383 w
= &hose
->mem_resources
[0];
3387 r
->start
= w
->start
;
3390 /* The 64KB 32-bits MSI region shouldn't be included in
3391 * the 32-bits bridge window. Otherwise, we can see strange
3392 * issues. One of them is EEH error observed on Garrison.
3394 * Exclude top 1MB region which is the minimal alignment of
3395 * 32-bits bridge window.
3404 static void pnv_pci_setup_bridge(struct pci_bus
*bus
, unsigned long type
)
3406 struct pci_controller
*hose
= pci_bus_to_host(bus
);
3407 struct pnv_phb
*phb
= hose
->private_data
;
3408 struct pci_dev
*bridge
= bus
->self
;
3409 struct pnv_ioda_pe
*pe
;
3410 bool all
= (pci_pcie_type(bridge
) == PCI_EXP_TYPE_PCI_BRIDGE
);
3412 /* Extend bridge's windows if necessary */
3413 pnv_pci_fixup_bridge_resources(bus
, type
);
3415 /* The PE for root bus should be realized before any one else */
3416 if (!phb
->ioda
.root_pe_populated
) {
3417 pe
= pnv_ioda_setup_bus_PE(phb
->hose
->bus
, false);
3419 phb
->ioda
.root_pe_idx
= pe
->pe_number
;
3420 phb
->ioda
.root_pe_populated
= true;
3424 /* Don't assign PE to PCI bus, which doesn't have subordinate devices */
3425 if (list_empty(&bus
->devices
))
3428 /* Reserve PEs according to used M64 resources */
3429 if (phb
->reserve_m64_pe
)
3430 phb
->reserve_m64_pe(bus
, NULL
, all
);
3433 * Assign PE. We might run here because of partial hotplug.
3434 * For the case, we just pick up the existing PE and should
3435 * not allocate resources again.
3437 pe
= pnv_ioda_setup_bus_PE(bus
, all
);
3441 pnv_ioda_setup_pe_seg(pe
);
3442 switch (phb
->type
) {
3444 pnv_pci_ioda1_setup_dma_pe(phb
, pe
);
3447 pnv_pci_ioda2_setup_dma_pe(phb
, pe
);
3450 pr_warn("%s: No DMA for PHB#%x (type %d)\n",
3451 __func__
, phb
->hose
->global_number
, phb
->type
);
3455 static resource_size_t
pnv_pci_default_alignment(void)
3460 #ifdef CONFIG_PCI_IOV
3461 static resource_size_t
pnv_pci_iov_resource_alignment(struct pci_dev
*pdev
,
3464 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
3465 struct pnv_phb
*phb
= hose
->private_data
;
3466 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
3467 resource_size_t align
;
3470 * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the
3471 * SR-IOV. While from hardware perspective, the range mapped by M64
3472 * BAR should be size aligned.
3474 * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra
3475 * powernv-specific hardware restriction is gone. But if just use the
3476 * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with
3477 * in one segment of M64 #15, which introduces the PE conflict between
3478 * PF and VF. Based on this, the minimum alignment of an IOV BAR is
3481 * This function returns the total IOV BAR size if M64 BAR is in
3482 * Shared PE mode or just VF BAR size if not.
3483 * If the M64 BAR is in Single PE mode, return the VF BAR size or
3484 * M64 segment size if IOV BAR size is less.
3486 align
= pci_iov_resource_size(pdev
, resno
);
3487 if (!pdn
->vfs_expanded
)
3489 if (pdn
->m64_single_mode
)
3490 return max(align
, (resource_size_t
)phb
->ioda
.m64_segsize
);
3492 return pdn
->vfs_expanded
* align
;
3494 #endif /* CONFIG_PCI_IOV */
3496 /* Prevent enabling devices for which we couldn't properly
3499 bool pnv_pci_enable_device_hook(struct pci_dev
*dev
)
3501 struct pci_controller
*hose
= pci_bus_to_host(dev
->bus
);
3502 struct pnv_phb
*phb
= hose
->private_data
;
3505 /* The function is probably called while the PEs have
3506 * not be created yet. For example, resource reassignment
3507 * during PCI probe period. We just skip the check if
3510 if (!phb
->initialized
)
3513 pdn
= pci_get_pdn(dev
);
3514 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
3520 static long pnv_pci_ioda1_unset_window(struct iommu_table_group
*table_group
,
3523 struct pnv_ioda_pe
*pe
= container_of(table_group
,
3524 struct pnv_ioda_pe
, table_group
);
3525 struct pnv_phb
*phb
= pe
->phb
;
3529 pe_info(pe
, "Removing DMA window #%d\n", num
);
3530 for (idx
= 0; idx
< phb
->ioda
.dma32_count
; idx
++) {
3531 if (phb
->ioda
.dma32_segmap
[idx
] != pe
->pe_number
)
3534 rc
= opal_pci_map_pe_dma_window(phb
->opal_id
, pe
->pe_number
,
3535 idx
, 0, 0ul, 0ul, 0ul);
3536 if (rc
!= OPAL_SUCCESS
) {
3537 pe_warn(pe
, "Failure %ld unmapping DMA32 segment#%d\n",
3542 phb
->ioda
.dma32_segmap
[idx
] = IODA_INVALID_PE
;
3545 pnv_pci_unlink_table_and_group(table_group
->tables
[num
], table_group
);
3546 return OPAL_SUCCESS
;
3549 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe
*pe
)
3551 unsigned int weight
= pnv_pci_ioda_pe_dma_weight(pe
);
3552 struct iommu_table
*tbl
= pe
->table_group
.tables
[0];
3558 rc
= pnv_pci_ioda1_unset_window(&pe
->table_group
, 0);
3559 if (rc
!= OPAL_SUCCESS
)
3562 pnv_pci_p7ioc_tce_invalidate(tbl
, tbl
->it_offset
, tbl
->it_size
, false);
3563 if (pe
->table_group
.group
) {
3564 iommu_group_put(pe
->table_group
.group
);
3565 WARN_ON(pe
->table_group
.group
);
3568 free_pages(tbl
->it_base
, get_order(tbl
->it_size
<< 3));
3569 iommu_tce_table_put(tbl
);
3572 static void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe
*pe
)
3574 struct iommu_table
*tbl
= pe
->table_group
.tables
[0];
3575 unsigned int weight
= pnv_pci_ioda_pe_dma_weight(pe
);
3576 #ifdef CONFIG_IOMMU_API
3583 #ifdef CONFIG_IOMMU_API
3584 rc
= pnv_pci_ioda2_unset_window(&pe
->table_group
, 0);
3586 pe_warn(pe
, "OPAL error %ld release DMA window\n", rc
);
3589 pnv_pci_ioda2_set_bypass(pe
, false);
3590 if (pe
->table_group
.group
) {
3591 iommu_group_put(pe
->table_group
.group
);
3592 WARN_ON(pe
->table_group
.group
);
3595 pnv_pci_ioda2_table_free_pages(tbl
);
3596 iommu_tce_table_put(tbl
);
3599 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe
*pe
,
3603 struct pnv_phb
*phb
= pe
->phb
;
3607 for (idx
= 0; idx
< phb
->ioda
.total_pe_num
; idx
++) {
3608 if (map
[idx
] != pe
->pe_number
)
3611 if (win
== OPAL_M64_WINDOW_TYPE
)
3612 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
3613 phb
->ioda
.reserved_pe_idx
, win
,
3614 idx
/ PNV_IODA1_M64_SEGS
,
3615 idx
% PNV_IODA1_M64_SEGS
);
3617 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
3618 phb
->ioda
.reserved_pe_idx
, win
, 0, idx
);
3620 if (rc
!= OPAL_SUCCESS
)
3621 pe_warn(pe
, "Error %ld unmapping (%d) segment#%d\n",
3624 map
[idx
] = IODA_INVALID_PE
;
3628 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe
*pe
)
3630 struct pnv_phb
*phb
= pe
->phb
;
3632 if (phb
->type
== PNV_PHB_IODA1
) {
3633 pnv_ioda_free_pe_seg(pe
, OPAL_IO_WINDOW_TYPE
,
3634 phb
->ioda
.io_segmap
);
3635 pnv_ioda_free_pe_seg(pe
, OPAL_M32_WINDOW_TYPE
,
3636 phb
->ioda
.m32_segmap
);
3637 pnv_ioda_free_pe_seg(pe
, OPAL_M64_WINDOW_TYPE
,
3638 phb
->ioda
.m64_segmap
);
3639 } else if (phb
->type
== PNV_PHB_IODA2
) {
3640 pnv_ioda_free_pe_seg(pe
, OPAL_M32_WINDOW_TYPE
,
3641 phb
->ioda
.m32_segmap
);
3645 static void pnv_ioda_release_pe(struct pnv_ioda_pe
*pe
)
3647 struct pnv_phb
*phb
= pe
->phb
;
3648 struct pnv_ioda_pe
*slave
, *tmp
;
3650 list_del(&pe
->list
);
3651 switch (phb
->type
) {
3653 pnv_pci_ioda1_release_pe_dma(pe
);
3656 pnv_pci_ioda2_release_pe_dma(pe
);
3662 pnv_ioda_release_pe_seg(pe
);
3663 pnv_ioda_deconfigure_pe(pe
->phb
, pe
);
3665 /* Release slave PEs in the compound PE */
3666 if (pe
->flags
& PNV_IODA_PE_MASTER
) {
3667 list_for_each_entry_safe(slave
, tmp
, &pe
->slaves
, list
) {
3668 list_del(&slave
->list
);
3669 pnv_ioda_free_pe(slave
);
3674 * The PE for root bus can be removed because of hotplug in EEH
3675 * recovery for fenced PHB error. We need to mark the PE dead so
3676 * that it can be populated again in PCI hot add path. The PE
3677 * shouldn't be destroyed as it's the global reserved resource.
3679 if (phb
->ioda
.root_pe_populated
&&
3680 phb
->ioda
.root_pe_idx
== pe
->pe_number
)
3681 phb
->ioda
.root_pe_populated
= false;
3683 pnv_ioda_free_pe(pe
);
3686 static void pnv_pci_release_device(struct pci_dev
*pdev
)
3688 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
3689 struct pnv_phb
*phb
= hose
->private_data
;
3690 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
3691 struct pnv_ioda_pe
*pe
;
3693 if (pdev
->is_virtfn
)
3696 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
3700 * PCI hotplug can happen as part of EEH error recovery. The @pdn
3701 * isn't removed and added afterwards in this scenario. We should
3702 * set the PE number in @pdn to an invalid one. Otherwise, the PE's
3703 * device count is decreased on removing devices while failing to
3704 * be increased on adding devices. It leads to unbalanced PE's device
3705 * count and eventually make normal PCI hotplug path broken.
3707 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
3708 pdn
->pe_number
= IODA_INVALID_PE
;
3710 WARN_ON(--pe
->device_count
< 0);
3711 if (pe
->device_count
== 0)
3712 pnv_ioda_release_pe(pe
);
3715 static void pnv_pci_ioda_shutdown(struct pci_controller
*hose
)
3717 struct pnv_phb
*phb
= hose
->private_data
;
3719 opal_pci_reset(phb
->opal_id
, OPAL_RESET_PCI_IODA_TABLE
,
3723 static const struct pci_controller_ops pnv_pci_ioda_controller_ops
= {
3724 .dma_dev_setup
= pnv_pci_dma_dev_setup
,
3725 .dma_bus_setup
= pnv_pci_dma_bus_setup
,
3726 #ifdef CONFIG_PCI_MSI
3727 .setup_msi_irqs
= pnv_setup_msi_irqs
,
3728 .teardown_msi_irqs
= pnv_teardown_msi_irqs
,
3730 .enable_device_hook
= pnv_pci_enable_device_hook
,
3731 .release_device
= pnv_pci_release_device
,
3732 .window_alignment
= pnv_pci_window_alignment
,
3733 .setup_bridge
= pnv_pci_setup_bridge
,
3734 .reset_secondary_bus
= pnv_pci_reset_secondary_bus
,
3735 .dma_set_mask
= pnv_pci_ioda_dma_set_mask
,
3736 .dma_get_required_mask
= pnv_pci_ioda_dma_get_required_mask
,
3737 .shutdown
= pnv_pci_ioda_shutdown
,
3740 static int pnv_npu_dma_set_mask(struct pci_dev
*npdev
, u64 dma_mask
)
3742 dev_err_once(&npdev
->dev
,
3743 "%s operation unsupported for NVLink devices\n",
3748 static const struct pci_controller_ops pnv_npu_ioda_controller_ops
= {
3749 .dma_dev_setup
= pnv_pci_dma_dev_setup
,
3750 #ifdef CONFIG_PCI_MSI
3751 .setup_msi_irqs
= pnv_setup_msi_irqs
,
3752 .teardown_msi_irqs
= pnv_teardown_msi_irqs
,
3754 .enable_device_hook
= pnv_pci_enable_device_hook
,
3755 .window_alignment
= pnv_pci_window_alignment
,
3756 .reset_secondary_bus
= pnv_pci_reset_secondary_bus
,
3757 .dma_set_mask
= pnv_npu_dma_set_mask
,
3758 .shutdown
= pnv_pci_ioda_shutdown
,
3761 #ifdef CONFIG_CXL_BASE
3762 const struct pci_controller_ops pnv_cxl_cx4_ioda_controller_ops
= {
3763 .dma_dev_setup
= pnv_pci_dma_dev_setup
,
3764 .dma_bus_setup
= pnv_pci_dma_bus_setup
,
3765 #ifdef CONFIG_PCI_MSI
3766 .setup_msi_irqs
= pnv_cxl_cx4_setup_msi_irqs
,
3767 .teardown_msi_irqs
= pnv_cxl_cx4_teardown_msi_irqs
,
3769 .enable_device_hook
= pnv_cxl_enable_device_hook
,
3770 .disable_device
= pnv_cxl_disable_device
,
3771 .release_device
= pnv_pci_release_device
,
3772 .window_alignment
= pnv_pci_window_alignment
,
3773 .setup_bridge
= pnv_pci_setup_bridge
,
3774 .reset_secondary_bus
= pnv_pci_reset_secondary_bus
,
3775 .dma_set_mask
= pnv_pci_ioda_dma_set_mask
,
3776 .dma_get_required_mask
= pnv_pci_ioda_dma_get_required_mask
,
3777 .shutdown
= pnv_pci_ioda_shutdown
,
3781 static void __init
pnv_pci_init_ioda_phb(struct device_node
*np
,
3782 u64 hub_id
, int ioda_type
)
3784 struct pci_controller
*hose
;
3785 struct pnv_phb
*phb
;
3786 unsigned long size
, m64map_off
, m32map_off
, pemap_off
;
3787 unsigned long iomap_off
= 0, dma32map_off
= 0;
3789 const __be64
*prop64
;
3790 const __be32
*prop32
;
3797 if (!of_device_is_available(np
))
3800 pr_info("Initializing %s PHB (%s)\n",
3801 pnv_phb_names
[ioda_type
], of_node_full_name(np
));
3803 prop64
= of_get_property(np
, "ibm,opal-phbid", NULL
);
3805 pr_err(" Missing \"ibm,opal-phbid\" property !\n");
3808 phb_id
= be64_to_cpup(prop64
);
3809 pr_debug(" PHB-ID : 0x%016llx\n", phb_id
);
3811 phb
= memblock_virt_alloc(sizeof(struct pnv_phb
), 0);
3813 /* Allocate PCI controller */
3814 phb
->hose
= hose
= pcibios_alloc_controller(np
);
3816 pr_err(" Can't allocate PCI controller for %s\n",
3818 memblock_free(__pa(phb
), sizeof(struct pnv_phb
));
3822 spin_lock_init(&phb
->lock
);
3823 prop32
= of_get_property(np
, "bus-range", &len
);
3824 if (prop32
&& len
== 8) {
3825 hose
->first_busno
= be32_to_cpu(prop32
[0]);
3826 hose
->last_busno
= be32_to_cpu(prop32
[1]);
3828 pr_warn(" Broken <bus-range> on %s\n", np
->full_name
);
3829 hose
->first_busno
= 0;
3830 hose
->last_busno
= 0xff;
3832 hose
->private_data
= phb
;
3833 phb
->hub_id
= hub_id
;
3834 phb
->opal_id
= phb_id
;
3835 phb
->type
= ioda_type
;
3836 mutex_init(&phb
->ioda
.pe_alloc_mutex
);
3838 /* Detect specific models for error handling */
3839 if (of_device_is_compatible(np
, "ibm,p7ioc-pciex"))
3840 phb
->model
= PNV_PHB_MODEL_P7IOC
;
3841 else if (of_device_is_compatible(np
, "ibm,power8-pciex"))
3842 phb
->model
= PNV_PHB_MODEL_PHB3
;
3843 else if (of_device_is_compatible(np
, "ibm,power8-npu-pciex"))
3844 phb
->model
= PNV_PHB_MODEL_NPU
;
3845 else if (of_device_is_compatible(np
, "ibm,power9-npu-pciex"))
3846 phb
->model
= PNV_PHB_MODEL_NPU2
;
3848 phb
->model
= PNV_PHB_MODEL_UNKNOWN
;
3850 /* Initialize diagnostic data buffer */
3851 prop32
= of_get_property(np
, "ibm,phb-diag-data-size", NULL
);
3853 phb
->diag_data_size
= be32_to_cpup(prop32
);
3855 phb
->diag_data_size
= PNV_PCI_DIAG_BUF_SIZE
;
3857 phb
->diag_data
= memblock_virt_alloc(phb
->diag_data_size
, 0);
3859 /* Parse 32-bit and IO ranges (if any) */
3860 pci_process_bridge_OF_ranges(hose
, np
, !hose
->global_number
);
3863 if (!of_address_to_resource(np
, 0, &r
)) {
3864 phb
->regs_phys
= r
.start
;
3865 phb
->regs
= ioremap(r
.start
, resource_size(&r
));
3866 if (phb
->regs
== NULL
)
3867 pr_err(" Failed to map registers !\n");
3870 /* Initialize more IODA stuff */
3871 phb
->ioda
.total_pe_num
= 1;
3872 prop32
= of_get_property(np
, "ibm,opal-num-pes", NULL
);
3874 phb
->ioda
.total_pe_num
= be32_to_cpup(prop32
);
3875 prop32
= of_get_property(np
, "ibm,opal-reserved-pe", NULL
);
3877 phb
->ioda
.reserved_pe_idx
= be32_to_cpup(prop32
);
3879 /* Invalidate RID to PE# mapping */
3880 for (segno
= 0; segno
< ARRAY_SIZE(phb
->ioda
.pe_rmap
); segno
++)
3881 phb
->ioda
.pe_rmap
[segno
] = IODA_INVALID_PE
;
3883 /* Parse 64-bit MMIO range */
3884 pnv_ioda_parse_m64_window(phb
);
3886 phb
->ioda
.m32_size
= resource_size(&hose
->mem_resources
[0]);
3887 /* FW Has already off top 64k of M32 space (MSI space) */
3888 phb
->ioda
.m32_size
+= 0x10000;
3890 phb
->ioda
.m32_segsize
= phb
->ioda
.m32_size
/ phb
->ioda
.total_pe_num
;
3891 phb
->ioda
.m32_pci_base
= hose
->mem_resources
[0].start
- hose
->mem_offset
[0];
3892 phb
->ioda
.io_size
= hose
->pci_io_size
;
3893 phb
->ioda
.io_segsize
= phb
->ioda
.io_size
/ phb
->ioda
.total_pe_num
;
3894 phb
->ioda
.io_pci_base
= 0; /* XXX calculate this ? */
3896 /* Calculate how many 32-bit TCE segments we have */
3897 phb
->ioda
.dma32_count
= phb
->ioda
.m32_pci_base
/
3898 PNV_IODA1_DMA32_SEGSIZE
;
3900 /* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3901 size
= _ALIGN_UP(max_t(unsigned, phb
->ioda
.total_pe_num
, 8) / 8,
3902 sizeof(unsigned long));
3904 size
+= phb
->ioda
.total_pe_num
* sizeof(phb
->ioda
.m64_segmap
[0]);
3906 size
+= phb
->ioda
.total_pe_num
* sizeof(phb
->ioda
.m32_segmap
[0]);
3907 if (phb
->type
== PNV_PHB_IODA1
) {
3909 size
+= phb
->ioda
.total_pe_num
* sizeof(phb
->ioda
.io_segmap
[0]);
3910 dma32map_off
= size
;
3911 size
+= phb
->ioda
.dma32_count
*
3912 sizeof(phb
->ioda
.dma32_segmap
[0]);
3915 size
+= phb
->ioda
.total_pe_num
* sizeof(struct pnv_ioda_pe
);
3916 aux
= memblock_virt_alloc(size
, 0);
3917 phb
->ioda
.pe_alloc
= aux
;
3918 phb
->ioda
.m64_segmap
= aux
+ m64map_off
;
3919 phb
->ioda
.m32_segmap
= aux
+ m32map_off
;
3920 for (segno
= 0; segno
< phb
->ioda
.total_pe_num
; segno
++) {
3921 phb
->ioda
.m64_segmap
[segno
] = IODA_INVALID_PE
;
3922 phb
->ioda
.m32_segmap
[segno
] = IODA_INVALID_PE
;
3924 if (phb
->type
== PNV_PHB_IODA1
) {
3925 phb
->ioda
.io_segmap
= aux
+ iomap_off
;
3926 for (segno
= 0; segno
< phb
->ioda
.total_pe_num
; segno
++)
3927 phb
->ioda
.io_segmap
[segno
] = IODA_INVALID_PE
;
3929 phb
->ioda
.dma32_segmap
= aux
+ dma32map_off
;
3930 for (segno
= 0; segno
< phb
->ioda
.dma32_count
; segno
++)
3931 phb
->ioda
.dma32_segmap
[segno
] = IODA_INVALID_PE
;
3933 phb
->ioda
.pe_array
= aux
+ pemap_off
;
3936 * Choose PE number for root bus, which shouldn't have
3937 * M64 resources consumed by its child devices. To pick
3938 * the PE number adjacent to the reserved one if possible.
3940 pnv_ioda_reserve_pe(phb
, phb
->ioda
.reserved_pe_idx
);
3941 if (phb
->ioda
.reserved_pe_idx
== 0) {
3942 phb
->ioda
.root_pe_idx
= 1;
3943 pnv_ioda_reserve_pe(phb
, phb
->ioda
.root_pe_idx
);
3944 } else if (phb
->ioda
.reserved_pe_idx
== (phb
->ioda
.total_pe_num
- 1)) {
3945 phb
->ioda
.root_pe_idx
= phb
->ioda
.reserved_pe_idx
- 1;
3946 pnv_ioda_reserve_pe(phb
, phb
->ioda
.root_pe_idx
);
3948 phb
->ioda
.root_pe_idx
= IODA_INVALID_PE
;
3951 INIT_LIST_HEAD(&phb
->ioda
.pe_list
);
3952 mutex_init(&phb
->ioda
.pe_list_mutex
);
3954 /* Calculate how many 32-bit TCE segments we have */
3955 phb
->ioda
.dma32_count
= phb
->ioda
.m32_pci_base
/
3956 PNV_IODA1_DMA32_SEGSIZE
;
3958 #if 0 /* We should really do that ... */
3959 rc
= opal_pci_set_phb_mem_window(opal
->phb_id
,
3962 starting_real_address
,
3963 starting_pci_address
,
3967 pr_info(" %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3968 phb
->ioda
.total_pe_num
, phb
->ioda
.reserved_pe_idx
,
3969 phb
->ioda
.m32_size
, phb
->ioda
.m32_segsize
);
3970 if (phb
->ioda
.m64_size
)
3971 pr_info(" M64: 0x%lx [segment=0x%lx]\n",
3972 phb
->ioda
.m64_size
, phb
->ioda
.m64_segsize
);
3973 if (phb
->ioda
.io_size
)
3974 pr_info(" IO: 0x%x [segment=0x%x]\n",
3975 phb
->ioda
.io_size
, phb
->ioda
.io_segsize
);
3978 phb
->hose
->ops
= &pnv_pci_ops
;
3979 phb
->get_pe_state
= pnv_ioda_get_pe_state
;
3980 phb
->freeze_pe
= pnv_ioda_freeze_pe
;
3981 phb
->unfreeze_pe
= pnv_ioda_unfreeze_pe
;
3983 /* Setup MSI support */
3984 pnv_pci_init_ioda_msis(phb
);
3987 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3988 * to let the PCI core do resource assignment. It's supposed
3989 * that the PCI core will do correct I/O and MMIO alignment
3990 * for the P2P bridge bars so that each PCI bus (excluding
3991 * the child P2P bridges) can form individual PE.
3993 ppc_md
.pcibios_fixup
= pnv_pci_ioda_fixup
;
3995 if (phb
->type
== PNV_PHB_NPU
) {
3996 hose
->controller_ops
= pnv_npu_ioda_controller_ops
;
3998 phb
->dma_dev_setup
= pnv_pci_ioda_dma_dev_setup
;
3999 hose
->controller_ops
= pnv_pci_ioda_controller_ops
;
4002 ppc_md
.pcibios_default_alignment
= pnv_pci_default_alignment
;
4004 #ifdef CONFIG_PCI_IOV
4005 ppc_md
.pcibios_fixup_sriov
= pnv_pci_ioda_fixup_iov_resources
;
4006 ppc_md
.pcibios_iov_resource_alignment
= pnv_pci_iov_resource_alignment
;
4009 pci_add_flags(PCI_REASSIGN_ALL_RSRC
);
4011 /* Reset IODA tables to a clean state */
4012 rc
= opal_pci_reset(phb_id
, OPAL_RESET_PCI_IODA_TABLE
, OPAL_ASSERT_RESET
);
4014 pr_warning(" OPAL Error %ld performing IODA table reset !\n", rc
);
4017 * If we're running in kdump kernel, the previous kernel never
4018 * shutdown PCI devices correctly. We already got IODA table
4019 * cleaned out. So we have to issue PHB reset to stop all PCI
4020 * transactions from previous kernel.
4022 if (is_kdump_kernel()) {
4023 pr_info(" Issue PHB reset ...\n");
4024 pnv_eeh_phb_reset(hose
, EEH_RESET_FUNDAMENTAL
);
4025 pnv_eeh_phb_reset(hose
, EEH_RESET_DEACTIVATE
);
4028 /* Remove M64 resource if we can't configure it successfully */
4029 if (!phb
->init_m64
|| phb
->init_m64(phb
))
4030 hose
->mem_resources
[1].flags
= 0;
4033 void __init
pnv_pci_init_ioda2_phb(struct device_node
*np
)
4035 pnv_pci_init_ioda_phb(np
, 0, PNV_PHB_IODA2
);
4038 void __init
pnv_pci_init_npu_phb(struct device_node
*np
)
4040 pnv_pci_init_ioda_phb(np
, 0, PNV_PHB_NPU
);
4043 void __init
pnv_pci_init_ioda_hub(struct device_node
*np
)
4045 struct device_node
*phbn
;
4046 const __be64
*prop64
;
4049 pr_info("Probing IODA IO-Hub %s\n", np
->full_name
);
4051 prop64
= of_get_property(np
, "ibm,opal-hubid", NULL
);
4053 pr_err(" Missing \"ibm,opal-hubid\" property !\n");
4056 hub_id
= be64_to_cpup(prop64
);
4057 pr_devel(" HUB-ID : 0x%016llx\n", hub_id
);
4059 /* Count child PHBs */
4060 for_each_child_of_node(np
, phbn
) {
4061 /* Look for IODA1 PHBs */
4062 if (of_device_is_compatible(phbn
, "ibm,ioda-phb"))
4063 pnv_pci_init_ioda_phb(phbn
, hub_id
, PNV_PHB_IODA1
);