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/debugfs.h>
18 #include <linux/delay.h>
19 #include <linux/string.h>
20 #include <linux/init.h>
21 #include <linux/bootmem.h>
22 #include <linux/irq.h>
24 #include <linux/msi.h>
25 #include <linux/memblock.h>
26 #include <linux/iommu.h>
27 #include <linux/rculist.h>
28 #include <linux/sizes.h>
30 #include <asm/sections.h>
33 #include <asm/pci-bridge.h>
34 #include <asm/machdep.h>
35 #include <asm/msi_bitmap.h>
36 #include <asm/ppc-pci.h>
38 #include <asm/iommu.h>
41 #include <asm/debug.h>
42 #include <asm/firmware.h>
43 #include <asm/pnv-pci.h>
44 #include <asm/mmzone.h>
46 #include <misc/cxl-base.h>
51 #define PNV_IODA1_M64_NUM 16 /* Number of M64 BARs */
52 #define PNV_IODA1_M64_SEGS 8 /* Segments per M64 BAR */
53 #define PNV_IODA1_DMA32_SEGSIZE 0x10000000
55 #define POWERNV_IOMMU_DEFAULT_LEVELS 1
56 #define POWERNV_IOMMU_MAX_LEVELS 5
58 static const char * const pnv_phb_names
[] = { "IODA1", "IODA2", "NPU" };
59 static void pnv_pci_ioda2_table_free_pages(struct iommu_table
*tbl
);
61 void pe_level_printk(const struct pnv_ioda_pe
*pe
, const char *level
,
73 if (pe
->flags
& PNV_IODA_PE_DEV
)
74 strlcpy(pfix
, dev_name(&pe
->pdev
->dev
), sizeof(pfix
));
75 else if (pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
))
76 sprintf(pfix
, "%04x:%02x ",
77 pci_domain_nr(pe
->pbus
), pe
->pbus
->number
);
79 else if (pe
->flags
& PNV_IODA_PE_VF
)
80 sprintf(pfix
, "%04x:%02x:%2x.%d",
81 pci_domain_nr(pe
->parent_dev
->bus
),
82 (pe
->rid
& 0xff00) >> 8,
83 PCI_SLOT(pe
->rid
), PCI_FUNC(pe
->rid
));
84 #endif /* CONFIG_PCI_IOV*/
86 printk("%spci %s: [PE# %.3d] %pV",
87 level
, pfix
, pe
->pe_number
, &vaf
);
92 static bool pnv_iommu_bypass_disabled __read_mostly
;
94 static int __init
iommu_setup(char *str
)
100 if (!strncmp(str
, "nobypass", 8)) {
101 pnv_iommu_bypass_disabled
= true;
102 pr_info("PowerNV: IOMMU bypass window disabled.\n");
105 str
+= strcspn(str
, ",");
112 early_param("iommu", iommu_setup
);
114 static inline bool pnv_pci_is_mem_pref_64(unsigned long flags
)
116 return ((flags
& (IORESOURCE_MEM_64
| IORESOURCE_PREFETCH
)) ==
117 (IORESOURCE_MEM_64
| IORESOURCE_PREFETCH
));
120 static struct pnv_ioda_pe
*pnv_ioda_init_pe(struct pnv_phb
*phb
, int pe_no
)
122 phb
->ioda
.pe_array
[pe_no
].phb
= phb
;
123 phb
->ioda
.pe_array
[pe_no
].pe_number
= pe_no
;
125 return &phb
->ioda
.pe_array
[pe_no
];
128 static void pnv_ioda_reserve_pe(struct pnv_phb
*phb
, int pe_no
)
130 if (!(pe_no
>= 0 && pe_no
< phb
->ioda
.total_pe_num
)) {
131 pr_warn("%s: Invalid PE %d on PHB#%x\n",
132 __func__
, pe_no
, phb
->hose
->global_number
);
136 if (test_and_set_bit(pe_no
, phb
->ioda
.pe_alloc
))
137 pr_debug("%s: PE %d was reserved on PHB#%x\n",
138 __func__
, pe_no
, phb
->hose
->global_number
);
140 pnv_ioda_init_pe(phb
, pe_no
);
143 static struct pnv_ioda_pe
*pnv_ioda_alloc_pe(struct pnv_phb
*phb
)
145 unsigned long pe
= phb
->ioda
.total_pe_num
- 1;
147 for (pe
= phb
->ioda
.total_pe_num
- 1; pe
>= 0; pe
--) {
148 if (!test_and_set_bit(pe
, phb
->ioda
.pe_alloc
))
149 return pnv_ioda_init_pe(phb
, pe
);
155 static void pnv_ioda_free_pe(struct pnv_ioda_pe
*pe
)
157 struct pnv_phb
*phb
= pe
->phb
;
161 memset(pe
, 0, sizeof(struct pnv_ioda_pe
));
162 clear_bit(pe
->pe_number
, phb
->ioda
.pe_alloc
);
165 /* The default M64 BAR is shared by all PEs */
166 static int pnv_ioda2_init_m64(struct pnv_phb
*phb
)
172 /* Configure the default M64 BAR */
173 rc
= opal_pci_set_phb_mem_window(phb
->opal_id
,
174 OPAL_M64_WINDOW_TYPE
,
175 phb
->ioda
.m64_bar_idx
,
179 if (rc
!= OPAL_SUCCESS
) {
180 desc
= "configuring";
184 /* Enable the default M64 BAR */
185 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
186 OPAL_M64_WINDOW_TYPE
,
187 phb
->ioda
.m64_bar_idx
,
188 OPAL_ENABLE_M64_SPLIT
);
189 if (rc
!= OPAL_SUCCESS
) {
194 /* Mark the M64 BAR assigned */
195 set_bit(phb
->ioda
.m64_bar_idx
, &phb
->ioda
.m64_bar_alloc
);
198 * Exclude the segments for reserved and root bus PE, which
199 * are first or last two PEs.
201 r
= &phb
->hose
->mem_resources
[1];
202 if (phb
->ioda
.reserved_pe_idx
== 0)
203 r
->start
+= (2 * phb
->ioda
.m64_segsize
);
204 else if (phb
->ioda
.reserved_pe_idx
== (phb
->ioda
.total_pe_num
- 1))
205 r
->end
-= (2 * phb
->ioda
.m64_segsize
);
207 pr_warn(" Cannot strip M64 segment for reserved PE#%d\n",
208 phb
->ioda
.reserved_pe_idx
);
213 pr_warn(" Failure %lld %s M64 BAR#%d\n",
214 rc
, desc
, phb
->ioda
.m64_bar_idx
);
215 opal_pci_phb_mmio_enable(phb
->opal_id
,
216 OPAL_M64_WINDOW_TYPE
,
217 phb
->ioda
.m64_bar_idx
,
222 static void pnv_ioda_reserve_dev_m64_pe(struct pci_dev
*pdev
,
223 unsigned long *pe_bitmap
)
225 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
226 struct pnv_phb
*phb
= hose
->private_data
;
228 resource_size_t base
, sgsz
, start
, end
;
231 base
= phb
->ioda
.m64_base
;
232 sgsz
= phb
->ioda
.m64_segsize
;
233 for (i
= 0; i
<= PCI_ROM_RESOURCE
; i
++) {
234 r
= &pdev
->resource
[i
];
235 if (!r
->parent
|| !pnv_pci_is_mem_pref_64(r
->flags
))
238 start
= _ALIGN_DOWN(r
->start
- base
, sgsz
);
239 end
= _ALIGN_UP(r
->end
- base
, sgsz
);
240 for (segno
= start
/ sgsz
; segno
< end
/ sgsz
; segno
++) {
242 set_bit(segno
, pe_bitmap
);
244 pnv_ioda_reserve_pe(phb
, segno
);
249 static int pnv_ioda1_init_m64(struct pnv_phb
*phb
)
255 * There are 16 M64 BARs, each of which has 8 segments. So
256 * there are as many M64 segments as the maximum number of
259 for (index
= 0; index
< PNV_IODA1_M64_NUM
; index
++) {
260 unsigned long base
, segsz
= phb
->ioda
.m64_segsize
;
263 base
= phb
->ioda
.m64_base
+
264 index
* PNV_IODA1_M64_SEGS
* segsz
;
265 rc
= opal_pci_set_phb_mem_window(phb
->opal_id
,
266 OPAL_M64_WINDOW_TYPE
, index
, base
, 0,
267 PNV_IODA1_M64_SEGS
* segsz
);
268 if (rc
!= OPAL_SUCCESS
) {
269 pr_warn(" Error %lld setting M64 PHB#%d-BAR#%d\n",
270 rc
, phb
->hose
->global_number
, index
);
274 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
275 OPAL_M64_WINDOW_TYPE
, index
,
276 OPAL_ENABLE_M64_SPLIT
);
277 if (rc
!= OPAL_SUCCESS
) {
278 pr_warn(" Error %lld enabling M64 PHB#%d-BAR#%d\n",
279 rc
, phb
->hose
->global_number
, index
);
285 * Exclude the segments for reserved and root bus PE, which
286 * are first or last two PEs.
288 r
= &phb
->hose
->mem_resources
[1];
289 if (phb
->ioda
.reserved_pe_idx
== 0)
290 r
->start
+= (2 * phb
->ioda
.m64_segsize
);
291 else if (phb
->ioda
.reserved_pe_idx
== (phb
->ioda
.total_pe_num
- 1))
292 r
->end
-= (2 * phb
->ioda
.m64_segsize
);
294 WARN(1, "Wrong reserved PE#%d on PHB#%d\n",
295 phb
->ioda
.reserved_pe_idx
, phb
->hose
->global_number
);
300 for ( ; index
>= 0; index
--)
301 opal_pci_phb_mmio_enable(phb
->opal_id
,
302 OPAL_M64_WINDOW_TYPE
, index
, OPAL_DISABLE_M64
);
307 static void pnv_ioda_reserve_m64_pe(struct pci_bus
*bus
,
308 unsigned long *pe_bitmap
,
311 struct pci_dev
*pdev
;
313 list_for_each_entry(pdev
, &bus
->devices
, bus_list
) {
314 pnv_ioda_reserve_dev_m64_pe(pdev
, pe_bitmap
);
316 if (all
&& pdev
->subordinate
)
317 pnv_ioda_reserve_m64_pe(pdev
->subordinate
,
322 static struct pnv_ioda_pe
*pnv_ioda_pick_m64_pe(struct pci_bus
*bus
, bool all
)
324 struct pci_controller
*hose
= pci_bus_to_host(bus
);
325 struct pnv_phb
*phb
= hose
->private_data
;
326 struct pnv_ioda_pe
*master_pe
, *pe
;
327 unsigned long size
, *pe_alloc
;
330 /* Root bus shouldn't use M64 */
331 if (pci_is_root_bus(bus
))
334 /* Allocate bitmap */
335 size
= _ALIGN_UP(phb
->ioda
.total_pe_num
/ 8, sizeof(unsigned long));
336 pe_alloc
= kzalloc(size
, GFP_KERNEL
);
338 pr_warn("%s: Out of memory !\n",
343 /* Figure out reserved PE numbers by the PE */
344 pnv_ioda_reserve_m64_pe(bus
, pe_alloc
, all
);
347 * the current bus might not own M64 window and that's all
348 * contributed by its child buses. For the case, we needn't
349 * pick M64 dependent PE#.
351 if (bitmap_empty(pe_alloc
, phb
->ioda
.total_pe_num
)) {
357 * Figure out the master PE and put all slave PEs to master
358 * PE's list to form compound PE.
362 while ((i
= find_next_bit(pe_alloc
, phb
->ioda
.total_pe_num
, i
+ 1)) <
363 phb
->ioda
.total_pe_num
) {
364 pe
= &phb
->ioda
.pe_array
[i
];
366 phb
->ioda
.m64_segmap
[pe
->pe_number
] = pe
->pe_number
;
368 pe
->flags
|= PNV_IODA_PE_MASTER
;
369 INIT_LIST_HEAD(&pe
->slaves
);
372 pe
->flags
|= PNV_IODA_PE_SLAVE
;
373 pe
->master
= master_pe
;
374 list_add_tail(&pe
->list
, &master_pe
->slaves
);
378 * P7IOC supports M64DT, which helps mapping M64 segment
379 * to one particular PE#. However, PHB3 has fixed mapping
380 * between M64 segment and PE#. In order to have same logic
381 * for P7IOC and PHB3, we enforce fixed mapping between M64
382 * segment and PE# on P7IOC.
384 if (phb
->type
== PNV_PHB_IODA1
) {
387 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
388 pe
->pe_number
, OPAL_M64_WINDOW_TYPE
,
389 pe
->pe_number
/ PNV_IODA1_M64_SEGS
,
390 pe
->pe_number
% PNV_IODA1_M64_SEGS
);
391 if (rc
!= OPAL_SUCCESS
)
392 pr_warn("%s: Error %lld mapping M64 for PHB#%d-PE#%d\n",
393 __func__
, rc
, phb
->hose
->global_number
,
402 static void __init
pnv_ioda_parse_m64_window(struct pnv_phb
*phb
)
404 struct pci_controller
*hose
= phb
->hose
;
405 struct device_node
*dn
= hose
->dn
;
406 struct resource
*res
;
410 if (phb
->type
!= PNV_PHB_IODA1
&& phb
->type
!= PNV_PHB_IODA2
) {
411 pr_info(" Not support M64 window\n");
415 if (!firmware_has_feature(FW_FEATURE_OPAL
)) {
416 pr_info(" Firmware too old to support M64 window\n");
420 r
= of_get_property(dn
, "ibm,opal-m64-window", NULL
);
422 pr_info(" No <ibm,opal-m64-window> on %s\n",
427 res
= &hose
->mem_resources
[1];
428 res
->name
= dn
->full_name
;
429 res
->start
= of_translate_address(dn
, r
+ 2);
430 res
->end
= res
->start
+ of_read_number(r
+ 4, 2) - 1;
431 res
->flags
= (IORESOURCE_MEM
| IORESOURCE_MEM_64
| IORESOURCE_PREFETCH
);
432 pci_addr
= of_read_number(r
, 2);
433 hose
->mem_offset
[1] = res
->start
- pci_addr
;
435 phb
->ioda
.m64_size
= resource_size(res
);
436 phb
->ioda
.m64_segsize
= phb
->ioda
.m64_size
/ phb
->ioda
.total_pe_num
;
437 phb
->ioda
.m64_base
= pci_addr
;
439 pr_info(" MEM64 0x%016llx..0x%016llx -> 0x%016llx\n",
440 res
->start
, res
->end
, pci_addr
);
442 /* Use last M64 BAR to cover M64 window */
443 phb
->ioda
.m64_bar_idx
= 15;
444 if (phb
->type
== PNV_PHB_IODA1
)
445 phb
->init_m64
= pnv_ioda1_init_m64
;
447 phb
->init_m64
= pnv_ioda2_init_m64
;
448 phb
->reserve_m64_pe
= pnv_ioda_reserve_m64_pe
;
449 phb
->pick_m64_pe
= pnv_ioda_pick_m64_pe
;
452 static void pnv_ioda_freeze_pe(struct pnv_phb
*phb
, int pe_no
)
454 struct pnv_ioda_pe
*pe
= &phb
->ioda
.pe_array
[pe_no
];
455 struct pnv_ioda_pe
*slave
;
458 /* Fetch master PE */
459 if (pe
->flags
& PNV_IODA_PE_SLAVE
) {
461 if (WARN_ON(!pe
|| !(pe
->flags
& PNV_IODA_PE_MASTER
)))
464 pe_no
= pe
->pe_number
;
467 /* Freeze master PE */
468 rc
= opal_pci_eeh_freeze_set(phb
->opal_id
,
470 OPAL_EEH_ACTION_SET_FREEZE_ALL
);
471 if (rc
!= OPAL_SUCCESS
) {
472 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
473 __func__
, rc
, phb
->hose
->global_number
, pe_no
);
477 /* Freeze slave PEs */
478 if (!(pe
->flags
& PNV_IODA_PE_MASTER
))
481 list_for_each_entry(slave
, &pe
->slaves
, list
) {
482 rc
= opal_pci_eeh_freeze_set(phb
->opal_id
,
484 OPAL_EEH_ACTION_SET_FREEZE_ALL
);
485 if (rc
!= OPAL_SUCCESS
)
486 pr_warn("%s: Failure %lld freezing PHB#%x-PE#%x\n",
487 __func__
, rc
, phb
->hose
->global_number
,
492 static int pnv_ioda_unfreeze_pe(struct pnv_phb
*phb
, int pe_no
, int opt
)
494 struct pnv_ioda_pe
*pe
, *slave
;
498 pe
= &phb
->ioda
.pe_array
[pe_no
];
499 if (pe
->flags
& PNV_IODA_PE_SLAVE
) {
501 WARN_ON(!pe
|| !(pe
->flags
& PNV_IODA_PE_MASTER
));
502 pe_no
= pe
->pe_number
;
505 /* Clear frozen state for master PE */
506 rc
= opal_pci_eeh_freeze_clear(phb
->opal_id
, pe_no
, opt
);
507 if (rc
!= OPAL_SUCCESS
) {
508 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
509 __func__
, rc
, opt
, phb
->hose
->global_number
, pe_no
);
513 if (!(pe
->flags
& PNV_IODA_PE_MASTER
))
516 /* Clear frozen state for slave PEs */
517 list_for_each_entry(slave
, &pe
->slaves
, list
) {
518 rc
= opal_pci_eeh_freeze_clear(phb
->opal_id
,
521 if (rc
!= OPAL_SUCCESS
) {
522 pr_warn("%s: Failure %lld clear %d on PHB#%x-PE#%x\n",
523 __func__
, rc
, opt
, phb
->hose
->global_number
,
532 static int pnv_ioda_get_pe_state(struct pnv_phb
*phb
, int pe_no
)
534 struct pnv_ioda_pe
*slave
, *pe
;
539 /* Sanity check on PE number */
540 if (pe_no
< 0 || pe_no
>= phb
->ioda
.total_pe_num
)
541 return OPAL_EEH_STOPPED_PERM_UNAVAIL
;
544 * Fetch the master PE and the PE instance might be
545 * not initialized yet.
547 pe
= &phb
->ioda
.pe_array
[pe_no
];
548 if (pe
->flags
& PNV_IODA_PE_SLAVE
) {
550 WARN_ON(!pe
|| !(pe
->flags
& PNV_IODA_PE_MASTER
));
551 pe_no
= pe
->pe_number
;
554 /* Check the master PE */
555 rc
= opal_pci_eeh_freeze_status(phb
->opal_id
, pe_no
,
556 &state
, &pcierr
, NULL
);
557 if (rc
!= OPAL_SUCCESS
) {
558 pr_warn("%s: Failure %lld getting "
559 "PHB#%x-PE#%x state\n",
561 phb
->hose
->global_number
, pe_no
);
562 return OPAL_EEH_STOPPED_TEMP_UNAVAIL
;
565 /* Check the slave PE */
566 if (!(pe
->flags
& PNV_IODA_PE_MASTER
))
569 list_for_each_entry(slave
, &pe
->slaves
, list
) {
570 rc
= opal_pci_eeh_freeze_status(phb
->opal_id
,
575 if (rc
!= OPAL_SUCCESS
) {
576 pr_warn("%s: Failure %lld getting "
577 "PHB#%x-PE#%x state\n",
579 phb
->hose
->global_number
, slave
->pe_number
);
580 return OPAL_EEH_STOPPED_TEMP_UNAVAIL
;
584 * Override the result based on the ascending
594 /* Currently those 2 are only used when MSIs are enabled, this will change
595 * but in the meantime, we need to protect them to avoid warnings
597 #ifdef CONFIG_PCI_MSI
598 struct pnv_ioda_pe
*pnv_ioda_get_pe(struct pci_dev
*dev
)
600 struct pci_controller
*hose
= pci_bus_to_host(dev
->bus
);
601 struct pnv_phb
*phb
= hose
->private_data
;
602 struct pci_dn
*pdn
= pci_get_pdn(dev
);
606 if (pdn
->pe_number
== IODA_INVALID_PE
)
608 return &phb
->ioda
.pe_array
[pdn
->pe_number
];
610 #endif /* CONFIG_PCI_MSI */
612 static int pnv_ioda_set_one_peltv(struct pnv_phb
*phb
,
613 struct pnv_ioda_pe
*parent
,
614 struct pnv_ioda_pe
*child
,
617 const char *desc
= is_add
? "adding" : "removing";
618 uint8_t op
= is_add
? OPAL_ADD_PE_TO_DOMAIN
:
619 OPAL_REMOVE_PE_FROM_DOMAIN
;
620 struct pnv_ioda_pe
*slave
;
623 /* Parent PE affects child PE */
624 rc
= opal_pci_set_peltv(phb
->opal_id
, parent
->pe_number
,
625 child
->pe_number
, op
);
626 if (rc
!= OPAL_SUCCESS
) {
627 pe_warn(child
, "OPAL error %ld %s to parent PELTV\n",
632 if (!(child
->flags
& PNV_IODA_PE_MASTER
))
635 /* Compound case: parent PE affects slave PEs */
636 list_for_each_entry(slave
, &child
->slaves
, list
) {
637 rc
= opal_pci_set_peltv(phb
->opal_id
, parent
->pe_number
,
638 slave
->pe_number
, op
);
639 if (rc
!= OPAL_SUCCESS
) {
640 pe_warn(slave
, "OPAL error %ld %s to parent PELTV\n",
649 static int pnv_ioda_set_peltv(struct pnv_phb
*phb
,
650 struct pnv_ioda_pe
*pe
,
653 struct pnv_ioda_pe
*slave
;
654 struct pci_dev
*pdev
= NULL
;
658 * Clear PE frozen state. If it's master PE, we need
659 * clear slave PE frozen state as well.
662 opal_pci_eeh_freeze_clear(phb
->opal_id
, pe
->pe_number
,
663 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
664 if (pe
->flags
& PNV_IODA_PE_MASTER
) {
665 list_for_each_entry(slave
, &pe
->slaves
, list
)
666 opal_pci_eeh_freeze_clear(phb
->opal_id
,
668 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
673 * Associate PE in PELT. We need add the PE into the
674 * corresponding PELT-V as well. Otherwise, the error
675 * originated from the PE might contribute to other
678 ret
= pnv_ioda_set_one_peltv(phb
, pe
, pe
, is_add
);
682 /* For compound PEs, any one affects all of them */
683 if (pe
->flags
& PNV_IODA_PE_MASTER
) {
684 list_for_each_entry(slave
, &pe
->slaves
, list
) {
685 ret
= pnv_ioda_set_one_peltv(phb
, slave
, pe
, is_add
);
691 if (pe
->flags
& (PNV_IODA_PE_BUS_ALL
| PNV_IODA_PE_BUS
))
692 pdev
= pe
->pbus
->self
;
693 else if (pe
->flags
& PNV_IODA_PE_DEV
)
694 pdev
= pe
->pdev
->bus
->self
;
695 #ifdef CONFIG_PCI_IOV
696 else if (pe
->flags
& PNV_IODA_PE_VF
)
697 pdev
= pe
->parent_dev
;
698 #endif /* CONFIG_PCI_IOV */
700 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
701 struct pnv_ioda_pe
*parent
;
703 if (pdn
&& pdn
->pe_number
!= IODA_INVALID_PE
) {
704 parent
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
705 ret
= pnv_ioda_set_one_peltv(phb
, parent
, pe
, is_add
);
710 pdev
= pdev
->bus
->self
;
716 static int pnv_ioda_deconfigure_pe(struct pnv_phb
*phb
, struct pnv_ioda_pe
*pe
)
718 struct pci_dev
*parent
;
719 uint8_t bcomp
, dcomp
, fcomp
;
723 /* Currently, we just deconfigure VF PE. Bus PE will always there.*/
727 dcomp
= OPAL_IGNORE_RID_DEVICE_NUMBER
;
728 fcomp
= OPAL_IGNORE_RID_FUNCTION_NUMBER
;
729 parent
= pe
->pbus
->self
;
730 if (pe
->flags
& PNV_IODA_PE_BUS_ALL
)
731 count
= pe
->pbus
->busn_res
.end
- pe
->pbus
->busn_res
.start
+ 1;
736 case 1: bcomp
= OpalPciBusAll
; break;
737 case 2: bcomp
= OpalPciBus7Bits
; break;
738 case 4: bcomp
= OpalPciBus6Bits
; break;
739 case 8: bcomp
= OpalPciBus5Bits
; break;
740 case 16: bcomp
= OpalPciBus4Bits
; break;
741 case 32: bcomp
= OpalPciBus3Bits
; break;
743 dev_err(&pe
->pbus
->dev
, "Number of subordinate buses %d unsupported\n",
745 /* Do an exact match only */
746 bcomp
= OpalPciBusAll
;
748 rid_end
= pe
->rid
+ (count
<< 8);
750 #ifdef CONFIG_PCI_IOV
751 if (pe
->flags
& PNV_IODA_PE_VF
)
752 parent
= pe
->parent_dev
;
755 parent
= pe
->pdev
->bus
->self
;
756 bcomp
= OpalPciBusAll
;
757 dcomp
= OPAL_COMPARE_RID_DEVICE_NUMBER
;
758 fcomp
= OPAL_COMPARE_RID_FUNCTION_NUMBER
;
759 rid_end
= pe
->rid
+ 1;
762 /* Clear the reverse map */
763 for (rid
= pe
->rid
; rid
< rid_end
; rid
++)
764 phb
->ioda
.pe_rmap
[rid
] = IODA_INVALID_PE
;
766 /* Release from all parents PELT-V */
768 struct pci_dn
*pdn
= pci_get_pdn(parent
);
769 if (pdn
&& pdn
->pe_number
!= IODA_INVALID_PE
) {
770 rc
= opal_pci_set_peltv(phb
->opal_id
, pdn
->pe_number
,
771 pe
->pe_number
, OPAL_REMOVE_PE_FROM_DOMAIN
);
772 /* XXX What to do in case of error ? */
774 parent
= parent
->bus
->self
;
777 opal_pci_eeh_freeze_clear(phb
->opal_id
, pe
->pe_number
,
778 OPAL_EEH_ACTION_CLEAR_FREEZE_ALL
);
780 /* Disassociate PE in PELT */
781 rc
= opal_pci_set_peltv(phb
->opal_id
, pe
->pe_number
,
782 pe
->pe_number
, OPAL_REMOVE_PE_FROM_DOMAIN
);
784 pe_warn(pe
, "OPAL error %ld remove self from PELTV\n", rc
);
785 rc
= opal_pci_set_pe(phb
->opal_id
, pe
->pe_number
, pe
->rid
,
786 bcomp
, dcomp
, fcomp
, OPAL_UNMAP_PE
);
788 pe_err(pe
, "OPAL error %ld trying to setup PELT table\n", rc
);
792 #ifdef CONFIG_PCI_IOV
793 pe
->parent_dev
= NULL
;
799 static int pnv_ioda_configure_pe(struct pnv_phb
*phb
, struct pnv_ioda_pe
*pe
)
801 struct pci_dev
*parent
;
802 uint8_t bcomp
, dcomp
, fcomp
;
803 long rc
, rid_end
, rid
;
805 /* Bus validation ? */
809 dcomp
= OPAL_IGNORE_RID_DEVICE_NUMBER
;
810 fcomp
= OPAL_IGNORE_RID_FUNCTION_NUMBER
;
811 parent
= pe
->pbus
->self
;
812 if (pe
->flags
& PNV_IODA_PE_BUS_ALL
)
813 count
= pe
->pbus
->busn_res
.end
- pe
->pbus
->busn_res
.start
+ 1;
818 case 1: bcomp
= OpalPciBusAll
; break;
819 case 2: bcomp
= OpalPciBus7Bits
; break;
820 case 4: bcomp
= OpalPciBus6Bits
; break;
821 case 8: bcomp
= OpalPciBus5Bits
; break;
822 case 16: bcomp
= OpalPciBus4Bits
; break;
823 case 32: bcomp
= OpalPciBus3Bits
; break;
825 dev_err(&pe
->pbus
->dev
, "Number of subordinate buses %d unsupported\n",
827 /* Do an exact match only */
828 bcomp
= OpalPciBusAll
;
830 rid_end
= pe
->rid
+ (count
<< 8);
832 #ifdef CONFIG_PCI_IOV
833 if (pe
->flags
& PNV_IODA_PE_VF
)
834 parent
= pe
->parent_dev
;
836 #endif /* CONFIG_PCI_IOV */
837 parent
= pe
->pdev
->bus
->self
;
838 bcomp
= OpalPciBusAll
;
839 dcomp
= OPAL_COMPARE_RID_DEVICE_NUMBER
;
840 fcomp
= OPAL_COMPARE_RID_FUNCTION_NUMBER
;
841 rid_end
= pe
->rid
+ 1;
845 * Associate PE in PELT. We need add the PE into the
846 * corresponding PELT-V as well. Otherwise, the error
847 * originated from the PE might contribute to other
850 rc
= opal_pci_set_pe(phb
->opal_id
, pe
->pe_number
, pe
->rid
,
851 bcomp
, dcomp
, fcomp
, OPAL_MAP_PE
);
853 pe_err(pe
, "OPAL error %ld trying to setup PELT table\n", rc
);
858 * Configure PELTV. NPUs don't have a PELTV table so skip
859 * configuration on them.
861 if (phb
->type
!= PNV_PHB_NPU
)
862 pnv_ioda_set_peltv(phb
, pe
, true);
864 /* Setup reverse map */
865 for (rid
= pe
->rid
; rid
< rid_end
; rid
++)
866 phb
->ioda
.pe_rmap
[rid
] = pe
->pe_number
;
868 /* Setup one MVTs on IODA1 */
869 if (phb
->type
!= PNV_PHB_IODA1
) {
874 pe
->mve_number
= pe
->pe_number
;
875 rc
= opal_pci_set_mve(phb
->opal_id
, pe
->mve_number
, pe
->pe_number
);
876 if (rc
!= OPAL_SUCCESS
) {
877 pe_err(pe
, "OPAL error %ld setting up MVE %d\n",
881 rc
= opal_pci_set_mve_enable(phb
->opal_id
,
882 pe
->mve_number
, OPAL_ENABLE_MVE
);
884 pe_err(pe
, "OPAL error %ld enabling MVE %d\n",
894 #ifdef CONFIG_PCI_IOV
895 static int pnv_pci_vf_resource_shift(struct pci_dev
*dev
, int offset
)
897 struct pci_dn
*pdn
= pci_get_pdn(dev
);
899 struct resource
*res
, res2
;
900 resource_size_t size
;
907 * "offset" is in VFs. The M64 windows are sized so that when they
908 * are segmented, each segment is the same size as the IOV BAR.
909 * Each segment is in a separate PE, and the high order bits of the
910 * address are the PE number. Therefore, each VF's BAR is in a
911 * separate PE, and changing the IOV BAR start address changes the
912 * range of PEs the VFs are in.
914 num_vfs
= pdn
->num_vfs
;
915 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
916 res
= &dev
->resource
[i
+ PCI_IOV_RESOURCES
];
917 if (!res
->flags
|| !res
->parent
)
921 * The actual IOV BAR range is determined by the start address
922 * and the actual size for num_vfs VFs BAR. This check is to
923 * make sure that after shifting, the range will not overlap
924 * with another device.
926 size
= pci_iov_resource_size(dev
, i
+ PCI_IOV_RESOURCES
);
927 res2
.flags
= res
->flags
;
928 res2
.start
= res
->start
+ (size
* offset
);
929 res2
.end
= res2
.start
+ (size
* num_vfs
) - 1;
931 if (res2
.end
> res
->end
) {
932 dev_err(&dev
->dev
, "VF BAR%d: %pR would extend past %pR (trying to enable %d VFs shifted by %d)\n",
933 i
, &res2
, res
, num_vfs
, offset
);
939 * After doing so, there would be a "hole" in the /proc/iomem when
940 * offset is a positive value. It looks like the device return some
941 * mmio back to the system, which actually no one could use it.
943 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
944 res
= &dev
->resource
[i
+ PCI_IOV_RESOURCES
];
945 if (!res
->flags
|| !res
->parent
)
948 size
= pci_iov_resource_size(dev
, i
+ PCI_IOV_RESOURCES
);
950 res
->start
+= size
* offset
;
952 dev_info(&dev
->dev
, "VF BAR%d: %pR shifted to %pR (%sabling %d VFs shifted by %d)\n",
953 i
, &res2
, res
, (offset
> 0) ? "En" : "Dis",
955 pci_update_resource(dev
, i
+ PCI_IOV_RESOURCES
);
959 #endif /* CONFIG_PCI_IOV */
961 static struct pnv_ioda_pe
*pnv_ioda_setup_dev_PE(struct pci_dev
*dev
)
963 struct pci_controller
*hose
= pci_bus_to_host(dev
->bus
);
964 struct pnv_phb
*phb
= hose
->private_data
;
965 struct pci_dn
*pdn
= pci_get_pdn(dev
);
966 struct pnv_ioda_pe
*pe
;
969 pr_err("%s: Device tree node not associated properly\n",
973 if (pdn
->pe_number
!= IODA_INVALID_PE
)
976 pe
= pnv_ioda_alloc_pe(phb
);
978 pr_warning("%s: Not enough PE# available, disabling device\n",
983 /* NOTE: We get only one ref to the pci_dev for the pdn, not for the
984 * pointer in the PE data structure, both should be destroyed at the
985 * same time. However, this needs to be looked at more closely again
986 * once we actually start removing things (Hotplug, SR-IOV, ...)
988 * At some point we want to remove the PDN completely anyways
992 pdn
->pe_number
= pe
->pe_number
;
993 pe
->flags
= PNV_IODA_PE_DEV
;
997 pe
->rid
= dev
->bus
->number
<< 8 | pdn
->devfn
;
999 pe_info(pe
, "Associated device to PE\n");
1001 if (pnv_ioda_configure_pe(phb
, pe
)) {
1002 /* XXX What do we do here ? */
1003 pnv_ioda_free_pe(pe
);
1004 pdn
->pe_number
= IODA_INVALID_PE
;
1010 /* Put PE to the list */
1011 list_add_tail(&pe
->list
, &phb
->ioda
.pe_list
);
1016 static void pnv_ioda_setup_same_PE(struct pci_bus
*bus
, struct pnv_ioda_pe
*pe
)
1018 struct pci_dev
*dev
;
1020 list_for_each_entry(dev
, &bus
->devices
, bus_list
) {
1021 struct pci_dn
*pdn
= pci_get_pdn(dev
);
1024 pr_warn("%s: No device node associated with device !\n",
1030 * In partial hotplug case, the PCI device might be still
1031 * associated with the PE and needn't attach it to the PE
1034 if (pdn
->pe_number
!= IODA_INVALID_PE
)
1039 pdn
->pe_number
= pe
->pe_number
;
1040 if ((pe
->flags
& PNV_IODA_PE_BUS_ALL
) && dev
->subordinate
)
1041 pnv_ioda_setup_same_PE(dev
->subordinate
, pe
);
1046 * There're 2 types of PCI bus sensitive PEs: One that is compromised of
1047 * single PCI bus. Another one that contains the primary PCI bus and its
1048 * subordinate PCI devices and buses. The second type of PE is normally
1049 * orgiriated by PCIe-to-PCI bridge or PLX switch downstream ports.
1051 static struct pnv_ioda_pe
*pnv_ioda_setup_bus_PE(struct pci_bus
*bus
, bool all
)
1053 struct pci_controller
*hose
= pci_bus_to_host(bus
);
1054 struct pnv_phb
*phb
= hose
->private_data
;
1055 struct pnv_ioda_pe
*pe
= NULL
;
1056 unsigned int pe_num
;
1059 * In partial hotplug case, the PE instance might be still alive.
1060 * We should reuse it instead of allocating a new one.
1062 pe_num
= phb
->ioda
.pe_rmap
[bus
->number
<< 8];
1063 if (pe_num
!= IODA_INVALID_PE
) {
1064 pe
= &phb
->ioda
.pe_array
[pe_num
];
1065 pnv_ioda_setup_same_PE(bus
, pe
);
1069 /* PE number for root bus should have been reserved */
1070 if (pci_is_root_bus(bus
) &&
1071 phb
->ioda
.root_pe_idx
!= IODA_INVALID_PE
)
1072 pe
= &phb
->ioda
.pe_array
[phb
->ioda
.root_pe_idx
];
1074 /* Check if PE is determined by M64 */
1075 if (!pe
&& phb
->pick_m64_pe
)
1076 pe
= phb
->pick_m64_pe(bus
, all
);
1078 /* The PE number isn't pinned by M64 */
1080 pe
= pnv_ioda_alloc_pe(phb
);
1083 pr_warning("%s: Not enough PE# available for PCI bus %04x:%02x\n",
1084 __func__
, pci_domain_nr(bus
), bus
->number
);
1088 pe
->flags
|= (all
? PNV_IODA_PE_BUS_ALL
: PNV_IODA_PE_BUS
);
1091 pe
->mve_number
= -1;
1092 pe
->rid
= bus
->busn_res
.start
<< 8;
1095 pe_info(pe
, "Secondary bus %d..%d associated with PE#%d\n",
1096 bus
->busn_res
.start
, bus
->busn_res
.end
, pe
->pe_number
);
1098 pe_info(pe
, "Secondary bus %d associated with PE#%d\n",
1099 bus
->busn_res
.start
, pe
->pe_number
);
1101 if (pnv_ioda_configure_pe(phb
, pe
)) {
1102 /* XXX What do we do here ? */
1103 pnv_ioda_free_pe(pe
);
1108 /* Associate it with all child devices */
1109 pnv_ioda_setup_same_PE(bus
, pe
);
1111 /* Put PE to the list */
1112 list_add_tail(&pe
->list
, &phb
->ioda
.pe_list
);
1117 static struct pnv_ioda_pe
*pnv_ioda_setup_npu_PE(struct pci_dev
*npu_pdev
)
1119 int pe_num
, found_pe
= false, rc
;
1121 struct pnv_ioda_pe
*pe
;
1122 struct pci_dev
*gpu_pdev
;
1123 struct pci_dn
*npu_pdn
;
1124 struct pci_controller
*hose
= pci_bus_to_host(npu_pdev
->bus
);
1125 struct pnv_phb
*phb
= hose
->private_data
;
1128 * Due to a hardware errata PE#0 on the NPU is reserved for
1129 * error handling. This means we only have three PEs remaining
1130 * which need to be assigned to four links, implying some
1131 * links must share PEs.
1133 * To achieve this we assign PEs such that NPUs linking the
1134 * same GPU get assigned the same PE.
1136 gpu_pdev
= pnv_pci_get_gpu_dev(npu_pdev
);
1137 for (pe_num
= 0; pe_num
< phb
->ioda
.total_pe_num
; pe_num
++) {
1138 pe
= &phb
->ioda
.pe_array
[pe_num
];
1142 if (pnv_pci_get_gpu_dev(pe
->pdev
) == gpu_pdev
) {
1144 * This device has the same peer GPU so should
1145 * be assigned the same PE as the existing
1148 dev_info(&npu_pdev
->dev
,
1149 "Associating to existing PE %d\n", pe_num
);
1150 pci_dev_get(npu_pdev
);
1151 npu_pdn
= pci_get_pdn(npu_pdev
);
1152 rid
= npu_pdev
->bus
->number
<< 8 | npu_pdn
->devfn
;
1153 npu_pdn
->pcidev
= npu_pdev
;
1154 npu_pdn
->pe_number
= pe_num
;
1155 phb
->ioda
.pe_rmap
[rid
] = pe
->pe_number
;
1157 /* Map the PE to this link */
1158 rc
= opal_pci_set_pe(phb
->opal_id
, pe_num
, rid
,
1160 OPAL_COMPARE_RID_DEVICE_NUMBER
,
1161 OPAL_COMPARE_RID_FUNCTION_NUMBER
,
1163 WARN_ON(rc
!= OPAL_SUCCESS
);
1171 * Could not find an existing PE so allocate a new
1174 return pnv_ioda_setup_dev_PE(npu_pdev
);
1179 static void pnv_ioda_setup_npu_PEs(struct pci_bus
*bus
)
1181 struct pci_dev
*pdev
;
1183 list_for_each_entry(pdev
, &bus
->devices
, bus_list
)
1184 pnv_ioda_setup_npu_PE(pdev
);
1187 static void pnv_pci_ioda_setup_PEs(void)
1189 struct pci_controller
*hose
, *tmp
;
1190 struct pnv_phb
*phb
;
1192 list_for_each_entry_safe(hose
, tmp
, &hose_list
, list_node
) {
1193 phb
= hose
->private_data
;
1194 if (phb
->type
== PNV_PHB_NPU
) {
1195 /* PE#0 is needed for error reporting */
1196 pnv_ioda_reserve_pe(phb
, 0);
1197 pnv_ioda_setup_npu_PEs(hose
->bus
);
1202 #ifdef CONFIG_PCI_IOV
1203 static int pnv_pci_vf_release_m64(struct pci_dev
*pdev
, u16 num_vfs
)
1205 struct pci_bus
*bus
;
1206 struct pci_controller
*hose
;
1207 struct pnv_phb
*phb
;
1213 hose
= pci_bus_to_host(bus
);
1214 phb
= hose
->private_data
;
1215 pdn
= pci_get_pdn(pdev
);
1217 if (pdn
->m64_single_mode
)
1222 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++)
1223 for (j
= 0; j
< m64_bars
; j
++) {
1224 if (pdn
->m64_map
[j
][i
] == IODA_INVALID_M64
)
1226 opal_pci_phb_mmio_enable(phb
->opal_id
,
1227 OPAL_M64_WINDOW_TYPE
, pdn
->m64_map
[j
][i
], 0);
1228 clear_bit(pdn
->m64_map
[j
][i
], &phb
->ioda
.m64_bar_alloc
);
1229 pdn
->m64_map
[j
][i
] = IODA_INVALID_M64
;
1232 kfree(pdn
->m64_map
);
1236 static int pnv_pci_vf_assign_m64(struct pci_dev
*pdev
, u16 num_vfs
)
1238 struct pci_bus
*bus
;
1239 struct pci_controller
*hose
;
1240 struct pnv_phb
*phb
;
1243 struct resource
*res
;
1247 resource_size_t size
, start
;
1252 hose
= pci_bus_to_host(bus
);
1253 phb
= hose
->private_data
;
1254 pdn
= pci_get_pdn(pdev
);
1255 total_vfs
= pci_sriov_get_totalvfs(pdev
);
1257 if (pdn
->m64_single_mode
)
1262 pdn
->m64_map
= kmalloc(sizeof(*pdn
->m64_map
) * m64_bars
, GFP_KERNEL
);
1265 /* Initialize the m64_map to IODA_INVALID_M64 */
1266 for (i
= 0; i
< m64_bars
; i
++)
1267 for (j
= 0; j
< PCI_SRIOV_NUM_BARS
; j
++)
1268 pdn
->m64_map
[i
][j
] = IODA_INVALID_M64
;
1271 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
1272 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
1273 if (!res
->flags
|| !res
->parent
)
1276 for (j
= 0; j
< m64_bars
; j
++) {
1278 win
= find_next_zero_bit(&phb
->ioda
.m64_bar_alloc
,
1279 phb
->ioda
.m64_bar_idx
+ 1, 0);
1281 if (win
>= phb
->ioda
.m64_bar_idx
+ 1)
1283 } while (test_and_set_bit(win
, &phb
->ioda
.m64_bar_alloc
));
1285 pdn
->m64_map
[j
][i
] = win
;
1287 if (pdn
->m64_single_mode
) {
1288 size
= pci_iov_resource_size(pdev
,
1289 PCI_IOV_RESOURCES
+ i
);
1290 start
= res
->start
+ size
* j
;
1292 size
= resource_size(res
);
1296 /* Map the M64 here */
1297 if (pdn
->m64_single_mode
) {
1298 pe_num
= pdn
->pe_num_map
[j
];
1299 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
1300 pe_num
, OPAL_M64_WINDOW_TYPE
,
1301 pdn
->m64_map
[j
][i
], 0);
1304 rc
= opal_pci_set_phb_mem_window(phb
->opal_id
,
1305 OPAL_M64_WINDOW_TYPE
,
1312 if (rc
!= OPAL_SUCCESS
) {
1313 dev_err(&pdev
->dev
, "Failed to map M64 window #%d: %lld\n",
1318 if (pdn
->m64_single_mode
)
1319 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
1320 OPAL_M64_WINDOW_TYPE
, pdn
->m64_map
[j
][i
], 2);
1322 rc
= opal_pci_phb_mmio_enable(phb
->opal_id
,
1323 OPAL_M64_WINDOW_TYPE
, pdn
->m64_map
[j
][i
], 1);
1325 if (rc
!= OPAL_SUCCESS
) {
1326 dev_err(&pdev
->dev
, "Failed to enable M64 window #%d: %llx\n",
1335 pnv_pci_vf_release_m64(pdev
, num_vfs
);
1339 static long pnv_pci_ioda2_unset_window(struct iommu_table_group
*table_group
,
1341 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe
*pe
, bool enable
);
1343 static void pnv_pci_ioda2_release_dma_pe(struct pci_dev
*dev
, struct pnv_ioda_pe
*pe
)
1345 struct iommu_table
*tbl
;
1348 tbl
= pe
->table_group
.tables
[0];
1349 rc
= pnv_pci_ioda2_unset_window(&pe
->table_group
, 0);
1351 pe_warn(pe
, "OPAL error %ld release DMA window\n", rc
);
1353 pnv_pci_ioda2_set_bypass(pe
, false);
1354 if (pe
->table_group
.group
) {
1355 iommu_group_put(pe
->table_group
.group
);
1356 BUG_ON(pe
->table_group
.group
);
1358 pnv_pci_ioda2_table_free_pages(tbl
);
1359 iommu_free_table(tbl
, of_node_full_name(dev
->dev
.of_node
));
1362 static void pnv_ioda_release_vf_PE(struct pci_dev
*pdev
)
1364 struct pci_bus
*bus
;
1365 struct pci_controller
*hose
;
1366 struct pnv_phb
*phb
;
1367 struct pnv_ioda_pe
*pe
, *pe_n
;
1371 hose
= pci_bus_to_host(bus
);
1372 phb
= hose
->private_data
;
1373 pdn
= pci_get_pdn(pdev
);
1375 if (!pdev
->is_physfn
)
1378 list_for_each_entry_safe(pe
, pe_n
, &phb
->ioda
.pe_list
, list
) {
1379 if (pe
->parent_dev
!= pdev
)
1382 pnv_pci_ioda2_release_dma_pe(pdev
, pe
);
1384 /* Remove from list */
1385 mutex_lock(&phb
->ioda
.pe_list_mutex
);
1386 list_del(&pe
->list
);
1387 mutex_unlock(&phb
->ioda
.pe_list_mutex
);
1389 pnv_ioda_deconfigure_pe(phb
, pe
);
1391 pnv_ioda_free_pe(pe
);
1395 void pnv_pci_sriov_disable(struct pci_dev
*pdev
)
1397 struct pci_bus
*bus
;
1398 struct pci_controller
*hose
;
1399 struct pnv_phb
*phb
;
1400 struct pnv_ioda_pe
*pe
;
1402 struct pci_sriov
*iov
;
1406 hose
= pci_bus_to_host(bus
);
1407 phb
= hose
->private_data
;
1408 pdn
= pci_get_pdn(pdev
);
1410 num_vfs
= pdn
->num_vfs
;
1412 /* Release VF PEs */
1413 pnv_ioda_release_vf_PE(pdev
);
1415 if (phb
->type
== PNV_PHB_IODA2
) {
1416 if (!pdn
->m64_single_mode
)
1417 pnv_pci_vf_resource_shift(pdev
, -*pdn
->pe_num_map
);
1419 /* Release M64 windows */
1420 pnv_pci_vf_release_m64(pdev
, num_vfs
);
1422 /* Release PE numbers */
1423 if (pdn
->m64_single_mode
) {
1424 for (i
= 0; i
< num_vfs
; i
++) {
1425 if (pdn
->pe_num_map
[i
] == IODA_INVALID_PE
)
1428 pe
= &phb
->ioda
.pe_array
[pdn
->pe_num_map
[i
]];
1429 pnv_ioda_free_pe(pe
);
1432 bitmap_clear(phb
->ioda
.pe_alloc
, *pdn
->pe_num_map
, num_vfs
);
1433 /* Releasing pe_num_map */
1434 kfree(pdn
->pe_num_map
);
1438 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb
*phb
,
1439 struct pnv_ioda_pe
*pe
);
1440 static void pnv_ioda_setup_vf_PE(struct pci_dev
*pdev
, u16 num_vfs
)
1442 struct pci_bus
*bus
;
1443 struct pci_controller
*hose
;
1444 struct pnv_phb
*phb
;
1445 struct pnv_ioda_pe
*pe
;
1451 hose
= pci_bus_to_host(bus
);
1452 phb
= hose
->private_data
;
1453 pdn
= pci_get_pdn(pdev
);
1455 if (!pdev
->is_physfn
)
1458 /* Reserve PE for each VF */
1459 for (vf_index
= 0; vf_index
< num_vfs
; vf_index
++) {
1460 if (pdn
->m64_single_mode
)
1461 pe_num
= pdn
->pe_num_map
[vf_index
];
1463 pe_num
= *pdn
->pe_num_map
+ vf_index
;
1465 pe
= &phb
->ioda
.pe_array
[pe_num
];
1466 pe
->pe_number
= pe_num
;
1468 pe
->flags
= PNV_IODA_PE_VF
;
1470 pe
->parent_dev
= pdev
;
1471 pe
->mve_number
= -1;
1472 pe
->rid
= (pci_iov_virtfn_bus(pdev
, vf_index
) << 8) |
1473 pci_iov_virtfn_devfn(pdev
, vf_index
);
1475 pe_info(pe
, "VF %04d:%02d:%02d.%d associated with PE#%d\n",
1476 hose
->global_number
, pdev
->bus
->number
,
1477 PCI_SLOT(pci_iov_virtfn_devfn(pdev
, vf_index
)),
1478 PCI_FUNC(pci_iov_virtfn_devfn(pdev
, vf_index
)), pe_num
);
1480 if (pnv_ioda_configure_pe(phb
, pe
)) {
1481 /* XXX What do we do here ? */
1482 pnv_ioda_free_pe(pe
);
1487 /* Put PE to the list */
1488 mutex_lock(&phb
->ioda
.pe_list_mutex
);
1489 list_add_tail(&pe
->list
, &phb
->ioda
.pe_list
);
1490 mutex_unlock(&phb
->ioda
.pe_list_mutex
);
1492 pnv_pci_ioda2_setup_dma_pe(phb
, pe
);
1496 int pnv_pci_sriov_enable(struct pci_dev
*pdev
, u16 num_vfs
)
1498 struct pci_bus
*bus
;
1499 struct pci_controller
*hose
;
1500 struct pnv_phb
*phb
;
1501 struct pnv_ioda_pe
*pe
;
1507 hose
= pci_bus_to_host(bus
);
1508 phb
= hose
->private_data
;
1509 pdn
= pci_get_pdn(pdev
);
1511 if (phb
->type
== PNV_PHB_IODA2
) {
1512 if (!pdn
->vfs_expanded
) {
1513 dev_info(&pdev
->dev
, "don't support this SRIOV device"
1514 " with non 64bit-prefetchable IOV BAR\n");
1519 * When M64 BARs functions in Single PE mode, the number of VFs
1520 * could be enabled must be less than the number of M64 BARs.
1522 if (pdn
->m64_single_mode
&& num_vfs
> phb
->ioda
.m64_bar_idx
) {
1523 dev_info(&pdev
->dev
, "Not enough M64 BAR for VFs\n");
1527 /* Allocating pe_num_map */
1528 if (pdn
->m64_single_mode
)
1529 pdn
->pe_num_map
= kmalloc(sizeof(*pdn
->pe_num_map
) * num_vfs
,
1532 pdn
->pe_num_map
= kmalloc(sizeof(*pdn
->pe_num_map
), GFP_KERNEL
);
1534 if (!pdn
->pe_num_map
)
1537 if (pdn
->m64_single_mode
)
1538 for (i
= 0; i
< num_vfs
; i
++)
1539 pdn
->pe_num_map
[i
] = IODA_INVALID_PE
;
1541 /* Calculate available PE for required VFs */
1542 if (pdn
->m64_single_mode
) {
1543 for (i
= 0; i
< num_vfs
; i
++) {
1544 pe
= pnv_ioda_alloc_pe(phb
);
1550 pdn
->pe_num_map
[i
] = pe
->pe_number
;
1553 mutex_lock(&phb
->ioda
.pe_alloc_mutex
);
1554 *pdn
->pe_num_map
= bitmap_find_next_zero_area(
1555 phb
->ioda
.pe_alloc
, phb
->ioda
.total_pe_num
,
1557 if (*pdn
->pe_num_map
>= phb
->ioda
.total_pe_num
) {
1558 mutex_unlock(&phb
->ioda
.pe_alloc_mutex
);
1559 dev_info(&pdev
->dev
, "Failed to enable VF%d\n", num_vfs
);
1560 kfree(pdn
->pe_num_map
);
1563 bitmap_set(phb
->ioda
.pe_alloc
, *pdn
->pe_num_map
, num_vfs
);
1564 mutex_unlock(&phb
->ioda
.pe_alloc_mutex
);
1566 pdn
->num_vfs
= num_vfs
;
1568 /* Assign M64 window accordingly */
1569 ret
= pnv_pci_vf_assign_m64(pdev
, num_vfs
);
1571 dev_info(&pdev
->dev
, "Not enough M64 window resources\n");
1576 * When using one M64 BAR to map one IOV BAR, we need to shift
1577 * the IOV BAR according to the PE# allocated to the VFs.
1578 * Otherwise, the PE# for the VF will conflict with others.
1580 if (!pdn
->m64_single_mode
) {
1581 ret
= pnv_pci_vf_resource_shift(pdev
, *pdn
->pe_num_map
);
1588 pnv_ioda_setup_vf_PE(pdev
, num_vfs
);
1593 if (pdn
->m64_single_mode
) {
1594 for (i
= 0; i
< num_vfs
; i
++) {
1595 if (pdn
->pe_num_map
[i
] == IODA_INVALID_PE
)
1598 pe
= &phb
->ioda
.pe_array
[pdn
->pe_num_map
[i
]];
1599 pnv_ioda_free_pe(pe
);
1602 bitmap_clear(phb
->ioda
.pe_alloc
, *pdn
->pe_num_map
, num_vfs
);
1604 /* Releasing pe_num_map */
1605 kfree(pdn
->pe_num_map
);
1610 int pcibios_sriov_disable(struct pci_dev
*pdev
)
1612 pnv_pci_sriov_disable(pdev
);
1614 /* Release PCI data */
1615 remove_dev_pci_data(pdev
);
1619 int pcibios_sriov_enable(struct pci_dev
*pdev
, u16 num_vfs
)
1621 /* Allocate PCI data */
1622 add_dev_pci_data(pdev
);
1624 return pnv_pci_sriov_enable(pdev
, num_vfs
);
1626 #endif /* CONFIG_PCI_IOV */
1628 static void pnv_pci_ioda_dma_dev_setup(struct pnv_phb
*phb
, struct pci_dev
*pdev
)
1630 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
1631 struct pnv_ioda_pe
*pe
;
1634 * The function can be called while the PE#
1635 * hasn't been assigned. Do nothing for the
1638 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
1641 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
1642 WARN_ON(get_dma_ops(&pdev
->dev
) != &dma_iommu_ops
);
1643 set_dma_offset(&pdev
->dev
, pe
->tce_bypass_base
);
1644 set_iommu_table_base(&pdev
->dev
, pe
->table_group
.tables
[0]);
1646 * Note: iommu_add_device() will fail here as
1647 * for physical PE: the device is already added by now;
1648 * for virtual PE: sysfs entries are not ready yet and
1649 * tce_iommu_bus_notifier will add the device to a group later.
1653 static int pnv_pci_ioda_dma_set_mask(struct pci_dev
*pdev
, u64 dma_mask
)
1655 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
1656 struct pnv_phb
*phb
= hose
->private_data
;
1657 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
1658 struct pnv_ioda_pe
*pe
;
1660 bool bypass
= false;
1662 if (WARN_ON(!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
))
1665 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
1666 if (pe
->tce_bypass_enabled
) {
1667 top
= pe
->tce_bypass_base
+ memblock_end_of_DRAM() - 1;
1668 bypass
= (dma_mask
>= top
);
1672 dev_info(&pdev
->dev
, "Using 64-bit DMA iommu bypass\n");
1673 set_dma_ops(&pdev
->dev
, &dma_direct_ops
);
1675 dev_info(&pdev
->dev
, "Using 32-bit DMA via iommu\n");
1676 set_dma_ops(&pdev
->dev
, &dma_iommu_ops
);
1678 *pdev
->dev
.dma_mask
= dma_mask
;
1680 /* Update peer npu devices */
1681 pnv_npu_try_dma_set_bypass(pdev
, bypass
);
1686 static u64
pnv_pci_ioda_dma_get_required_mask(struct pci_dev
*pdev
)
1688 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
1689 struct pnv_phb
*phb
= hose
->private_data
;
1690 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
1691 struct pnv_ioda_pe
*pe
;
1694 if (WARN_ON(!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
))
1697 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
1698 if (!pe
->tce_bypass_enabled
)
1699 return __dma_get_required_mask(&pdev
->dev
);
1702 end
= pe
->tce_bypass_base
+ memblock_end_of_DRAM();
1703 mask
= 1ULL << (fls64(end
) - 1);
1709 static void pnv_ioda_setup_bus_dma(struct pnv_ioda_pe
*pe
,
1710 struct pci_bus
*bus
)
1712 struct pci_dev
*dev
;
1714 list_for_each_entry(dev
, &bus
->devices
, bus_list
) {
1715 set_iommu_table_base(&dev
->dev
, pe
->table_group
.tables
[0]);
1716 set_dma_offset(&dev
->dev
, pe
->tce_bypass_base
);
1717 iommu_add_device(&dev
->dev
);
1719 if ((pe
->flags
& PNV_IODA_PE_BUS_ALL
) && dev
->subordinate
)
1720 pnv_ioda_setup_bus_dma(pe
, dev
->subordinate
);
1724 static void pnv_pci_ioda1_tce_invalidate(struct iommu_table
*tbl
,
1725 unsigned long index
, unsigned long npages
, bool rm
)
1727 struct iommu_table_group_link
*tgl
= list_first_entry_or_null(
1728 &tbl
->it_group_list
, struct iommu_table_group_link
,
1730 struct pnv_ioda_pe
*pe
= container_of(tgl
->table_group
,
1731 struct pnv_ioda_pe
, table_group
);
1732 __be64 __iomem
*invalidate
= rm
?
1733 (__be64 __iomem
*)pe
->phb
->ioda
.tce_inval_reg_phys
:
1734 pe
->phb
->ioda
.tce_inval_reg
;
1735 unsigned long start
, end
, inc
;
1736 const unsigned shift
= tbl
->it_page_shift
;
1738 start
= __pa(((__be64
*)tbl
->it_base
) + index
- tbl
->it_offset
);
1739 end
= __pa(((__be64
*)tbl
->it_base
) + index
- tbl
->it_offset
+
1742 /* BML uses this case for p6/p7/galaxy2: Shift addr and put in node */
1743 if (tbl
->it_busno
) {
1746 inc
= 128ull << shift
;
1747 start
|= tbl
->it_busno
;
1748 end
|= tbl
->it_busno
;
1749 } else if (tbl
->it_type
& TCE_PCI_SWINV_PAIR
) {
1750 /* p7ioc-style invalidation, 2 TCEs per write */
1751 start
|= (1ull << 63);
1752 end
|= (1ull << 63);
1755 /* Default (older HW) */
1759 end
|= inc
- 1; /* round up end to be different than start */
1761 mb(); /* Ensure above stores are visible */
1762 while (start
<= end
) {
1764 __raw_rm_writeq(cpu_to_be64(start
), invalidate
);
1766 __raw_writeq(cpu_to_be64(start
), invalidate
);
1771 * The iommu layer will do another mb() for us on build()
1772 * and we don't care on free()
1776 static int pnv_ioda1_tce_build(struct iommu_table
*tbl
, long index
,
1777 long npages
, unsigned long uaddr
,
1778 enum dma_data_direction direction
,
1779 struct dma_attrs
*attrs
)
1781 int ret
= pnv_tce_build(tbl
, index
, npages
, uaddr
, direction
,
1784 if (!ret
&& (tbl
->it_type
& TCE_PCI_SWINV_CREATE
))
1785 pnv_pci_ioda1_tce_invalidate(tbl
, index
, npages
, false);
1790 #ifdef CONFIG_IOMMU_API
1791 static int pnv_ioda1_tce_xchg(struct iommu_table
*tbl
, long index
,
1792 unsigned long *hpa
, enum dma_data_direction
*direction
)
1794 long ret
= pnv_tce_xchg(tbl
, index
, hpa
, direction
);
1796 if (!ret
&& (tbl
->it_type
&
1797 (TCE_PCI_SWINV_CREATE
| TCE_PCI_SWINV_FREE
)))
1798 pnv_pci_ioda1_tce_invalidate(tbl
, index
, 1, false);
1804 static void pnv_ioda1_tce_free(struct iommu_table
*tbl
, long index
,
1807 pnv_tce_free(tbl
, index
, npages
);
1809 if (tbl
->it_type
& TCE_PCI_SWINV_FREE
)
1810 pnv_pci_ioda1_tce_invalidate(tbl
, index
, npages
, false);
1813 static struct iommu_table_ops pnv_ioda1_iommu_ops
= {
1814 .set
= pnv_ioda1_tce_build
,
1815 #ifdef CONFIG_IOMMU_API
1816 .exchange
= pnv_ioda1_tce_xchg
,
1818 .clear
= pnv_ioda1_tce_free
,
1822 #define TCE_KILL_INVAL_ALL PPC_BIT(0)
1823 #define TCE_KILL_INVAL_PE PPC_BIT(1)
1824 #define TCE_KILL_INVAL_TCE PPC_BIT(2)
1826 void pnv_pci_ioda2_tce_invalidate_entire(struct pnv_phb
*phb
, bool rm
)
1828 const unsigned long val
= TCE_KILL_INVAL_ALL
;
1830 mb(); /* Ensure previous TCE table stores are visible */
1832 __raw_rm_writeq(cpu_to_be64(val
),
1834 phb
->ioda
.tce_inval_reg_phys
);
1836 __raw_writeq(cpu_to_be64(val
), phb
->ioda
.tce_inval_reg
);
1839 static inline void pnv_pci_ioda2_tce_invalidate_pe(struct pnv_ioda_pe
*pe
)
1841 /* 01xb - invalidate TCEs that match the specified PE# */
1842 unsigned long val
= TCE_KILL_INVAL_PE
| (pe
->pe_number
& 0xFF);
1843 struct pnv_phb
*phb
= pe
->phb
;
1845 if (!phb
->ioda
.tce_inval_reg
)
1848 mb(); /* Ensure above stores are visible */
1849 __raw_writeq(cpu_to_be64(val
), phb
->ioda
.tce_inval_reg
);
1852 static void pnv_pci_ioda2_do_tce_invalidate(unsigned pe_number
, bool rm
,
1853 __be64 __iomem
*invalidate
, unsigned shift
,
1854 unsigned long index
, unsigned long npages
)
1856 unsigned long start
, end
, inc
;
1858 /* We'll invalidate DMA address in PE scope */
1859 start
= TCE_KILL_INVAL_TCE
;
1860 start
|= (pe_number
& 0xFF);
1863 /* Figure out the start, end and step */
1864 start
|= (index
<< shift
);
1865 end
|= ((index
+ npages
- 1) << shift
);
1866 inc
= (0x1ull
<< shift
);
1869 while (start
<= end
) {
1871 __raw_rm_writeq(cpu_to_be64(start
), invalidate
);
1873 __raw_writeq(cpu_to_be64(start
), invalidate
);
1878 static void pnv_pci_ioda2_tce_invalidate(struct iommu_table
*tbl
,
1879 unsigned long index
, unsigned long npages
, bool rm
)
1881 struct iommu_table_group_link
*tgl
;
1883 list_for_each_entry_rcu(tgl
, &tbl
->it_group_list
, next
) {
1884 struct pnv_ioda_pe
*pe
= container_of(tgl
->table_group
,
1885 struct pnv_ioda_pe
, table_group
);
1886 __be64 __iomem
*invalidate
= rm
?
1887 (__be64 __iomem
*)pe
->phb
->ioda
.tce_inval_reg_phys
:
1888 pe
->phb
->ioda
.tce_inval_reg
;
1890 if (pe
->phb
->type
== PNV_PHB_NPU
) {
1892 * The NVLink hardware does not support TCE kill
1893 * per TCE entry so we have to invalidate
1894 * the entire cache for it.
1896 pnv_pci_ioda2_tce_invalidate_entire(pe
->phb
, rm
);
1899 pnv_pci_ioda2_do_tce_invalidate(pe
->pe_number
, rm
,
1900 invalidate
, tbl
->it_page_shift
,
1905 static int pnv_ioda2_tce_build(struct iommu_table
*tbl
, long index
,
1906 long npages
, unsigned long uaddr
,
1907 enum dma_data_direction direction
,
1908 struct dma_attrs
*attrs
)
1910 int ret
= pnv_tce_build(tbl
, index
, npages
, uaddr
, direction
,
1913 if (!ret
&& (tbl
->it_type
& TCE_PCI_SWINV_CREATE
))
1914 pnv_pci_ioda2_tce_invalidate(tbl
, index
, npages
, false);
1919 #ifdef CONFIG_IOMMU_API
1920 static int pnv_ioda2_tce_xchg(struct iommu_table
*tbl
, long index
,
1921 unsigned long *hpa
, enum dma_data_direction
*direction
)
1923 long ret
= pnv_tce_xchg(tbl
, index
, hpa
, direction
);
1925 if (!ret
&& (tbl
->it_type
&
1926 (TCE_PCI_SWINV_CREATE
| TCE_PCI_SWINV_FREE
)))
1927 pnv_pci_ioda2_tce_invalidate(tbl
, index
, 1, false);
1933 static void pnv_ioda2_tce_free(struct iommu_table
*tbl
, long index
,
1936 pnv_tce_free(tbl
, index
, npages
);
1938 if (tbl
->it_type
& TCE_PCI_SWINV_FREE
)
1939 pnv_pci_ioda2_tce_invalidate(tbl
, index
, npages
, false);
1942 static void pnv_ioda2_table_free(struct iommu_table
*tbl
)
1944 pnv_pci_ioda2_table_free_pages(tbl
);
1945 iommu_free_table(tbl
, "pnv");
1948 static struct iommu_table_ops pnv_ioda2_iommu_ops
= {
1949 .set
= pnv_ioda2_tce_build
,
1950 #ifdef CONFIG_IOMMU_API
1951 .exchange
= pnv_ioda2_tce_xchg
,
1953 .clear
= pnv_ioda2_tce_free
,
1955 .free
= pnv_ioda2_table_free
,
1958 static int pnv_pci_ioda_dev_dma_weight(struct pci_dev
*dev
, void *data
)
1960 unsigned int *weight
= (unsigned int *)data
;
1962 /* This is quite simplistic. The "base" weight of a device
1963 * is 10. 0 means no DMA is to be accounted for it.
1965 if (dev
->hdr_type
!= PCI_HEADER_TYPE_NORMAL
)
1968 if (dev
->class == PCI_CLASS_SERIAL_USB_UHCI
||
1969 dev
->class == PCI_CLASS_SERIAL_USB_OHCI
||
1970 dev
->class == PCI_CLASS_SERIAL_USB_EHCI
)
1972 else if ((dev
->class >> 8) == PCI_CLASS_STORAGE_RAID
)
1980 static unsigned int pnv_pci_ioda_pe_dma_weight(struct pnv_ioda_pe
*pe
)
1982 unsigned int weight
= 0;
1984 /* SRIOV VF has same DMA32 weight as its PF */
1985 #ifdef CONFIG_PCI_IOV
1986 if ((pe
->flags
& PNV_IODA_PE_VF
) && pe
->parent_dev
) {
1987 pnv_pci_ioda_dev_dma_weight(pe
->parent_dev
, &weight
);
1992 if ((pe
->flags
& PNV_IODA_PE_DEV
) && pe
->pdev
) {
1993 pnv_pci_ioda_dev_dma_weight(pe
->pdev
, &weight
);
1994 } else if ((pe
->flags
& PNV_IODA_PE_BUS
) && pe
->pbus
) {
1995 struct pci_dev
*pdev
;
1997 list_for_each_entry(pdev
, &pe
->pbus
->devices
, bus_list
)
1998 pnv_pci_ioda_dev_dma_weight(pdev
, &weight
);
1999 } else if ((pe
->flags
& PNV_IODA_PE_BUS_ALL
) && pe
->pbus
) {
2000 pci_walk_bus(pe
->pbus
, pnv_pci_ioda_dev_dma_weight
, &weight
);
2006 static void pnv_pci_ioda1_setup_dma_pe(struct pnv_phb
*phb
,
2007 struct pnv_ioda_pe
*pe
)
2010 struct page
*tce_mem
= NULL
;
2011 struct iommu_table
*tbl
;
2012 unsigned int weight
, total_weight
= 0;
2013 unsigned int tce32_segsz
, base
, segs
, avail
, i
;
2017 /* XXX FIXME: Handle 64-bit only DMA devices */
2018 /* XXX FIXME: Provide 64-bit DMA facilities & non-4K TCE tables etc.. */
2019 /* XXX FIXME: Allocate multi-level tables on PHB3 */
2020 weight
= pnv_pci_ioda_pe_dma_weight(pe
);
2024 pci_walk_bus(phb
->hose
->bus
, pnv_pci_ioda_dev_dma_weight
,
2026 segs
= (weight
* phb
->ioda
.dma32_count
) / total_weight
;
2031 * Allocate contiguous DMA32 segments. We begin with the expected
2032 * number of segments. With one more attempt, the number of DMA32
2033 * segments to be allocated is decreased by one until one segment
2034 * is allocated successfully.
2037 for (base
= 0; base
<= phb
->ioda
.dma32_count
- segs
; base
++) {
2038 for (avail
= 0, i
= base
; i
< base
+ segs
; i
++) {
2039 if (phb
->ioda
.dma32_segmap
[i
] ==
2050 pe_warn(pe
, "No available DMA32 segments\n");
2055 tbl
= pnv_pci_table_alloc(phb
->hose
->node
);
2056 iommu_register_group(&pe
->table_group
, phb
->hose
->global_number
,
2058 pnv_pci_link_table_and_group(phb
->hose
->node
, 0, tbl
, &pe
->table_group
);
2060 /* Grab a 32-bit TCE table */
2061 pe_info(pe
, "DMA weight %d (%d), assigned (%d) %d DMA32 segments\n",
2062 weight
, total_weight
, base
, segs
);
2063 pe_info(pe
, " Setting up 32-bit TCE table at %08x..%08x\n",
2064 base
* PNV_IODA1_DMA32_SEGSIZE
,
2065 (base
+ segs
) * PNV_IODA1_DMA32_SEGSIZE
- 1);
2067 /* XXX Currently, we allocate one big contiguous table for the
2068 * TCEs. We only really need one chunk per 256M of TCE space
2069 * (ie per segment) but that's an optimization for later, it
2070 * requires some added smarts with our get/put_tce implementation
2072 * Each TCE page is 4KB in size and each TCE entry occupies 8
2075 tce32_segsz
= PNV_IODA1_DMA32_SEGSIZE
>> (IOMMU_PAGE_SHIFT_4K
- 3);
2076 tce_mem
= alloc_pages_node(phb
->hose
->node
, GFP_KERNEL
,
2077 get_order(tce32_segsz
* segs
));
2079 pe_err(pe
, " Failed to allocate a 32-bit TCE memory\n");
2082 addr
= page_address(tce_mem
);
2083 memset(addr
, 0, tce32_segsz
* segs
);
2086 for (i
= 0; i
< segs
; i
++) {
2087 rc
= opal_pci_map_pe_dma_window(phb
->opal_id
,
2090 __pa(addr
) + tce32_segsz
* i
,
2091 tce32_segsz
, IOMMU_PAGE_SIZE_4K
);
2093 pe_err(pe
, " Failed to configure 32-bit TCE table,"
2099 /* Setup DMA32 segment mapping */
2100 for (i
= base
; i
< base
+ segs
; i
++)
2101 phb
->ioda
.dma32_segmap
[i
] = pe
->pe_number
;
2103 /* Setup linux iommu table */
2104 pnv_pci_setup_iommu_table(tbl
, addr
, tce32_segsz
* segs
,
2105 base
* PNV_IODA1_DMA32_SEGSIZE
,
2106 IOMMU_PAGE_SHIFT_4K
);
2108 /* OPAL variant of P7IOC SW invalidated TCEs */
2109 if (phb
->ioda
.tce_inval_reg
)
2110 tbl
->it_type
|= (TCE_PCI_SWINV_CREATE
|
2111 TCE_PCI_SWINV_FREE
|
2112 TCE_PCI_SWINV_PAIR
);
2114 tbl
->it_ops
= &pnv_ioda1_iommu_ops
;
2115 pe
->table_group
.tce32_start
= tbl
->it_offset
<< tbl
->it_page_shift
;
2116 pe
->table_group
.tce32_size
= tbl
->it_size
<< tbl
->it_page_shift
;
2117 iommu_init_table(tbl
, phb
->hose
->node
);
2119 if (pe
->flags
& PNV_IODA_PE_DEV
) {
2121 * Setting table base here only for carrying iommu_group
2122 * further down to let iommu_add_device() do the job.
2123 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2125 set_iommu_table_base(&pe
->pdev
->dev
, tbl
);
2126 iommu_add_device(&pe
->pdev
->dev
);
2127 } else if (pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
))
2128 pnv_ioda_setup_bus_dma(pe
, pe
->pbus
);
2132 /* XXX Failure: Try to fallback to 64-bit only ? */
2134 __free_pages(tce_mem
, get_order(tce32_segsz
* segs
));
2136 pnv_pci_unlink_table_and_group(tbl
, &pe
->table_group
);
2137 iommu_free_table(tbl
, "pnv");
2141 static long pnv_pci_ioda2_set_window(struct iommu_table_group
*table_group
,
2142 int num
, struct iommu_table
*tbl
)
2144 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2146 struct pnv_phb
*phb
= pe
->phb
;
2148 const unsigned long size
= tbl
->it_indirect_levels
?
2149 tbl
->it_level_size
: tbl
->it_size
;
2150 const __u64 start_addr
= tbl
->it_offset
<< tbl
->it_page_shift
;
2151 const __u64 win_size
= tbl
->it_size
<< tbl
->it_page_shift
;
2153 pe_info(pe
, "Setting up window#%d %llx..%llx pg=%x\n", num
,
2154 start_addr
, start_addr
+ win_size
- 1,
2155 IOMMU_PAGE_SIZE(tbl
));
2158 * Map TCE table through TVT. The TVE index is the PE number
2159 * shifted by 1 bit for 32-bits DMA space.
2161 rc
= opal_pci_map_pe_dma_window(phb
->opal_id
,
2163 (pe
->pe_number
<< 1) + num
,
2164 tbl
->it_indirect_levels
+ 1,
2167 IOMMU_PAGE_SIZE(tbl
));
2169 pe_err(pe
, "Failed to configure TCE table, err %ld\n", rc
);
2173 pnv_pci_link_table_and_group(phb
->hose
->node
, num
,
2174 tbl
, &pe
->table_group
);
2175 pnv_pci_ioda2_tce_invalidate_pe(pe
);
2180 static void pnv_pci_ioda2_set_bypass(struct pnv_ioda_pe
*pe
, bool enable
)
2182 uint16_t window_id
= (pe
->pe_number
<< 1 ) + 1;
2185 pe_info(pe
, "%sabling 64-bit DMA bypass\n", enable
? "En" : "Dis");
2187 phys_addr_t top
= memblock_end_of_DRAM();
2189 top
= roundup_pow_of_two(top
);
2190 rc
= opal_pci_map_pe_dma_window_real(pe
->phb
->opal_id
,
2193 pe
->tce_bypass_base
,
2196 rc
= opal_pci_map_pe_dma_window_real(pe
->phb
->opal_id
,
2199 pe
->tce_bypass_base
,
2203 pe_err(pe
, "OPAL error %lld configuring bypass window\n", rc
);
2205 pe
->tce_bypass_enabled
= enable
;
2208 static long pnv_pci_ioda2_table_alloc_pages(int nid
, __u64 bus_offset
,
2209 __u32 page_shift
, __u64 window_size
, __u32 levels
,
2210 struct iommu_table
*tbl
);
2212 static long pnv_pci_ioda2_create_table(struct iommu_table_group
*table_group
,
2213 int num
, __u32 page_shift
, __u64 window_size
, __u32 levels
,
2214 struct iommu_table
**ptbl
)
2216 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2218 int nid
= pe
->phb
->hose
->node
;
2219 __u64 bus_offset
= num
? pe
->tce_bypass_base
: table_group
->tce32_start
;
2221 struct iommu_table
*tbl
;
2223 tbl
= pnv_pci_table_alloc(nid
);
2227 ret
= pnv_pci_ioda2_table_alloc_pages(nid
,
2228 bus_offset
, page_shift
, window_size
,
2231 iommu_free_table(tbl
, "pnv");
2235 tbl
->it_ops
= &pnv_ioda2_iommu_ops
;
2236 if (pe
->phb
->ioda
.tce_inval_reg
)
2237 tbl
->it_type
|= (TCE_PCI_SWINV_CREATE
| TCE_PCI_SWINV_FREE
);
2244 static long pnv_pci_ioda2_setup_default_config(struct pnv_ioda_pe
*pe
)
2246 struct iommu_table
*tbl
= NULL
;
2250 * crashkernel= specifies the kdump kernel's maximum memory at
2251 * some offset and there is no guaranteed the result is a power
2252 * of 2, which will cause errors later.
2254 const u64 max_memory
= __rounddown_pow_of_two(memory_hotplug_max());
2257 * In memory constrained environments, e.g. kdump kernel, the
2258 * DMA window can be larger than available memory, which will
2259 * cause errors later.
2261 const u64 window_size
= min((u64
)pe
->table_group
.tce32_size
, max_memory
);
2263 rc
= pnv_pci_ioda2_create_table(&pe
->table_group
, 0,
2264 IOMMU_PAGE_SHIFT_4K
,
2266 POWERNV_IOMMU_DEFAULT_LEVELS
, &tbl
);
2268 pe_err(pe
, "Failed to create 32-bit TCE table, err %ld",
2273 iommu_init_table(tbl
, pe
->phb
->hose
->node
);
2275 rc
= pnv_pci_ioda2_set_window(&pe
->table_group
, 0, tbl
);
2277 pe_err(pe
, "Failed to configure 32-bit TCE table, err %ld\n",
2279 pnv_ioda2_table_free(tbl
);
2283 if (!pnv_iommu_bypass_disabled
)
2284 pnv_pci_ioda2_set_bypass(pe
, true);
2286 /* OPAL variant of PHB3 invalidated TCEs */
2287 if (pe
->phb
->ioda
.tce_inval_reg
)
2288 tbl
->it_type
|= (TCE_PCI_SWINV_CREATE
| TCE_PCI_SWINV_FREE
);
2291 * Setting table base here only for carrying iommu_group
2292 * further down to let iommu_add_device() do the job.
2293 * pnv_pci_ioda_dma_dev_setup will override it later anyway.
2295 if (pe
->flags
& PNV_IODA_PE_DEV
)
2296 set_iommu_table_base(&pe
->pdev
->dev
, tbl
);
2301 #if defined(CONFIG_IOMMU_API) || defined(CONFIG_PCI_IOV)
2302 static long pnv_pci_ioda2_unset_window(struct iommu_table_group
*table_group
,
2305 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2307 struct pnv_phb
*phb
= pe
->phb
;
2310 pe_info(pe
, "Removing DMA window #%d\n", num
);
2312 ret
= opal_pci_map_pe_dma_window(phb
->opal_id
, pe
->pe_number
,
2313 (pe
->pe_number
<< 1) + num
,
2314 0/* levels */, 0/* table address */,
2315 0/* table size */, 0/* page size */);
2317 pe_warn(pe
, "Unmapping failed, ret = %ld\n", ret
);
2319 pnv_pci_ioda2_tce_invalidate_pe(pe
);
2321 pnv_pci_unlink_table_and_group(table_group
->tables
[num
], table_group
);
2327 #ifdef CONFIG_IOMMU_API
2328 static unsigned long pnv_pci_ioda2_get_table_size(__u32 page_shift
,
2329 __u64 window_size
, __u32 levels
)
2331 unsigned long bytes
= 0;
2332 const unsigned window_shift
= ilog2(window_size
);
2333 unsigned entries_shift
= window_shift
- page_shift
;
2334 unsigned table_shift
= entries_shift
+ 3;
2335 unsigned long tce_table_size
= max(0x1000UL
, 1UL << table_shift
);
2336 unsigned long direct_table_size
;
2338 if (!levels
|| (levels
> POWERNV_IOMMU_MAX_LEVELS
) ||
2339 (window_size
> memory_hotplug_max()) ||
2340 !is_power_of_2(window_size
))
2343 /* Calculate a direct table size from window_size and levels */
2344 entries_shift
= (entries_shift
+ levels
- 1) / levels
;
2345 table_shift
= entries_shift
+ 3;
2346 table_shift
= max_t(unsigned, table_shift
, PAGE_SHIFT
);
2347 direct_table_size
= 1UL << table_shift
;
2349 for ( ; levels
; --levels
) {
2350 bytes
+= _ALIGN_UP(tce_table_size
, direct_table_size
);
2352 tce_table_size
/= direct_table_size
;
2353 tce_table_size
<<= 3;
2354 tce_table_size
= _ALIGN_UP(tce_table_size
, direct_table_size
);
2360 static void pnv_ioda2_take_ownership(struct iommu_table_group
*table_group
)
2362 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2364 /* Store @tbl as pnv_pci_ioda2_unset_window() resets it */
2365 struct iommu_table
*tbl
= pe
->table_group
.tables
[0];
2367 pnv_pci_ioda2_set_bypass(pe
, false);
2368 pnv_pci_ioda2_unset_window(&pe
->table_group
, 0);
2369 pnv_ioda2_table_free(tbl
);
2372 static void pnv_ioda2_release_ownership(struct iommu_table_group
*table_group
)
2374 struct pnv_ioda_pe
*pe
= container_of(table_group
, struct pnv_ioda_pe
,
2377 pnv_pci_ioda2_setup_default_config(pe
);
2380 static struct iommu_table_group_ops pnv_pci_ioda2_ops
= {
2381 .get_table_size
= pnv_pci_ioda2_get_table_size
,
2382 .create_table
= pnv_pci_ioda2_create_table
,
2383 .set_window
= pnv_pci_ioda2_set_window
,
2384 .unset_window
= pnv_pci_ioda2_unset_window
,
2385 .take_ownership
= pnv_ioda2_take_ownership
,
2386 .release_ownership
= pnv_ioda2_release_ownership
,
2389 static int gpe_table_group_to_npe_cb(struct device
*dev
, void *opaque
)
2391 struct pci_controller
*hose
;
2392 struct pnv_phb
*phb
;
2393 struct pnv_ioda_pe
**ptmppe
= opaque
;
2394 struct pci_dev
*pdev
= container_of(dev
, struct pci_dev
, dev
);
2395 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
2397 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
2400 hose
= pci_bus_to_host(pdev
->bus
);
2401 phb
= hose
->private_data
;
2402 if (phb
->type
!= PNV_PHB_NPU
)
2405 *ptmppe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
2411 * This returns PE of associated NPU.
2412 * This assumes that NPU is in the same IOMMU group with GPU and there is
2415 static struct pnv_ioda_pe
*gpe_table_group_to_npe(
2416 struct iommu_table_group
*table_group
)
2418 struct pnv_ioda_pe
*npe
= NULL
;
2419 int ret
= iommu_group_for_each_dev(table_group
->group
, &npe
,
2420 gpe_table_group_to_npe_cb
);
2422 BUG_ON(!ret
|| !npe
);
2427 static long pnv_pci_ioda2_npu_set_window(struct iommu_table_group
*table_group
,
2428 int num
, struct iommu_table
*tbl
)
2430 long ret
= pnv_pci_ioda2_set_window(table_group
, num
, tbl
);
2435 ret
= pnv_npu_set_window(gpe_table_group_to_npe(table_group
), num
, tbl
);
2437 pnv_pci_ioda2_unset_window(table_group
, num
);
2442 static long pnv_pci_ioda2_npu_unset_window(
2443 struct iommu_table_group
*table_group
,
2446 long ret
= pnv_pci_ioda2_unset_window(table_group
, num
);
2451 return pnv_npu_unset_window(gpe_table_group_to_npe(table_group
), num
);
2454 static void pnv_ioda2_npu_take_ownership(struct iommu_table_group
*table_group
)
2457 * Detach NPU first as pnv_ioda2_take_ownership() will destroy
2458 * the iommu_table if 32bit DMA is enabled.
2460 pnv_npu_take_ownership(gpe_table_group_to_npe(table_group
));
2461 pnv_ioda2_take_ownership(table_group
);
2464 static struct iommu_table_group_ops pnv_pci_ioda2_npu_ops
= {
2465 .get_table_size
= pnv_pci_ioda2_get_table_size
,
2466 .create_table
= pnv_pci_ioda2_create_table
,
2467 .set_window
= pnv_pci_ioda2_npu_set_window
,
2468 .unset_window
= pnv_pci_ioda2_npu_unset_window
,
2469 .take_ownership
= pnv_ioda2_npu_take_ownership
,
2470 .release_ownership
= pnv_ioda2_release_ownership
,
2473 static void pnv_pci_ioda_setup_iommu_api(void)
2475 struct pci_controller
*hose
, *tmp
;
2476 struct pnv_phb
*phb
;
2477 struct pnv_ioda_pe
*pe
, *gpe
;
2480 * Now we have all PHBs discovered, time to add NPU devices to
2481 * the corresponding IOMMU groups.
2483 list_for_each_entry_safe(hose
, tmp
, &hose_list
, list_node
) {
2484 phb
= hose
->private_data
;
2486 if (phb
->type
!= PNV_PHB_NPU
)
2489 list_for_each_entry(pe
, &phb
->ioda
.pe_list
, list
) {
2490 gpe
= pnv_pci_npu_setup_iommu(pe
);
2492 gpe
->table_group
.ops
= &pnv_pci_ioda2_npu_ops
;
2496 #else /* !CONFIG_IOMMU_API */
2497 static void pnv_pci_ioda_setup_iommu_api(void) { };
2500 static void pnv_pci_ioda_setup_opal_tce_kill(struct pnv_phb
*phb
)
2502 const __be64
*swinvp
;
2504 /* OPAL variant of PHB3 invalidated TCEs */
2505 swinvp
= of_get_property(phb
->hose
->dn
, "ibm,opal-tce-kill", NULL
);
2509 phb
->ioda
.tce_inval_reg_phys
= be64_to_cpup(swinvp
);
2510 phb
->ioda
.tce_inval_reg
= ioremap(phb
->ioda
.tce_inval_reg_phys
, 8);
2513 static __be64
*pnv_pci_ioda2_table_do_alloc_pages(int nid
, unsigned shift
,
2514 unsigned levels
, unsigned long limit
,
2515 unsigned long *current_offset
, unsigned long *total_allocated
)
2517 struct page
*tce_mem
= NULL
;
2519 unsigned order
= max_t(unsigned, shift
, PAGE_SHIFT
) - PAGE_SHIFT
;
2520 unsigned long allocated
= 1UL << (order
+ PAGE_SHIFT
);
2521 unsigned entries
= 1UL << (shift
- 3);
2524 tce_mem
= alloc_pages_node(nid
, GFP_KERNEL
, order
);
2526 pr_err("Failed to allocate a TCE memory, order=%d\n", order
);
2529 addr
= page_address(tce_mem
);
2530 memset(addr
, 0, allocated
);
2531 *total_allocated
+= allocated
;
2535 *current_offset
+= allocated
;
2539 for (i
= 0; i
< entries
; ++i
) {
2540 tmp
= pnv_pci_ioda2_table_do_alloc_pages(nid
, shift
,
2541 levels
, limit
, current_offset
, total_allocated
);
2545 addr
[i
] = cpu_to_be64(__pa(tmp
) |
2546 TCE_PCI_READ
| TCE_PCI_WRITE
);
2548 if (*current_offset
>= limit
)
2555 static void pnv_pci_ioda2_table_do_free_pages(__be64
*addr
,
2556 unsigned long size
, unsigned level
);
2558 static long pnv_pci_ioda2_table_alloc_pages(int nid
, __u64 bus_offset
,
2559 __u32 page_shift
, __u64 window_size
, __u32 levels
,
2560 struct iommu_table
*tbl
)
2563 unsigned long offset
= 0, level_shift
, total_allocated
= 0;
2564 const unsigned window_shift
= ilog2(window_size
);
2565 unsigned entries_shift
= window_shift
- page_shift
;
2566 unsigned table_shift
= max_t(unsigned, entries_shift
+ 3, PAGE_SHIFT
);
2567 const unsigned long tce_table_size
= 1UL << table_shift
;
2569 if (!levels
|| (levels
> POWERNV_IOMMU_MAX_LEVELS
))
2572 if ((window_size
> memory_hotplug_max()) || !is_power_of_2(window_size
))
2575 /* Adjust direct table size from window_size and levels */
2576 entries_shift
= (entries_shift
+ levels
- 1) / levels
;
2577 level_shift
= entries_shift
+ 3;
2578 level_shift
= max_t(unsigned, level_shift
, PAGE_SHIFT
);
2580 /* Allocate TCE table */
2581 addr
= pnv_pci_ioda2_table_do_alloc_pages(nid
, level_shift
,
2582 levels
, tce_table_size
, &offset
, &total_allocated
);
2584 /* addr==NULL means that the first level allocation failed */
2589 * First level was allocated but some lower level failed as
2590 * we did not allocate as much as we wanted,
2591 * release partially allocated table.
2593 if (offset
< tce_table_size
) {
2594 pnv_pci_ioda2_table_do_free_pages(addr
,
2595 1ULL << (level_shift
- 3), levels
- 1);
2599 /* Setup linux iommu table */
2600 pnv_pci_setup_iommu_table(tbl
, addr
, tce_table_size
, bus_offset
,
2602 tbl
->it_level_size
= 1ULL << (level_shift
- 3);
2603 tbl
->it_indirect_levels
= levels
- 1;
2604 tbl
->it_allocated_size
= total_allocated
;
2606 pr_devel("Created TCE table: ws=%08llx ts=%lx @%08llx\n",
2607 window_size
, tce_table_size
, bus_offset
);
2612 static void pnv_pci_ioda2_table_do_free_pages(__be64
*addr
,
2613 unsigned long size
, unsigned level
)
2615 const unsigned long addr_ul
= (unsigned long) addr
&
2616 ~(TCE_PCI_READ
| TCE_PCI_WRITE
);
2620 u64
*tmp
= (u64
*) addr_ul
;
2622 for (i
= 0; i
< size
; ++i
) {
2623 unsigned long hpa
= be64_to_cpu(tmp
[i
]);
2625 if (!(hpa
& (TCE_PCI_READ
| TCE_PCI_WRITE
)))
2628 pnv_pci_ioda2_table_do_free_pages(__va(hpa
), size
,
2633 free_pages(addr_ul
, get_order(size
<< 3));
2636 static void pnv_pci_ioda2_table_free_pages(struct iommu_table
*tbl
)
2638 const unsigned long size
= tbl
->it_indirect_levels
?
2639 tbl
->it_level_size
: tbl
->it_size
;
2644 pnv_pci_ioda2_table_do_free_pages((__be64
*)tbl
->it_base
, size
,
2645 tbl
->it_indirect_levels
);
2648 static void pnv_pci_ioda2_setup_dma_pe(struct pnv_phb
*phb
,
2649 struct pnv_ioda_pe
*pe
)
2653 if (!pnv_pci_ioda_pe_dma_weight(pe
))
2656 /* TVE #1 is selected by PCI address bit 59 */
2657 pe
->tce_bypass_base
= 1ull << 59;
2659 iommu_register_group(&pe
->table_group
, phb
->hose
->global_number
,
2662 /* The PE will reserve all possible 32-bits space */
2663 pe_info(pe
, "Setting up 32-bit TCE table at 0..%08x\n",
2664 phb
->ioda
.m32_pci_base
);
2666 /* Setup linux iommu table */
2667 pe
->table_group
.tce32_start
= 0;
2668 pe
->table_group
.tce32_size
= phb
->ioda
.m32_pci_base
;
2669 pe
->table_group
.max_dynamic_windows_supported
=
2670 IOMMU_TABLE_GROUP_MAX_TABLES
;
2671 pe
->table_group
.max_levels
= POWERNV_IOMMU_MAX_LEVELS
;
2672 pe
->table_group
.pgsizes
= SZ_4K
| SZ_64K
| SZ_16M
;
2673 #ifdef CONFIG_IOMMU_API
2674 pe
->table_group
.ops
= &pnv_pci_ioda2_ops
;
2677 rc
= pnv_pci_ioda2_setup_default_config(pe
);
2681 if (pe
->flags
& PNV_IODA_PE_DEV
)
2682 iommu_add_device(&pe
->pdev
->dev
);
2683 else if (pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
))
2684 pnv_ioda_setup_bus_dma(pe
, pe
->pbus
);
2687 #ifdef CONFIG_PCI_MSI
2688 static void pnv_ioda2_msi_eoi(struct irq_data
*d
)
2690 unsigned int hw_irq
= (unsigned int)irqd_to_hwirq(d
);
2691 struct irq_chip
*chip
= irq_data_get_irq_chip(d
);
2692 struct pnv_phb
*phb
= container_of(chip
, struct pnv_phb
,
2696 rc
= opal_pci_msi_eoi(phb
->opal_id
, hw_irq
);
2703 void pnv_set_msi_irq_chip(struct pnv_phb
*phb
, unsigned int virq
)
2705 struct irq_data
*idata
;
2706 struct irq_chip
*ichip
;
2708 if (phb
->type
!= PNV_PHB_IODA2
)
2711 if (!phb
->ioda
.irq_chip_init
) {
2713 * First time we setup an MSI IRQ, we need to setup the
2714 * corresponding IRQ chip to route correctly.
2716 idata
= irq_get_irq_data(virq
);
2717 ichip
= irq_data_get_irq_chip(idata
);
2718 phb
->ioda
.irq_chip_init
= 1;
2719 phb
->ioda
.irq_chip
= *ichip
;
2720 phb
->ioda
.irq_chip
.irq_eoi
= pnv_ioda2_msi_eoi
;
2722 irq_set_chip(virq
, &phb
->ioda
.irq_chip
);
2725 static int pnv_pci_ioda_msi_setup(struct pnv_phb
*phb
, struct pci_dev
*dev
,
2726 unsigned int hwirq
, unsigned int virq
,
2727 unsigned int is_64
, struct msi_msg
*msg
)
2729 struct pnv_ioda_pe
*pe
= pnv_ioda_get_pe(dev
);
2730 unsigned int xive_num
= hwirq
- phb
->msi_base
;
2734 /* No PE assigned ? bail out ... no MSI for you ! */
2738 /* Check if we have an MVE */
2739 if (pe
->mve_number
< 0)
2742 /* Force 32-bit MSI on some broken devices */
2743 if (dev
->no_64bit_msi
)
2746 /* Assign XIVE to PE */
2747 rc
= opal_pci_set_xive_pe(phb
->opal_id
, pe
->pe_number
, xive_num
);
2749 pr_warn("%s: OPAL error %d setting XIVE %d PE\n",
2750 pci_name(dev
), rc
, xive_num
);
2757 rc
= opal_get_msi_64(phb
->opal_id
, pe
->mve_number
, xive_num
, 1,
2760 pr_warn("%s: OPAL error %d getting 64-bit MSI data\n",
2764 msg
->address_hi
= be64_to_cpu(addr64
) >> 32;
2765 msg
->address_lo
= be64_to_cpu(addr64
) & 0xfffffffful
;
2769 rc
= opal_get_msi_32(phb
->opal_id
, pe
->mve_number
, xive_num
, 1,
2772 pr_warn("%s: OPAL error %d getting 32-bit MSI data\n",
2776 msg
->address_hi
= 0;
2777 msg
->address_lo
= be32_to_cpu(addr32
);
2779 msg
->data
= be32_to_cpu(data
);
2781 pnv_set_msi_irq_chip(phb
, virq
);
2783 pr_devel("%s: %s-bit MSI on hwirq %x (xive #%d),"
2784 " address=%x_%08x data=%x PE# %d\n",
2785 pci_name(dev
), is_64
? "64" : "32", hwirq
, xive_num
,
2786 msg
->address_hi
, msg
->address_lo
, data
, pe
->pe_number
);
2791 static void pnv_pci_init_ioda_msis(struct pnv_phb
*phb
)
2794 const __be32
*prop
= of_get_property(phb
->hose
->dn
,
2795 "ibm,opal-msi-ranges", NULL
);
2798 prop
= of_get_property(phb
->hose
->dn
, "msi-ranges", NULL
);
2803 phb
->msi_base
= be32_to_cpup(prop
);
2804 count
= be32_to_cpup(prop
+ 1);
2805 if (msi_bitmap_alloc(&phb
->msi_bmp
, count
, phb
->hose
->dn
)) {
2806 pr_err("PCI %d: Failed to allocate MSI bitmap !\n",
2807 phb
->hose
->global_number
);
2811 phb
->msi_setup
= pnv_pci_ioda_msi_setup
;
2812 phb
->msi32_support
= 1;
2813 pr_info(" Allocated bitmap for %d MSIs (base IRQ 0x%x)\n",
2814 count
, phb
->msi_base
);
2817 static void pnv_pci_init_ioda_msis(struct pnv_phb
*phb
) { }
2818 #endif /* CONFIG_PCI_MSI */
2820 #ifdef CONFIG_PCI_IOV
2821 static void pnv_pci_ioda_fixup_iov_resources(struct pci_dev
*pdev
)
2823 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
2824 struct pnv_phb
*phb
= hose
->private_data
;
2825 const resource_size_t gate
= phb
->ioda
.m64_segsize
>> 2;
2826 struct resource
*res
;
2828 resource_size_t size
, total_vf_bar_sz
;
2832 if (!pdev
->is_physfn
|| pdev
->is_added
)
2835 pdn
= pci_get_pdn(pdev
);
2836 pdn
->vfs_expanded
= 0;
2837 pdn
->m64_single_mode
= false;
2839 total_vfs
= pci_sriov_get_totalvfs(pdev
);
2840 mul
= phb
->ioda
.total_pe_num
;
2841 total_vf_bar_sz
= 0;
2843 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
2844 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
2845 if (!res
->flags
|| res
->parent
)
2847 if (!pnv_pci_is_mem_pref_64(res
->flags
)) {
2848 dev_warn(&pdev
->dev
, "Don't support SR-IOV with"
2849 " non M64 VF BAR%d: %pR. \n",
2854 total_vf_bar_sz
+= pci_iov_resource_size(pdev
,
2855 i
+ PCI_IOV_RESOURCES
);
2858 * If bigger than quarter of M64 segment size, just round up
2861 * Generally, one M64 BAR maps one IOV BAR. To avoid conflict
2862 * with other devices, IOV BAR size is expanded to be
2863 * (total_pe * VF_BAR_size). When VF_BAR_size is half of M64
2864 * segment size , the expanded size would equal to half of the
2865 * whole M64 space size, which will exhaust the M64 Space and
2866 * limit the system flexibility. This is a design decision to
2867 * set the boundary to quarter of the M64 segment size.
2869 if (total_vf_bar_sz
> gate
) {
2870 mul
= roundup_pow_of_two(total_vfs
);
2871 dev_info(&pdev
->dev
,
2872 "VF BAR Total IOV size %llx > %llx, roundup to %d VFs\n",
2873 total_vf_bar_sz
, gate
, mul
);
2874 pdn
->m64_single_mode
= true;
2879 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
2880 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
2881 if (!res
->flags
|| res
->parent
)
2884 size
= pci_iov_resource_size(pdev
, i
+ PCI_IOV_RESOURCES
);
2886 * On PHB3, the minimum size alignment of M64 BAR in single
2889 if (pdn
->m64_single_mode
&& (size
< SZ_32M
))
2891 dev_dbg(&pdev
->dev
, " Fixing VF BAR%d: %pR to\n", i
, res
);
2892 res
->end
= res
->start
+ size
* mul
- 1;
2893 dev_dbg(&pdev
->dev
, " %pR\n", res
);
2894 dev_info(&pdev
->dev
, "VF BAR%d: %pR (expanded to %d VFs for PE alignment)",
2897 pdn
->vfs_expanded
= mul
;
2902 /* To save MMIO space, IOV BAR is truncated. */
2903 for (i
= 0; i
< PCI_SRIOV_NUM_BARS
; i
++) {
2904 res
= &pdev
->resource
[i
+ PCI_IOV_RESOURCES
];
2906 res
->end
= res
->start
- 1;
2909 #endif /* CONFIG_PCI_IOV */
2911 static void pnv_ioda_setup_pe_res(struct pnv_ioda_pe
*pe
,
2912 struct resource
*res
)
2914 struct pnv_phb
*phb
= pe
->phb
;
2915 struct pci_bus_region region
;
2919 if (!res
|| !res
->flags
|| res
->start
> res
->end
)
2922 if (res
->flags
& IORESOURCE_IO
) {
2923 region
.start
= res
->start
- phb
->ioda
.io_pci_base
;
2924 region
.end
= res
->end
- phb
->ioda
.io_pci_base
;
2925 index
= region
.start
/ phb
->ioda
.io_segsize
;
2927 while (index
< phb
->ioda
.total_pe_num
&&
2928 region
.start
<= region
.end
) {
2929 phb
->ioda
.io_segmap
[index
] = pe
->pe_number
;
2930 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
2931 pe
->pe_number
, OPAL_IO_WINDOW_TYPE
, 0, index
);
2932 if (rc
!= OPAL_SUCCESS
) {
2933 pr_err("%s: Error %lld mapping IO segment#%d to PE#%d\n",
2934 __func__
, rc
, index
, pe
->pe_number
);
2938 region
.start
+= phb
->ioda
.io_segsize
;
2941 } else if ((res
->flags
& IORESOURCE_MEM
) &&
2942 !pnv_pci_is_mem_pref_64(res
->flags
)) {
2943 region
.start
= res
->start
-
2944 phb
->hose
->mem_offset
[0] -
2945 phb
->ioda
.m32_pci_base
;
2946 region
.end
= res
->end
-
2947 phb
->hose
->mem_offset
[0] -
2948 phb
->ioda
.m32_pci_base
;
2949 index
= region
.start
/ phb
->ioda
.m32_segsize
;
2951 while (index
< phb
->ioda
.total_pe_num
&&
2952 region
.start
<= region
.end
) {
2953 phb
->ioda
.m32_segmap
[index
] = pe
->pe_number
;
2954 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
2955 pe
->pe_number
, OPAL_M32_WINDOW_TYPE
, 0, index
);
2956 if (rc
!= OPAL_SUCCESS
) {
2957 pr_err("%s: Error %lld mapping M32 segment#%d to PE#%d",
2958 __func__
, rc
, index
, pe
->pe_number
);
2962 region
.start
+= phb
->ioda
.m32_segsize
;
2969 * This function is supposed to be called on basis of PE from top
2970 * to bottom style. So the the I/O or MMIO segment assigned to
2971 * parent PE could be overrided by its child PEs if necessary.
2973 static void pnv_ioda_setup_pe_seg(struct pnv_ioda_pe
*pe
)
2975 struct pci_dev
*pdev
;
2979 * NOTE: We only care PCI bus based PE for now. For PCI
2980 * device based PE, for example SRIOV sensitive VF should
2981 * be figured out later.
2983 BUG_ON(!(pe
->flags
& (PNV_IODA_PE_BUS
| PNV_IODA_PE_BUS_ALL
)));
2985 list_for_each_entry(pdev
, &pe
->pbus
->devices
, bus_list
) {
2986 for (i
= 0; i
<= PCI_ROM_RESOURCE
; i
++)
2987 pnv_ioda_setup_pe_res(pe
, &pdev
->resource
[i
]);
2990 * If the PE contains all subordinate PCI buses, the
2991 * windows of the child bridges should be mapped to
2994 if (!(pe
->flags
& PNV_IODA_PE_BUS_ALL
) || !pci_is_bridge(pdev
))
2996 for (i
= 0; i
< PCI_BRIDGE_RESOURCE_NUM
; i
++)
2997 pnv_ioda_setup_pe_res(pe
,
2998 &pdev
->resource
[PCI_BRIDGE_RESOURCES
+ i
]);
3002 static void pnv_pci_ioda_create_dbgfs(void)
3004 #ifdef CONFIG_DEBUG_FS
3005 struct pci_controller
*hose
, *tmp
;
3006 struct pnv_phb
*phb
;
3009 list_for_each_entry_safe(hose
, tmp
, &hose_list
, list_node
) {
3010 phb
= hose
->private_data
;
3012 /* Notify initialization of PHB done */
3013 phb
->initialized
= 1;
3015 sprintf(name
, "PCI%04x", hose
->global_number
);
3016 phb
->dbgfs
= debugfs_create_dir(name
, powerpc_debugfs_root
);
3018 pr_warning("%s: Error on creating debugfs on PHB#%x\n",
3019 __func__
, hose
->global_number
);
3021 #endif /* CONFIG_DEBUG_FS */
3024 static void pnv_pci_ioda_fixup(void)
3026 pnv_pci_ioda_setup_PEs();
3027 pnv_pci_ioda_setup_iommu_api();
3028 pnv_pci_ioda_create_dbgfs();
3032 eeh_addr_cache_build();
3037 * Returns the alignment for I/O or memory windows for P2P
3038 * bridges. That actually depends on how PEs are segmented.
3039 * For now, we return I/O or M32 segment size for PE sensitive
3040 * P2P bridges. Otherwise, the default values (4KiB for I/O,
3041 * 1MiB for memory) will be returned.
3043 * The current PCI bus might be put into one PE, which was
3044 * create against the parent PCI bridge. For that case, we
3045 * needn't enlarge the alignment so that we can save some
3048 static resource_size_t
pnv_pci_window_alignment(struct pci_bus
*bus
,
3051 struct pci_dev
*bridge
;
3052 struct pci_controller
*hose
= pci_bus_to_host(bus
);
3053 struct pnv_phb
*phb
= hose
->private_data
;
3054 int num_pci_bridges
= 0;
3058 if (pci_pcie_type(bridge
) == PCI_EXP_TYPE_PCI_BRIDGE
) {
3060 if (num_pci_bridges
>= 2)
3064 bridge
= bridge
->bus
->self
;
3067 /* We fail back to M32 if M64 isn't supported */
3068 if (phb
->ioda
.m64_segsize
&&
3069 pnv_pci_is_mem_pref_64(type
))
3070 return phb
->ioda
.m64_segsize
;
3071 if (type
& IORESOURCE_MEM
)
3072 return phb
->ioda
.m32_segsize
;
3074 return phb
->ioda
.io_segsize
;
3078 * We are updating root port or the upstream port of the
3079 * bridge behind the root port with PHB's windows in order
3080 * to accommodate the changes on required resources during
3081 * PCI (slot) hotplug, which is connected to either root
3082 * port or the downstream ports of PCIe switch behind the
3085 static void pnv_pci_fixup_bridge_resources(struct pci_bus
*bus
,
3088 struct pci_controller
*hose
= pci_bus_to_host(bus
);
3089 struct pnv_phb
*phb
= hose
->private_data
;
3090 struct pci_dev
*bridge
= bus
->self
;
3091 struct resource
*r
, *w
;
3092 bool msi_region
= false;
3095 /* Check if we need apply fixup to the bridge's windows */
3096 if (!pci_is_root_bus(bridge
->bus
) &&
3097 !pci_is_root_bus(bridge
->bus
->self
->bus
))
3100 /* Fixup the resources */
3101 for (i
= 0; i
< PCI_BRIDGE_RESOURCE_NUM
; i
++) {
3102 r
= &bridge
->resource
[PCI_BRIDGE_RESOURCES
+ i
];
3103 if (!r
->flags
|| !r
->parent
)
3107 if (r
->flags
& type
& IORESOURCE_IO
)
3108 w
= &hose
->io_resource
;
3109 else if (pnv_pci_is_mem_pref_64(r
->flags
) &&
3110 (type
& IORESOURCE_PREFETCH
) &&
3111 phb
->ioda
.m64_segsize
)
3112 w
= &hose
->mem_resources
[1];
3113 else if (r
->flags
& type
& IORESOURCE_MEM
) {
3114 w
= &hose
->mem_resources
[0];
3118 r
->start
= w
->start
;
3121 /* The 64KB 32-bits MSI region shouldn't be included in
3122 * the 32-bits bridge window. Otherwise, we can see strange
3123 * issues. One of them is EEH error observed on Garrison.
3125 * Exclude top 1MB region which is the minimal alignment of
3126 * 32-bits bridge window.
3135 static void pnv_pci_setup_bridge(struct pci_bus
*bus
, unsigned long type
)
3137 struct pci_controller
*hose
= pci_bus_to_host(bus
);
3138 struct pnv_phb
*phb
= hose
->private_data
;
3139 struct pci_dev
*bridge
= bus
->self
;
3140 struct pnv_ioda_pe
*pe
;
3141 bool all
= (pci_pcie_type(bridge
) == PCI_EXP_TYPE_PCI_BRIDGE
);
3143 /* Extend bridge's windows if necessary */
3144 pnv_pci_fixup_bridge_resources(bus
, type
);
3146 /* The PE for root bus should be realized before any one else */
3147 if (!phb
->ioda
.root_pe_populated
) {
3148 pe
= pnv_ioda_setup_bus_PE(phb
->hose
->bus
, false);
3150 phb
->ioda
.root_pe_idx
= pe
->pe_number
;
3151 phb
->ioda
.root_pe_populated
= true;
3155 /* Don't assign PE to PCI bus, which doesn't have subordinate devices */
3156 if (list_empty(&bus
->devices
))
3159 /* Reserve PEs according to used M64 resources */
3160 if (phb
->reserve_m64_pe
)
3161 phb
->reserve_m64_pe(bus
, NULL
, all
);
3164 * Assign PE. We might run here because of partial hotplug.
3165 * For the case, we just pick up the existing PE and should
3166 * not allocate resources again.
3168 pe
= pnv_ioda_setup_bus_PE(bus
, all
);
3172 pnv_ioda_setup_pe_seg(pe
);
3173 switch (phb
->type
) {
3175 pnv_pci_ioda1_setup_dma_pe(phb
, pe
);
3178 pnv_pci_ioda2_setup_dma_pe(phb
, pe
);
3181 pr_warn("%s: No DMA for PHB#%d (type %d)\n",
3182 __func__
, phb
->hose
->global_number
, phb
->type
);
3186 #ifdef CONFIG_PCI_IOV
3187 static resource_size_t
pnv_pci_iov_resource_alignment(struct pci_dev
*pdev
,
3190 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
3191 struct pnv_phb
*phb
= hose
->private_data
;
3192 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
3193 resource_size_t align
;
3196 * On PowerNV platform, IOV BAR is mapped by M64 BAR to enable the
3197 * SR-IOV. While from hardware perspective, the range mapped by M64
3198 * BAR should be size aligned.
3200 * When IOV BAR is mapped with M64 BAR in Single PE mode, the extra
3201 * powernv-specific hardware restriction is gone. But if just use the
3202 * VF BAR size as the alignment, PF BAR / VF BAR may be allocated with
3203 * in one segment of M64 #15, which introduces the PE conflict between
3204 * PF and VF. Based on this, the minimum alignment of an IOV BAR is
3207 * This function returns the total IOV BAR size if M64 BAR is in
3208 * Shared PE mode or just VF BAR size if not.
3209 * If the M64 BAR is in Single PE mode, return the VF BAR size or
3210 * M64 segment size if IOV BAR size is less.
3212 align
= pci_iov_resource_size(pdev
, resno
);
3213 if (!pdn
->vfs_expanded
)
3215 if (pdn
->m64_single_mode
)
3216 return max(align
, (resource_size_t
)phb
->ioda
.m64_segsize
);
3218 return pdn
->vfs_expanded
* align
;
3220 #endif /* CONFIG_PCI_IOV */
3222 /* Prevent enabling devices for which we couldn't properly
3225 static bool pnv_pci_enable_device_hook(struct pci_dev
*dev
)
3227 struct pci_controller
*hose
= pci_bus_to_host(dev
->bus
);
3228 struct pnv_phb
*phb
= hose
->private_data
;
3231 /* The function is probably called while the PEs have
3232 * not be created yet. For example, resource reassignment
3233 * during PCI probe period. We just skip the check if
3236 if (!phb
->initialized
)
3239 pdn
= pci_get_pdn(dev
);
3240 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
3246 static long pnv_pci_ioda1_unset_window(struct iommu_table_group
*table_group
,
3249 struct pnv_ioda_pe
*pe
= container_of(table_group
,
3250 struct pnv_ioda_pe
, table_group
);
3251 struct pnv_phb
*phb
= pe
->phb
;
3255 pe_info(pe
, "Removing DMA window #%d\n", num
);
3256 for (idx
= 0; idx
< phb
->ioda
.dma32_count
; idx
++) {
3257 if (phb
->ioda
.dma32_segmap
[idx
] != pe
->pe_number
)
3260 rc
= opal_pci_map_pe_dma_window(phb
->opal_id
, pe
->pe_number
,
3261 idx
, 0, 0ul, 0ul, 0ul);
3262 if (rc
!= OPAL_SUCCESS
) {
3263 pe_warn(pe
, "Failure %ld unmapping DMA32 segment#%d\n",
3268 phb
->ioda
.dma32_segmap
[idx
] = IODA_INVALID_PE
;
3271 pnv_pci_unlink_table_and_group(table_group
->tables
[num
], table_group
);
3272 return OPAL_SUCCESS
;
3275 static void pnv_pci_ioda1_release_pe_dma(struct pnv_ioda_pe
*pe
)
3277 unsigned int weight
= pnv_pci_ioda_pe_dma_weight(pe
);
3278 struct iommu_table
*tbl
= pe
->table_group
.tables
[0];
3284 rc
= pnv_pci_ioda1_unset_window(&pe
->table_group
, 0);
3285 if (rc
!= OPAL_SUCCESS
)
3288 pnv_pci_ioda1_tce_invalidate(tbl
, tbl
->it_offset
, tbl
->it_size
, false);
3289 if (pe
->table_group
.group
) {
3290 iommu_group_put(pe
->table_group
.group
);
3291 WARN_ON(pe
->table_group
.group
);
3294 free_pages(tbl
->it_base
, get_order(tbl
->it_size
<< 3));
3295 iommu_free_table(tbl
, "pnv");
3298 static void pnv_pci_ioda2_release_pe_dma(struct pnv_ioda_pe
*pe
)
3300 struct iommu_table
*tbl
= pe
->table_group
.tables
[0];
3301 unsigned int weight
= pnv_pci_ioda_pe_dma_weight(pe
);
3302 #ifdef CONFIG_IOMMU_API
3309 #ifdef CONFIG_IOMMU_API
3310 rc
= pnv_pci_ioda2_unset_window(&pe
->table_group
, 0);
3312 pe_warn(pe
, "OPAL error %ld release DMA window\n", rc
);
3315 pnv_pci_ioda2_set_bypass(pe
, false);
3316 if (pe
->table_group
.group
) {
3317 iommu_group_put(pe
->table_group
.group
);
3318 WARN_ON(pe
->table_group
.group
);
3321 pnv_pci_ioda2_table_free_pages(tbl
);
3322 iommu_free_table(tbl
, "pnv");
3325 static void pnv_ioda_free_pe_seg(struct pnv_ioda_pe
*pe
,
3329 struct pnv_phb
*phb
= pe
->phb
;
3333 for (idx
= 0; idx
< phb
->ioda
.total_pe_num
; idx
++) {
3334 if (map
[idx
] != pe
->pe_number
)
3337 if (win
== OPAL_M64_WINDOW_TYPE
)
3338 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
3339 phb
->ioda
.reserved_pe_idx
, win
,
3340 idx
/ PNV_IODA1_M64_SEGS
,
3341 idx
% PNV_IODA1_M64_SEGS
);
3343 rc
= opal_pci_map_pe_mmio_window(phb
->opal_id
,
3344 phb
->ioda
.reserved_pe_idx
, win
, 0, idx
);
3346 if (rc
!= OPAL_SUCCESS
)
3347 pe_warn(pe
, "Error %ld unmapping (%d) segment#%d\n",
3350 map
[idx
] = IODA_INVALID_PE
;
3354 static void pnv_ioda_release_pe_seg(struct pnv_ioda_pe
*pe
)
3356 struct pnv_phb
*phb
= pe
->phb
;
3358 if (phb
->type
== PNV_PHB_IODA1
) {
3359 pnv_ioda_free_pe_seg(pe
, OPAL_IO_WINDOW_TYPE
,
3360 phb
->ioda
.io_segmap
);
3361 pnv_ioda_free_pe_seg(pe
, OPAL_M32_WINDOW_TYPE
,
3362 phb
->ioda
.m32_segmap
);
3363 pnv_ioda_free_pe_seg(pe
, OPAL_M64_WINDOW_TYPE
,
3364 phb
->ioda
.m64_segmap
);
3365 } else if (phb
->type
== PNV_PHB_IODA2
) {
3366 pnv_ioda_free_pe_seg(pe
, OPAL_M32_WINDOW_TYPE
,
3367 phb
->ioda
.m32_segmap
);
3371 static void pnv_ioda_release_pe(struct pnv_ioda_pe
*pe
)
3373 struct pnv_phb
*phb
= pe
->phb
;
3374 struct pnv_ioda_pe
*slave
, *tmp
;
3376 /* Release slave PEs in compound PE */
3377 if (pe
->flags
& PNV_IODA_PE_MASTER
) {
3378 list_for_each_entry_safe(slave
, tmp
, &pe
->slaves
, list
)
3379 pnv_ioda_release_pe(slave
);
3382 list_del(&pe
->list
);
3383 switch (phb
->type
) {
3385 pnv_pci_ioda1_release_pe_dma(pe
);
3388 pnv_pci_ioda2_release_pe_dma(pe
);
3394 pnv_ioda_release_pe_seg(pe
);
3395 pnv_ioda_deconfigure_pe(pe
->phb
, pe
);
3396 pnv_ioda_free_pe(pe
);
3399 static void pnv_pci_release_device(struct pci_dev
*pdev
)
3401 struct pci_controller
*hose
= pci_bus_to_host(pdev
->bus
);
3402 struct pnv_phb
*phb
= hose
->private_data
;
3403 struct pci_dn
*pdn
= pci_get_pdn(pdev
);
3404 struct pnv_ioda_pe
*pe
;
3406 if (pdev
->is_virtfn
)
3409 if (!pdn
|| pdn
->pe_number
== IODA_INVALID_PE
)
3412 pe
= &phb
->ioda
.pe_array
[pdn
->pe_number
];
3413 WARN_ON(--pe
->device_count
< 0);
3414 if (pe
->device_count
== 0)
3415 pnv_ioda_release_pe(pe
);
3418 static void pnv_pci_ioda_shutdown(struct pci_controller
*hose
)
3420 struct pnv_phb
*phb
= hose
->private_data
;
3422 opal_pci_reset(phb
->opal_id
, OPAL_RESET_PCI_IODA_TABLE
,
3426 static const struct pci_controller_ops pnv_pci_ioda_controller_ops
= {
3427 .dma_dev_setup
= pnv_pci_dma_dev_setup
,
3428 .dma_bus_setup
= pnv_pci_dma_bus_setup
,
3429 #ifdef CONFIG_PCI_MSI
3430 .setup_msi_irqs
= pnv_setup_msi_irqs
,
3431 .teardown_msi_irqs
= pnv_teardown_msi_irqs
,
3433 .enable_device_hook
= pnv_pci_enable_device_hook
,
3434 .release_device
= pnv_pci_release_device
,
3435 .window_alignment
= pnv_pci_window_alignment
,
3436 .setup_bridge
= pnv_pci_setup_bridge
,
3437 .reset_secondary_bus
= pnv_pci_reset_secondary_bus
,
3438 .dma_set_mask
= pnv_pci_ioda_dma_set_mask
,
3439 .dma_get_required_mask
= pnv_pci_ioda_dma_get_required_mask
,
3440 .shutdown
= pnv_pci_ioda_shutdown
,
3443 static int pnv_npu_dma_set_mask(struct pci_dev
*npdev
, u64 dma_mask
)
3445 dev_err_once(&npdev
->dev
,
3446 "%s operation unsupported for NVLink devices\n",
3451 static const struct pci_controller_ops pnv_npu_ioda_controller_ops
= {
3452 .dma_dev_setup
= pnv_pci_dma_dev_setup
,
3453 #ifdef CONFIG_PCI_MSI
3454 .setup_msi_irqs
= pnv_setup_msi_irqs
,
3455 .teardown_msi_irqs
= pnv_teardown_msi_irqs
,
3457 .enable_device_hook
= pnv_pci_enable_device_hook
,
3458 .window_alignment
= pnv_pci_window_alignment
,
3459 .reset_secondary_bus
= pnv_pci_reset_secondary_bus
,
3460 .dma_set_mask
= pnv_npu_dma_set_mask
,
3461 .shutdown
= pnv_pci_ioda_shutdown
,
3464 static void __init
pnv_pci_init_ioda_phb(struct device_node
*np
,
3465 u64 hub_id
, int ioda_type
)
3467 struct pci_controller
*hose
;
3468 struct pnv_phb
*phb
;
3469 unsigned long size
, m64map_off
, m32map_off
, pemap_off
;
3470 unsigned long iomap_off
= 0, dma32map_off
= 0;
3471 const __be64
*prop64
;
3472 const __be32
*prop32
;
3479 pr_info("Initializing %s PHB (%s)\n",
3480 pnv_phb_names
[ioda_type
], of_node_full_name(np
));
3482 prop64
= of_get_property(np
, "ibm,opal-phbid", NULL
);
3484 pr_err(" Missing \"ibm,opal-phbid\" property !\n");
3487 phb_id
= be64_to_cpup(prop64
);
3488 pr_debug(" PHB-ID : 0x%016llx\n", phb_id
);
3490 phb
= memblock_virt_alloc(sizeof(struct pnv_phb
), 0);
3492 /* Allocate PCI controller */
3493 phb
->hose
= hose
= pcibios_alloc_controller(np
);
3495 pr_err(" Can't allocate PCI controller for %s\n",
3497 memblock_free(__pa(phb
), sizeof(struct pnv_phb
));
3501 spin_lock_init(&phb
->lock
);
3502 prop32
= of_get_property(np
, "bus-range", &len
);
3503 if (prop32
&& len
== 8) {
3504 hose
->first_busno
= be32_to_cpu(prop32
[0]);
3505 hose
->last_busno
= be32_to_cpu(prop32
[1]);
3507 pr_warn(" Broken <bus-range> on %s\n", np
->full_name
);
3508 hose
->first_busno
= 0;
3509 hose
->last_busno
= 0xff;
3511 hose
->private_data
= phb
;
3512 phb
->hub_id
= hub_id
;
3513 phb
->opal_id
= phb_id
;
3514 phb
->type
= ioda_type
;
3515 mutex_init(&phb
->ioda
.pe_alloc_mutex
);
3517 /* Detect specific models for error handling */
3518 if (of_device_is_compatible(np
, "ibm,p7ioc-pciex"))
3519 phb
->model
= PNV_PHB_MODEL_P7IOC
;
3520 else if (of_device_is_compatible(np
, "ibm,power8-pciex"))
3521 phb
->model
= PNV_PHB_MODEL_PHB3
;
3522 else if (of_device_is_compatible(np
, "ibm,power8-npu-pciex"))
3523 phb
->model
= PNV_PHB_MODEL_NPU
;
3525 phb
->model
= PNV_PHB_MODEL_UNKNOWN
;
3527 /* Parse 32-bit and IO ranges (if any) */
3528 pci_process_bridge_OF_ranges(hose
, np
, !hose
->global_number
);
3531 phb
->regs
= of_iomap(np
, 0);
3532 if (phb
->regs
== NULL
)
3533 pr_err(" Failed to map registers !\n");
3535 /* Initialize TCE kill register */
3536 pnv_pci_ioda_setup_opal_tce_kill(phb
);
3538 /* Initialize more IODA stuff */
3539 phb
->ioda
.total_pe_num
= 1;
3540 prop32
= of_get_property(np
, "ibm,opal-num-pes", NULL
);
3542 phb
->ioda
.total_pe_num
= be32_to_cpup(prop32
);
3543 prop32
= of_get_property(np
, "ibm,opal-reserved-pe", NULL
);
3545 phb
->ioda
.reserved_pe_idx
= be32_to_cpup(prop32
);
3547 /* Invalidate RID to PE# mapping */
3548 for (segno
= 0; segno
< ARRAY_SIZE(phb
->ioda
.pe_rmap
); segno
++)
3549 phb
->ioda
.pe_rmap
[segno
] = IODA_INVALID_PE
;
3551 /* Parse 64-bit MMIO range */
3552 pnv_ioda_parse_m64_window(phb
);
3554 phb
->ioda
.m32_size
= resource_size(&hose
->mem_resources
[0]);
3555 /* FW Has already off top 64k of M32 space (MSI space) */
3556 phb
->ioda
.m32_size
+= 0x10000;
3558 phb
->ioda
.m32_segsize
= phb
->ioda
.m32_size
/ phb
->ioda
.total_pe_num
;
3559 phb
->ioda
.m32_pci_base
= hose
->mem_resources
[0].start
- hose
->mem_offset
[0];
3560 phb
->ioda
.io_size
= hose
->pci_io_size
;
3561 phb
->ioda
.io_segsize
= phb
->ioda
.io_size
/ phb
->ioda
.total_pe_num
;
3562 phb
->ioda
.io_pci_base
= 0; /* XXX calculate this ? */
3564 /* Calculate how many 32-bit TCE segments we have */
3565 phb
->ioda
.dma32_count
= phb
->ioda
.m32_pci_base
/
3566 PNV_IODA1_DMA32_SEGSIZE
;
3568 /* Allocate aux data & arrays. We don't have IO ports on PHB3 */
3569 size
= _ALIGN_UP(max_t(unsigned, phb
->ioda
.total_pe_num
, 8) / 8,
3570 sizeof(unsigned long));
3572 size
+= phb
->ioda
.total_pe_num
* sizeof(phb
->ioda
.m64_segmap
[0]);
3574 size
+= phb
->ioda
.total_pe_num
* sizeof(phb
->ioda
.m32_segmap
[0]);
3575 if (phb
->type
== PNV_PHB_IODA1
) {
3577 size
+= phb
->ioda
.total_pe_num
* sizeof(phb
->ioda
.io_segmap
[0]);
3578 dma32map_off
= size
;
3579 size
+= phb
->ioda
.dma32_count
*
3580 sizeof(phb
->ioda
.dma32_segmap
[0]);
3583 size
+= phb
->ioda
.total_pe_num
* sizeof(struct pnv_ioda_pe
);
3584 aux
= memblock_virt_alloc(size
, 0);
3585 phb
->ioda
.pe_alloc
= aux
;
3586 phb
->ioda
.m64_segmap
= aux
+ m64map_off
;
3587 phb
->ioda
.m32_segmap
= aux
+ m32map_off
;
3588 for (segno
= 0; segno
< phb
->ioda
.total_pe_num
; segno
++) {
3589 phb
->ioda
.m64_segmap
[segno
] = IODA_INVALID_PE
;
3590 phb
->ioda
.m32_segmap
[segno
] = IODA_INVALID_PE
;
3592 if (phb
->type
== PNV_PHB_IODA1
) {
3593 phb
->ioda
.io_segmap
= aux
+ iomap_off
;
3594 for (segno
= 0; segno
< phb
->ioda
.total_pe_num
; segno
++)
3595 phb
->ioda
.io_segmap
[segno
] = IODA_INVALID_PE
;
3597 phb
->ioda
.dma32_segmap
= aux
+ dma32map_off
;
3598 for (segno
= 0; segno
< phb
->ioda
.dma32_count
; segno
++)
3599 phb
->ioda
.dma32_segmap
[segno
] = IODA_INVALID_PE
;
3601 phb
->ioda
.pe_array
= aux
+ pemap_off
;
3604 * Choose PE number for root bus, which shouldn't have
3605 * M64 resources consumed by its child devices. To pick
3606 * the PE number adjacent to the reserved one if possible.
3608 pnv_ioda_reserve_pe(phb
, phb
->ioda
.reserved_pe_idx
);
3609 if (phb
->ioda
.reserved_pe_idx
== 0) {
3610 phb
->ioda
.root_pe_idx
= 1;
3611 pnv_ioda_reserve_pe(phb
, phb
->ioda
.root_pe_idx
);
3612 } else if (phb
->ioda
.reserved_pe_idx
== (phb
->ioda
.total_pe_num
- 1)) {
3613 phb
->ioda
.root_pe_idx
= phb
->ioda
.reserved_pe_idx
- 1;
3614 pnv_ioda_reserve_pe(phb
, phb
->ioda
.root_pe_idx
);
3616 phb
->ioda
.root_pe_idx
= IODA_INVALID_PE
;
3619 INIT_LIST_HEAD(&phb
->ioda
.pe_list
);
3620 mutex_init(&phb
->ioda
.pe_list_mutex
);
3622 /* Calculate how many 32-bit TCE segments we have */
3623 phb
->ioda
.dma32_count
= phb
->ioda
.m32_pci_base
/
3624 PNV_IODA1_DMA32_SEGSIZE
;
3626 #if 0 /* We should really do that ... */
3627 rc
= opal_pci_set_phb_mem_window(opal
->phb_id
,
3630 starting_real_address
,
3631 starting_pci_address
,
3635 pr_info(" %03d (%03d) PE's M32: 0x%x [segment=0x%x]\n",
3636 phb
->ioda
.total_pe_num
, phb
->ioda
.reserved_pe_idx
,
3637 phb
->ioda
.m32_size
, phb
->ioda
.m32_segsize
);
3638 if (phb
->ioda
.m64_size
)
3639 pr_info(" M64: 0x%lx [segment=0x%lx]\n",
3640 phb
->ioda
.m64_size
, phb
->ioda
.m64_segsize
);
3641 if (phb
->ioda
.io_size
)
3642 pr_info(" IO: 0x%x [segment=0x%x]\n",
3643 phb
->ioda
.io_size
, phb
->ioda
.io_segsize
);
3646 phb
->hose
->ops
= &pnv_pci_ops
;
3647 phb
->get_pe_state
= pnv_ioda_get_pe_state
;
3648 phb
->freeze_pe
= pnv_ioda_freeze_pe
;
3649 phb
->unfreeze_pe
= pnv_ioda_unfreeze_pe
;
3651 /* Setup MSI support */
3652 pnv_pci_init_ioda_msis(phb
);
3655 * We pass the PCI probe flag PCI_REASSIGN_ALL_RSRC here
3656 * to let the PCI core do resource assignment. It's supposed
3657 * that the PCI core will do correct I/O and MMIO alignment
3658 * for the P2P bridge bars so that each PCI bus (excluding
3659 * the child P2P bridges) can form individual PE.
3661 ppc_md
.pcibios_fixup
= pnv_pci_ioda_fixup
;
3663 if (phb
->type
== PNV_PHB_NPU
) {
3664 hose
->controller_ops
= pnv_npu_ioda_controller_ops
;
3666 phb
->dma_dev_setup
= pnv_pci_ioda_dma_dev_setup
;
3667 hose
->controller_ops
= pnv_pci_ioda_controller_ops
;
3670 #ifdef CONFIG_PCI_IOV
3671 ppc_md
.pcibios_fixup_sriov
= pnv_pci_ioda_fixup_iov_resources
;
3672 ppc_md
.pcibios_iov_resource_alignment
= pnv_pci_iov_resource_alignment
;
3675 pci_add_flags(PCI_REASSIGN_ALL_RSRC
);
3677 /* Reset IODA tables to a clean state */
3678 rc
= opal_pci_reset(phb_id
, OPAL_RESET_PCI_IODA_TABLE
, OPAL_ASSERT_RESET
);
3680 pr_warning(" OPAL Error %ld performing IODA table reset !\n", rc
);
3682 /* If we're running in kdump kerenl, the previous kerenl never
3683 * shutdown PCI devices correctly. We already got IODA table
3684 * cleaned out. So we have to issue PHB reset to stop all PCI
3685 * transactions from previous kerenl.
3687 if (is_kdump_kernel()) {
3688 pr_info(" Issue PHB reset ...\n");
3689 pnv_eeh_phb_reset(hose
, EEH_RESET_FUNDAMENTAL
);
3690 pnv_eeh_phb_reset(hose
, EEH_RESET_DEACTIVATE
);
3693 /* Remove M64 resource if we can't configure it successfully */
3694 if (!phb
->init_m64
|| phb
->init_m64(phb
))
3695 hose
->mem_resources
[1].flags
= 0;
3698 void __init
pnv_pci_init_ioda2_phb(struct device_node
*np
)
3700 pnv_pci_init_ioda_phb(np
, 0, PNV_PHB_IODA2
);
3703 void __init
pnv_pci_init_npu_phb(struct device_node
*np
)
3705 pnv_pci_init_ioda_phb(np
, 0, PNV_PHB_NPU
);
3708 void __init
pnv_pci_init_ioda_hub(struct device_node
*np
)
3710 struct device_node
*phbn
;
3711 const __be64
*prop64
;
3714 pr_info("Probing IODA IO-Hub %s\n", np
->full_name
);
3716 prop64
= of_get_property(np
, "ibm,opal-hubid", NULL
);
3718 pr_err(" Missing \"ibm,opal-hubid\" property !\n");
3721 hub_id
= be64_to_cpup(prop64
);
3722 pr_devel(" HUB-ID : 0x%016llx\n", hub_id
);
3724 /* Count child PHBs */
3725 for_each_child_of_node(np
, phbn
) {
3726 /* Look for IODA1 PHBs */
3727 if (of_device_is_compatible(phbn
, "ibm,ioda-phb"))
3728 pnv_pci_init_ioda_phb(phbn
, hub_id
, PNV_PHB_IODA1
);