1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * Copyright (C) 2001 Mike Corrigan & Dave Engebretsen, IBM Corporation
7 * Copyright (C) 2004 Olof Johansson <olof@lixom.net>, IBM Corporation
8 * Copyright (C) 2006 Olof Johansson <olof@lixom.net>
10 * Dynamic DMA mapping support, pSeries-specific parts, both SMP and LPAR.
13 #include <linux/init.h>
14 #include <linux/types.h>
15 #include <linux/slab.h>
17 #include <linux/memblock.h>
18 #include <linux/spinlock.h>
19 #include <linux/string.h>
20 #include <linux/pci.h>
21 #include <linux/dma-mapping.h>
22 #include <linux/crash_dump.h>
23 #include <linux/memory.h>
25 #include <linux/iommu.h>
26 #include <linux/rculist.h>
30 #include <asm/iommu.h>
31 #include <asm/pci-bridge.h>
32 #include <asm/machdep.h>
33 #include <asm/firmware.h>
35 #include <asm/ppc-pci.h>
37 #include <asm/mmzone.h>
38 #include <asm/plpar_wrappers.h>
43 DDW_QUERY_PE_DMA_WIN = 0,
44 DDW_CREATE_PE_DMA_WIN = 1,
45 DDW_REMOVE_PE_DMA_WIN = 2,
52 DDW_EXT_RESET_DMA_WIN = 1,
53 DDW_EXT_QUERY_OUT_SIZE = 2
56 static struct iommu_table_group *iommu_pseries_alloc_group(int node)
58 struct iommu_table_group *table_group;
59 struct iommu_table *tbl;
61 table_group = kzalloc_node(sizeof(struct iommu_table_group), GFP_KERNEL,
66 tbl = kzalloc_node(sizeof(struct iommu_table), GFP_KERNEL, node);
70 INIT_LIST_HEAD_RCU(&tbl->it_group_list);
71 kref_init(&tbl->it_kref);
73 table_group->tables[0] = tbl;
82 static void iommu_pseries_free_group(struct iommu_table_group *table_group,
83 const char *node_name)
85 struct iommu_table *tbl;
90 tbl = table_group->tables[0];
91 #ifdef CONFIG_IOMMU_API
92 if (table_group->group) {
93 iommu_group_put(table_group->group);
94 BUG_ON(table_group->group);
97 iommu_tce_table_put(tbl);
102 static int tce_build_pSeries(struct iommu_table *tbl, long index,
103 long npages, unsigned long uaddr,
104 enum dma_data_direction direction,
111 proto_tce = TCE_PCI_READ; // Read allowed
113 if (direction != DMA_TO_DEVICE)
114 proto_tce |= TCE_PCI_WRITE;
116 tcep = ((__be64 *)tbl->it_base) + index;
119 /* can't move this out since we might cross MEMBLOCK boundary */
120 rpn = __pa(uaddr) >> TCE_SHIFT;
121 *tcep = cpu_to_be64(proto_tce | (rpn & TCE_RPN_MASK) << TCE_RPN_SHIFT);
123 uaddr += TCE_PAGE_SIZE;
130 static void tce_free_pSeries(struct iommu_table *tbl, long index, long npages)
134 tcep = ((__be64 *)tbl->it_base) + index;
140 static unsigned long tce_get_pseries(struct iommu_table *tbl, long index)
144 tcep = ((__be64 *)tbl->it_base) + index;
146 return be64_to_cpu(*tcep);
149 static void tce_free_pSeriesLP(unsigned long liobn, long, long);
150 static void tce_freemulti_pSeriesLP(struct iommu_table*, long, long);
152 static int tce_build_pSeriesLP(unsigned long liobn, long tcenum, long tceshift,
153 long npages, unsigned long uaddr,
154 enum dma_data_direction direction,
161 long tcenum_start = tcenum, npages_start = npages;
163 rpn = __pa(uaddr) >> tceshift;
164 proto_tce = TCE_PCI_READ;
165 if (direction != DMA_TO_DEVICE)
166 proto_tce |= TCE_PCI_WRITE;
169 tce = proto_tce | (rpn & TCE_RPN_MASK) << tceshift;
170 rc = plpar_tce_put((u64)liobn, (u64)tcenum << tceshift, tce);
172 if (unlikely(rc == H_NOT_ENOUGH_RESOURCES)) {
174 tce_free_pSeriesLP(liobn, tcenum_start,
175 (npages_start - (npages + 1)));
179 if (rc && printk_ratelimit()) {
180 printk("tce_build_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
181 printk("\tindex = 0x%llx\n", (u64)liobn);
182 printk("\ttcenum = 0x%llx\n", (u64)tcenum);
183 printk("\ttce val = 0x%llx\n", tce );
193 static DEFINE_PER_CPU(__be64 *, tce_page);
195 static int tce_buildmulti_pSeriesLP(struct iommu_table *tbl, long tcenum,
196 long npages, unsigned long uaddr,
197 enum dma_data_direction direction,
205 long tcenum_start = tcenum, npages_start = npages;
209 if ((npages == 1) || !firmware_has_feature(FW_FEATURE_PUT_TCE_IND)) {
210 return tce_build_pSeriesLP(tbl->it_index, tcenum,
211 tbl->it_page_shift, npages, uaddr,
215 local_irq_save(flags); /* to protect tcep and the page behind it */
217 tcep = __this_cpu_read(tce_page);
219 /* This is safe to do since interrupts are off when we're called
220 * from iommu_alloc{,_sg}()
223 tcep = (__be64 *)__get_free_page(GFP_ATOMIC);
224 /* If allocation fails, fall back to the loop implementation */
226 local_irq_restore(flags);
227 return tce_build_pSeriesLP(tbl->it_index, tcenum,
229 npages, uaddr, direction, attrs);
231 __this_cpu_write(tce_page, tcep);
234 rpn = __pa(uaddr) >> TCE_SHIFT;
235 proto_tce = TCE_PCI_READ;
236 if (direction != DMA_TO_DEVICE)
237 proto_tce |= TCE_PCI_WRITE;
239 /* We can map max one pageful of TCEs at a time */
242 * Set up the page with TCE data, looping through and setting
245 limit = min_t(long, npages, 4096/TCE_ENTRY_SIZE);
247 for (l = 0; l < limit; l++) {
248 tcep[l] = cpu_to_be64(proto_tce | (rpn & TCE_RPN_MASK) << TCE_RPN_SHIFT);
252 rc = plpar_tce_put_indirect((u64)tbl->it_index,
259 } while (npages > 0 && !rc);
261 local_irq_restore(flags);
263 if (unlikely(rc == H_NOT_ENOUGH_RESOURCES)) {
265 tce_freemulti_pSeriesLP(tbl, tcenum_start,
266 (npages_start - (npages + limit)));
270 if (rc && printk_ratelimit()) {
271 printk("tce_buildmulti_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
272 printk("\tindex = 0x%llx\n", (u64)tbl->it_index);
273 printk("\tnpages = 0x%llx\n", (u64)npages);
274 printk("\ttce[0] val = 0x%llx\n", tcep[0]);
280 static void tce_free_pSeriesLP(unsigned long liobn, long tcenum, long npages)
285 rc = plpar_tce_put((u64)liobn, (u64)tcenum << 12, 0);
287 if (rc && printk_ratelimit()) {
288 printk("tce_free_pSeriesLP: plpar_tce_put failed. rc=%lld\n", rc);
289 printk("\tindex = 0x%llx\n", (u64)liobn);
290 printk("\ttcenum = 0x%llx\n", (u64)tcenum);
299 static void tce_freemulti_pSeriesLP(struct iommu_table *tbl, long tcenum, long npages)
303 if (!firmware_has_feature(FW_FEATURE_STUFF_TCE))
304 return tce_free_pSeriesLP(tbl->it_index, tcenum, npages);
306 rc = plpar_tce_stuff((u64)tbl->it_index, (u64)tcenum << 12, 0, npages);
308 if (rc && printk_ratelimit()) {
309 printk("tce_freemulti_pSeriesLP: plpar_tce_stuff failed\n");
310 printk("\trc = %lld\n", rc);
311 printk("\tindex = 0x%llx\n", (u64)tbl->it_index);
312 printk("\tnpages = 0x%llx\n", (u64)npages);
317 static unsigned long tce_get_pSeriesLP(struct iommu_table *tbl, long tcenum)
320 unsigned long tce_ret;
322 rc = plpar_tce_get((u64)tbl->it_index, (u64)tcenum << 12, &tce_ret);
324 if (rc && printk_ratelimit()) {
325 printk("tce_get_pSeriesLP: plpar_tce_get failed. rc=%lld\n", rc);
326 printk("\tindex = 0x%llx\n", (u64)tbl->it_index);
327 printk("\ttcenum = 0x%llx\n", (u64)tcenum);
334 /* this is compatible with cells for the device tree property */
335 struct dynamic_dma_window_prop {
336 __be32 liobn; /* tce table number */
337 __be64 dma_base; /* address hi,lo */
338 __be32 tce_shift; /* ilog2(tce_page_size) */
339 __be32 window_shift; /* ilog2(tce_window_size) */
342 struct direct_window {
343 struct device_node *device;
344 const struct dynamic_dma_window_prop *prop;
345 struct list_head list;
348 /* Dynamic DMA Window support */
349 struct ddw_query_response {
350 u32 windows_available;
351 u64 largest_available_block;
353 u32 migration_capable;
356 struct ddw_create_response {
362 static LIST_HEAD(direct_window_list);
363 /* prevents races between memory on/offline and window creation */
364 static DEFINE_SPINLOCK(direct_window_list_lock);
365 /* protects initializing window twice for same device */
366 static DEFINE_MUTEX(direct_window_init_mutex);
367 #define DIRECT64_PROPNAME "linux,direct64-ddr-window-info"
369 static int tce_clearrange_multi_pSeriesLP(unsigned long start_pfn,
370 unsigned long num_pfn, const void *arg)
372 const struct dynamic_dma_window_prop *maprange = arg;
374 u64 tce_size, num_tce, dma_offset, next;
378 tce_shift = be32_to_cpu(maprange->tce_shift);
379 tce_size = 1ULL << tce_shift;
380 next = start_pfn << PAGE_SHIFT;
381 num_tce = num_pfn << PAGE_SHIFT;
383 /* round back to the beginning of the tce page size */
384 num_tce += next & (tce_size - 1);
385 next &= ~(tce_size - 1);
387 /* covert to number of tces */
388 num_tce |= tce_size - 1;
389 num_tce >>= tce_shift;
393 * Set up the page with TCE data, looping through and setting
396 limit = min_t(long, num_tce, 512);
397 dma_offset = next + be64_to_cpu(maprange->dma_base);
399 rc = plpar_tce_stuff((u64)be32_to_cpu(maprange->liobn),
402 next += limit * tce_size;
404 } while (num_tce > 0 && !rc);
409 static int tce_setrange_multi_pSeriesLP(unsigned long start_pfn,
410 unsigned long num_pfn, const void *arg)
412 const struct dynamic_dma_window_prop *maprange = arg;
413 u64 tce_size, num_tce, dma_offset, next, proto_tce, liobn;
419 if (!firmware_has_feature(FW_FEATURE_PUT_TCE_IND)) {
420 unsigned long tceshift = be32_to_cpu(maprange->tce_shift);
421 unsigned long dmastart = (start_pfn << PAGE_SHIFT) +
422 be64_to_cpu(maprange->dma_base);
423 unsigned long tcenum = dmastart >> tceshift;
424 unsigned long npages = num_pfn << PAGE_SHIFT >> tceshift;
425 void *uaddr = __va(start_pfn << PAGE_SHIFT);
427 return tce_build_pSeriesLP(be32_to_cpu(maprange->liobn),
428 tcenum, tceshift, npages, (unsigned long) uaddr,
429 DMA_BIDIRECTIONAL, 0);
432 local_irq_disable(); /* to protect tcep and the page behind it */
433 tcep = __this_cpu_read(tce_page);
436 tcep = (__be64 *)__get_free_page(GFP_ATOMIC);
441 __this_cpu_write(tce_page, tcep);
444 proto_tce = TCE_PCI_READ | TCE_PCI_WRITE;
446 liobn = (u64)be32_to_cpu(maprange->liobn);
447 tce_shift = be32_to_cpu(maprange->tce_shift);
448 tce_size = 1ULL << tce_shift;
449 next = start_pfn << PAGE_SHIFT;
450 num_tce = num_pfn << PAGE_SHIFT;
452 /* round back to the beginning of the tce page size */
453 num_tce += next & (tce_size - 1);
454 next &= ~(tce_size - 1);
456 /* covert to number of tces */
457 num_tce |= tce_size - 1;
458 num_tce >>= tce_shift;
460 /* We can map max one pageful of TCEs at a time */
463 * Set up the page with TCE data, looping through and setting
466 limit = min_t(long, num_tce, 4096/TCE_ENTRY_SIZE);
467 dma_offset = next + be64_to_cpu(maprange->dma_base);
469 for (l = 0; l < limit; l++) {
470 tcep[l] = cpu_to_be64(proto_tce | next);
474 rc = plpar_tce_put_indirect(liobn,
480 } while (num_tce > 0 && !rc);
482 /* error cleanup: caller will clear whole range */
488 static int tce_setrange_multi_pSeriesLP_walk(unsigned long start_pfn,
489 unsigned long num_pfn, void *arg)
491 return tce_setrange_multi_pSeriesLP(start_pfn, num_pfn, arg);
494 static void iommu_table_setparms(struct pci_controller *phb,
495 struct device_node *dn,
496 struct iommu_table *tbl)
498 struct device_node *node;
499 const unsigned long *basep;
504 basep = of_get_property(node, "linux,tce-base", NULL);
505 sizep = of_get_property(node, "linux,tce-size", NULL);
506 if (basep == NULL || sizep == NULL) {
507 printk(KERN_ERR "PCI_DMA: iommu_table_setparms: %pOF has "
508 "missing tce entries !\n", dn);
512 tbl->it_base = (unsigned long)__va(*basep);
514 if (!is_kdump_kernel())
515 memset((void *)tbl->it_base, 0, *sizep);
517 tbl->it_busno = phb->bus->number;
518 tbl->it_page_shift = IOMMU_PAGE_SHIFT_4K;
520 /* Units of tce entries */
521 tbl->it_offset = phb->dma_window_base_cur >> tbl->it_page_shift;
523 /* Test if we are going over 2GB of DMA space */
524 if (phb->dma_window_base_cur + phb->dma_window_size > 0x80000000ul) {
525 udbg_printf("PCI_DMA: Unexpected number of IOAs under this PHB.\n");
526 panic("PCI_DMA: Unexpected number of IOAs under this PHB.\n");
529 phb->dma_window_base_cur += phb->dma_window_size;
531 /* Set the tce table size - measured in entries */
532 tbl->it_size = phb->dma_window_size >> tbl->it_page_shift;
535 tbl->it_blocksize = 16;
536 tbl->it_type = TCE_PCI;
540 * iommu_table_setparms_lpar
542 * Function: On pSeries LPAR systems, return TCE table info, given a pci bus.
544 static void iommu_table_setparms_lpar(struct pci_controller *phb,
545 struct device_node *dn,
546 struct iommu_table *tbl,
547 struct iommu_table_group *table_group,
548 const __be32 *dma_window)
550 unsigned long offset, size;
552 of_parse_dma_window(dn, dma_window, &tbl->it_index, &offset, &size);
554 tbl->it_busno = phb->bus->number;
555 tbl->it_page_shift = IOMMU_PAGE_SHIFT_4K;
557 tbl->it_blocksize = 16;
558 tbl->it_type = TCE_PCI;
559 tbl->it_offset = offset >> tbl->it_page_shift;
560 tbl->it_size = size >> tbl->it_page_shift;
562 table_group->tce32_start = offset;
563 table_group->tce32_size = size;
566 struct iommu_table_ops iommu_table_pseries_ops = {
567 .set = tce_build_pSeries,
568 .clear = tce_free_pSeries,
569 .get = tce_get_pseries
572 static void pci_dma_bus_setup_pSeries(struct pci_bus *bus)
574 struct device_node *dn;
575 struct iommu_table *tbl;
576 struct device_node *isa_dn, *isa_dn_orig;
577 struct device_node *tmp;
581 dn = pci_bus_to_OF_node(bus);
583 pr_debug("pci_dma_bus_setup_pSeries: setting up bus %pOF\n", dn);
586 /* This is not a root bus, any setup will be done for the
587 * device-side of the bridge in iommu_dev_setup_pSeries().
593 /* Check if the ISA bus on the system is under
596 isa_dn = isa_dn_orig = of_find_node_by_type(NULL, "isa");
598 while (isa_dn && isa_dn != dn)
599 isa_dn = isa_dn->parent;
601 of_node_put(isa_dn_orig);
603 /* Count number of direct PCI children of the PHB. */
604 for (children = 0, tmp = dn->child; tmp; tmp = tmp->sibling)
607 pr_debug("Children: %d\n", children);
609 /* Calculate amount of DMA window per slot. Each window must be
610 * a power of two (due to pci_alloc_consistent requirements).
612 * Keep 256MB aside for PHBs with ISA.
616 /* No ISA/IDE - just set window size and return */
617 pci->phb->dma_window_size = 0x80000000ul; /* To be divided */
619 while (pci->phb->dma_window_size * children > 0x80000000ul)
620 pci->phb->dma_window_size >>= 1;
621 pr_debug("No ISA/IDE, window size is 0x%llx\n",
622 pci->phb->dma_window_size);
623 pci->phb->dma_window_base_cur = 0;
628 /* If we have ISA, then we probably have an IDE
629 * controller too. Allocate a 128MB table but
630 * skip the first 128MB to avoid stepping on ISA
633 pci->phb->dma_window_size = 0x8000000ul;
634 pci->phb->dma_window_base_cur = 0x8000000ul;
636 pci->table_group = iommu_pseries_alloc_group(pci->phb->node);
637 tbl = pci->table_group->tables[0];
639 iommu_table_setparms(pci->phb, dn, tbl);
640 tbl->it_ops = &iommu_table_pseries_ops;
641 if (!iommu_init_table(tbl, pci->phb->node, 0, 0))
642 panic("Failed to initialize iommu table");
644 /* Divide the rest (1.75GB) among the children */
645 pci->phb->dma_window_size = 0x80000000ul;
646 while (pci->phb->dma_window_size * children > 0x70000000ul)
647 pci->phb->dma_window_size >>= 1;
649 pr_debug("ISA/IDE, window size is 0x%llx\n", pci->phb->dma_window_size);
652 #ifdef CONFIG_IOMMU_API
653 static int tce_exchange_pseries(struct iommu_table *tbl, long index, unsigned
654 long *tce, enum dma_data_direction *direction,
658 unsigned long ioba = (unsigned long) index << tbl->it_page_shift;
659 unsigned long flags, oldtce = 0;
660 u64 proto_tce = iommu_direction_to_tce_perm(*direction);
661 unsigned long newtce = *tce | proto_tce;
663 spin_lock_irqsave(&tbl->large_pool.lock, flags);
665 rc = plpar_tce_get((u64)tbl->it_index, ioba, &oldtce);
667 rc = plpar_tce_put((u64)tbl->it_index, ioba, newtce);
670 *direction = iommu_tce_direction(oldtce);
671 *tce = oldtce & ~(TCE_PCI_READ | TCE_PCI_WRITE);
674 spin_unlock_irqrestore(&tbl->large_pool.lock, flags);
680 struct iommu_table_ops iommu_table_lpar_multi_ops = {
681 .set = tce_buildmulti_pSeriesLP,
682 #ifdef CONFIG_IOMMU_API
683 .xchg_no_kill = tce_exchange_pseries,
685 .clear = tce_freemulti_pSeriesLP,
686 .get = tce_get_pSeriesLP
689 static void pci_dma_bus_setup_pSeriesLP(struct pci_bus *bus)
691 struct iommu_table *tbl;
692 struct device_node *dn, *pdn;
694 const __be32 *dma_window = NULL;
696 dn = pci_bus_to_OF_node(bus);
698 pr_debug("pci_dma_bus_setup_pSeriesLP: setting up bus %pOF\n",
701 /* Find nearest ibm,dma-window, walking up the device tree */
702 for (pdn = dn; pdn != NULL; pdn = pdn->parent) {
703 dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
704 if (dma_window != NULL)
708 if (dma_window == NULL) {
709 pr_debug(" no ibm,dma-window property !\n");
715 pr_debug(" parent is %pOF, iommu_table: 0x%p\n",
716 pdn, ppci->table_group);
718 if (!ppci->table_group) {
719 ppci->table_group = iommu_pseries_alloc_group(ppci->phb->node);
720 tbl = ppci->table_group->tables[0];
721 iommu_table_setparms_lpar(ppci->phb, pdn, tbl,
722 ppci->table_group, dma_window);
723 tbl->it_ops = &iommu_table_lpar_multi_ops;
724 if (!iommu_init_table(tbl, ppci->phb->node, 0, 0))
725 panic("Failed to initialize iommu table");
726 iommu_register_group(ppci->table_group,
727 pci_domain_nr(bus), 0);
728 pr_debug(" created table: %p\n", ppci->table_group);
733 static void pci_dma_dev_setup_pSeries(struct pci_dev *dev)
735 struct device_node *dn;
736 struct iommu_table *tbl;
738 pr_debug("pci_dma_dev_setup_pSeries: %s\n", pci_name(dev));
740 dn = dev->dev.of_node;
742 /* If we're the direct child of a root bus, then we need to allocate
743 * an iommu table ourselves. The bus setup code should have setup
744 * the window sizes already.
746 if (!dev->bus->self) {
747 struct pci_controller *phb = PCI_DN(dn)->phb;
749 pr_debug(" --> first child, no bridge. Allocating iommu table.\n");
750 PCI_DN(dn)->table_group = iommu_pseries_alloc_group(phb->node);
751 tbl = PCI_DN(dn)->table_group->tables[0];
752 iommu_table_setparms(phb, dn, tbl);
753 tbl->it_ops = &iommu_table_pseries_ops;
754 if (!iommu_init_table(tbl, phb->node, 0, 0))
755 panic("Failed to initialize iommu table");
757 set_iommu_table_base(&dev->dev, tbl);
761 /* If this device is further down the bus tree, search upwards until
762 * an already allocated iommu table is found and use that.
765 while (dn && PCI_DN(dn) && PCI_DN(dn)->table_group == NULL)
768 if (dn && PCI_DN(dn))
769 set_iommu_table_base(&dev->dev,
770 PCI_DN(dn)->table_group->tables[0]);
772 printk(KERN_WARNING "iommu: Device %s has no iommu table\n",
776 static int __read_mostly disable_ddw;
778 static int __init disable_ddw_setup(char *str)
781 printk(KERN_INFO "ppc iommu: disabling ddw.\n");
786 early_param("disable_ddw", disable_ddw_setup);
788 static void remove_dma_window(struct device_node *np, u32 *ddw_avail,
789 struct property *win)
791 struct dynamic_dma_window_prop *dwp;
796 liobn = (u64)be32_to_cpu(dwp->liobn);
798 /* clear the whole window, note the arg is in kernel pages */
799 ret = tce_clearrange_multi_pSeriesLP(0,
800 1ULL << (be32_to_cpu(dwp->window_shift) - PAGE_SHIFT), dwp);
802 pr_warn("%pOF failed to clear tces in window.\n",
805 pr_debug("%pOF successfully cleared tces in window.\n",
808 ret = rtas_call(ddw_avail[DDW_REMOVE_PE_DMA_WIN], 1, 1, NULL, liobn);
810 pr_warn("%pOF: failed to remove direct window: rtas returned "
811 "%d to ibm,remove-pe-dma-window(%x) %llx\n",
812 np, ret, ddw_avail[DDW_REMOVE_PE_DMA_WIN], liobn);
814 pr_debug("%pOF: successfully removed direct window: rtas returned "
815 "%d to ibm,remove-pe-dma-window(%x) %llx\n",
816 np, ret, ddw_avail[DDW_REMOVE_PE_DMA_WIN], liobn);
819 static void remove_ddw(struct device_node *np, bool remove_prop)
821 struct property *win;
822 u32 ddw_avail[DDW_APPLICABLE_SIZE];
825 ret = of_property_read_u32_array(np, "ibm,ddw-applicable",
826 &ddw_avail[0], DDW_APPLICABLE_SIZE);
830 win = of_find_property(np, DIRECT64_PROPNAME, NULL);
834 if (win->length >= sizeof(struct dynamic_dma_window_prop))
835 remove_dma_window(np, ddw_avail, win);
840 ret = of_remove_property(np, win);
842 pr_warn("%pOF: failed to remove direct window property: %d\n",
846 static u64 find_existing_ddw(struct device_node *pdn, int *window_shift)
848 struct direct_window *window;
849 const struct dynamic_dma_window_prop *direct64;
852 spin_lock(&direct_window_list_lock);
853 /* check if we already created a window and dupe that config if so */
854 list_for_each_entry(window, &direct_window_list, list) {
855 if (window->device == pdn) {
856 direct64 = window->prop;
857 dma_addr = be64_to_cpu(direct64->dma_base);
858 *window_shift = be32_to_cpu(direct64->window_shift);
862 spin_unlock(&direct_window_list_lock);
867 static int find_existing_ddw_windows(void)
870 struct device_node *pdn;
871 struct direct_window *window;
872 const struct dynamic_dma_window_prop *direct64;
874 if (!firmware_has_feature(FW_FEATURE_LPAR))
877 for_each_node_with_property(pdn, DIRECT64_PROPNAME) {
878 direct64 = of_get_property(pdn, DIRECT64_PROPNAME, &len);
882 window = kzalloc(sizeof(*window), GFP_KERNEL);
883 if (!window || len < sizeof(struct dynamic_dma_window_prop)) {
885 remove_ddw(pdn, true);
889 window->device = pdn;
890 window->prop = direct64;
891 spin_lock(&direct_window_list_lock);
892 list_add(&window->list, &direct_window_list);
893 spin_unlock(&direct_window_list_lock);
898 machine_arch_initcall(pseries, find_existing_ddw_windows);
901 * ddw_read_ext - Get the value of an DDW extension
902 * @np: device node from which the extension value is to be read.
903 * @extnum: index number of the extension.
904 * @value: pointer to return value, modified when extension is available.
906 * Checks if "ibm,ddw-extensions" exists for this node, and get the value
908 * It can be used only to check if a property exists, passing value == NULL.
911 * 0 if extension successfully read
912 * -EINVAL if the "ibm,ddw-extensions" does not exist,
913 * -ENODATA if "ibm,ddw-extensions" does not have a value, and
914 * -EOVERFLOW if "ibm,ddw-extensions" does not contain this extension.
916 static inline int ddw_read_ext(const struct device_node *np, int extnum,
919 static const char propname[] = "ibm,ddw-extensions";
923 ret = of_property_read_u32_index(np, propname, DDW_EXT_SIZE, &count);
933 return of_property_read_u32_index(np, propname, extnum, value);
936 static int query_ddw(struct pci_dev *dev, const u32 *ddw_avail,
937 struct ddw_query_response *query,
938 struct device_node *parent)
940 struct device_node *dn;
942 u32 cfg_addr, ext_query, query_out[5];
947 * From LoPAR level 2.8, "ibm,ddw-extensions" index 3 can rule how many
948 * output parameters ibm,query-pe-dma-windows will have, ranging from
951 ret = ddw_read_ext(parent, DDW_EXT_QUERY_OUT_SIZE, &ext_query);
952 if (!ret && ext_query == 1)
958 * Get the config address and phb buid of the PE window.
959 * Rely on eeh to retrieve this for us.
960 * Retrieve them from the pci device, not the node with the
961 * dma-window property
963 dn = pci_device_to_OF_node(dev);
965 buid = pdn->phb->buid;
966 cfg_addr = ((pdn->busno << 16) | (pdn->devfn << 8));
968 ret = rtas_call(ddw_avail[DDW_QUERY_PE_DMA_WIN], 3, out_sz, query_out,
969 cfg_addr, BUID_HI(buid), BUID_LO(buid));
970 dev_info(&dev->dev, "ibm,query-pe-dma-windows(%x) %x %x %x returned %d\n",
971 ddw_avail[DDW_QUERY_PE_DMA_WIN], cfg_addr, BUID_HI(buid),
976 query->windows_available = query_out[0];
977 query->largest_available_block = query_out[1];
978 query->page_size = query_out[2];
979 query->migration_capable = query_out[3];
982 query->windows_available = query_out[0];
983 query->largest_available_block = ((u64)query_out[1] << 32) |
985 query->page_size = query_out[3];
986 query->migration_capable = query_out[4];
993 static int create_ddw(struct pci_dev *dev, const u32 *ddw_avail,
994 struct ddw_create_response *create, int page_shift,
997 struct device_node *dn;
1004 * Get the config address and phb buid of the PE window.
1005 * Rely on eeh to retrieve this for us.
1006 * Retrieve them from the pci device, not the node with the
1007 * dma-window property
1009 dn = pci_device_to_OF_node(dev);
1011 buid = pdn->phb->buid;
1012 cfg_addr = ((pdn->busno << 16) | (pdn->devfn << 8));
1015 /* extra outputs are LIOBN and dma-addr (hi, lo) */
1016 ret = rtas_call(ddw_avail[DDW_CREATE_PE_DMA_WIN], 5, 4,
1017 (u32 *)create, cfg_addr, BUID_HI(buid),
1018 BUID_LO(buid), page_shift, window_shift);
1019 } while (rtas_busy_delay(ret));
1021 "ibm,create-pe-dma-window(%x) %x %x %x %x %x returned %d "
1022 "(liobn = 0x%x starting addr = %x %x)\n",
1023 ddw_avail[DDW_CREATE_PE_DMA_WIN], cfg_addr, BUID_HI(buid),
1024 BUID_LO(buid), page_shift, window_shift, ret, create->liobn,
1025 create->addr_hi, create->addr_lo);
1030 struct failed_ddw_pdn {
1031 struct device_node *pdn;
1032 struct list_head list;
1035 static LIST_HEAD(failed_ddw_pdn_list);
1037 static phys_addr_t ddw_memory_hotplug_max(void)
1039 phys_addr_t max_addr = memory_hotplug_max();
1040 struct device_node *memory;
1043 * The "ibm,pmemory" can appear anywhere in the address space.
1044 * Assuming it is still backed by page structs, set the upper limit
1045 * for the huge DMA window as MAX_PHYSMEM_BITS.
1047 if (of_find_node_by_type(NULL, "ibm,pmemory"))
1048 return (sizeof(phys_addr_t) * 8 <= MAX_PHYSMEM_BITS) ?
1049 (phys_addr_t) -1 : (1ULL << MAX_PHYSMEM_BITS);
1051 for_each_node_by_type(memory, "memory") {
1052 unsigned long start, size;
1053 int n_mem_addr_cells, n_mem_size_cells, len;
1054 const __be32 *memcell_buf;
1056 memcell_buf = of_get_property(memory, "reg", &len);
1057 if (!memcell_buf || len <= 0)
1060 n_mem_addr_cells = of_n_addr_cells(memory);
1061 n_mem_size_cells = of_n_size_cells(memory);
1063 start = of_read_number(memcell_buf, n_mem_addr_cells);
1064 memcell_buf += n_mem_addr_cells;
1065 size = of_read_number(memcell_buf, n_mem_size_cells);
1066 memcell_buf += n_mem_size_cells;
1068 max_addr = max_t(phys_addr_t, max_addr, start + size);
1075 * Platforms supporting the DDW option starting with LoPAR level 2.7 implement
1076 * ibm,ddw-extensions, which carries the rtas token for
1077 * ibm,reset-pe-dma-windows.
1078 * That rtas-call can be used to restore the default DMA window for the device.
1080 static void reset_dma_window(struct pci_dev *dev, struct device_node *par_dn)
1083 u32 cfg_addr, reset_dma_win;
1085 struct device_node *dn;
1088 ret = ddw_read_ext(par_dn, DDW_EXT_RESET_DMA_WIN, &reset_dma_win);
1092 dn = pci_device_to_OF_node(dev);
1094 buid = pdn->phb->buid;
1095 cfg_addr = (pdn->busno << 16) | (pdn->devfn << 8);
1097 ret = rtas_call(reset_dma_win, 3, 1, NULL, cfg_addr, BUID_HI(buid),
1101 "ibm,reset-pe-dma-windows(%x) %x %x %x returned %d ",
1102 reset_dma_win, cfg_addr, BUID_HI(buid), BUID_LO(buid),
1106 /* Return largest page shift based on "IO Page Sizes" output of ibm,query-pe-dma-window. */
1107 static int iommu_get_page_shift(u32 query_page_size)
1109 /* Supported IO page-sizes according to LoPAR */
1110 const int shift[] = {
1111 __builtin_ctzll(SZ_4K), __builtin_ctzll(SZ_64K), __builtin_ctzll(SZ_16M),
1112 __builtin_ctzll(SZ_32M), __builtin_ctzll(SZ_64M), __builtin_ctzll(SZ_128M),
1113 __builtin_ctzll(SZ_256M), __builtin_ctzll(SZ_16G)
1116 int i = ARRAY_SIZE(shift) - 1;
1119 * On LoPAR, ibm,query-pe-dma-window outputs "IO Page Sizes" using a bit field:
1120 * - bit 31 means 4k pages are supported,
1121 * - bit 30 means 64k pages are supported, and so on.
1122 * Larger pagesizes map more memory with the same amount of TCEs, so start probing them.
1124 for (; i >= 0 ; i--) {
1125 if (query_page_size & (1 << i))
1129 /* No valid page size found. */
1134 * If the PE supports dynamic dma windows, and there is space for a table
1135 * that can map all pages in a linear offset, then setup such a table,
1136 * and record the dma-offset in the struct device.
1138 * dev: the pci device we are checking
1139 * pdn: the parent pe node with the ibm,dma_window property
1140 * Future: also check if we can remap the base window for our base page size
1142 * returns the dma offset for use by the direct mapped DMA code.
1144 static u64 enable_ddw(struct pci_dev *dev, struct device_node *pdn)
1147 int max_ram_len = order_base_2(ddw_memory_hotplug_max());
1148 struct ddw_query_response query;
1149 struct ddw_create_response create;
1152 struct device_node *dn;
1153 u32 ddw_avail[DDW_APPLICABLE_SIZE];
1154 struct direct_window *window;
1155 struct property *win64;
1156 struct dynamic_dma_window_prop *ddwprop;
1157 struct failed_ddw_pdn *fpdn;
1158 bool default_win_removed = false;
1161 dn = of_find_node_by_type(NULL, "ibm,pmemory");
1162 pmem_present = dn != NULL;
1165 mutex_lock(&direct_window_init_mutex);
1167 dma_addr = find_existing_ddw(pdn, &len);
1172 * If we already went through this for a previous function of
1173 * the same device and failed, we don't want to muck with the
1174 * DMA window again, as it will race with in-flight operations
1175 * and can lead to EEHs. The above mutex protects access to the
1178 list_for_each_entry(fpdn, &failed_ddw_pdn_list, list) {
1179 if (fpdn->pdn == pdn)
1184 * the ibm,ddw-applicable property holds the tokens for:
1185 * ibm,query-pe-dma-window
1186 * ibm,create-pe-dma-window
1187 * ibm,remove-pe-dma-window
1188 * for the given node in that order.
1189 * the property is actually in the parent, not the PE
1191 ret = of_property_read_u32_array(pdn, "ibm,ddw-applicable",
1192 &ddw_avail[0], DDW_APPLICABLE_SIZE);
1197 * Query if there is a second window of size to map the
1198 * whole partition. Query returns number of windows, largest
1199 * block assigned to PE (partition endpoint), and two bitmasks
1200 * of page sizes: supported and supported for migrate-dma.
1202 dn = pci_device_to_OF_node(dev);
1203 ret = query_ddw(dev, ddw_avail, &query, pdn);
1208 * If there is no window available, remove the default DMA window,
1209 * if it's present. This will make all the resources available to the
1211 * If anything fails after this, we need to restore it, so also check
1212 * for extensions presence.
1214 if (query.windows_available == 0) {
1215 struct property *default_win;
1218 default_win = of_find_property(pdn, "ibm,dma-window", NULL);
1222 reset_win_ext = ddw_read_ext(pdn, DDW_EXT_RESET_DMA_WIN, NULL);
1226 remove_dma_window(pdn, ddw_avail, default_win);
1227 default_win_removed = true;
1229 /* Query again, to check if the window is available */
1230 ret = query_ddw(dev, ddw_avail, &query, pdn);
1234 if (query.windows_available == 0) {
1235 /* no windows are available for this device. */
1236 dev_dbg(&dev->dev, "no free dynamic windows");
1241 page_shift = iommu_get_page_shift(query.page_size);
1243 dev_dbg(&dev->dev, "no supported direct page size in mask %x",
1247 /* verify the window * number of ptes will map the partition */
1248 /* check largest block * page size > max memory hotplug addr */
1250 * The "ibm,pmemory" can appear anywhere in the address space.
1251 * Assuming it is still backed by page structs, try MAX_PHYSMEM_BITS
1252 * for the upper limit and fallback to max RAM otherwise but this
1253 * disables device::dma_ops_bypass.
1257 if (query.largest_available_block >=
1258 (1ULL << (MAX_PHYSMEM_BITS - page_shift)))
1259 len = MAX_PHYSMEM_BITS;
1261 dev_info(&dev->dev, "Skipping ibm,pmemory");
1264 if (query.largest_available_block < (1ULL << (len - page_shift))) {
1266 "can't map partition max 0x%llx with %llu %llu-sized pages\n",
1268 query.largest_available_block,
1269 1ULL << page_shift);
1272 win64 = kzalloc(sizeof(struct property), GFP_KERNEL);
1275 "couldn't allocate property for 64bit dma window\n");
1278 win64->name = kstrdup(DIRECT64_PROPNAME, GFP_KERNEL);
1279 win64->value = ddwprop = kmalloc(sizeof(*ddwprop), GFP_KERNEL);
1280 win64->length = sizeof(*ddwprop);
1281 if (!win64->name || !win64->value) {
1283 "couldn't allocate property name and value\n");
1287 ret = create_ddw(dev, ddw_avail, &create, page_shift, len);
1291 ddwprop->liobn = cpu_to_be32(create.liobn);
1292 ddwprop->dma_base = cpu_to_be64(((u64)create.addr_hi << 32) |
1294 ddwprop->tce_shift = cpu_to_be32(page_shift);
1295 ddwprop->window_shift = cpu_to_be32(len);
1297 dev_dbg(&dev->dev, "created tce table LIOBN 0x%x for %pOF\n",
1300 window = kzalloc(sizeof(*window), GFP_KERNEL);
1302 goto out_clear_window;
1304 ret = walk_system_ram_range(0, memblock_end_of_DRAM() >> PAGE_SHIFT,
1305 win64->value, tce_setrange_multi_pSeriesLP_walk);
1307 dev_info(&dev->dev, "failed to map direct window for %pOF: %d\n",
1309 goto out_free_window;
1312 ret = of_add_property(pdn, win64);
1314 dev_err(&dev->dev, "unable to add dma window property for %pOF: %d",
1316 goto out_free_window;
1319 window->device = pdn;
1320 window->prop = ddwprop;
1321 spin_lock(&direct_window_list_lock);
1322 list_add(&window->list, &direct_window_list);
1323 spin_unlock(&direct_window_list_lock);
1325 dma_addr = be64_to_cpu(ddwprop->dma_base);
1332 remove_ddw(pdn, true);
1336 kfree(win64->value);
1340 if (default_win_removed)
1341 reset_dma_window(dev, pdn);
1343 fpdn = kzalloc(sizeof(*fpdn), GFP_KERNEL);
1347 list_add(&fpdn->list, &failed_ddw_pdn_list);
1350 mutex_unlock(&direct_window_init_mutex);
1353 * If we have persistent memory and the window size is only as big
1354 * as RAM, then we failed to create a window to cover persistent
1355 * memory and need to set the DMA limit.
1357 if (pmem_present && dma_addr && (len == max_ram_len))
1358 dev->dev.bus_dma_limit = dma_addr + (1ULL << len);
1363 static void pci_dma_dev_setup_pSeriesLP(struct pci_dev *dev)
1365 struct device_node *pdn, *dn;
1366 struct iommu_table *tbl;
1367 const __be32 *dma_window = NULL;
1370 pr_debug("pci_dma_dev_setup_pSeriesLP: %s\n", pci_name(dev));
1372 /* dev setup for LPAR is a little tricky, since the device tree might
1373 * contain the dma-window properties per-device and not necessarily
1374 * for the bus. So we need to search upwards in the tree until we
1375 * either hit a dma-window property, OR find a parent with a table
1376 * already allocated.
1378 dn = pci_device_to_OF_node(dev);
1379 pr_debug(" node is %pOF\n", dn);
1381 for (pdn = dn; pdn && PCI_DN(pdn) && !PCI_DN(pdn)->table_group;
1382 pdn = pdn->parent) {
1383 dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
1388 if (!pdn || !PCI_DN(pdn)) {
1389 printk(KERN_WARNING "pci_dma_dev_setup_pSeriesLP: "
1390 "no DMA window found for pci dev=%s dn=%pOF\n",
1394 pr_debug(" parent is %pOF\n", pdn);
1397 if (!pci->table_group) {
1398 pci->table_group = iommu_pseries_alloc_group(pci->phb->node);
1399 tbl = pci->table_group->tables[0];
1400 iommu_table_setparms_lpar(pci->phb, pdn, tbl,
1401 pci->table_group, dma_window);
1402 tbl->it_ops = &iommu_table_lpar_multi_ops;
1403 iommu_init_table(tbl, pci->phb->node, 0, 0);
1404 iommu_register_group(pci->table_group,
1405 pci_domain_nr(pci->phb->bus), 0);
1406 pr_debug(" created table: %p\n", pci->table_group);
1408 pr_debug(" found DMA window, table: %p\n", pci->table_group);
1411 set_iommu_table_base(&dev->dev, pci->table_group->tables[0]);
1412 iommu_add_device(pci->table_group, &dev->dev);
1415 static bool iommu_bypass_supported_pSeriesLP(struct pci_dev *pdev, u64 dma_mask)
1417 struct device_node *dn = pci_device_to_OF_node(pdev), *pdn;
1418 const __be32 *dma_window = NULL;
1420 /* only attempt to use a new window if 64-bit DMA is requested */
1421 if (dma_mask < DMA_BIT_MASK(64))
1424 dev_dbg(&pdev->dev, "node is %pOF\n", dn);
1427 * the device tree might contain the dma-window properties
1428 * per-device and not necessarily for the bus. So we need to
1429 * search upwards in the tree until we either hit a dma-window
1430 * property, OR find a parent with a table already allocated.
1432 for (pdn = dn; pdn && PCI_DN(pdn) && !PCI_DN(pdn)->table_group;
1433 pdn = pdn->parent) {
1434 dma_window = of_get_property(pdn, "ibm,dma-window", NULL);
1439 if (pdn && PCI_DN(pdn)) {
1440 pdev->dev.archdata.dma_offset = enable_ddw(pdev, pdn);
1441 if (pdev->dev.archdata.dma_offset)
1448 static int iommu_mem_notifier(struct notifier_block *nb, unsigned long action,
1451 struct direct_window *window;
1452 struct memory_notify *arg = data;
1456 case MEM_GOING_ONLINE:
1457 spin_lock(&direct_window_list_lock);
1458 list_for_each_entry(window, &direct_window_list, list) {
1459 ret |= tce_setrange_multi_pSeriesLP(arg->start_pfn,
1460 arg->nr_pages, window->prop);
1463 spin_unlock(&direct_window_list_lock);
1465 case MEM_CANCEL_ONLINE:
1467 spin_lock(&direct_window_list_lock);
1468 list_for_each_entry(window, &direct_window_list, list) {
1469 ret |= tce_clearrange_multi_pSeriesLP(arg->start_pfn,
1470 arg->nr_pages, window->prop);
1473 spin_unlock(&direct_window_list_lock);
1478 if (ret && action != MEM_CANCEL_ONLINE)
1484 static struct notifier_block iommu_mem_nb = {
1485 .notifier_call = iommu_mem_notifier,
1488 static int iommu_reconfig_notifier(struct notifier_block *nb, unsigned long action, void *data)
1490 int err = NOTIFY_OK;
1491 struct of_reconfig_data *rd = data;
1492 struct device_node *np = rd->dn;
1493 struct pci_dn *pci = PCI_DN(np);
1494 struct direct_window *window;
1497 case OF_RECONFIG_DETACH_NODE:
1499 * Removing the property will invoke the reconfig
1500 * notifier again, which causes dead-lock on the
1501 * read-write semaphore of the notifier chain. So
1502 * we have to remove the property when releasing
1505 remove_ddw(np, false);
1506 if (pci && pci->table_group)
1507 iommu_pseries_free_group(pci->table_group,
1510 spin_lock(&direct_window_list_lock);
1511 list_for_each_entry(window, &direct_window_list, list) {
1512 if (window->device == np) {
1513 list_del(&window->list);
1518 spin_unlock(&direct_window_list_lock);
1527 static struct notifier_block iommu_reconfig_nb = {
1528 .notifier_call = iommu_reconfig_notifier,
1531 /* These are called very early. */
1532 void iommu_init_early_pSeries(void)
1534 if (of_chosen && of_get_property(of_chosen, "linux,iommu-off", NULL))
1537 if (firmware_has_feature(FW_FEATURE_LPAR)) {
1538 pseries_pci_controller_ops.dma_bus_setup = pci_dma_bus_setup_pSeriesLP;
1539 pseries_pci_controller_ops.dma_dev_setup = pci_dma_dev_setup_pSeriesLP;
1541 pseries_pci_controller_ops.iommu_bypass_supported =
1542 iommu_bypass_supported_pSeriesLP;
1544 pseries_pci_controller_ops.dma_bus_setup = pci_dma_bus_setup_pSeries;
1545 pseries_pci_controller_ops.dma_dev_setup = pci_dma_dev_setup_pSeries;
1549 of_reconfig_notifier_register(&iommu_reconfig_nb);
1550 register_memory_notifier(&iommu_mem_nb);
1552 set_pci_dma_ops(&dma_iommu_ops);
1555 static int __init disable_multitce(char *str)
1557 if (strcmp(str, "off") == 0 &&
1558 firmware_has_feature(FW_FEATURE_LPAR) &&
1559 (firmware_has_feature(FW_FEATURE_PUT_TCE_IND) ||
1560 firmware_has_feature(FW_FEATURE_STUFF_TCE))) {
1561 printk(KERN_INFO "Disabling MULTITCE firmware feature\n");
1562 powerpc_firmware_features &=
1563 ~(FW_FEATURE_PUT_TCE_IND | FW_FEATURE_STUFF_TCE);
1568 __setup("multitce=", disable_multitce);
1570 static int tce_iommu_bus_notifier(struct notifier_block *nb,
1571 unsigned long action, void *data)
1573 struct device *dev = data;
1576 case BUS_NOTIFY_DEL_DEVICE:
1577 iommu_del_device(dev);
1584 static struct notifier_block tce_iommu_bus_nb = {
1585 .notifier_call = tce_iommu_bus_notifier,
1588 static int __init tce_iommu_bus_notifier_init(void)
1590 bus_register_notifier(&pci_bus_type, &tce_iommu_bus_nb);
1593 machine_subsys_initcall_sync(pseries, tce_iommu_bus_notifier_init);