2 * IOMMU implementation for Cell Broadband Processor Architecture
4 * (C) Copyright IBM Corporation 2006-2008
6 * Author: Jeremy Kerr <jk@ozlabs.org>
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License as published by
10 * the Free Software Foundation; either version 2, or (at your option)
13 * This program is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with this program; if not, write to the Free Software
20 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
25 #include <linux/kernel.h>
26 #include <linux/init.h>
27 #include <linux/interrupt.h>
28 #include <linux/notifier.h>
30 #include <linux/of_platform.h>
33 #include <asm/iommu.h>
34 #include <asm/machdep.h>
35 #include <asm/pci-bridge.h>
38 #include <asm/firmware.h>
39 #include <asm/cell-regs.h>
41 #include "interrupt.h"
43 /* Define CELL_IOMMU_REAL_UNMAP to actually unmap non-used pages
44 * instead of leaving them mapped to some dummy page. This can be
45 * enabled once the appropriate workarounds for spider bugs have
48 #define CELL_IOMMU_REAL_UNMAP
50 /* Define CELL_IOMMU_STRICT_PROTECTION to enforce protection of
51 * IO PTEs based on the transfer direction. That can be enabled
52 * once spider-net has been fixed to pass the correct direction
53 * to the DMA mapping functions
55 #define CELL_IOMMU_STRICT_PROTECTION
60 /* IOC mmap registers */
61 #define IOC_Reg_Size 0x2000
63 #define IOC_IOPT_CacheInvd 0x908
64 #define IOC_IOPT_CacheInvd_NE_Mask 0xffe0000000000000ul
65 #define IOC_IOPT_CacheInvd_IOPTE_Mask 0x000003fffffffff8ul
66 #define IOC_IOPT_CacheInvd_Busy 0x0000000000000001ul
68 #define IOC_IOST_Origin 0x918
69 #define IOC_IOST_Origin_E 0x8000000000000000ul
70 #define IOC_IOST_Origin_HW 0x0000000000000800ul
71 #define IOC_IOST_Origin_HL 0x0000000000000400ul
73 #define IOC_IO_ExcpStat 0x920
74 #define IOC_IO_ExcpStat_V 0x8000000000000000ul
75 #define IOC_IO_ExcpStat_SPF_Mask 0x6000000000000000ul
76 #define IOC_IO_ExcpStat_SPF_S 0x6000000000000000ul
77 #define IOC_IO_ExcpStat_SPF_P 0x4000000000000000ul
78 #define IOC_IO_ExcpStat_ADDR_Mask 0x00000007fffff000ul
79 #define IOC_IO_ExcpStat_RW_Mask 0x0000000000000800ul
80 #define IOC_IO_ExcpStat_IOID_Mask 0x00000000000007fful
82 #define IOC_IO_ExcpMask 0x928
83 #define IOC_IO_ExcpMask_SFE 0x4000000000000000ul
84 #define IOC_IO_ExcpMask_PFE 0x2000000000000000ul
86 #define IOC_IOCmd_Offset 0x1000
88 #define IOC_IOCmd_Cfg 0xc00
89 #define IOC_IOCmd_Cfg_TE 0x0000800000000000ul
92 /* Segment table entries */
93 #define IOSTE_V 0x8000000000000000ul /* valid */
94 #define IOSTE_H 0x4000000000000000ul /* cache hint */
95 #define IOSTE_PT_Base_RPN_Mask 0x3ffffffffffff000ul /* base RPN of IOPT */
96 #define IOSTE_NPPT_Mask 0x0000000000000fe0ul /* no. pages in IOPT */
97 #define IOSTE_PS_Mask 0x0000000000000007ul /* page size */
98 #define IOSTE_PS_4K 0x0000000000000001ul /* - 4kB */
99 #define IOSTE_PS_64K 0x0000000000000003ul /* - 64kB */
100 #define IOSTE_PS_1M 0x0000000000000005ul /* - 1MB */
101 #define IOSTE_PS_16M 0x0000000000000007ul /* - 16MB */
103 /* Page table entries */
104 #define IOPTE_PP_W 0x8000000000000000ul /* protection: write */
105 #define IOPTE_PP_R 0x4000000000000000ul /* protection: read */
106 #define IOPTE_M 0x2000000000000000ul /* coherency required */
107 #define IOPTE_SO_R 0x1000000000000000ul /* ordering: writes */
108 #define IOPTE_SO_RW 0x1800000000000000ul /* ordering: r & w */
109 #define IOPTE_RPN_Mask 0x07fffffffffff000ul /* RPN */
110 #define IOPTE_H 0x0000000000000800ul /* cache hint */
111 #define IOPTE_IOID_Mask 0x00000000000007fful /* ioid */
115 #define IO_SEGMENT_SHIFT 28
116 #define IO_PAGENO_BITS (IO_SEGMENT_SHIFT - IOMMU_PAGE_SHIFT)
118 /* The high bit needs to be set on every DMA address */
119 #define SPIDER_DMA_OFFSET 0x80000000ul
121 struct iommu_window {
122 struct list_head list;
123 struct cbe_iommu *iommu;
124 unsigned long offset;
126 unsigned long pte_offset;
128 struct iommu_table table;
135 void __iomem *xlate_regs;
136 void __iomem *cmd_regs;
140 struct list_head windows;
143 /* Static array of iommus, one per node
144 * each contains a list of windows, keyed from dma_window property
145 * - on bus setup, look for a matching window, or create one
146 * - on dev setup, assign iommu_table ptr
148 static struct cbe_iommu iommus[NR_IOMMUS];
149 static int cbe_nr_iommus;
151 static void invalidate_tce_cache(struct cbe_iommu *iommu, unsigned long *pte,
154 unsigned long __iomem *reg;
158 reg = iommu->xlate_regs + IOC_IOPT_CacheInvd;
161 /* we can invalidate up to 1 << 11 PTEs at once */
162 n = min(n_ptes, 1l << 11);
163 val = (((n /*- 1*/) << 53) & IOC_IOPT_CacheInvd_NE_Mask)
164 | (__pa(pte) & IOC_IOPT_CacheInvd_IOPTE_Mask)
165 | IOC_IOPT_CacheInvd_Busy;
168 while (in_be64(reg) & IOC_IOPT_CacheInvd_Busy)
176 static void tce_build_cell(struct iommu_table *tbl, long index, long npages,
177 unsigned long uaddr, enum dma_data_direction direction)
180 unsigned long *io_pte, base_pte;
181 struct iommu_window *window =
182 container_of(tbl, struct iommu_window, table);
184 /* implementing proper protection causes problems with the spidernet
185 * driver - check mapping directions later, but allow read & write by
187 #ifdef CELL_IOMMU_STRICT_PROTECTION
188 /* to avoid referencing a global, we use a trick here to setup the
189 * protection bit. "prot" is setup to be 3 fields of 4 bits apprended
190 * together for each of the 3 supported direction values. It is then
191 * shifted left so that the fields matching the desired direction
192 * lands on the appropriate bits, and other bits are masked out.
194 const unsigned long prot = 0xc48;
196 ((prot << (52 + 4 * direction)) & (IOPTE_PP_W | IOPTE_PP_R))
197 | IOPTE_M | IOPTE_SO_RW | (window->ioid & IOPTE_IOID_Mask);
199 base_pte = IOPTE_PP_W | IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW |
200 (window->ioid & IOPTE_IOID_Mask);
203 io_pte = (unsigned long *)tbl->it_base + (index - window->pte_offset);
205 for (i = 0; i < npages; i++, uaddr += IOMMU_PAGE_SIZE)
206 io_pte[i] = base_pte | (__pa(uaddr) & IOPTE_RPN_Mask);
210 invalidate_tce_cache(window->iommu, io_pte, npages);
212 pr_debug("tce_build_cell(index=%lx,n=%lx,dir=%d,base_pte=%lx)\n",
213 index, npages, direction, base_pte);
216 static void tce_free_cell(struct iommu_table *tbl, long index, long npages)
220 unsigned long *io_pte, pte;
221 struct iommu_window *window =
222 container_of(tbl, struct iommu_window, table);
224 pr_debug("tce_free_cell(index=%lx,n=%lx)\n", index, npages);
226 #ifdef CELL_IOMMU_REAL_UNMAP
229 /* spider bridge does PCI reads after freeing - insert a mapping
230 * to a scratch page instead of an invalid entry */
231 pte = IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW | __pa(window->iommu->pad_page)
232 | (window->ioid & IOPTE_IOID_Mask);
235 io_pte = (unsigned long *)tbl->it_base + (index - window->pte_offset);
237 for (i = 0; i < npages; i++)
242 invalidate_tce_cache(window->iommu, io_pte, npages);
245 static irqreturn_t ioc_interrupt(int irq, void *data)
248 struct cbe_iommu *iommu = data;
250 stat = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
252 /* Might want to rate limit it */
253 printk(KERN_ERR "iommu: DMA exception 0x%016lx\n", stat);
254 printk(KERN_ERR " V=%d, SPF=[%c%c], RW=%s, IOID=0x%04x\n",
255 !!(stat & IOC_IO_ExcpStat_V),
256 (stat & IOC_IO_ExcpStat_SPF_S) ? 'S' : ' ',
257 (stat & IOC_IO_ExcpStat_SPF_P) ? 'P' : ' ',
258 (stat & IOC_IO_ExcpStat_RW_Mask) ? "Read" : "Write",
259 (unsigned int)(stat & IOC_IO_ExcpStat_IOID_Mask));
260 printk(KERN_ERR " page=0x%016lx\n",
261 stat & IOC_IO_ExcpStat_ADDR_Mask);
263 /* clear interrupt */
264 stat &= ~IOC_IO_ExcpStat_V;
265 out_be64(iommu->xlate_regs + IOC_IO_ExcpStat, stat);
270 static int cell_iommu_find_ioc(int nid, unsigned long *base)
272 struct device_node *np;
277 /* First look for new style /be nodes */
278 for_each_node_by_name(np, "ioc") {
279 if (of_node_to_nid(np) != nid)
281 if (of_address_to_resource(np, 0, &r)) {
282 printk(KERN_ERR "iommu: can't get address for %s\n",
291 /* Ok, let's try the old way */
292 for_each_node_by_type(np, "cpu") {
293 const unsigned int *nidp;
294 const unsigned long *tmp;
296 nidp = of_get_property(np, "node-id", NULL);
297 if (nidp && *nidp == nid) {
298 tmp = of_get_property(np, "ioc-translation", NULL);
310 static void cell_iommu_setup_page_tables(struct cbe_iommu *iommu,
311 unsigned long dbase, unsigned long dsize,
312 unsigned long fbase, unsigned long fsize)
316 unsigned long reg, segments, pages_per_segment, ptab_size, stab_size,
321 base = min(fbase, dbase);
323 segments = max(dbase + dsize, fbase + fsize) >> IO_SEGMENT_SHIFT;
324 pages_per_segment = 1ull << IO_PAGENO_BITS;
326 pr_debug("%s: iommu[%d]: segments: %lu, pages per segment: %lu\n",
327 __FUNCTION__, iommu->nid, segments, pages_per_segment);
329 /* set up the segment table */
330 stab_size = segments * sizeof(unsigned long);
331 page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(stab_size));
333 iommu->stab = page_address(page);
334 clear_page(iommu->stab);
336 /* ... and the page tables. Since these are contiguous, we can treat
337 * the page tables as one array of ptes, like pSeries does.
339 ptab_size = segments * pages_per_segment * sizeof(unsigned long);
340 pr_debug("%s: iommu[%d]: ptab_size: %lu, order: %d\n", __FUNCTION__,
341 iommu->nid, ptab_size, get_order(ptab_size));
342 page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(ptab_size));
345 iommu->ptab = page_address(page);
346 memset(iommu->ptab, 0, ptab_size);
348 /* allocate a bogus page for the end of each mapping */
349 page = alloc_pages_node(iommu->nid, GFP_KERNEL, 0);
351 iommu->pad_page = page_address(page);
352 clear_page(iommu->pad_page);
354 /* number of pages needed for a page table */
355 n_pte_pages = (pages_per_segment *
356 sizeof(unsigned long)) >> IOMMU_PAGE_SHIFT;
358 pr_debug("%s: iommu[%d]: stab at %p, ptab at %p, n_pte_pages: %lu\n",
359 __FUNCTION__, iommu->nid, iommu->stab, iommu->ptab,
362 /* initialise the STEs */
363 reg = IOSTE_V | ((n_pte_pages - 1) << 5);
365 if (IOMMU_PAGE_SIZE == 0x1000)
367 else if (IOMMU_PAGE_SIZE == 0x10000)
370 extern void __unknown_page_size_error(void);
371 __unknown_page_size_error();
374 pr_debug("Setting up IOMMU stab:\n");
375 for (i = base >> IO_SEGMENT_SHIFT; i < segments; i++) {
376 iommu->stab[i] = reg |
377 (__pa(iommu->ptab) + n_pte_pages * IOMMU_PAGE_SIZE * i);
378 pr_debug("\t[%d] 0x%016lx\n", i, iommu->stab[i]);
382 static void cell_iommu_enable_hardware(struct cbe_iommu *iommu)
385 unsigned long reg, xlate_base;
388 if (cell_iommu_find_ioc(iommu->nid, &xlate_base))
389 panic("%s: missing IOC register mappings for node %d\n",
390 __FUNCTION__, iommu->nid);
392 iommu->xlate_regs = ioremap(xlate_base, IOC_Reg_Size);
393 iommu->cmd_regs = iommu->xlate_regs + IOC_IOCmd_Offset;
395 /* ensure that the STEs have updated */
398 /* setup interrupts for the iommu. */
399 reg = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
400 out_be64(iommu->xlate_regs + IOC_IO_ExcpStat,
401 reg & ~IOC_IO_ExcpStat_V);
402 out_be64(iommu->xlate_regs + IOC_IO_ExcpMask,
403 IOC_IO_ExcpMask_PFE | IOC_IO_ExcpMask_SFE);
405 virq = irq_create_mapping(NULL,
406 IIC_IRQ_IOEX_ATI | (iommu->nid << IIC_IRQ_NODE_SHIFT));
407 BUG_ON(virq == NO_IRQ);
409 ret = request_irq(virq, ioc_interrupt, IRQF_DISABLED,
413 /* set the IOC segment table origin register (and turn on the iommu) */
414 reg = IOC_IOST_Origin_E | __pa(iommu->stab) | IOC_IOST_Origin_HW;
415 out_be64(iommu->xlate_regs + IOC_IOST_Origin, reg);
416 in_be64(iommu->xlate_regs + IOC_IOST_Origin);
418 /* turn on IO translation */
419 reg = in_be64(iommu->cmd_regs + IOC_IOCmd_Cfg) | IOC_IOCmd_Cfg_TE;
420 out_be64(iommu->cmd_regs + IOC_IOCmd_Cfg, reg);
423 static void cell_iommu_setup_hardware(struct cbe_iommu *iommu,
424 unsigned long base, unsigned long size)
426 cell_iommu_setup_page_tables(iommu, base, size, 0, 0);
427 cell_iommu_enable_hardware(iommu);
430 #if 0/* Unused for now */
431 static struct iommu_window *find_window(struct cbe_iommu *iommu,
432 unsigned long offset, unsigned long size)
434 struct iommu_window *window;
436 /* todo: check for overlapping (but not equal) windows) */
438 list_for_each_entry(window, &(iommu->windows), list) {
439 if (window->offset == offset && window->size == size)
447 static inline u32 cell_iommu_get_ioid(struct device_node *np)
451 ioid = of_get_property(np, "ioid", NULL);
453 printk(KERN_WARNING "iommu: missing ioid for %s using 0\n",
461 static struct iommu_window * __init
462 cell_iommu_setup_window(struct cbe_iommu *iommu, struct device_node *np,
463 unsigned long offset, unsigned long size,
464 unsigned long pte_offset)
466 struct iommu_window *window;
469 ioid = cell_iommu_get_ioid(np);
471 window = kmalloc_node(sizeof(*window), GFP_KERNEL, iommu->nid);
472 BUG_ON(window == NULL);
474 window->offset = offset;
477 window->iommu = iommu;
478 window->pte_offset = pte_offset;
480 window->table.it_blocksize = 16;
481 window->table.it_base = (unsigned long)iommu->ptab;
482 window->table.it_index = iommu->nid;
483 window->table.it_offset = (offset >> IOMMU_PAGE_SHIFT) +
485 window->table.it_size = size >> IOMMU_PAGE_SHIFT;
487 iommu_init_table(&window->table, iommu->nid);
489 pr_debug("\tioid %d\n", window->ioid);
490 pr_debug("\tblocksize %ld\n", window->table.it_blocksize);
491 pr_debug("\tbase 0x%016lx\n", window->table.it_base);
492 pr_debug("\toffset 0x%lx\n", window->table.it_offset);
493 pr_debug("\tsize %ld\n", window->table.it_size);
495 list_add(&window->list, &iommu->windows);
500 /* We need to map and reserve the first IOMMU page since it's used
501 * by the spider workaround. In theory, we only need to do that when
502 * running on spider but it doesn't really matter.
504 * This code also assumes that we have a window that starts at 0,
505 * which is the case on all spider based blades.
507 __set_bit(0, window->table.it_map);
508 tce_build_cell(&window->table, window->table.it_offset, 1,
509 (unsigned long)iommu->pad_page, DMA_TO_DEVICE);
510 window->table.it_hint = window->table.it_blocksize;
515 static struct cbe_iommu *cell_iommu_for_node(int nid)
519 for (i = 0; i < cbe_nr_iommus; i++)
520 if (iommus[i].nid == nid)
525 static unsigned long cell_dma_direct_offset;
527 static unsigned long dma_iommu_fixed_base;
528 struct dma_mapping_ops dma_iommu_fixed_ops;
530 static void cell_dma_dev_setup_iommu(struct device *dev)
532 struct iommu_window *window;
533 struct cbe_iommu *iommu;
534 struct dev_archdata *archdata = &dev->archdata;
536 /* Current implementation uses the first window available in that
537 * node's iommu. We -might- do something smarter later though it may
540 iommu = cell_iommu_for_node(archdata->numa_node);
541 if (iommu == NULL || list_empty(&iommu->windows)) {
542 printk(KERN_ERR "iommu: missing iommu for %s (node %d)\n",
543 archdata->of_node ? archdata->of_node->full_name : "?",
544 archdata->numa_node);
547 window = list_entry(iommu->windows.next, struct iommu_window, list);
549 archdata->dma_data = &window->table;
552 static void cell_dma_dev_setup_static(struct device *dev);
554 static void cell_dma_dev_setup(struct device *dev)
556 struct dev_archdata *archdata = &dev->archdata;
558 /* Order is important here, these are not mutually exclusive */
559 if (get_dma_ops(dev) == &dma_iommu_fixed_ops)
560 cell_dma_dev_setup_static(dev);
561 else if (get_pci_dma_ops() == &dma_iommu_ops)
562 cell_dma_dev_setup_iommu(dev);
563 else if (get_pci_dma_ops() == &dma_direct_ops)
564 archdata->dma_data = (void *)cell_dma_direct_offset;
569 static void cell_pci_dma_dev_setup(struct pci_dev *dev)
571 cell_dma_dev_setup(&dev->dev);
574 static int cell_of_bus_notify(struct notifier_block *nb, unsigned long action,
577 struct device *dev = data;
579 /* We are only intereted in device addition */
580 if (action != BUS_NOTIFY_ADD_DEVICE)
583 /* We use the PCI DMA ops */
584 dev->archdata.dma_ops = get_pci_dma_ops();
586 cell_dma_dev_setup(dev);
591 static struct notifier_block cell_of_bus_notifier = {
592 .notifier_call = cell_of_bus_notify
595 static int __init cell_iommu_get_window(struct device_node *np,
599 const void *dma_window;
602 /* Use ibm,dma-window if available, else, hard code ! */
603 dma_window = of_get_property(np, "ibm,dma-window", NULL);
604 if (dma_window == NULL) {
610 of_parse_dma_window(np, dma_window, &index, base, size);
614 static struct cbe_iommu * __init cell_iommu_alloc(struct device_node *np)
616 struct cbe_iommu *iommu;
620 nid = of_node_to_nid(np);
622 printk(KERN_ERR "iommu: failed to get node for %s\n",
626 pr_debug("iommu: setting up iommu for node %d (%s)\n",
629 /* XXX todo: If we can have multiple windows on the same IOMMU, which
630 * isn't the case today, we probably want here to check wether the
631 * iommu for that node is already setup.
632 * However, there might be issue with getting the size right so let's
633 * ignore that for now. We might want to completely get rid of the
634 * multiple window support since the cell iommu supports per-page ioids
637 if (cbe_nr_iommus >= NR_IOMMUS) {
638 printk(KERN_ERR "iommu: too many IOMMUs detected ! (%s)\n",
643 /* Init base fields */
648 snprintf(iommu->name, sizeof(iommu->name), "iommu%d", i);
649 INIT_LIST_HEAD(&iommu->windows);
654 static void __init cell_iommu_init_one(struct device_node *np,
655 unsigned long offset)
657 struct cbe_iommu *iommu;
658 unsigned long base, size;
660 iommu = cell_iommu_alloc(np);
664 /* Obtain a window for it */
665 cell_iommu_get_window(np, &base, &size);
667 pr_debug("\ttranslating window 0x%lx...0x%lx\n",
668 base, base + size - 1);
670 /* Initialize the hardware */
671 cell_iommu_setup_hardware(iommu, base, size);
673 /* Setup the iommu_table */
674 cell_iommu_setup_window(iommu, np, base, size,
675 offset >> IOMMU_PAGE_SHIFT);
678 static void __init cell_disable_iommus(void)
681 unsigned long base, val;
682 void __iomem *xregs, *cregs;
684 /* Make sure IOC translation is disabled on all nodes */
685 for_each_online_node(node) {
686 if (cell_iommu_find_ioc(node, &base))
688 xregs = ioremap(base, IOC_Reg_Size);
691 cregs = xregs + IOC_IOCmd_Offset;
693 pr_debug("iommu: cleaning up iommu on node %d\n", node);
695 out_be64(xregs + IOC_IOST_Origin, 0);
696 (void)in_be64(xregs + IOC_IOST_Origin);
697 val = in_be64(cregs + IOC_IOCmd_Cfg);
698 val &= ~IOC_IOCmd_Cfg_TE;
699 out_be64(cregs + IOC_IOCmd_Cfg, val);
700 (void)in_be64(cregs + IOC_IOCmd_Cfg);
706 static int __init cell_iommu_init_disabled(void)
708 struct device_node *np = NULL;
709 unsigned long base = 0, size;
711 /* When no iommu is present, we use direct DMA ops */
712 set_pci_dma_ops(&dma_direct_ops);
714 /* First make sure all IOC translation is turned off */
715 cell_disable_iommus();
717 /* If we have no Axon, we set up the spider DMA magic offset */
718 if (of_find_node_by_name(NULL, "axon") == NULL)
719 cell_dma_direct_offset = SPIDER_DMA_OFFSET;
721 /* Now we need to check to see where the memory is mapped
722 * in PCI space. We assume that all busses use the same dma
723 * window which is always the case so far on Cell, thus we
724 * pick up the first pci-internal node we can find and check
725 * the DMA window from there.
727 for_each_node_by_name(np, "axon") {
728 if (np->parent == NULL || np->parent->parent != NULL)
730 if (cell_iommu_get_window(np, &base, &size) == 0)
734 for_each_node_by_name(np, "pci-internal") {
735 if (np->parent == NULL || np->parent->parent != NULL)
737 if (cell_iommu_get_window(np, &base, &size) == 0)
743 /* If we found a DMA window, we check if it's big enough to enclose
744 * all of physical memory. If not, we force enable IOMMU
746 if (np && size < lmb_end_of_DRAM()) {
747 printk(KERN_WARNING "iommu: force-enabled, dma window"
748 " (%ldMB) smaller than total memory (%ldMB)\n",
749 size >> 20, lmb_end_of_DRAM() >> 20);
753 cell_dma_direct_offset += base;
755 if (cell_dma_direct_offset != 0)
756 ppc_md.pci_dma_dev_setup = cell_pci_dma_dev_setup;
758 printk("iommu: disabled, direct DMA offset is 0x%lx\n",
759 cell_dma_direct_offset);
765 * Fixed IOMMU mapping support
767 * This code adds support for setting up a fixed IOMMU mapping on certain
768 * cell machines. For 64-bit devices this avoids the performance overhead of
769 * mapping and unmapping pages at runtime. 32-bit devices are unable to use
772 * The fixed mapping is established at boot, and maps all of physical memory
773 * 1:1 into device space at some offset. On machines with < 30 GB of memory
774 * we setup the fixed mapping immediately above the normal IOMMU window.
776 * For example a machine with 4GB of memory would end up with the normal
777 * IOMMU window from 0-2GB and the fixed mapping window from 2GB to 6GB. In
778 * this case a 64-bit device wishing to DMA to 1GB would be told to DMA to
779 * 3GB, plus any offset required by firmware. The firmware offset is encoded
780 * in the "dma-ranges" property.
782 * On machines with 30GB or more of memory, we are unable to place the fixed
783 * mapping above the normal IOMMU window as we would run out of address space.
784 * Instead we move the normal IOMMU window to coincide with the hash page
785 * table, this region does not need to be part of the fixed mapping as no
786 * device should ever be DMA'ing to it. We then setup the fixed mapping
790 static u64 cell_iommu_get_fixed_address(struct device *dev)
792 u64 cpu_addr, size, best_size, pci_addr = OF_BAD_ADDR;
793 struct device_node *np;
794 const u32 *ranges = NULL;
797 np = of_node_get(dev->archdata.of_node);
799 ranges = of_get_property(np, "dma-ranges", &len);
802 np = of_get_next_parent(np);
806 dev_dbg(dev, "iommu: no dma-ranges found\n");
812 /* dma-ranges format:
814 * 2 cells: pci address
815 * 2 cells: parent address
818 for (i = 0, best = -1, best_size = 0; i < len; i += 7) {
819 cpu_addr = of_translate_dma_address(np, ranges +i + 3);
820 size = of_read_number(ranges + i + 5, 2);
822 if (cpu_addr == 0 && size > best_size) {
829 pci_addr = of_read_number(ranges + best + 1, 2);
831 dev_dbg(dev, "iommu: no suitable range found!\n");
839 static int dma_set_mask_and_switch(struct device *dev, u64 dma_mask)
841 if (!dev->dma_mask || !dma_supported(dev, dma_mask))
844 if (dma_mask == DMA_BIT_MASK(64) &&
845 cell_iommu_get_fixed_address(dev) != OF_BAD_ADDR)
847 dev_dbg(dev, "iommu: 64-bit OK, using fixed ops\n");
848 set_dma_ops(dev, &dma_iommu_fixed_ops);
850 dev_dbg(dev, "iommu: not 64-bit, using default ops\n");
851 set_dma_ops(dev, get_pci_dma_ops());
854 cell_dma_dev_setup(dev);
856 *dev->dma_mask = dma_mask;
861 static void cell_dma_dev_setup_static(struct device *dev)
863 struct dev_archdata *archdata = &dev->archdata;
866 addr = cell_iommu_get_fixed_address(dev) + dma_iommu_fixed_base;
867 archdata->dma_data = (void *)addr;
869 dev_dbg(dev, "iommu: fixed addr = %lx\n", addr);
872 static void cell_iommu_setup_fixed_ptab(struct cbe_iommu *iommu,
873 struct device_node *np, unsigned long dbase, unsigned long dsize,
874 unsigned long fbase, unsigned long fsize)
876 unsigned long base_pte, uaddr, *io_pte;
879 dma_iommu_fixed_base = fbase;
881 /* convert from bytes into page table indices */
882 dbase = dbase >> IOMMU_PAGE_SHIFT;
883 dsize = dsize >> IOMMU_PAGE_SHIFT;
884 fbase = fbase >> IOMMU_PAGE_SHIFT;
885 fsize = fsize >> IOMMU_PAGE_SHIFT;
887 pr_debug("iommu: mapping 0x%lx pages from 0x%lx\n", fsize, fbase);
889 io_pte = iommu->ptab;
890 base_pte = IOPTE_PP_W | IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW
891 | (cell_iommu_get_ioid(np) & IOPTE_IOID_Mask);
894 for (i = fbase; i < fbase + fsize; i++, uaddr += IOMMU_PAGE_SIZE) {
895 /* Don't touch the dynamic region */
896 if (i >= dbase && i < (dbase + dsize)) {
897 pr_debug("iommu: static/dynamic overlap, skipping\n");
900 io_pte[i] = base_pte | (__pa(uaddr) & IOPTE_RPN_Mask);
906 static int __init cell_iommu_fixed_mapping_init(void)
908 unsigned long dbase, dsize, fbase, fsize, hbase, hend;
909 struct cbe_iommu *iommu;
910 struct device_node *np;
912 /* The fixed mapping is only supported on axon machines */
913 np = of_find_node_by_name(NULL, "axon");
915 pr_debug("iommu: fixed mapping disabled, no axons found\n");
919 /* We must have dma-ranges properties for fixed mapping to work */
920 for (np = NULL; (np = of_find_all_nodes(np));) {
921 if (of_find_property(np, "dma-ranges", NULL))
927 pr_debug("iommu: no dma-ranges found, no fixed mapping\n");
931 /* The default setup is to have the fixed mapping sit after the
932 * dynamic region, so find the top of the largest IOMMU window
933 * on any axon, then add the size of RAM and that's our max value.
934 * If that is > 32GB we have to do other shennanigans.
937 for_each_node_by_name(np, "axon") {
938 cell_iommu_get_window(np, &dbase, &dsize);
939 fbase = max(fbase, dbase + dsize);
942 fbase = _ALIGN_UP(fbase, 1 << IO_SEGMENT_SHIFT);
943 fsize = lmb_phys_mem_size();
945 if ((fbase + fsize) <= 0x800000000)
946 hbase = 0; /* use the device tree window */
948 /* If we're over 32 GB we need to cheat. We can't map all of
949 * RAM with the fixed mapping, and also fit the dynamic
950 * region. So try to place the dynamic region where the hash
951 * table sits, drivers never need to DMA to it, we don't
952 * need a fixed mapping for that area.
955 pr_debug("iommu: htab is NULL, on LPAR? Huh?\n");
958 hbase = __pa(htab_address);
959 hend = hbase + htab_size_bytes;
961 /* The window must start and end on a segment boundary */
962 if ((hbase != _ALIGN_UP(hbase, 1 << IO_SEGMENT_SHIFT)) ||
963 (hend != _ALIGN_UP(hend, 1 << IO_SEGMENT_SHIFT))) {
964 pr_debug("iommu: hash window not segment aligned\n");
968 /* Check the hash window fits inside the real DMA window */
969 for_each_node_by_name(np, "axon") {
970 cell_iommu_get_window(np, &dbase, &dsize);
972 if (hbase < dbase || (hend > (dbase + dsize))) {
973 pr_debug("iommu: hash window doesn't fit in"
974 "real DMA window\n");
982 /* Setup the dynamic regions */
983 for_each_node_by_name(np, "axon") {
984 iommu = cell_iommu_alloc(np);
988 cell_iommu_get_window(np, &dbase, &dsize);
991 dsize = htab_size_bytes;
994 printk(KERN_DEBUG "iommu: node %d, dynamic window 0x%lx-0x%lx "
995 "fixed window 0x%lx-0x%lx\n", iommu->nid, dbase,
996 dbase + dsize, fbase, fbase + fsize);
998 cell_iommu_setup_page_tables(iommu, dbase, dsize, fbase, fsize);
999 cell_iommu_setup_fixed_ptab(iommu, np, dbase, dsize,
1001 cell_iommu_enable_hardware(iommu);
1002 cell_iommu_setup_window(iommu, np, dbase, dsize, 0);
1005 dma_iommu_fixed_ops = dma_direct_ops;
1006 dma_iommu_fixed_ops.set_dma_mask = dma_set_mask_and_switch;
1008 dma_iommu_ops.set_dma_mask = dma_set_mask_and_switch;
1009 set_pci_dma_ops(&dma_iommu_ops);
1014 static int iommu_fixed_disabled;
1016 static int __init setup_iommu_fixed(char *str)
1018 if (strcmp(str, "off") == 0)
1019 iommu_fixed_disabled = 1;
1023 __setup("iommu_fixed=", setup_iommu_fixed);
1025 static int __init cell_iommu_init(void)
1027 struct device_node *np;
1029 /* If IOMMU is disabled or we have little enough RAM to not need
1030 * to enable it, we setup a direct mapping.
1032 * Note: should we make sure we have the IOMMU actually disabled ?
1035 (!iommu_force_on && lmb_end_of_DRAM() <= 0x80000000ull))
1036 if (cell_iommu_init_disabled() == 0)
1039 /* Setup various ppc_md. callbacks */
1040 ppc_md.pci_dma_dev_setup = cell_pci_dma_dev_setup;
1041 ppc_md.tce_build = tce_build_cell;
1042 ppc_md.tce_free = tce_free_cell;
1044 if (!iommu_fixed_disabled && cell_iommu_fixed_mapping_init() == 0)
1047 /* Create an iommu for each /axon node. */
1048 for_each_node_by_name(np, "axon") {
1049 if (np->parent == NULL || np->parent->parent != NULL)
1051 cell_iommu_init_one(np, 0);
1054 /* Create an iommu for each toplevel /pci-internal node for
1055 * old hardware/firmware
1057 for_each_node_by_name(np, "pci-internal") {
1058 if (np->parent == NULL || np->parent->parent != NULL)
1060 cell_iommu_init_one(np, SPIDER_DMA_OFFSET);
1063 /* Setup default PCI iommu ops */
1064 set_pci_dma_ops(&dma_iommu_ops);
1067 /* Register callbacks on OF platform device addition/removal
1068 * to handle linking them to the right DMA operations
1070 bus_register_notifier(&of_platform_bus_type, &cell_of_bus_notifier);
1074 machine_arch_initcall(cell, cell_iommu_init);
1075 machine_arch_initcall(celleb_native, cell_iommu_init);