2 * IOMMU implementation for Cell Broadband Processor Architecture
4 * (C) Copyright IBM Corporation 2006
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>
31 #include <asm/iommu.h>
32 #include <asm/machdep.h>
33 #include <asm/pci-bridge.h>
35 #include <asm/of_platform.h>
39 #include "interrupt.h"
41 /* Define CELL_IOMMU_REAL_UNMAP to actually unmap non-used pages
42 * instead of leaving them mapped to some dummy page. This can be
43 * enabled once the appropriate workarounds for spider bugs have
46 #define CELL_IOMMU_REAL_UNMAP
48 /* Define CELL_IOMMU_STRICT_PROTECTION to enforce protection of
49 * IO PTEs based on the transfer direction. That can be enabled
50 * once spider-net has been fixed to pass the correct direction
51 * to the DMA mapping functions
53 #define CELL_IOMMU_STRICT_PROTECTION
58 /* IOC mmap registers */
59 #define IOC_Reg_Size 0x2000
61 #define IOC_IOPT_CacheInvd 0x908
62 #define IOC_IOPT_CacheInvd_NE_Mask 0xffe0000000000000ul
63 #define IOC_IOPT_CacheInvd_IOPTE_Mask 0x000003fffffffff8ul
64 #define IOC_IOPT_CacheInvd_Busy 0x0000000000000001ul
66 #define IOC_IOST_Origin 0x918
67 #define IOC_IOST_Origin_E 0x8000000000000000ul
68 #define IOC_IOST_Origin_HW 0x0000000000000800ul
69 #define IOC_IOST_Origin_HL 0x0000000000000400ul
71 #define IOC_IO_ExcpStat 0x920
72 #define IOC_IO_ExcpStat_V 0x8000000000000000ul
73 #define IOC_IO_ExcpStat_SPF_Mask 0x6000000000000000ul
74 #define IOC_IO_ExcpStat_SPF_S 0x6000000000000000ul
75 #define IOC_IO_ExcpStat_SPF_P 0x4000000000000000ul
76 #define IOC_IO_ExcpStat_ADDR_Mask 0x00000007fffff000ul
77 #define IOC_IO_ExcpStat_RW_Mask 0x0000000000000800ul
78 #define IOC_IO_ExcpStat_IOID_Mask 0x00000000000007fful
80 #define IOC_IO_ExcpMask 0x928
81 #define IOC_IO_ExcpMask_SFE 0x4000000000000000ul
82 #define IOC_IO_ExcpMask_PFE 0x2000000000000000ul
84 #define IOC_IOCmd_Offset 0x1000
86 #define IOC_IOCmd_Cfg 0xc00
87 #define IOC_IOCmd_Cfg_TE 0x0000800000000000ul
90 /* Segment table entries */
91 #define IOSTE_V 0x8000000000000000ul /* valid */
92 #define IOSTE_H 0x4000000000000000ul /* cache hint */
93 #define IOSTE_PT_Base_RPN_Mask 0x3ffffffffffff000ul /* base RPN of IOPT */
94 #define IOSTE_NPPT_Mask 0x0000000000000fe0ul /* no. pages in IOPT */
95 #define IOSTE_PS_Mask 0x0000000000000007ul /* page size */
96 #define IOSTE_PS_4K 0x0000000000000001ul /* - 4kB */
97 #define IOSTE_PS_64K 0x0000000000000003ul /* - 64kB */
98 #define IOSTE_PS_1M 0x0000000000000005ul /* - 1MB */
99 #define IOSTE_PS_16M 0x0000000000000007ul /* - 16MB */
101 /* Page table entries */
102 #define IOPTE_PP_W 0x8000000000000000ul /* protection: write */
103 #define IOPTE_PP_R 0x4000000000000000ul /* protection: read */
104 #define IOPTE_M 0x2000000000000000ul /* coherency required */
105 #define IOPTE_SO_R 0x1000000000000000ul /* ordering: writes */
106 #define IOPTE_SO_RW 0x1800000000000000ul /* ordering: r & w */
107 #define IOPTE_RPN_Mask 0x07fffffffffff000ul /* RPN */
108 #define IOPTE_H 0x0000000000000800ul /* cache hint */
109 #define IOPTE_IOID_Mask 0x00000000000007fful /* ioid */
113 #define IO_SEGMENT_SHIFT 28
114 #define IO_PAGENO_BITS (IO_SEGMENT_SHIFT - IOMMU_PAGE_SHIFT)
116 /* The high bit needs to be set on every DMA address */
117 #define SPIDER_DMA_OFFSET 0x80000000ul
119 struct iommu_window {
120 struct list_head list;
121 struct cbe_iommu *iommu;
122 unsigned long offset;
124 unsigned long pte_offset;
126 struct iommu_table table;
133 void __iomem *xlate_regs;
134 void __iomem *cmd_regs;
138 struct list_head windows;
141 /* Static array of iommus, one per node
142 * each contains a list of windows, keyed from dma_window property
143 * - on bus setup, look for a matching window, or create one
144 * - on dev setup, assign iommu_table ptr
146 static struct cbe_iommu iommus[NR_IOMMUS];
147 static int cbe_nr_iommus;
149 static void invalidate_tce_cache(struct cbe_iommu *iommu, unsigned long *pte,
152 unsigned long *reg, val;
155 reg = iommu->xlate_regs + IOC_IOPT_CacheInvd;
158 /* we can invalidate up to 1 << 11 PTEs at once */
159 n = min(n_ptes, 1l << 11);
160 val = (((n /*- 1*/) << 53) & IOC_IOPT_CacheInvd_NE_Mask)
161 | (__pa(pte) & IOC_IOPT_CacheInvd_IOPTE_Mask)
162 | IOC_IOPT_CacheInvd_Busy;
165 while (in_be64(reg) & IOC_IOPT_CacheInvd_Busy)
173 static void tce_build_cell(struct iommu_table *tbl, long index, long npages,
174 unsigned long uaddr, enum dma_data_direction direction)
177 unsigned long *io_pte, base_pte;
178 struct iommu_window *window =
179 container_of(tbl, struct iommu_window, table);
181 /* implementing proper protection causes problems with the spidernet
182 * driver - check mapping directions later, but allow read & write by
184 #ifdef CELL_IOMMU_STRICT_PROTECTION
185 /* to avoid referencing a global, we use a trick here to setup the
186 * protection bit. "prot" is setup to be 3 fields of 4 bits apprended
187 * together for each of the 3 supported direction values. It is then
188 * shifted left so that the fields matching the desired direction
189 * lands on the appropriate bits, and other bits are masked out.
191 const unsigned long prot = 0xc48;
193 ((prot << (52 + 4 * direction)) & (IOPTE_PP_W | IOPTE_PP_R))
194 | IOPTE_M | IOPTE_SO_RW | (window->ioid & IOPTE_IOID_Mask);
196 base_pte = IOPTE_PP_W | IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW |
197 (window->ioid & IOPTE_IOID_Mask);
200 io_pte = (unsigned long *)tbl->it_base + (index - window->pte_offset);
202 for (i = 0; i < npages; i++, uaddr += IOMMU_PAGE_SIZE)
203 io_pte[i] = base_pte | (__pa(uaddr) & IOPTE_RPN_Mask);
207 invalidate_tce_cache(window->iommu, io_pte, npages);
209 pr_debug("tce_build_cell(index=%lx,n=%lx,dir=%d,base_pte=%lx)\n",
210 index, npages, direction, base_pte);
213 static void tce_free_cell(struct iommu_table *tbl, long index, long npages)
217 unsigned long *io_pte, pte;
218 struct iommu_window *window =
219 container_of(tbl, struct iommu_window, table);
221 pr_debug("tce_free_cell(index=%lx,n=%lx)\n", index, npages);
223 #ifdef CELL_IOMMU_REAL_UNMAP
226 /* spider bridge does PCI reads after freeing - insert a mapping
227 * to a scratch page instead of an invalid entry */
228 pte = IOPTE_PP_R | IOPTE_M | IOPTE_SO_RW | __pa(window->iommu->pad_page)
229 | (window->ioid & IOPTE_IOID_Mask);
232 io_pte = (unsigned long *)tbl->it_base + (index - window->pte_offset);
234 for (i = 0; i < npages; i++)
239 invalidate_tce_cache(window->iommu, io_pte, npages);
242 static irqreturn_t ioc_interrupt(int irq, void *data)
245 struct cbe_iommu *iommu = data;
247 stat = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
249 /* Might want to rate limit it */
250 printk(KERN_ERR "iommu: DMA exception 0x%016lx\n", stat);
251 printk(KERN_ERR " V=%d, SPF=[%c%c], RW=%s, IOID=0x%04x\n",
252 !!(stat & IOC_IO_ExcpStat_V),
253 (stat & IOC_IO_ExcpStat_SPF_S) ? 'S' : ' ',
254 (stat & IOC_IO_ExcpStat_SPF_P) ? 'P' : ' ',
255 (stat & IOC_IO_ExcpStat_RW_Mask) ? "Read" : "Write",
256 (unsigned int)(stat & IOC_IO_ExcpStat_IOID_Mask));
257 printk(KERN_ERR " page=0x%016lx\n",
258 stat & IOC_IO_ExcpStat_ADDR_Mask);
260 /* clear interrupt */
261 stat &= ~IOC_IO_ExcpStat_V;
262 out_be64(iommu->xlate_regs + IOC_IO_ExcpStat, stat);
267 static int cell_iommu_find_ioc(int nid, unsigned long *base)
269 struct device_node *np;
274 /* First look for new style /be nodes */
275 for_each_node_by_name(np, "ioc") {
276 if (of_node_to_nid(np) != nid)
278 if (of_address_to_resource(np, 0, &r)) {
279 printk(KERN_ERR "iommu: can't get address for %s\n",
288 /* Ok, let's try the old way */
289 for_each_node_by_type(np, "cpu") {
290 const unsigned int *nidp;
291 const unsigned long *tmp;
293 nidp = get_property(np, "node-id", NULL);
294 if (nidp && *nidp == nid) {
295 tmp = get_property(np, "ioc-translation", NULL);
307 static void cell_iommu_setup_hardware(struct cbe_iommu *iommu, unsigned long size)
311 unsigned long reg, segments, pages_per_segment, ptab_size, n_pte_pages;
312 unsigned long xlate_base;
315 if (cell_iommu_find_ioc(iommu->nid, &xlate_base))
316 panic("%s: missing IOC register mappings for node %d\n",
317 __FUNCTION__, iommu->nid);
319 iommu->xlate_regs = ioremap(xlate_base, IOC_Reg_Size);
320 iommu->cmd_regs = iommu->xlate_regs + IOC_IOCmd_Offset;
322 segments = size >> IO_SEGMENT_SHIFT;
323 pages_per_segment = 1ull << IO_PAGENO_BITS;
325 pr_debug("%s: iommu[%d]: segments: %lu, pages per segment: %lu\n",
326 __FUNCTION__, iommu->nid, segments, pages_per_segment);
328 /* set up the segment table */
329 page = alloc_pages_node(iommu->nid, GFP_KERNEL, 0);
331 iommu->stab = page_address(page);
332 clear_page(iommu->stab);
334 /* ... and the page tables. Since these are contiguous, we can treat
335 * the page tables as one array of ptes, like pSeries does.
337 ptab_size = segments * pages_per_segment * sizeof(unsigned long);
338 pr_debug("%s: iommu[%d]: ptab_size: %lu, order: %d\n", __FUNCTION__,
339 iommu->nid, ptab_size, get_order(ptab_size));
340 page = alloc_pages_node(iommu->nid, GFP_KERNEL, get_order(ptab_size));
343 iommu->ptab = page_address(page);
344 memset(iommu->ptab, 0, ptab_size);
346 /* allocate a bogus page for the end of each mapping */
347 page = alloc_pages_node(iommu->nid, GFP_KERNEL, 0);
349 iommu->pad_page = page_address(page);
350 clear_page(iommu->pad_page);
352 /* number of pages needed for a page table */
353 n_pte_pages = (pages_per_segment *
354 sizeof(unsigned long)) >> IOMMU_PAGE_SHIFT;
356 pr_debug("%s: iommu[%d]: stab at %p, ptab at %p, n_pte_pages: %lu\n",
357 __FUNCTION__, iommu->nid, iommu->stab, iommu->ptab,
360 /* initialise the STEs */
361 reg = IOSTE_V | ((n_pte_pages - 1) << 5);
363 if (IOMMU_PAGE_SIZE == 0x1000)
365 else if (IOMMU_PAGE_SIZE == 0x10000)
368 extern void __unknown_page_size_error(void);
369 __unknown_page_size_error();
372 pr_debug("Setting up IOMMU stab:\n");
373 for (i = 0; i * (1ul << IO_SEGMENT_SHIFT) < size; i++) {
374 iommu->stab[i] = reg |
375 (__pa(iommu->ptab) + n_pte_pages * IOMMU_PAGE_SIZE * i);
376 pr_debug("\t[%d] 0x%016lx\n", i, iommu->stab[i]);
379 /* ensure that the STEs have updated */
382 /* setup interrupts for the iommu. */
383 reg = in_be64(iommu->xlate_regs + IOC_IO_ExcpStat);
384 out_be64(iommu->xlate_regs + IOC_IO_ExcpStat,
385 reg & ~IOC_IO_ExcpStat_V);
386 out_be64(iommu->xlate_regs + IOC_IO_ExcpMask,
387 IOC_IO_ExcpMask_PFE | IOC_IO_ExcpMask_SFE);
389 virq = irq_create_mapping(NULL,
390 IIC_IRQ_IOEX_ATI | (iommu->nid << IIC_IRQ_NODE_SHIFT));
391 BUG_ON(virq == NO_IRQ);
393 ret = request_irq(virq, ioc_interrupt, IRQF_DISABLED,
397 /* set the IOC segment table origin register (and turn on the iommu) */
398 reg = IOC_IOST_Origin_E | __pa(iommu->stab) | IOC_IOST_Origin_HW;
399 out_be64(iommu->xlate_regs + IOC_IOST_Origin, reg);
400 in_be64(iommu->xlate_regs + IOC_IOST_Origin);
402 /* turn on IO translation */
403 reg = in_be64(iommu->cmd_regs + IOC_IOCmd_Cfg) | IOC_IOCmd_Cfg_TE;
404 out_be64(iommu->cmd_regs + IOC_IOCmd_Cfg, reg);
407 #if 0/* Unused for now */
408 static struct iommu_window *find_window(struct cbe_iommu *iommu,
409 unsigned long offset, unsigned long size)
411 struct iommu_window *window;
413 /* todo: check for overlapping (but not equal) windows) */
415 list_for_each_entry(window, &(iommu->windows), list) {
416 if (window->offset == offset && window->size == size)
424 static struct iommu_window * __init
425 cell_iommu_setup_window(struct cbe_iommu *iommu, struct device_node *np,
426 unsigned long offset, unsigned long size,
427 unsigned long pte_offset)
429 struct iommu_window *window;
430 const unsigned int *ioid;
432 ioid = get_property(np, "ioid", NULL);
434 printk(KERN_WARNING "iommu: missing ioid for %s using 0\n",
437 window = kmalloc_node(sizeof(*window), GFP_KERNEL, iommu->nid);
438 BUG_ON(window == NULL);
440 window->offset = offset;
442 window->ioid = ioid ? *ioid : 0;
443 window->iommu = iommu;
444 window->pte_offset = pte_offset;
446 window->table.it_blocksize = 16;
447 window->table.it_base = (unsigned long)iommu->ptab;
448 window->table.it_index = iommu->nid;
449 window->table.it_offset = (offset >> IOMMU_PAGE_SHIFT) +
451 window->table.it_size = size >> IOMMU_PAGE_SHIFT;
453 iommu_init_table(&window->table, iommu->nid);
455 pr_debug("\tioid %d\n", window->ioid);
456 pr_debug("\tblocksize %ld\n", window->table.it_blocksize);
457 pr_debug("\tbase 0x%016lx\n", window->table.it_base);
458 pr_debug("\toffset 0x%lx\n", window->table.it_offset);
459 pr_debug("\tsize %ld\n", window->table.it_size);
461 list_add(&window->list, &iommu->windows);
466 /* We need to map and reserve the first IOMMU page since it's used
467 * by the spider workaround. In theory, we only need to do that when
468 * running on spider but it doesn't really matter.
470 * This code also assumes that we have a window that starts at 0,
471 * which is the case on all spider based blades.
473 __set_bit(0, window->table.it_map);
474 tce_build_cell(&window->table, window->table.it_offset, 1,
475 (unsigned long)iommu->pad_page, DMA_TO_DEVICE);
476 window->table.it_hint = window->table.it_blocksize;
481 static struct cbe_iommu *cell_iommu_for_node(int nid)
485 for (i = 0; i < cbe_nr_iommus; i++)
486 if (iommus[i].nid == nid)
491 static void cell_dma_dev_setup(struct device *dev)
493 struct iommu_window *window;
494 struct cbe_iommu *iommu;
495 struct dev_archdata *archdata = &dev->archdata;
497 /* If we run without iommu, no need to do anything */
498 if (pci_dma_ops == &dma_direct_ops)
501 /* Current implementation uses the first window available in that
502 * node's iommu. We -might- do something smarter later though it may
505 iommu = cell_iommu_for_node(archdata->numa_node);
506 if (iommu == NULL || list_empty(&iommu->windows)) {
507 printk(KERN_ERR "iommu: missing iommu for %s (node %d)\n",
508 archdata->of_node ? archdata->of_node->full_name : "?",
509 archdata->numa_node);
512 window = list_entry(iommu->windows.next, struct iommu_window, list);
514 archdata->dma_data = &window->table;
517 static void cell_pci_dma_dev_setup(struct pci_dev *dev)
519 cell_dma_dev_setup(&dev->dev);
522 static int cell_of_bus_notify(struct notifier_block *nb, unsigned long action,
525 struct device *dev = data;
527 /* We are only intereted in device addition */
528 if (action != BUS_NOTIFY_ADD_DEVICE)
531 /* We use the PCI DMA ops */
532 dev->archdata.dma_ops = pci_dma_ops;
534 cell_dma_dev_setup(dev);
539 static struct notifier_block cell_of_bus_notifier = {
540 .notifier_call = cell_of_bus_notify
543 static int __init cell_iommu_get_window(struct device_node *np,
547 const void *dma_window;
550 /* Use ibm,dma-window if available, else, hard code ! */
551 dma_window = get_property(np, "ibm,dma-window", NULL);
552 if (dma_window == NULL) {
558 of_parse_dma_window(np, dma_window, &index, base, size);
562 static void __init cell_iommu_init_one(struct device_node *np, unsigned long offset)
564 struct cbe_iommu *iommu;
565 unsigned long base, size;
569 nid = of_node_to_nid(np);
571 printk(KERN_ERR "iommu: failed to get node for %s\n",
575 pr_debug("iommu: setting up iommu for node %d (%s)\n",
578 /* XXX todo: If we can have multiple windows on the same IOMMU, which
579 * isn't the case today, we probably want here to check wether the
580 * iommu for that node is already setup.
581 * However, there might be issue with getting the size right so let's
582 * ignore that for now. We might want to completely get rid of the
583 * multiple window support since the cell iommu supports per-page ioids
586 if (cbe_nr_iommus >= NR_IOMMUS) {
587 printk(KERN_ERR "iommu: too many IOMMUs detected ! (%s)\n",
592 /* Init base fields */
597 snprintf(iommu->name, sizeof(iommu->name), "iommu%d", i);
598 INIT_LIST_HEAD(&iommu->windows);
600 /* Obtain a window for it */
601 cell_iommu_get_window(np, &base, &size);
603 pr_debug("\ttranslating window 0x%lx...0x%lx\n",
604 base, base + size - 1);
606 /* Initialize the hardware */
607 cell_iommu_setup_hardware(iommu, size);
609 /* Setup the iommu_table */
610 cell_iommu_setup_window(iommu, np, base, size,
611 offset >> IOMMU_PAGE_SHIFT);
614 static void __init cell_disable_iommus(void)
617 unsigned long base, val;
618 void __iomem *xregs, *cregs;
620 /* Make sure IOC translation is disabled on all nodes */
621 for_each_online_node(node) {
622 if (cell_iommu_find_ioc(node, &base))
624 xregs = ioremap(base, IOC_Reg_Size);
627 cregs = xregs + IOC_IOCmd_Offset;
629 pr_debug("iommu: cleaning up iommu on node %d\n", node);
631 out_be64(xregs + IOC_IOST_Origin, 0);
632 (void)in_be64(xregs + IOC_IOST_Origin);
633 val = in_be64(cregs + IOC_IOCmd_Cfg);
634 val &= ~IOC_IOCmd_Cfg_TE;
635 out_be64(cregs + IOC_IOCmd_Cfg, val);
636 (void)in_be64(cregs + IOC_IOCmd_Cfg);
642 static int __init cell_iommu_init_disabled(void)
644 struct device_node *np = NULL;
645 unsigned long base = 0, size;
647 /* When no iommu is present, we use direct DMA ops */
648 pci_dma_ops = &dma_direct_ops;
650 /* First make sure all IOC translation is turned off */
651 cell_disable_iommus();
653 /* If we have no Axon, we set up the spider DMA magic offset */
654 if (of_find_node_by_name(NULL, "axon") == NULL)
655 dma_direct_offset = SPIDER_DMA_OFFSET;
657 /* Now we need to check to see where the memory is mapped
658 * in PCI space. We assume that all busses use the same dma
659 * window which is always the case so far on Cell, thus we
660 * pick up the first pci-internal node we can find and check
661 * the DMA window from there.
663 for_each_node_by_name(np, "axon") {
664 if (np->parent == NULL || np->parent->parent != NULL)
666 if (cell_iommu_get_window(np, &base, &size) == 0)
670 for_each_node_by_name(np, "pci-internal") {
671 if (np->parent == NULL || np->parent->parent != NULL)
673 if (cell_iommu_get_window(np, &base, &size) == 0)
679 /* If we found a DMA window, we check if it's big enough to enclose
680 * all of physical memory. If not, we force enable IOMMU
682 if (np && size < lmb_end_of_DRAM()) {
683 printk(KERN_WARNING "iommu: force-enabled, dma window"
684 " (%ldMB) smaller than total memory (%ldMB)\n",
685 size >> 20, lmb_end_of_DRAM() >> 20);
689 dma_direct_offset += base;
691 printk("iommu: disabled, direct DMA offset is 0x%lx\n",
697 static int __init cell_iommu_init(void)
699 struct device_node *np;
701 if (!machine_is(cell))
704 /* If IOMMU is disabled or we have little enough RAM to not need
705 * to enable it, we setup a direct mapping.
707 * Note: should we make sure we have the IOMMU actually disabled ?
710 (!iommu_force_on && lmb_end_of_DRAM() <= 0x80000000ull))
711 if (cell_iommu_init_disabled() == 0)
714 /* Setup various ppc_md. callbacks */
715 ppc_md.pci_dma_dev_setup = cell_pci_dma_dev_setup;
716 ppc_md.tce_build = tce_build_cell;
717 ppc_md.tce_free = tce_free_cell;
719 /* Create an iommu for each /axon node. */
720 for_each_node_by_name(np, "axon") {
721 if (np->parent == NULL || np->parent->parent != NULL)
723 cell_iommu_init_one(np, 0);
726 /* Create an iommu for each toplevel /pci-internal node for
727 * old hardware/firmware
729 for_each_node_by_name(np, "pci-internal") {
730 if (np->parent == NULL || np->parent->parent != NULL)
732 cell_iommu_init_one(np, SPIDER_DMA_OFFSET);
735 /* Setup default PCI iommu ops */
736 pci_dma_ops = &dma_iommu_ops;
739 /* Register callbacks on OF platform device addition/removal
740 * to handle linking them to the right DMA operations
742 bus_register_notifier(&of_platform_bus_type, &cell_of_bus_notifier);
746 arch_initcall(cell_iommu_init);