xref: /openbmc/linux/arch/powerpc/sysdev/dart_iommu.c (revision e23feb16)
1 /*
2  * arch/powerpc/sysdev/dart_iommu.c
3  *
4  * Copyright (C) 2004 Olof Johansson <olof@lixom.net>, IBM Corporation
5  * Copyright (C) 2005 Benjamin Herrenschmidt <benh@kernel.crashing.org>,
6  *                    IBM Corporation
7  *
8  * Based on pSeries_iommu.c:
9  * Copyright (C) 2001 Mike Corrigan & Dave Engebretsen, IBM Corporation
10  * Copyright (C) 2004 Olof Johansson <olof@lixom.net>, IBM Corporation
11  *
12  * Dynamic DMA mapping support, Apple U3, U4 & IBM CPC925 "DART" iommu.
13  *
14  *
15  * This program is free software; you can redistribute it and/or modify
16  * it under the terms of the GNU General Public License as published by
17  * the Free Software Foundation; either version 2 of the License, or
18  * (at your option) any later version.
19  *
20  * This program is distributed in the hope that it will be useful,
21  * but WITHOUT ANY WARRANTY; without even the implied warranty of
22  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
23  * GNU General Public License for more details.
24  *
25  * You should have received a copy of the GNU General Public License
26  * along with this program; if not, write to the Free Software
27  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
28  */
29 
30 #include <linux/init.h>
31 #include <linux/types.h>
32 #include <linux/mm.h>
33 #include <linux/spinlock.h>
34 #include <linux/string.h>
35 #include <linux/pci.h>
36 #include <linux/dma-mapping.h>
37 #include <linux/vmalloc.h>
38 #include <linux/suspend.h>
39 #include <linux/memblock.h>
40 #include <linux/gfp.h>
41 #include <asm/io.h>
42 #include <asm/prom.h>
43 #include <asm/iommu.h>
44 #include <asm/pci-bridge.h>
45 #include <asm/machdep.h>
46 #include <asm/cacheflush.h>
47 #include <asm/ppc-pci.h>
48 
49 #include "dart.h"
50 
51 /* Physical base address and size of the DART table */
52 unsigned long dart_tablebase; /* exported to htab_initialize */
53 static unsigned long dart_tablesize;
54 
55 /* Virtual base address of the DART table */
56 static u32 *dart_vbase;
57 #ifdef CONFIG_PM
58 static u32 *dart_copy;
59 #endif
60 
61 /* Mapped base address for the dart */
62 static unsigned int __iomem *dart;
63 
64 /* Dummy val that entries are set to when unused */
65 static unsigned int dart_emptyval;
66 
67 static struct iommu_table iommu_table_dart;
68 static int iommu_table_dart_inited;
69 static int dart_dirty;
70 static int dart_is_u4;
71 
72 #define DART_U4_BYPASS_BASE	0x8000000000ull
73 
74 #define DBG(...)
75 
76 static DEFINE_SPINLOCK(invalidate_lock);
77 
78 static inline void dart_tlb_invalidate_all(void)
79 {
80 	unsigned long l = 0;
81 	unsigned int reg, inv_bit;
82 	unsigned long limit;
83 	unsigned long flags;
84 
85 	spin_lock_irqsave(&invalidate_lock, flags);
86 
87 	DBG("dart: flush\n");
88 
89 	/* To invalidate the DART, set the DARTCNTL_FLUSHTLB bit in the
90 	 * control register and wait for it to clear.
91 	 *
92 	 * Gotcha: Sometimes, the DART won't detect that the bit gets
93 	 * set. If so, clear it and set it again.
94 	 */
95 
96 	limit = 0;
97 
98 	inv_bit = dart_is_u4 ? DART_CNTL_U4_FLUSHTLB : DART_CNTL_U3_FLUSHTLB;
99 retry:
100 	l = 0;
101 	reg = DART_IN(DART_CNTL);
102 	reg |= inv_bit;
103 	DART_OUT(DART_CNTL, reg);
104 
105 	while ((DART_IN(DART_CNTL) & inv_bit) && l < (1L << limit))
106 		l++;
107 	if (l == (1L << limit)) {
108 		if (limit < 4) {
109 			limit++;
110 			reg = DART_IN(DART_CNTL);
111 			reg &= ~inv_bit;
112 			DART_OUT(DART_CNTL, reg);
113 			goto retry;
114 		} else
115 			panic("DART: TLB did not flush after waiting a long "
116 			      "time. Buggy U3 ?");
117 	}
118 
119 	spin_unlock_irqrestore(&invalidate_lock, flags);
120 }
121 
122 static inline void dart_tlb_invalidate_one(unsigned long bus_rpn)
123 {
124 	unsigned int reg;
125 	unsigned int l, limit;
126 	unsigned long flags;
127 
128 	spin_lock_irqsave(&invalidate_lock, flags);
129 
130 	reg = DART_CNTL_U4_ENABLE | DART_CNTL_U4_IONE |
131 		(bus_rpn & DART_CNTL_U4_IONE_MASK);
132 	DART_OUT(DART_CNTL, reg);
133 
134 	limit = 0;
135 wait_more:
136 	l = 0;
137 	while ((DART_IN(DART_CNTL) & DART_CNTL_U4_IONE) && l < (1L << limit)) {
138 		rmb();
139 		l++;
140 	}
141 
142 	if (l == (1L << limit)) {
143 		if (limit < 4) {
144 			limit++;
145 			goto wait_more;
146 		} else
147 			panic("DART: TLB did not flush after waiting a long "
148 			      "time. Buggy U4 ?");
149 	}
150 
151 	spin_unlock_irqrestore(&invalidate_lock, flags);
152 }
153 
154 static void dart_flush(struct iommu_table *tbl)
155 {
156 	mb();
157 	if (dart_dirty) {
158 		dart_tlb_invalidate_all();
159 		dart_dirty = 0;
160 	}
161 }
162 
163 static int dart_build(struct iommu_table *tbl, long index,
164 		       long npages, unsigned long uaddr,
165 		       enum dma_data_direction direction,
166 		       struct dma_attrs *attrs)
167 {
168 	unsigned int *dp;
169 	unsigned int rpn;
170 	long l;
171 
172 	DBG("dart: build at: %lx, %lx, addr: %x\n", index, npages, uaddr);
173 
174 	dp = ((unsigned int*)tbl->it_base) + index;
175 
176 	/* On U3, all memory is contiguous, so we can move this
177 	 * out of the loop.
178 	 */
179 	l = npages;
180 	while (l--) {
181 		rpn = __pa(uaddr) >> DART_PAGE_SHIFT;
182 
183 		*(dp++) = DARTMAP_VALID | (rpn & DARTMAP_RPNMASK);
184 
185 		uaddr += DART_PAGE_SIZE;
186 	}
187 
188 	/* make sure all updates have reached memory */
189 	mb();
190 	in_be32((unsigned __iomem *)dp);
191 	mb();
192 
193 	if (dart_is_u4) {
194 		rpn = index;
195 		while (npages--)
196 			dart_tlb_invalidate_one(rpn++);
197 	} else {
198 		dart_dirty = 1;
199 	}
200 	return 0;
201 }
202 
203 
204 static void dart_free(struct iommu_table *tbl, long index, long npages)
205 {
206 	unsigned int *dp;
207 
208 	/* We don't worry about flushing the TLB cache. The only drawback of
209 	 * not doing it is that we won't catch buggy device drivers doing
210 	 * bad DMAs, but then no 32-bit architecture ever does either.
211 	 */
212 
213 	DBG("dart: free at: %lx, %lx\n", index, npages);
214 
215 	dp  = ((unsigned int *)tbl->it_base) + index;
216 
217 	while (npages--)
218 		*(dp++) = dart_emptyval;
219 }
220 
221 
222 static int __init dart_init(struct device_node *dart_node)
223 {
224 	unsigned int i;
225 	unsigned long tmp, base, size;
226 	struct resource r;
227 
228 	if (dart_tablebase == 0 || dart_tablesize == 0) {
229 		printk(KERN_INFO "DART: table not allocated, using "
230 		       "direct DMA\n");
231 		return -ENODEV;
232 	}
233 
234 	if (of_address_to_resource(dart_node, 0, &r))
235 		panic("DART: can't get register base ! ");
236 
237 	/* Make sure nothing from the DART range remains in the CPU cache
238 	 * from a previous mapping that existed before the kernel took
239 	 * over
240 	 */
241 	flush_dcache_phys_range(dart_tablebase,
242 				dart_tablebase + dart_tablesize);
243 
244 	/* Allocate a spare page to map all invalid DART pages. We need to do
245 	 * that to work around what looks like a problem with the HT bridge
246 	 * prefetching into invalid pages and corrupting data
247 	 */
248 	tmp = memblock_alloc(DART_PAGE_SIZE, DART_PAGE_SIZE);
249 	dart_emptyval = DARTMAP_VALID | ((tmp >> DART_PAGE_SHIFT) &
250 					 DARTMAP_RPNMASK);
251 
252 	/* Map in DART registers */
253 	dart = ioremap(r.start, resource_size(&r));
254 	if (dart == NULL)
255 		panic("DART: Cannot map registers!");
256 
257 	/* Map in DART table */
258 	dart_vbase = ioremap(__pa(dart_tablebase), dart_tablesize);
259 
260 	/* Fill initial table */
261 	for (i = 0; i < dart_tablesize/4; i++)
262 		dart_vbase[i] = dart_emptyval;
263 
264 	/* Initialize DART with table base and enable it. */
265 	base = dart_tablebase >> DART_PAGE_SHIFT;
266 	size = dart_tablesize >> DART_PAGE_SHIFT;
267 	if (dart_is_u4) {
268 		size &= DART_SIZE_U4_SIZE_MASK;
269 		DART_OUT(DART_BASE_U4, base);
270 		DART_OUT(DART_SIZE_U4, size);
271 		DART_OUT(DART_CNTL, DART_CNTL_U4_ENABLE);
272 	} else {
273 		size &= DART_CNTL_U3_SIZE_MASK;
274 		DART_OUT(DART_CNTL,
275 			 DART_CNTL_U3_ENABLE |
276 			 (base << DART_CNTL_U3_BASE_SHIFT) |
277 			 (size << DART_CNTL_U3_SIZE_SHIFT));
278 	}
279 
280 	/* Invalidate DART to get rid of possible stale TLBs */
281 	dart_tlb_invalidate_all();
282 
283 	printk(KERN_INFO "DART IOMMU initialized for %s type chipset\n",
284 	       dart_is_u4 ? "U4" : "U3");
285 
286 	return 0;
287 }
288 
289 static void iommu_table_dart_setup(void)
290 {
291 	iommu_table_dart.it_busno = 0;
292 	iommu_table_dart.it_offset = 0;
293 	/* it_size is in number of entries */
294 	iommu_table_dart.it_size = dart_tablesize / sizeof(u32);
295 
296 	/* Initialize the common IOMMU code */
297 	iommu_table_dart.it_base = (unsigned long)dart_vbase;
298 	iommu_table_dart.it_index = 0;
299 	iommu_table_dart.it_blocksize = 1;
300 	iommu_init_table(&iommu_table_dart, -1);
301 
302 	/* Reserve the last page of the DART to avoid possible prefetch
303 	 * past the DART mapped area
304 	 */
305 	set_bit(iommu_table_dart.it_size - 1, iommu_table_dart.it_map);
306 }
307 
308 static void dma_dev_setup_dart(struct device *dev)
309 {
310 	/* We only have one iommu table on the mac for now, which makes
311 	 * things simple. Setup all PCI devices to point to this table
312 	 */
313 	if (get_dma_ops(dev) == &dma_direct_ops)
314 		set_dma_offset(dev, DART_U4_BYPASS_BASE);
315 	else
316 		set_iommu_table_base(dev, &iommu_table_dart);
317 }
318 
319 static void pci_dma_dev_setup_dart(struct pci_dev *dev)
320 {
321 	dma_dev_setup_dart(&dev->dev);
322 }
323 
324 static void pci_dma_bus_setup_dart(struct pci_bus *bus)
325 {
326 	if (!iommu_table_dart_inited) {
327 		iommu_table_dart_inited = 1;
328 		iommu_table_dart_setup();
329 	}
330 }
331 
332 static bool dart_device_on_pcie(struct device *dev)
333 {
334 	struct device_node *np = of_node_get(dev->of_node);
335 
336 	while(np) {
337 		if (of_device_is_compatible(np, "U4-pcie") ||
338 		    of_device_is_compatible(np, "u4-pcie")) {
339 			of_node_put(np);
340 			return true;
341 		}
342 		np = of_get_next_parent(np);
343 	}
344 	return false;
345 }
346 
347 static int dart_dma_set_mask(struct device *dev, u64 dma_mask)
348 {
349 	if (!dev->dma_mask || !dma_supported(dev, dma_mask))
350 		return -EIO;
351 
352 	/* U4 supports a DART bypass, we use it for 64-bit capable
353 	 * devices to improve performances. However, that only works
354 	 * for devices connected to U4 own PCIe interface, not bridged
355 	 * through hypertransport. We need the device to support at
356 	 * least 40 bits of addresses.
357 	 */
358 	if (dart_device_on_pcie(dev) && dma_mask >= DMA_BIT_MASK(40)) {
359 		dev_info(dev, "Using 64-bit DMA iommu bypass\n");
360 		set_dma_ops(dev, &dma_direct_ops);
361 	} else {
362 		dev_info(dev, "Using 32-bit DMA via iommu\n");
363 		set_dma_ops(dev, &dma_iommu_ops);
364 	}
365 	dma_dev_setup_dart(dev);
366 
367 	*dev->dma_mask = dma_mask;
368 	return 0;
369 }
370 
371 void __init iommu_init_early_dart(void)
372 {
373 	struct device_node *dn;
374 
375 	/* Find the DART in the device-tree */
376 	dn = of_find_compatible_node(NULL, "dart", "u3-dart");
377 	if (dn == NULL) {
378 		dn = of_find_compatible_node(NULL, "dart", "u4-dart");
379 		if (dn == NULL)
380 			return;	/* use default direct_dma_ops */
381 		dart_is_u4 = 1;
382 	}
383 
384 	/* Initialize the DART HW */
385 	if (dart_init(dn) != 0)
386 		goto bail;
387 
388 	/* Setup low level TCE operations for the core IOMMU code */
389 	ppc_md.tce_build = dart_build;
390 	ppc_md.tce_free  = dart_free;
391 	ppc_md.tce_flush = dart_flush;
392 
393 	/* Setup bypass if supported */
394 	if (dart_is_u4)
395 		ppc_md.dma_set_mask = dart_dma_set_mask;
396 
397 	ppc_md.pci_dma_dev_setup = pci_dma_dev_setup_dart;
398 	ppc_md.pci_dma_bus_setup = pci_dma_bus_setup_dart;
399 
400 	/* Setup pci_dma ops */
401 	set_pci_dma_ops(&dma_iommu_ops);
402 	return;
403 
404  bail:
405 	/* If init failed, use direct iommu and null setup functions */
406 	ppc_md.pci_dma_dev_setup = NULL;
407 	ppc_md.pci_dma_bus_setup = NULL;
408 
409 	/* Setup pci_dma ops */
410 	set_pci_dma_ops(&dma_direct_ops);
411 }
412 
413 #ifdef CONFIG_PM
414 static void iommu_dart_save(void)
415 {
416 	memcpy(dart_copy, dart_vbase, 2*1024*1024);
417 }
418 
419 static void iommu_dart_restore(void)
420 {
421 	memcpy(dart_vbase, dart_copy, 2*1024*1024);
422 	dart_tlb_invalidate_all();
423 }
424 
425 static int __init iommu_init_late_dart(void)
426 {
427 	unsigned long tbasepfn;
428 	struct page *p;
429 
430 	/* if no dart table exists then we won't need to save it
431 	 * and the area has also not been reserved */
432 	if (!dart_tablebase)
433 		return 0;
434 
435 	tbasepfn = __pa(dart_tablebase) >> PAGE_SHIFT;
436 	register_nosave_region_late(tbasepfn,
437 				    tbasepfn + ((1<<24) >> PAGE_SHIFT));
438 
439 	/* For suspend we need to copy the dart contents because
440 	 * it is not part of the regular mapping (see above) and
441 	 * thus not saved automatically. The memory for this copy
442 	 * must be allocated early because we need 2 MB. */
443 	p = alloc_pages(GFP_KERNEL, 21 - PAGE_SHIFT);
444 	BUG_ON(!p);
445 	dart_copy = page_address(p);
446 
447 	ppc_md.iommu_save = iommu_dart_save;
448 	ppc_md.iommu_restore = iommu_dart_restore;
449 
450 	return 0;
451 }
452 
453 late_initcall(iommu_init_late_dart);
454 #endif
455 
456 void __init alloc_dart_table(void)
457 {
458 	/* Only reserve DART space if machine has more than 1GB of RAM
459 	 * or if requested with iommu=on on cmdline.
460 	 *
461 	 * 1GB of RAM is picked as limit because some default devices
462 	 * (i.e. Airport Extreme) have 30 bit address range limits.
463 	 */
464 
465 	if (iommu_is_off)
466 		return;
467 
468 	if (!iommu_force_on && memblock_end_of_DRAM() <= 0x40000000ull)
469 		return;
470 
471 	/* 512 pages (2MB) is max DART tablesize. */
472 	dart_tablesize = 1UL << 21;
473 	/* 16MB (1 << 24) alignment. We allocate a full 16Mb chuck since we
474 	 * will blow up an entire large page anyway in the kernel mapping
475 	 */
476 	dart_tablebase = (unsigned long)
477 		__va(memblock_alloc_base(1UL<<24, 1UL<<24, 0x80000000L));
478 
479 	printk(KERN_INFO "DART table allocated at: %lx\n", dart_tablebase);
480 }
481