page allocator: do not check NUMA node ID when the caller knows the node is valid
[deliverable/linux.git] / arch / ia64 / sn / pci / pci_dma.c
1 /*
2 * This file is subject to the terms and conditions of the GNU General Public
3 * License. See the file "COPYING" in the main directory of this archive
4 * for more details.
5 *
6 * Copyright (C) 2000,2002-2005 Silicon Graphics, Inc. All rights reserved.
7 *
8 * Routines for PCI DMA mapping. See Documentation/DMA-API.txt for
9 * a description of how these routines should be used.
10 */
11
12 #include <linux/module.h>
13 #include <linux/dma-mapping.h>
14 #include <asm/dma.h>
15 #include <asm/sn/intr.h>
16 #include <asm/sn/pcibus_provider_defs.h>
17 #include <asm/sn/pcidev.h>
18 #include <asm/sn/sn_sal.h>
19
20 #define SG_ENT_VIRT_ADDRESS(sg) (sg_virt((sg)))
21 #define SG_ENT_PHYS_ADDRESS(SG) virt_to_phys(SG_ENT_VIRT_ADDRESS(SG))
22
23 /**
24 * sn_dma_supported - test a DMA mask
25 * @dev: device to test
26 * @mask: DMA mask to test
27 *
28 * Return whether the given PCI device DMA address mask can be supported
29 * properly. For example, if your device can only drive the low 24-bits
30 * during PCI bus mastering, then you would pass 0x00ffffff as the mask to
31 * this function. Of course, SN only supports devices that have 32 or more
32 * address bits when using the PMU.
33 */
34 static int sn_dma_supported(struct device *dev, u64 mask)
35 {
36 BUG_ON(dev->bus != &pci_bus_type);
37
38 if (mask < 0x7fffffff)
39 return 0;
40 return 1;
41 }
42
43 /**
44 * sn_dma_set_mask - set the DMA mask
45 * @dev: device to set
46 * @dma_mask: new mask
47 *
48 * Set @dev's DMA mask if the hw supports it.
49 */
50 int sn_dma_set_mask(struct device *dev, u64 dma_mask)
51 {
52 BUG_ON(dev->bus != &pci_bus_type);
53
54 if (!sn_dma_supported(dev, dma_mask))
55 return 0;
56
57 *dev->dma_mask = dma_mask;
58 return 1;
59 }
60 EXPORT_SYMBOL(sn_dma_set_mask);
61
62 /**
63 * sn_dma_alloc_coherent - allocate memory for coherent DMA
64 * @dev: device to allocate for
65 * @size: size of the region
66 * @dma_handle: DMA (bus) address
67 * @flags: memory allocation flags
68 *
69 * dma_alloc_coherent() returns a pointer to a memory region suitable for
70 * coherent DMA traffic to/from a PCI device. On SN platforms, this means
71 * that @dma_handle will have the %PCIIO_DMA_CMD flag set.
72 *
73 * This interface is usually used for "command" streams (e.g. the command
74 * queue for a SCSI controller). See Documentation/DMA-API.txt for
75 * more information.
76 */
77 static void *sn_dma_alloc_coherent(struct device *dev, size_t size,
78 dma_addr_t * dma_handle, gfp_t flags)
79 {
80 void *cpuaddr;
81 unsigned long phys_addr;
82 int node;
83 struct pci_dev *pdev = to_pci_dev(dev);
84 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
85
86 BUG_ON(dev->bus != &pci_bus_type);
87
88 /*
89 * Allocate the memory.
90 */
91 node = pcibus_to_node(pdev->bus);
92 if (likely(node >=0)) {
93 struct page *p = alloc_pages_exact_node(node,
94 flags, get_order(size));
95
96 if (likely(p))
97 cpuaddr = page_address(p);
98 else
99 return NULL;
100 } else
101 cpuaddr = (void *)__get_free_pages(flags, get_order(size));
102
103 if (unlikely(!cpuaddr))
104 return NULL;
105
106 memset(cpuaddr, 0x0, size);
107
108 /* physical addr. of the memory we just got */
109 phys_addr = __pa(cpuaddr);
110
111 /*
112 * 64 bit address translations should never fail.
113 * 32 bit translations can fail if there are insufficient mapping
114 * resources.
115 */
116
117 *dma_handle = provider->dma_map_consistent(pdev, phys_addr, size,
118 SN_DMA_ADDR_PHYS);
119 if (!*dma_handle) {
120 printk(KERN_ERR "%s: out of ATEs\n", __func__);
121 free_pages((unsigned long)cpuaddr, get_order(size));
122 return NULL;
123 }
124
125 return cpuaddr;
126 }
127
128 /**
129 * sn_pci_free_coherent - free memory associated with coherent DMAable region
130 * @dev: device to free for
131 * @size: size to free
132 * @cpu_addr: kernel virtual address to free
133 * @dma_handle: DMA address associated with this region
134 *
135 * Frees the memory allocated by dma_alloc_coherent(), potentially unmapping
136 * any associated IOMMU mappings.
137 */
138 static void sn_dma_free_coherent(struct device *dev, size_t size, void *cpu_addr,
139 dma_addr_t dma_handle)
140 {
141 struct pci_dev *pdev = to_pci_dev(dev);
142 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
143
144 BUG_ON(dev->bus != &pci_bus_type);
145
146 provider->dma_unmap(pdev, dma_handle, 0);
147 free_pages((unsigned long)cpu_addr, get_order(size));
148 }
149
150 /**
151 * sn_dma_map_single_attrs - map a single page for DMA
152 * @dev: device to map for
153 * @cpu_addr: kernel virtual address of the region to map
154 * @size: size of the region
155 * @direction: DMA direction
156 * @attrs: optional dma attributes
157 *
158 * Map the region pointed to by @cpu_addr for DMA and return the
159 * DMA address.
160 *
161 * We map this to the one step pcibr_dmamap_trans interface rather than
162 * the two step pcibr_dmamap_alloc/pcibr_dmamap_addr because we have
163 * no way of saving the dmamap handle from the alloc to later free
164 * (which is pretty much unacceptable).
165 *
166 * mappings with the DMA_ATTR_WRITE_BARRIER get mapped with
167 * dma_map_consistent() so that writes force a flush of pending DMA.
168 * (See "SGI Altix Architecture Considerations for Linux Device Drivers",
169 * Document Number: 007-4763-001)
170 *
171 * TODO: simplify our interface;
172 * figure out how to save dmamap handle so can use two step.
173 */
174 static dma_addr_t sn_dma_map_page(struct device *dev, struct page *page,
175 unsigned long offset, size_t size,
176 enum dma_data_direction dir,
177 struct dma_attrs *attrs)
178 {
179 void *cpu_addr = page_address(page) + offset;
180 dma_addr_t dma_addr;
181 unsigned long phys_addr;
182 struct pci_dev *pdev = to_pci_dev(dev);
183 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
184 int dmabarr;
185
186 dmabarr = dma_get_attr(DMA_ATTR_WRITE_BARRIER, attrs);
187
188 BUG_ON(dev->bus != &pci_bus_type);
189
190 phys_addr = __pa(cpu_addr);
191 if (dmabarr)
192 dma_addr = provider->dma_map_consistent(pdev, phys_addr,
193 size, SN_DMA_ADDR_PHYS);
194 else
195 dma_addr = provider->dma_map(pdev, phys_addr, size,
196 SN_DMA_ADDR_PHYS);
197
198 if (!dma_addr) {
199 printk(KERN_ERR "%s: out of ATEs\n", __func__);
200 return 0;
201 }
202 return dma_addr;
203 }
204
205 /**
206 * sn_dma_unmap_single_attrs - unamp a DMA mapped page
207 * @dev: device to sync
208 * @dma_addr: DMA address to sync
209 * @size: size of region
210 * @direction: DMA direction
211 * @attrs: optional dma attributes
212 *
213 * This routine is supposed to sync the DMA region specified
214 * by @dma_handle into the coherence domain. On SN, we're always cache
215 * coherent, so we just need to free any ATEs associated with this mapping.
216 */
217 static void sn_dma_unmap_page(struct device *dev, dma_addr_t dma_addr,
218 size_t size, enum dma_data_direction dir,
219 struct dma_attrs *attrs)
220 {
221 struct pci_dev *pdev = to_pci_dev(dev);
222 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
223
224 BUG_ON(dev->bus != &pci_bus_type);
225
226 provider->dma_unmap(pdev, dma_addr, dir);
227 }
228
229 /**
230 * sn_dma_unmap_sg - unmap a DMA scatterlist
231 * @dev: device to unmap
232 * @sg: scatterlist to unmap
233 * @nhwentries: number of scatterlist entries
234 * @direction: DMA direction
235 * @attrs: optional dma attributes
236 *
237 * Unmap a set of streaming mode DMA translations.
238 */
239 static void sn_dma_unmap_sg(struct device *dev, struct scatterlist *sgl,
240 int nhwentries, enum dma_data_direction dir,
241 struct dma_attrs *attrs)
242 {
243 int i;
244 struct pci_dev *pdev = to_pci_dev(dev);
245 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
246 struct scatterlist *sg;
247
248 BUG_ON(dev->bus != &pci_bus_type);
249
250 for_each_sg(sgl, sg, nhwentries, i) {
251 provider->dma_unmap(pdev, sg->dma_address, dir);
252 sg->dma_address = (dma_addr_t) NULL;
253 sg->dma_length = 0;
254 }
255 }
256
257 /**
258 * sn_dma_map_sg - map a scatterlist for DMA
259 * @dev: device to map for
260 * @sg: scatterlist to map
261 * @nhwentries: number of entries
262 * @direction: direction of the DMA transaction
263 * @attrs: optional dma attributes
264 *
265 * mappings with the DMA_ATTR_WRITE_BARRIER get mapped with
266 * dma_map_consistent() so that writes force a flush of pending DMA.
267 * (See "SGI Altix Architecture Considerations for Linux Device Drivers",
268 * Document Number: 007-4763-001)
269 *
270 * Maps each entry of @sg for DMA.
271 */
272 static int sn_dma_map_sg(struct device *dev, struct scatterlist *sgl,
273 int nhwentries, enum dma_data_direction dir,
274 struct dma_attrs *attrs)
275 {
276 unsigned long phys_addr;
277 struct scatterlist *saved_sg = sgl, *sg;
278 struct pci_dev *pdev = to_pci_dev(dev);
279 struct sn_pcibus_provider *provider = SN_PCIDEV_BUSPROVIDER(pdev);
280 int i;
281 int dmabarr;
282
283 dmabarr = dma_get_attr(DMA_ATTR_WRITE_BARRIER, attrs);
284
285 BUG_ON(dev->bus != &pci_bus_type);
286
287 /*
288 * Setup a DMA address for each entry in the scatterlist.
289 */
290 for_each_sg(sgl, sg, nhwentries, i) {
291 dma_addr_t dma_addr;
292 phys_addr = SG_ENT_PHYS_ADDRESS(sg);
293 if (dmabarr)
294 dma_addr = provider->dma_map_consistent(pdev,
295 phys_addr,
296 sg->length,
297 SN_DMA_ADDR_PHYS);
298 else
299 dma_addr = provider->dma_map(pdev, phys_addr,
300 sg->length,
301 SN_DMA_ADDR_PHYS);
302
303 sg->dma_address = dma_addr;
304 if (!sg->dma_address) {
305 printk(KERN_ERR "%s: out of ATEs\n", __func__);
306
307 /*
308 * Free any successfully allocated entries.
309 */
310 if (i > 0)
311 sn_dma_unmap_sg(dev, saved_sg, i, dir, attrs);
312 return 0;
313 }
314
315 sg->dma_length = sg->length;
316 }
317
318 return nhwentries;
319 }
320
321 static void sn_dma_sync_single_for_cpu(struct device *dev, dma_addr_t dma_handle,
322 size_t size, enum dma_data_direction dir)
323 {
324 BUG_ON(dev->bus != &pci_bus_type);
325 }
326
327 static void sn_dma_sync_single_for_device(struct device *dev, dma_addr_t dma_handle,
328 size_t size,
329 enum dma_data_direction dir)
330 {
331 BUG_ON(dev->bus != &pci_bus_type);
332 }
333
334 static void sn_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
335 int nelems, enum dma_data_direction dir)
336 {
337 BUG_ON(dev->bus != &pci_bus_type);
338 }
339
340 static void sn_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
341 int nelems, enum dma_data_direction dir)
342 {
343 BUG_ON(dev->bus != &pci_bus_type);
344 }
345
346 static int sn_dma_mapping_error(struct device *dev, dma_addr_t dma_addr)
347 {
348 return 0;
349 }
350
351 u64 sn_dma_get_required_mask(struct device *dev)
352 {
353 return DMA_BIT_MASK(64);
354 }
355 EXPORT_SYMBOL_GPL(sn_dma_get_required_mask);
356
357 char *sn_pci_get_legacy_mem(struct pci_bus *bus)
358 {
359 if (!SN_PCIBUS_BUSSOFT(bus))
360 return ERR_PTR(-ENODEV);
361
362 return (char *)(SN_PCIBUS_BUSSOFT(bus)->bs_legacy_mem | __IA64_UNCACHED_OFFSET);
363 }
364
365 int sn_pci_legacy_read(struct pci_bus *bus, u16 port, u32 *val, u8 size)
366 {
367 unsigned long addr;
368 int ret;
369 struct ia64_sal_retval isrv;
370
371 /*
372 * First, try the SN_SAL_IOIF_PCI_SAFE SAL call which can work
373 * around hw issues at the pci bus level. SGI proms older than
374 * 4.10 don't implement this.
375 */
376
377 SAL_CALL(isrv, SN_SAL_IOIF_PCI_SAFE,
378 pci_domain_nr(bus), bus->number,
379 0, /* io */
380 0, /* read */
381 port, size, __pa(val));
382
383 if (isrv.status == 0)
384 return size;
385
386 /*
387 * If the above failed, retry using the SAL_PROBE call which should
388 * be present in all proms (but which cannot work round PCI chipset
389 * bugs). This code is retained for compatibility with old
390 * pre-4.10 proms, and should be removed at some point in the future.
391 */
392
393 if (!SN_PCIBUS_BUSSOFT(bus))
394 return -ENODEV;
395
396 addr = SN_PCIBUS_BUSSOFT(bus)->bs_legacy_io | __IA64_UNCACHED_OFFSET;
397 addr += port;
398
399 ret = ia64_sn_probe_mem(addr, (long)size, (void *)val);
400
401 if (ret == 2)
402 return -EINVAL;
403
404 if (ret == 1)
405 *val = -1;
406
407 return size;
408 }
409
410 int sn_pci_legacy_write(struct pci_bus *bus, u16 port, u32 val, u8 size)
411 {
412 int ret = size;
413 unsigned long paddr;
414 unsigned long *addr;
415 struct ia64_sal_retval isrv;
416
417 /*
418 * First, try the SN_SAL_IOIF_PCI_SAFE SAL call which can work
419 * around hw issues at the pci bus level. SGI proms older than
420 * 4.10 don't implement this.
421 */
422
423 SAL_CALL(isrv, SN_SAL_IOIF_PCI_SAFE,
424 pci_domain_nr(bus), bus->number,
425 0, /* io */
426 1, /* write */
427 port, size, __pa(&val));
428
429 if (isrv.status == 0)
430 return size;
431
432 /*
433 * If the above failed, retry using the SAL_PROBE call which should
434 * be present in all proms (but which cannot work round PCI chipset
435 * bugs). This code is retained for compatibility with old
436 * pre-4.10 proms, and should be removed at some point in the future.
437 */
438
439 if (!SN_PCIBUS_BUSSOFT(bus)) {
440 ret = -ENODEV;
441 goto out;
442 }
443
444 /* Put the phys addr in uncached space */
445 paddr = SN_PCIBUS_BUSSOFT(bus)->bs_legacy_io | __IA64_UNCACHED_OFFSET;
446 paddr += port;
447 addr = (unsigned long *)paddr;
448
449 switch (size) {
450 case 1:
451 *(volatile u8 *)(addr) = (u8)(val);
452 break;
453 case 2:
454 *(volatile u16 *)(addr) = (u16)(val);
455 break;
456 case 4:
457 *(volatile u32 *)(addr) = (u32)(val);
458 break;
459 default:
460 ret = -EINVAL;
461 break;
462 }
463 out:
464 return ret;
465 }
466
467 static struct dma_map_ops sn_dma_ops = {
468 .alloc_coherent = sn_dma_alloc_coherent,
469 .free_coherent = sn_dma_free_coherent,
470 .map_page = sn_dma_map_page,
471 .unmap_page = sn_dma_unmap_page,
472 .map_sg = sn_dma_map_sg,
473 .unmap_sg = sn_dma_unmap_sg,
474 .sync_single_for_cpu = sn_dma_sync_single_for_cpu,
475 .sync_sg_for_cpu = sn_dma_sync_sg_for_cpu,
476 .sync_single_for_device = sn_dma_sync_single_for_device,
477 .sync_sg_for_device = sn_dma_sync_sg_for_device,
478 .mapping_error = sn_dma_mapping_error,
479 .dma_supported = sn_dma_supported,
480 };
481
482 void sn_dma_init(void)
483 {
484 dma_ops = &sn_dma_ops;
485 }
This page took 0.106941 seconds and 5 git commands to generate.