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) 2001-2005 Silicon Graphics, Inc. All rights reserved.
7 */
8
9#include <linux/types.h>
10#include <linux/pci.h>
11#include <linux/export.h>
12#include <asm/sn/addrs.h>
13#include <asm/sn/geo.h>
14#include <asm/sn/pcibr_provider.h>
15#include <asm/sn/pcibus_provider_defs.h>
16#include <asm/sn/pcidev.h>
17#include <asm/sn/pic.h>
18#include <asm/sn/sn_sal.h>
19#include <asm/sn/tiocp.h>
20#include "tio.h"
21#include "xtalk/xwidgetdev.h"
22#include "xtalk/hubdev.h"
23
24extern int sn_ioif_inited;
25
26/* =====================================================================
27 *    DMA MANAGEMENT
28 *
29 *      The Bridge ASIC provides three methods of doing DMA: via a "direct map"
30 *      register available in 32-bit PCI space (which selects a contiguous 2G
31 *	address space on some other widget), via "direct" addressing via 64-bit
32 *      PCI space (all destination information comes from the PCI address,
33 *      including transfer attributes), and via a "mapped" region that allows
34 *      a bunch of different small mappings to be established with the PMU.
35 *
36 *      For efficiency, we most prefer to use the 32bit direct mapping facility,
37 *      since it requires no resource allocations. The advantage of using the
38 *      PMU over the 64-bit direct is that single-cycle PCI addressing can be
39 *      used; the advantage of using 64-bit direct over PMU addressing is that
40 *      we do not have to allocate entries in the PMU.
41 */
42
43static dma_addr_t
44pcibr_dmamap_ate32(struct pcidev_info *info,
45		   u64 paddr, size_t req_size, u64 flags, int dma_flags)
46{
47
48	struct pcidev_info *pcidev_info = info->pdi_host_pcidev_info;
49	struct pcibus_info *pcibus_info = (struct pcibus_info *)pcidev_info->
50	    pdi_pcibus_info;
51	u8 internal_device = (PCI_SLOT(pcidev_info->pdi_host_pcidev_info->
52					    pdi_linux_pcidev->devfn)) - 1;
53	int ate_count;
54	int ate_index;
55	u64 ate_flags = flags | PCI32_ATE_V;
56	u64 ate;
57	u64 pci_addr;
58	u64 xio_addr;
59	u64 offset;
60
61	/* PIC in PCI-X mode does not supports 32bit PageMap mode */
62	if (IS_PIC_SOFT(pcibus_info) && IS_PCIX(pcibus_info)) {
63		return 0;
64	}
65
66	/* Calculate the number of ATEs needed. */
67	if (!(MINIMAL_ATE_FLAG(paddr, req_size))) {
68		ate_count = IOPG((IOPGSIZE - 1)	/* worst case start offset */
69				 +req_size	/* max mapping bytes */
70				 - 1) + 1;	/* round UP */
71	} else {		/* assume requested target is page aligned */
72		ate_count = IOPG(req_size	/* max mapping bytes */
73				 - 1) + 1;	/* round UP */
74	}
75
76	/* Get the number of ATEs required. */
77	ate_index = pcibr_ate_alloc(pcibus_info, ate_count);
78	if (ate_index < 0)
79		return 0;
80
81	/* In PCI-X mode, Prefetch not supported */
82	if (IS_PCIX(pcibus_info))
83		ate_flags &= ~(PCI32_ATE_PREF);
84
85	if (SN_DMA_ADDRTYPE(dma_flags == SN_DMA_ADDR_PHYS))
86		xio_addr = IS_PIC_SOFT(pcibus_info) ? PHYS_TO_DMA(paddr) :
87	    					      PHYS_TO_TIODMA(paddr);
88	else
89		xio_addr = paddr;
90
91	offset = IOPGOFF(xio_addr);
92	ate = ate_flags | (xio_addr - offset);
93
94	/* If PIC, put the targetid in the ATE */
95	if (IS_PIC_SOFT(pcibus_info)) {
96		ate |= (pcibus_info->pbi_hub_xid << PIC_ATE_TARGETID_SHFT);
97	}
98
99	/*
100	 * If we're mapping for MSI, set the MSI bit in the ATE.  If it's a
101	 * TIOCP based pci bus, we also need to set the PIO bit in the ATE.
102	 */
103	if (dma_flags & SN_DMA_MSI) {
104		ate |= PCI32_ATE_MSI;
105		if (IS_TIOCP_SOFT(pcibus_info))
106			ate |= PCI32_ATE_PIO;
107	}
108
109	ate_write(pcibus_info, ate_index, ate_count, ate);
110
111	/*
112	 * Set up the DMA mapped Address.
113	 */
114	pci_addr = PCI32_MAPPED_BASE + offset + IOPGSIZE * ate_index;
115
116	/*
117	 * If swap was set in device in pcibr_endian_set()
118	 * we need to turn swapping on.
119	 */
120	if (pcibus_info->pbi_devreg[internal_device] & PCIBR_DEV_SWAP_DIR)
121		ATE_SWAP_ON(pci_addr);
122
123
124	return pci_addr;
125}
126
127static dma_addr_t
128pcibr_dmatrans_direct64(struct pcidev_info * info, u64 paddr,
129			u64 dma_attributes, int dma_flags)
130{
131	struct pcibus_info *pcibus_info = (struct pcibus_info *)
132	    ((info->pdi_host_pcidev_info)->pdi_pcibus_info);
133	u64 pci_addr;
134
135	/* Translate to Crosstalk View of Physical Address */
136	if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS)
137		pci_addr = IS_PIC_SOFT(pcibus_info) ?
138				PHYS_TO_DMA(paddr) :
139				PHYS_TO_TIODMA(paddr);
140	else
141		pci_addr = paddr;
142	pci_addr |= dma_attributes;
143
144	/* Handle Bus mode */
145	if (IS_PCIX(pcibus_info))
146		pci_addr &= ~PCI64_ATTR_PREF;
147
148	/* Handle Bridge Chipset differences */
149	if (IS_PIC_SOFT(pcibus_info)) {
150		pci_addr |=
151		    ((u64) pcibus_info->
152		     pbi_hub_xid << PIC_PCI64_ATTR_TARG_SHFT);
153	} else
154		pci_addr |= (dma_flags & SN_DMA_MSI) ?
155				TIOCP_PCI64_CMDTYPE_MSI :
156				TIOCP_PCI64_CMDTYPE_MEM;
157
158	/* If PCI mode, func zero uses VCHAN0, every other func uses VCHAN1 */
159	if (!IS_PCIX(pcibus_info) && PCI_FUNC(info->pdi_linux_pcidev->devfn))
160		pci_addr |= PCI64_ATTR_VIRTUAL;
161
162	return pci_addr;
163}
164
165static dma_addr_t
166pcibr_dmatrans_direct32(struct pcidev_info * info,
167			u64 paddr, size_t req_size, u64 flags, int dma_flags)
168{
169	struct pcidev_info *pcidev_info = info->pdi_host_pcidev_info;
170	struct pcibus_info *pcibus_info = (struct pcibus_info *)pcidev_info->
171	    pdi_pcibus_info;
172	u64 xio_addr;
173
174	u64 xio_base;
175	u64 offset;
176	u64 endoff;
177
178	if (IS_PCIX(pcibus_info)) {
179		return 0;
180	}
181
182	if (dma_flags & SN_DMA_MSI)
183		return 0;
184
185	if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS)
186		xio_addr = IS_PIC_SOFT(pcibus_info) ? PHYS_TO_DMA(paddr) :
187	    					      PHYS_TO_TIODMA(paddr);
188	else
189		xio_addr = paddr;
190
191	xio_base = pcibus_info->pbi_dir_xbase;
192	offset = xio_addr - xio_base;
193	endoff = req_size + offset;
194	if ((req_size > (1ULL << 31)) ||	/* Too Big */
195	    (xio_addr < xio_base) ||	/* Out of range for mappings */
196	    (endoff > (1ULL << 31))) {	/* Too Big */
197		return 0;
198	}
199
200	return PCI32_DIRECT_BASE | offset;
201}
202
203/*
204 * Wrapper routine for freeing DMA maps
205 * DMA mappings for Direct 64 and 32 do not have any DMA maps.
206 */
207void
208pcibr_dma_unmap(struct pci_dev *hwdev, dma_addr_t dma_handle, int direction)
209{
210	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);
211	struct pcibus_info *pcibus_info =
212	    (struct pcibus_info *)pcidev_info->pdi_pcibus_info;
213
214	if (IS_PCI32_MAPPED(dma_handle)) {
215		int ate_index;
216
217		ate_index =
218		    IOPG((ATE_SWAP_OFF(dma_handle) - PCI32_MAPPED_BASE));
219		pcibr_ate_free(pcibus_info, ate_index);
220	}
221}
222
223/*
224 * On SN systems there is a race condition between a PIO read response and
225 * DMA's.  In rare cases, the read response may beat the DMA, causing the
226 * driver to think that data in memory is complete and meaningful.  This code
227 * eliminates that race.  This routine is called by the PIO read routines
228 * after doing the read.  For PIC this routine then forces a fake interrupt
229 * on another line, which is logically associated with the slot that the PIO
230 * is addressed to.  It then spins while watching the memory location that
231 * the interrupt is targeted to.  When the interrupt response arrives, we
232 * are sure that the DMA has landed in memory and it is safe for the driver
233 * to proceed.	For TIOCP use the Device(x) Write Request Buffer Flush
234 * Bridge register since it ensures the data has entered the coherence domain,
235 * unlike the PIC Device(x) Write Request Buffer Flush register.
236 */
237
238void sn_dma_flush(u64 addr)
239{
240	nasid_t nasid;
241	int is_tio;
242	int wid_num;
243	int i, j;
244	unsigned long flags;
245	u64 itte;
246	struct hubdev_info *hubinfo;
247	struct sn_flush_device_kernel *p;
248	struct sn_flush_device_common *common;
249	struct sn_flush_nasid_entry *flush_nasid_list;
250
251	if (!sn_ioif_inited)
252		return;
253
254	nasid = NASID_GET(addr);
255	if (-1 == nasid_to_cnodeid(nasid))
256		return;
257
258	hubinfo = (NODEPDA(nasid_to_cnodeid(nasid)))->pdinfo;
259
260	BUG_ON(!hubinfo);
261
262	flush_nasid_list = &hubinfo->hdi_flush_nasid_list;
263	if (flush_nasid_list->widget_p == NULL)
264		return;
265
266	is_tio = (nasid & 1);
267	if (is_tio) {
268		int itte_index;
269
270		if (TIO_HWIN(addr))
271			itte_index = 0;
272		else if (TIO_BWIN_WINDOWNUM(addr))
273			itte_index = TIO_BWIN_WINDOWNUM(addr);
274		else
275			itte_index = -1;
276
277		if (itte_index >= 0) {
278			itte = flush_nasid_list->iio_itte[itte_index];
279			if (! TIO_ITTE_VALID(itte))
280				return;
281			wid_num = TIO_ITTE_WIDGET(itte);
282		} else
283			wid_num = TIO_SWIN_WIDGETNUM(addr);
284	} else {
285		if (BWIN_WINDOWNUM(addr)) {
286			itte = flush_nasid_list->iio_itte[BWIN_WINDOWNUM(addr)];
287			wid_num = IIO_ITTE_WIDGET(itte);
288		} else
289			wid_num = SWIN_WIDGETNUM(addr);
290	}
291	if (flush_nasid_list->widget_p[wid_num] == NULL)
292		return;
293	p = &flush_nasid_list->widget_p[wid_num][0];
294
295	/* find a matching BAR */
296	for (i = 0; i < DEV_PER_WIDGET; i++,p++) {
297		common = p->common;
298		for (j = 0; j < PCI_ROM_RESOURCE; j++) {
299			if (common->sfdl_bar_list[j].start == 0)
300				break;
301			if (addr >= common->sfdl_bar_list[j].start
302			    && addr <= common->sfdl_bar_list[j].end)
303				break;
304		}
305		if (j < PCI_ROM_RESOURCE && common->sfdl_bar_list[j].start != 0)
306			break;
307	}
308
309	/* if no matching BAR, return without doing anything. */
310	if (i == DEV_PER_WIDGET)
311		return;
312
313	/*
314	 * For TIOCP use the Device(x) Write Request Buffer Flush Bridge
315	 * register since it ensures the data has entered the coherence
316	 * domain, unlike PIC.
317	 */
318	if (is_tio) {
319		/*
320	 	 * Note:  devices behind TIOCE should never be matched in the
321		 * above code, and so the following code is PIC/CP centric.
322		 * If CE ever needs the sn_dma_flush mechanism, we will have
323		 * to account for that here and in tioce_bus_fixup().
324	 	 */
325		u32 tio_id = HUB_L(TIO_IOSPACE_ADDR(nasid, TIO_NODE_ID));
326		u32 revnum = XWIDGET_PART_REV_NUM(tio_id);
327
328		/* TIOCP BRINGUP WAR (PV907516): Don't write buffer flush reg */
329		if ((1 << XWIDGET_PART_REV_NUM_REV(revnum)) & PV907516) {
330			return;
331		} else {
332			pcireg_wrb_flush_get(common->sfdl_pcibus_info,
333					     (common->sfdl_slot - 1));
334		}
335	} else {
336		spin_lock_irqsave(&p->sfdl_flush_lock, flags);
337		*common->sfdl_flush_addr = 0;
338
339		/* force an interrupt. */
340		*(volatile u32 *)(common->sfdl_force_int_addr) = 1;
341
342		/* wait for the interrupt to come back. */
343		while (*(common->sfdl_flush_addr) != 0x10f)
344			cpu_relax();
345
346		/* okay, everything is synched up. */
347		spin_unlock_irqrestore(&p->sfdl_flush_lock, flags);
348	}
349	return;
350}
351
352/*
353 * DMA interfaces.  Called from pci_dma.c routines.
354 */
355
356dma_addr_t
357pcibr_dma_map(struct pci_dev * hwdev, unsigned long phys_addr, size_t size, int dma_flags)
358{
359	dma_addr_t dma_handle;
360	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);
361
362	/* SN cannot support DMA addresses smaller than 32 bits. */
363	if (hwdev->dma_mask < 0x7fffffff) {
364		return 0;
365	}
366
367	if (hwdev->dma_mask == ~0UL) {
368		/*
369		 * Handle the most common case: 64 bit cards.  This
370		 * call should always succeed.
371		 */
372
373		dma_handle = pcibr_dmatrans_direct64(pcidev_info, phys_addr,
374						     PCI64_ATTR_PREF, dma_flags);
375	} else {
376		/* Handle 32-63 bit cards via direct mapping */
377		dma_handle = pcibr_dmatrans_direct32(pcidev_info, phys_addr,
378						     size, 0, dma_flags);
379		if (!dma_handle) {
380			/*
381			 * It is a 32 bit card and we cannot do direct mapping,
382			 * so we use an ATE.
383			 */
384
385			dma_handle = pcibr_dmamap_ate32(pcidev_info, phys_addr,
386							size, PCI32_ATE_PREF,
387							dma_flags);
388		}
389	}
390
391	return dma_handle;
392}
393
394dma_addr_t
395pcibr_dma_map_consistent(struct pci_dev * hwdev, unsigned long phys_addr,
396			 size_t size, int dma_flags)
397{
398	dma_addr_t dma_handle;
399	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);
400
401	if (hwdev->dev.coherent_dma_mask == ~0UL) {
402		dma_handle = pcibr_dmatrans_direct64(pcidev_info, phys_addr,
403					    PCI64_ATTR_BAR, dma_flags);
404	} else {
405		dma_handle = (dma_addr_t) pcibr_dmamap_ate32(pcidev_info,
406						    phys_addr, size,
407						    PCI32_ATE_BAR, dma_flags);
408	}
409
410	return dma_handle;
411}
412
413EXPORT_SYMBOL(sn_dma_flush);
414