[linux-block.git] / arch / ia64 / sn / pci / pcibr / pcibr_dma.c

/*
 * This file is subject to the terms and conditions of the GNU General Public
 * License.  See the file "COPYING" in the main directory of this archive
 * for more details.
 *
 * Copyright (C) 2001-2005 Silicon Graphics, Inc. All rights reserved.
 */

#include <linux/types.h>
#include <linux/pci.h>
#include <asm/sn/addrs.h>
#include <asm/sn/geo.h>
#include <asm/sn/pcibr_provider.h>
#include <asm/sn/pcibus_provider_defs.h>
#include <asm/sn/pcidev.h>
#include <asm/sn/pic.h>
#include <asm/sn/sn_sal.h>
#include <asm/sn/tiocp.h>
#include "tio.h"
#include "xtalk/xwidgetdev.h"
#include "xtalk/hubdev.h"

extern int sn_ioif_inited;

/* =====================================================================
 *    DMA MANAGEMENT
 *
 *      The Bridge ASIC provides three methods of doing DMA: via a "direct map"
 *      register available in 32-bit PCI space (which selects a contiguous 2G
 *	address space on some other widget), via "direct" addressing via 64-bit
 *      PCI space (all destination information comes from the PCI address,
 *      including transfer attributes), and via a "mapped" region that allows 
 *      a bunch of different small mappings to be established with the PMU.
 *
 *      For efficiency, we most prefer to use the 32bit direct mapping facility,
 *      since it requires no resource allocations. The advantage of using the
 *      PMU over the 64-bit direct is that single-cycle PCI addressing can be
 *      used; the advantage of using 64-bit direct over PMU addressing is that
 *      we do not have to allocate entries in the PMU.
 */

static dma_addr_t
pcibr_dmamap_ate32(struct pcidev_info *info,
		   u64 paddr, size_t req_size, u64 flags, int dma_flags)
{

	struct pcidev_info *pcidev_info = info->pdi_host_pcidev_info;
	struct pcibus_info *pcibus_info = (struct pcibus_info *)pcidev_info->
	    pdi_pcibus_info;
	u8 internal_device = (PCI_SLOT(pcidev_info->pdi_host_pcidev_info->
					    pdi_linux_pcidev->devfn)) - 1;
	int ate_count;
	int ate_index;
	u64 ate_flags = flags | PCI32_ATE_V;
	u64 ate;
	u64 pci_addr;
	u64 xio_addr;
	u64 offset;

	/* PIC in PCI-X mode does not supports 32bit PageMap mode */
	if (IS_PIC_SOFT(pcibus_info) && IS_PCIX(pcibus_info)) {
		return 0;
	}

	/* Calculate the number of ATEs needed. */
	if (!(MINIMAL_ATE_FLAG(paddr, req_size))) {
		ate_count = IOPG((IOPGSIZE - 1)	/* worst case start offset */
				 +req_size	/* max mapping bytes */
				 - 1) + 1;	/* round UP */
	} else {		/* assume requested target is page aligned */
		ate_count = IOPG(req_size	/* max mapping bytes */
				 - 1) + 1;	/* round UP */
	}

	/* Get the number of ATEs required. */
	ate_index = pcibr_ate_alloc(pcibus_info, ate_count);
	if (ate_index < 0)
		return 0;

	/* In PCI-X mode, Prefetch not supported */
	if (IS_PCIX(pcibus_info))
		ate_flags &= ~(PCI32_ATE_PREF);

	if (SN_DMA_ADDRTYPE(dma_flags == SN_DMA_ADDR_PHYS))
		xio_addr = IS_PIC_SOFT(pcibus_info) ? PHYS_TO_DMA(paddr) :
	    					      PHYS_TO_TIODMA(paddr);
	else
		xio_addr = paddr;

	offset = IOPGOFF(xio_addr);
	ate = ate_flags | (xio_addr - offset);

	/* If PIC, put the targetid in the ATE */
	if (IS_PIC_SOFT(pcibus_info)) {
		ate |= (pcibus_info->pbi_hub_xid << PIC_ATE_TARGETID_SHFT);
	}

	/*
	 * If we're mapping for MSI, set the MSI bit in the ATE.  If it's a
	 * TIOCP based pci bus, we also need to set the PIO bit in the ATE.
	 */
	if (dma_flags & SN_DMA_MSI) {
		ate |= PCI32_ATE_MSI;
		if (IS_TIOCP_SOFT(pcibus_info))
			ate |= PCI32_ATE_PIO;
	}

	ate_write(pcibus_info, ate_index, ate_count, ate);

	/*
	 * Set up the DMA mapped Address.
	 */
	pci_addr = PCI32_MAPPED_BASE + offset + IOPGSIZE * ate_index;

	/*
	 * If swap was set in device in pcibr_endian_set()
	 * we need to turn swapping on.
	 */
	if (pcibus_info->pbi_devreg[internal_device] & PCIBR_DEV_SWAP_DIR)
		ATE_SWAP_ON(pci_addr);


	return pci_addr;
}

static dma_addr_t
pcibr_dmatrans_direct64(struct pcidev_info * info, u64 paddr,
			u64 dma_attributes, int dma_flags)
{
	struct pcibus_info *pcibus_info = (struct pcibus_info *)
	    ((info->pdi_host_pcidev_info)->pdi_pcibus_info);
	u64 pci_addr;

	/* Translate to Crosstalk View of Physical Address */
	if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS)
		pci_addr = IS_PIC_SOFT(pcibus_info) ?
				PHYS_TO_DMA(paddr) :
				PHYS_TO_TIODMA(paddr);
	else
		pci_addr = paddr;
	pci_addr |= dma_attributes;

	/* Handle Bus mode */
	if (IS_PCIX(pcibus_info))
		pci_addr &= ~PCI64_ATTR_PREF;

	/* Handle Bridge Chipset differences */
	if (IS_PIC_SOFT(pcibus_info)) {
		pci_addr |=
		    ((u64) pcibus_info->
		     pbi_hub_xid << PIC_PCI64_ATTR_TARG_SHFT);
	} else
		pci_addr |= (dma_flags & SN_DMA_MSI) ?
				TIOCP_PCI64_CMDTYPE_MSI :
				TIOCP_PCI64_CMDTYPE_MEM;

	/* If PCI mode, func zero uses VCHAN0, every other func uses VCHAN1 */
	if (!IS_PCIX(pcibus_info) && PCI_FUNC(info->pdi_linux_pcidev->devfn))
		pci_addr |= PCI64_ATTR_VIRTUAL;

	return pci_addr;
}

static dma_addr_t
pcibr_dmatrans_direct32(struct pcidev_info * info,
			u64 paddr, size_t req_size, u64 flags, int dma_flags)
{
	struct pcidev_info *pcidev_info = info->pdi_host_pcidev_info;
	struct pcibus_info *pcibus_info = (struct pcibus_info *)pcidev_info->
	    pdi_pcibus_info;
	u64 xio_addr;

	u64 xio_base;
	u64 offset;
	u64 endoff;

	if (IS_PCIX(pcibus_info)) {
		return 0;
	}

	if (dma_flags & SN_DMA_MSI)
		return 0;

	if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS)
		xio_addr = IS_PIC_SOFT(pcibus_info) ? PHYS_TO_DMA(paddr) :
	    					      PHYS_TO_TIODMA(paddr);
	else
		xio_addr = paddr;

	xio_base = pcibus_info->pbi_dir_xbase;
	offset = xio_addr - xio_base;
	endoff = req_size + offset;
	if ((req_size > (1ULL << 31)) ||	/* Too Big */
	    (xio_addr < xio_base) ||	/* Out of range for mappings */
	    (endoff > (1ULL << 31))) {	/* Too Big */
		return 0;
	}

	return PCI32_DIRECT_BASE | offset;
}

/*
 * Wrapper routine for freeing DMA maps
 * DMA mappings for Direct 64 and 32 do not have any DMA maps.
 */
void
pcibr_dma_unmap(struct pci_dev *hwdev, dma_addr_t dma_handle, int direction)
{
	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);
	struct pcibus_info *pcibus_info =
	    (struct pcibus_info *)pcidev_info->pdi_pcibus_info;

	if (IS_PCI32_MAPPED(dma_handle)) {
		int ate_index;

		ate_index =
		    IOPG((ATE_SWAP_OFF(dma_handle) - PCI32_MAPPED_BASE));
		pcibr_ate_free(pcibus_info, ate_index);
	}
}

/*
 * On SN systems there is a race condition between a PIO read response and 
 * DMA's.  In rare cases, the read response may beat the DMA, causing the
 * driver to think that data in memory is complete and meaningful.  This code
 * eliminates that race.  This routine is called by the PIO read routines
 * after doing the read.  For PIC this routine then forces a fake interrupt
 * on another line, which is logically associated with the slot that the PIO
 * is addressed to.  It then spins while watching the memory location that
 * the interrupt is targetted to.  When the interrupt response arrives, we 
 * are sure that the DMA has landed in memory and it is safe for the driver
 * to proceed.	For TIOCP use the Device(x) Write Request Buffer Flush 
 * Bridge register since it ensures the data has entered the coherence domain,
 * unlike the PIC Device(x) Write Request Buffer Flush register.
 */

void sn_dma_flush(u64 addr)
{
	nasid_t nasid;
	int is_tio;
	int wid_num;
	int i, j;
	unsigned long flags;
	u64 itte;
	struct hubdev_info *hubinfo;
	struct sn_flush_device_kernel *p;
	struct sn_flush_device_common *common;
	struct sn_flush_nasid_entry *flush_nasid_list;

	if (!sn_ioif_inited)
		return;

	nasid = NASID_GET(addr);
	if (-1 == nasid_to_cnodeid(nasid))
		return;

	hubinfo = (NODEPDA(nasid_to_cnodeid(nasid)))->pdinfo;

	if (!hubinfo) {
		BUG();
	}

	flush_nasid_list = &hubinfo->hdi_flush_nasid_list;
	if (flush_nasid_list->widget_p == NULL)
		return;

	is_tio = (nasid & 1);
	if (is_tio) {
		int itte_index;

		if (TIO_HWIN(addr))
			itte_index = 0;
		else if (TIO_BWIN_WINDOWNUM(addr))
			itte_index = TIO_BWIN_WINDOWNUM(addr);
		else
			itte_index = -1;

		if (itte_index >= 0) {
			itte = flush_nasid_list->iio_itte[itte_index];
			if (! TIO_ITTE_VALID(itte))
				return;
			wid_num = TIO_ITTE_WIDGET(itte);
		} else
			wid_num = TIO_SWIN_WIDGETNUM(addr);
	} else {
		if (BWIN_WINDOWNUM(addr)) {
			itte = flush_nasid_list->iio_itte[BWIN_WINDOWNUM(addr)];
			wid_num = IIO_ITTE_WIDGET(itte);
		} else
			wid_num = SWIN_WIDGETNUM(addr);
	}
	if (flush_nasid_list->widget_p[wid_num] == NULL)
		return;
	p = &flush_nasid_list->widget_p[wid_num][0];

	/* find a matching BAR */
	for (i = 0; i < DEV_PER_WIDGET; i++,p++) {
		common = p->common;
		for (j = 0; j < PCI_ROM_RESOURCE; j++) {
			if (common->sfdl_bar_list[j].start == 0)
				break;
			if (addr >= common->sfdl_bar_list[j].start
			    && addr <= common->sfdl_bar_list[j].end)
				break;
		}
		if (j < PCI_ROM_RESOURCE && common->sfdl_bar_list[j].start != 0)
			break;
	}

	/* if no matching BAR, return without doing anything. */
	if (i == DEV_PER_WIDGET)
		return;

	/*
	 * For TIOCP use the Device(x) Write Request Buffer Flush Bridge
	 * register since it ensures the data has entered the coherence
	 * domain, unlike PIC.
	 */
	if (is_tio) {
		/*
	 	 * Note:  devices behind TIOCE should never be matched in the
		 * above code, and so the following code is PIC/CP centric.
		 * If CE ever needs the sn_dma_flush mechanism, we will have
		 * to account for that here and in tioce_bus_fixup().
	 	 */
		u32 tio_id = HUB_L(TIO_IOSPACE_ADDR(nasid, TIO_NODE_ID));
		u32 revnum = XWIDGET_PART_REV_NUM(tio_id);

		/* TIOCP BRINGUP WAR (PV907516): Don't write buffer flush reg */
		if ((1 << XWIDGET_PART_REV_NUM_REV(revnum)) & PV907516) {
			return;
		} else {
			pcireg_wrb_flush_get(common->sfdl_pcibus_info,
					     (common->sfdl_slot - 1));
		}
	} else {
		spin_lock_irqsave(&p->sfdl_flush_lock, flags);
		*common->sfdl_flush_addr = 0;

		/* force an interrupt. */
		*(volatile u32 *)(common->sfdl_force_int_addr) = 1;

		/* wait for the interrupt to come back. */
		while (*(common->sfdl_flush_addr) != 0x10f)
			cpu_relax();

		/* okay, everything is synched up. */
		spin_unlock_irqrestore(&p->sfdl_flush_lock, flags);
	}
	return;
}

/*
 * DMA interfaces.  Called from pci_dma.c routines.
 */

dma_addr_t
pcibr_dma_map(struct pci_dev * hwdev, unsigned long phys_addr, size_t size, int dma_flags)
{
	dma_addr_t dma_handle;
	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);

	/* SN cannot support DMA addresses smaller than 32 bits. */
	if (hwdev->dma_mask < 0x7fffffff) {
		return 0;
	}

	if (hwdev->dma_mask == ~0UL) {
		/*
		 * Handle the most common case: 64 bit cards.  This
		 * call should always succeed.
		 */

		dma_handle = pcibr_dmatrans_direct64(pcidev_info, phys_addr,
						     PCI64_ATTR_PREF, dma_flags);
	} else {
		/* Handle 32-63 bit cards via direct mapping */
		dma_handle = pcibr_dmatrans_direct32(pcidev_info, phys_addr,
						     size, 0, dma_flags);
		if (!dma_handle) {
			/*
			 * It is a 32 bit card and we cannot do direct mapping,
			 * so we use an ATE.
			 */

			dma_handle = pcibr_dmamap_ate32(pcidev_info, phys_addr,
							size, PCI32_ATE_PREF,
							dma_flags);
		}
	}

	return dma_handle;
}

dma_addr_t
pcibr_dma_map_consistent(struct pci_dev * hwdev, unsigned long phys_addr,
			 size_t size, int dma_flags)
{
	dma_addr_t dma_handle;
	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);

	if (hwdev->dev.coherent_dma_mask == ~0UL) {
		dma_handle = pcibr_dmatrans_direct64(pcidev_info, phys_addr,
					    PCI64_ATTR_BAR, dma_flags);
	} else {
		dma_handle = (dma_addr_t) pcibr_dmamap_ate32(pcidev_info,
						    phys_addr, size,
						    PCI32_ATE_BAR, dma_flags);
	}

	return dma_handle;
}

EXPORT_SYMBOL(sn_dma_flush);
Commit	Line	Data
1da177e4 LT	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	*
5b53ed1f	6	* Copyright (C) 2001-2005 Silicon Graphics, Inc. All rights reserved.
1da177e4 LT	7	*/
	8
	9	#include <linux/types.h>
	10	#include <linux/pci.h>
c13cf371	11	#include <asm/sn/addrs.h>
1da177e4	12	#include <asm/sn/geo.h>
c13cf371	13	#include <asm/sn/pcibr_provider.h>
9b08ebd1 MM	14	#include <asm/sn/pcibus_provider_defs.h>
9b08ebd1 MM	15	#include <asm/sn/pcidev.h>
c13cf371 PB	16	#include <asm/sn/pic.h>
	17	#include <asm/sn/sn_sal.h>
	18	#include <asm/sn/tiocp.h>
1da177e4	19	#include "tio.h"
c13cf371 PB	20	#include "xtalk/xwidgetdev.h"
c13cf371 PB	21	#include "xtalk/hubdev.h"
1da177e4 LT	22
	23	extern int sn_ioif_inited;
	24
	25	/* =====================================================================
	26	* DMA MANAGEMENT
	27	*
	28	* The Bridge ASIC provides three methods of doing DMA: via a "direct map"
	29	* register available in 32-bit PCI space (which selects a contiguous 2G
	30	* address space on some other widget), via "direct" addressing via 64-bit
	31	* PCI space (all destination information comes from the PCI address,
	32	* including transfer attributes), and via a "mapped" region that allows
	33	* a bunch of different small mappings to be established with the PMU.
	34	*
	35	* For efficiency, we most prefer to use the 32bit direct mapping facility,
	36	* since it requires no resource allocations. The advantage of using the
	37	* PMU over the 64-bit direct is that single-cycle PCI addressing can be
	38	* used; the advantage of using 64-bit direct over PMU addressing is that
	39	* we do not have to allocate entries in the PMU.
	40	*/
	41
e955d825	42	static dma_addr_t
1da177e4	43	pcibr_dmamap_ate32(struct pcidev_info *info,
83821d3f	44	u64 paddr, size_t req_size, u64 flags, int dma_flags)
1da177e4 LT	45	{
	46
	47	struct pcidev_info *pcidev_info = info->pdi_host_pcidev_info;
	48	struct pcibus_info pcibus_info = (struct pcibus_info )pcidev_info->
	49	pdi_pcibus_info;
53493dcf	50	u8 internal_device = (PCI_SLOT(pcidev_info->pdi_host_pcidev_info->
1da177e4 LT	51	pdi_linux_pcidev->devfn)) - 1;
	52	int ate_count;
	53	int ate_index;
53493dcf PB	54	u64 ate_flags = flags \| PCI32_ATE_V;
	55	u64 ate;
	56	u64 pci_addr;
	57	u64 xio_addr;
	58	u64 offset;
1da177e4 LT	59
	60	/* PIC in PCI-X mode does not supports 32bit PageMap mode */
	61	if (IS_PIC_SOFT(pcibus_info) && IS_PCIX(pcibus_info)) {
	62	return 0;
	63	}
	64
	65	/* Calculate the number of ATEs needed. */
	66	if (!(MINIMAL_ATE_FLAG(paddr, req_size))) {
	67	ate_count = IOPG((IOPGSIZE - 1) /* worst case start offset */
	68	+req_size /* max mapping bytes */
	69	- 1) + 1; /* round UP */
	70	} else { /* assume requested target is page aligned */
	71	ate_count = IOPG(req_size /* max mapping bytes */
	72	- 1) + 1; /* round UP */
	73	}
	74
	75	/* Get the number of ATEs required. */
	76	ate_index = pcibr_ate_alloc(pcibus_info, ate_count);
	77	if (ate_index < 0)
	78	return 0;
	79
	80	/* In PCI-X mode, Prefetch not supported */
	81	if (IS_PCIX(pcibus_info))
	82	ate_flags &= ~(PCI32_ATE_PREF);
	83
83821d3f MM	84	if (SN_DMA_ADDRTYPE(dma_flags == SN_DMA_ADDR_PHYS))
	85	xio_addr = IS_PIC_SOFT(pcibus_info) ? PHYS_TO_DMA(paddr) :
	86	PHYS_TO_TIODMA(paddr);
	87	else
	88	xio_addr = paddr;
	89
1da177e4 LT	90	offset = IOPGOFF(xio_addr);
	91	ate = ate_flags \| (xio_addr - offset);
	92
	93	/* If PIC, put the targetid in the ATE */
	94	if (IS_PIC_SOFT(pcibus_info)) {
	95	ate \|= (pcibus_info->pbi_hub_xid << PIC_ATE_TARGETID_SHFT);
	96	}
83821d3f MM	97
83821d3f MM	98	/*
2e0d232b MH	99	* If we're mapping for MSI, set the MSI bit in the ATE. If it's a
2e0d232b MH	100	* TIOCP based pci bus, we also need to set the PIO bit in the ATE.
83821d3f	101	*/
2e0d232b	102	if (dma_flags & SN_DMA_MSI) {
83821d3f	103	ate \|= PCI32_ATE_MSI;
2e0d232b MH	104	if (IS_TIOCP_SOFT(pcibus_info))
	105	ate \|= PCI32_ATE_PIO;
	106	}
83821d3f	107
1da177e4 LT	108	ate_write(pcibus_info, ate_index, ate_count, ate);
	109
	110	/*
	111	* Set up the DMA mapped Address.
	112	*/
	113	pci_addr = PCI32_MAPPED_BASE + offset + IOPGSIZE * ate_index;
	114
	115	/*
	116	* If swap was set in device in pcibr_endian_set()
	117	* we need to turn swapping on.
	118	*/
	119	if (pcibus_info->pbi_devreg[internal_device] & PCIBR_DEV_SWAP_DIR)
	120	ATE_SWAP_ON(pci_addr);
	121
83821d3f	122
1da177e4 LT	123	return pci_addr;
	124	}
	125
e955d825	126	static dma_addr_t
53493dcf	127	pcibr_dmatrans_direct64(struct pcidev_info * info, u64 paddr,
83821d3f	128	u64 dma_attributes, int dma_flags)
1da177e4 LT	129	{
	130	struct pcibus_info pcibus_info = (struct pcibus_info )
	131	((info->pdi_host_pcidev_info)->pdi_pcibus_info);
53493dcf	132	u64 pci_addr;
1da177e4 LT	133
1da177e4 LT	134	/* Translate to Crosstalk View of Physical Address */
83821d3f MM	135	if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS)
	136	pci_addr = IS_PIC_SOFT(pcibus_info) ?
	137	PHYS_TO_DMA(paddr) :
c63c5805	138	PHYS_TO_TIODMA(paddr);
83821d3f	139	else
c63c5805 JH	140	pci_addr = paddr;
c63c5805 JH	141	pci_addr \|= dma_attributes;
1da177e4 LT	142
	143	/* Handle Bus mode */
	144	if (IS_PCIX(pcibus_info))
	145	pci_addr &= ~PCI64_ATTR_PREF;
	146
	147	/* Handle Bridge Chipset differences */
	148	if (IS_PIC_SOFT(pcibus_info)) {
	149	pci_addr \|=
53493dcf	150	((u64) pcibus_info->
1da177e4 LT	151	pbi_hub_xid << PIC_PCI64_ATTR_TARG_SHFT);
1da177e4 LT	152	} else
83821d3f MM	153	pci_addr \|= (dma_flags & SN_DMA_MSI) ?
	154	TIOCP_PCI64_CMDTYPE_MSI :
	155	TIOCP_PCI64_CMDTYPE_MEM;
1da177e4 LT	156
	157	/* If PCI mode, func zero uses VCHAN0, every other func uses VCHAN1 */
	158	if (!IS_PCIX(pcibus_info) && PCI_FUNC(info->pdi_linux_pcidev->devfn))
	159	pci_addr \|= PCI64_ATTR_VIRTUAL;
	160
	161	return pci_addr;
1da177e4 LT	162	}
1da177e4 LT	163
e955d825	164	static dma_addr_t
1da177e4	165	pcibr_dmatrans_direct32(struct pcidev_info * info,
83821d3f	166	u64 paddr, size_t req_size, u64 flags, int dma_flags)
1da177e4	167	{
1da177e4 LT	168	struct pcidev_info *pcidev_info = info->pdi_host_pcidev_info;
	169	struct pcibus_info pcibus_info = (struct pcibus_info )pcidev_info->
	170	pdi_pcibus_info;
53493dcf	171	u64 xio_addr;
1da177e4	172
53493dcf PB	173	u64 xio_base;
	174	u64 offset;
	175	u64 endoff;
1da177e4 LT	176
	177	if (IS_PCIX(pcibus_info)) {
	178	return 0;
	179	}
	180
83821d3f MM	181	if (dma_flags & SN_DMA_MSI)
	182	return 0;
	183
	184	if (SN_DMA_ADDRTYPE(dma_flags) == SN_DMA_ADDR_PHYS)
	185	xio_addr = IS_PIC_SOFT(pcibus_info) ? PHYS_TO_DMA(paddr) :
	186	PHYS_TO_TIODMA(paddr);
	187	else
	188	xio_addr = paddr;
1da177e4 LT	189
	190	xio_base = pcibus_info->pbi_dir_xbase;
	191	offset = xio_addr - xio_base;
	192	endoff = req_size + offset;
	193	if ((req_size > (1ULL << 31)) \|\| /* Too Big */
	194	(xio_addr < xio_base) \|\| /* Out of range for mappings */
	195	(endoff > (1ULL << 31))) { /* Too Big */
	196	return 0;
	197	}
	198
	199	return PCI32_DIRECT_BASE \| offset;
1da177e4 LT	200	}
	201
	202	/*
72fdbdce	203	* Wrapper routine for freeing DMA maps
1da177e4 LT	204	* DMA mappings for Direct 64 and 32 do not have any DMA maps.
	205	*/
	206	void
e955d825	207	pcibr_dma_unmap(struct pci_dev *hwdev, dma_addr_t dma_handle, int direction)
1da177e4	208	{
e955d825 MM	209	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);
	210	struct pcibus_info *pcibus_info =
	211	(struct pcibus_info *)pcidev_info->pdi_pcibus_info;
1da177e4 LT	212
	213	if (IS_PCI32_MAPPED(dma_handle)) {
	214	int ate_index;
	215
	216	ate_index =
	217	IOPG((ATE_SWAP_OFF(dma_handle) - PCI32_MAPPED_BASE));
	218	pcibr_ate_free(pcibus_info, ate_index);
	219	}
	220	}
	221
	222	/*
	223	* On SN systems there is a race condition between a PIO read response and
	224	* DMA's. In rare cases, the read response may beat the DMA, causing the
	225	* driver to think that data in memory is complete and meaningful. This code
	226	* eliminates that race. This routine is called by the PIO read routines
	227	* after doing the read. For PIC this routine then forces a fake interrupt
	228	* on another line, which is logically associated with the slot that the PIO
	229	* is addressed to. It then spins while watching the memory location that
	230	* the interrupt is targetted to. When the interrupt response arrives, we
	231	* are sure that the DMA has landed in memory and it is safe for the driver
	232	* to proceed. For TIOCP use the Device(x) Write Request Buffer Flush
	233	* Bridge register since it ensures the data has entered the coherence domain,
	234	* unlike the PIC Device(x) Write Request Buffer Flush register.
	235	*/
	236
53493dcf	237	void sn_dma_flush(u64 addr)
1da177e4 LT	238	{
	239	nasid_t nasid;
	240	int is_tio;
	241	int wid_num;
	242	int i, j;
c53421b1	243	unsigned long flags;
53493dcf	244	u64 itte;
1da177e4	245	struct hubdev_info *hubinfo;
8ed9b2c7 JS	246	struct sn_flush_device_kernel *p;
8ed9b2c7 JS	247	struct sn_flush_device_common *common;
1da177e4 LT	248	struct sn_flush_nasid_entry *flush_nasid_list;
	249
	250	if (!sn_ioif_inited)
	251	return;
	252
	253	nasid = NASID_GET(addr);
	254	if (-1 == nasid_to_cnodeid(nasid))
	255	return;
	256
	257	hubinfo = (NODEPDA(nasid_to_cnodeid(nasid)))->pdinfo;
	258
	259	if (!hubinfo) {
	260	BUG();
	261	}
1da177e4 LT	262
	263	flush_nasid_list = &hubinfo->hdi_flush_nasid_list;
	264	if (flush_nasid_list->widget_p == NULL)
	265	return;
5b53ed1f MM	266
	267	is_tio = (nasid & 1);
	268	if (is_tio) {
	269	int itte_index;
	270
	271	if (TIO_HWIN(addr))
	272	itte_index = 0;
	273	else if (TIO_BWIN_WINDOWNUM(addr))
	274	itte_index = TIO_BWIN_WINDOWNUM(addr);
	275	else
	276	itte_index = -1;
	277
	278	if (itte_index >= 0) {
	279	itte = flush_nasid_list->iio_itte[itte_index];
	280	if (! TIO_ITTE_VALID(itte))
	281	return;
	282	wid_num = TIO_ITTE_WIDGET(itte);
	283	} else
	284	wid_num = TIO_SWIN_WIDGETNUM(addr);
	285	} else {
	286	if (BWIN_WINDOWNUM(addr)) {
	287	itte = flush_nasid_list->iio_itte[BWIN_WINDOWNUM(addr)];
	288	wid_num = IIO_ITTE_WIDGET(itte);
	289	} else
	290	wid_num = SWIN_WIDGETNUM(addr);
1da177e4	291	}
1da177e4 LT	292	if (flush_nasid_list->widget_p[wid_num] == NULL)
	293	return;
	294	p = &flush_nasid_list->widget_p[wid_num][0];
	295
	296	/* find a matching BAR */
6d6e4200 PB	297	for (i = 0; i < DEV_PER_WIDGET; i++,p++) {
6d6e4200 PB	298	common = p->common;
1da177e4	299	for (j = 0; j < PCI_ROM_RESOURCE; j++) {
6d6e4200	300	if (common->sfdl_bar_list[j].start == 0)
1da177e4	301	break;
6d6e4200 PB	302	if (addr >= common->sfdl_bar_list[j].start
6d6e4200 PB	303	&& addr <= common->sfdl_bar_list[j].end)
1da177e4 LT	304	break;
1da177e4 LT	305	}
6d6e4200	306	if (j < PCI_ROM_RESOURCE && common->sfdl_bar_list[j].start != 0)
1da177e4	307	break;
1da177e4 LT	308	}
	309
	310	/* if no matching BAR, return without doing anything. */
	311	if (i == DEV_PER_WIDGET)
	312	return;
	313
	314	/*
	315	* For TIOCP use the Device(x) Write Request Buffer Flush Bridge
	316	* register since it ensures the data has entered the coherence
5b53ed1f	317	* domain, unlike PIC.
1da177e4 LT	318	*/
1da177e4 LT	319	if (is_tio) {
5b53ed1f MM	320	/*
	321	* Note: devices behind TIOCE should never be matched in the
	322	* above code, and so the following code is PIC/CP centric.
	323	* If CE ever needs the sn_dma_flush mechanism, we will have
	324	* to account for that here and in tioce_bus_fixup().
	325	*/
53493dcf PB	326	u32 tio_id = HUB_L(TIO_IOSPACE_ADDR(nasid, TIO_NODE_ID));
53493dcf PB	327	u32 revnum = XWIDGET_PART_REV_NUM(tio_id);
1da177e4 LT	328
	329	/* TIOCP BRINGUP WAR (PV907516): Don't write buffer flush reg */
	330	if ((1 << XWIDGET_PART_REV_NUM_REV(revnum)) & PV907516) {
	331	return;
	332	} else {
6d6e4200 PB	333	pcireg_wrb_flush_get(common->sfdl_pcibus_info,
6d6e4200 PB	334	(common->sfdl_slot - 1));
1da177e4 LT	335	}
1da177e4 LT	336	} else {
8ed9b2c7	337	spin_lock_irqsave(&p->sfdl_flush_lock, flags);
6d6e4200	338	*common->sfdl_flush_addr = 0;
1da177e4 LT	339
1da177e4 LT	340	/* force an interrupt. */
53493dcf	341	(volatile u32 )(common->sfdl_force_int_addr) = 1;
1da177e4 LT	342
1da177e4 LT	343	/* wait for the interrupt to come back. */
6d6e4200	344	while (*(common->sfdl_flush_addr) != 0x10f)
68b9753f	345	cpu_relax();
1da177e4 LT	346
1da177e4 LT	347	/* okay, everything is synched up. */
8ed9b2c7	348	spin_unlock_irqrestore(&p->sfdl_flush_lock, flags);
1da177e4 LT	349	}
	350	return;
	351	}
	352
	353	/*
e955d825	354	* DMA interfaces. Called from pci_dma.c routines.
1da177e4 LT	355	*/
1da177e4 LT	356
e955d825	357	dma_addr_t
83821d3f	358	pcibr_dma_map(struct pci_dev * hwdev, unsigned long phys_addr, size_t size, int dma_flags)
1da177e4 LT	359	{
1da177e4 LT	360	dma_addr_t dma_handle;
e955d825	361	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);
1da177e4	362
e955d825 MM	363	/* SN cannot support DMA addresses smaller than 32 bits. */
	364	if (hwdev->dma_mask < 0x7fffffff) {
	365	return 0;
	366	}
1da177e4	367
e955d825 MM	368	if (hwdev->dma_mask == ~0UL) {
	369	/*
	370	* Handle the most common case: 64 bit cards. This
	371	* call should always succeed.
	372	*/
	373
	374	dma_handle = pcibr_dmatrans_direct64(pcidev_info, phys_addr,
83821d3f	375	PCI64_ATTR_PREF, dma_flags);
e955d825 MM	376	} else {
	377	/* Handle 32-63 bit cards via direct mapping */
	378	dma_handle = pcibr_dmatrans_direct32(pcidev_info, phys_addr,
83821d3f	379	size, 0, dma_flags);
e955d825	380	if (!dma_handle) {
1da177e4	381	/*
e955d825 MM	382	* It is a 32 bit card and we cannot do direct mapping,
e955d825 MM	383	* so we use an ATE.
1da177e4 LT	384	*/
1da177e4 LT	385
e955d825	386	dma_handle = pcibr_dmamap_ate32(pcidev_info, phys_addr,
83821d3f MM	387	size, PCI32_ATE_PREF,
83821d3f MM	388	dma_flags);
1da177e4 LT	389	}
	390	}
	391
	392	return dma_handle;
	393	}
	394
e955d825 MM	395	dma_addr_t
e955d825 MM	396	pcibr_dma_map_consistent(struct pci_dev * hwdev, unsigned long phys_addr,
83821d3f	397	size_t size, int dma_flags)
e955d825 MM	398	{
	399	dma_addr_t dma_handle;
	400	struct pcidev_info *pcidev_info = SN_PCIDEV_INFO(hwdev);
	401
	402	if (hwdev->dev.coherent_dma_mask == ~0UL) {
	403	dma_handle = pcibr_dmatrans_direct64(pcidev_info, phys_addr,
83821d3f	404	PCI64_ATTR_BAR, dma_flags);
e955d825 MM	405	} else {
	406	dma_handle = (dma_addr_t) pcibr_dmamap_ate32(pcidev_info,
	407	phys_addr, size,
83821d3f	408	PCI32_ATE_BAR, dma_flags);
e955d825 MM	409	}
	410
	411	return dma_handle;
	412	}
	413
1da177e4	414	EXPORT_SYMBOL(sn_dma_flush);