1 /*
2  * Copyright (c) 2012 Intel Corporation. All rights reserved.
3  * Copyright (c) 2006 - 2012 QLogic Corporation. All rights reserved.
4  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 /*
36  * This file is conditionally built on x86_64 only.  Otherwise weak symbol
37  * versions of the functions exported from here are used.
38  */
39 
40 #include <linux/pci.h>
41 #include <asm/mtrr.h>
42 #include <asm/processor.h>
43 
44 #include "qib.h"
45 
46 /**
47  * qib_enable_wc - enable write combining for MMIO writes to the device
48  * @dd: qlogic_ib device
49  *
50  * This routine is x86_64-specific; it twiddles the CPU's MTRRs to enable
51  * write combining.
52  */
qib_enable_wc(struct qib_devdata * dd)53 int qib_enable_wc(struct qib_devdata *dd)
54 {
55 	int ret = 0;
56 	u64 pioaddr, piolen;
57 	unsigned bits;
58 	const unsigned long addr = pci_resource_start(dd->pcidev, 0);
59 	const size_t len = pci_resource_len(dd->pcidev, 0);
60 
61 	/*
62 	 * Set the PIO buffers to be WCCOMB, so we get HT bursts to the
63 	 * chip.  Linux (possibly the hardware) requires it to be on a power
64 	 * of 2 address matching the length (which has to be a power of 2).
65 	 * For rev1, that means the base address, for rev2, it will be just
66 	 * the PIO buffers themselves.
67 	 * For chips with two sets of buffers, the calculations are
68 	 * somewhat more complicated; we need to sum, and the piobufbase
69 	 * register has both offsets, 2K in low 32 bits, 4K in high 32 bits.
70 	 * The buffers are still packed, so a single range covers both.
71 	 */
72 	if (dd->piobcnt2k && dd->piobcnt4k) {
73 		/* 2 sizes for chip */
74 		unsigned long pio2kbase, pio4kbase;
75 
76 		pio2kbase = dd->piobufbase & 0xffffffffUL;
77 		pio4kbase = (dd->piobufbase >> 32) & 0xffffffffUL;
78 		if (pio2kbase < pio4kbase) {
79 			/* all current chips */
80 			pioaddr = addr + pio2kbase;
81 			piolen = pio4kbase - pio2kbase +
82 				dd->piobcnt4k * dd->align4k;
83 		} else {
84 			pioaddr = addr + pio4kbase;
85 			piolen = pio2kbase - pio4kbase +
86 				dd->piobcnt2k * dd->palign;
87 		}
88 	} else {  /* single buffer size (2K, currently) */
89 		pioaddr = addr + dd->piobufbase;
90 		piolen = dd->piobcnt2k * dd->palign +
91 			dd->piobcnt4k * dd->align4k;
92 	}
93 
94 	for (bits = 0; !(piolen & (1ULL << bits)); bits++)
95 		; /* do nothing */
96 
97 	if (piolen != (1ULL << bits)) {
98 		piolen >>= bits;
99 		while (piolen >>= 1)
100 			bits++;
101 		piolen = 1ULL << (bits + 1);
102 	}
103 	if (pioaddr & (piolen - 1)) {
104 		u64 atmp = pioaddr & ~(piolen - 1);
105 
106 		if (atmp < addr || (atmp + piolen) > (addr + len)) {
107 			qib_dev_err(dd,
108 				"No way to align address/size (%llx/%llx), no WC mtrr\n",
109 				(unsigned long long) atmp,
110 				(unsigned long long) piolen << 1);
111 			ret = -ENODEV;
112 		} else {
113 			pioaddr = atmp;
114 			piolen <<= 1;
115 		}
116 	}
117 
118 	if (!ret) {
119 		dd->wc_cookie = arch_phys_wc_add(pioaddr, piolen);
120 		if (dd->wc_cookie < 0)
121 			/* use error from routine */
122 			ret = dd->wc_cookie;
123 	}
124 
125 	return ret;
126 }
127 
128 /**
129  * qib_disable_wc - disable write combining for MMIO writes to the device
130  * @dd: qlogic_ib device
131  */
qib_disable_wc(struct qib_devdata * dd)132 void qib_disable_wc(struct qib_devdata *dd)
133 {
134 	arch_phys_wc_del(dd->wc_cookie);
135 }
136 
137 /**
138  * qib_unordered_wc - indicate whether write combining is ordered
139  *
140  * Because our performance depends on our ability to do write combining mmio
141  * writes in the most efficient way, we need to know if we are on an Intel
142  * or AMD x86_64 processor.  AMD x86_64 processors flush WC buffers out in
143  * the order completed, and so no special flushing is required to get
144  * correct ordering.  Intel processors, however, will flush write buffers
145  * out in "random" orders, and so explicit ordering is needed at times.
146  */
qib_unordered_wc(void)147 int qib_unordered_wc(void)
148 {
149 	return boot_cpu_data.x86_vendor != X86_VENDOR_AMD;
150 }
151