Make vfio MSI interrupt be non-threaded.
[kvmfornfv.git] / kernel / drivers / vfio / pci / vfio_pci_rdwr.c
1 /*
2  * VFIO PCI I/O Port & MMIO access
3  *
4  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
5  *     Author: Alex Williamson <alex.williamson@redhat.com>
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License version 2 as
9  * published by the Free Software Foundation.
10  *
11  * Derived from original vfio:
12  * Copyright 2010 Cisco Systems, Inc.  All rights reserved.
13  * Author: Tom Lyon, pugs@cisco.com
14  */
15
16 #include <linux/fs.h>
17 #include <linux/pci.h>
18 #include <linux/uaccess.h>
19 #include <linux/io.h>
20 #include <linux/vgaarb.h>
21
22 #include "vfio_pci_private.h"
23
24 /*
25  * Read or write from an __iomem region (MMIO or I/O port) with an excluded
26  * range which is inaccessible.  The excluded range drops writes and fills
27  * reads with -1.  This is intended for handling MSI-X vector tables and
28  * leftover space for ROM BARs.
29  */
30 static ssize_t do_io_rw(void __iomem *io, char __user *buf,
31                         loff_t off, size_t count, size_t x_start,
32                         size_t x_end, bool iswrite)
33 {
34         ssize_t done = 0;
35
36         while (count) {
37                 size_t fillable, filled;
38
39                 if (off < x_start)
40                         fillable = min(count, (size_t)(x_start - off));
41                 else if (off >= x_end)
42                         fillable = count;
43                 else
44                         fillable = 0;
45
46                 if (fillable >= 4 && !(off % 4)) {
47                         __le32 val;
48
49                         if (iswrite) {
50                                 if (copy_from_user(&val, buf, 4))
51                                         return -EFAULT;
52
53                                 iowrite32(le32_to_cpu(val), io + off);
54                         } else {
55                                 val = cpu_to_le32(ioread32(io + off));
56
57                                 if (copy_to_user(buf, &val, 4))
58                                         return -EFAULT;
59                         }
60
61                         filled = 4;
62                 } else if (fillable >= 2 && !(off % 2)) {
63                         __le16 val;
64
65                         if (iswrite) {
66                                 if (copy_from_user(&val, buf, 2))
67                                         return -EFAULT;
68
69                                 iowrite16(le16_to_cpu(val), io + off);
70                         } else {
71                                 val = cpu_to_le16(ioread16(io + off));
72
73                                 if (copy_to_user(buf, &val, 2))
74                                         return -EFAULT;
75                         }
76
77                         filled = 2;
78                 } else if (fillable) {
79                         u8 val;
80
81                         if (iswrite) {
82                                 if (copy_from_user(&val, buf, 1))
83                                         return -EFAULT;
84
85                                 iowrite8(val, io + off);
86                         } else {
87                                 val = ioread8(io + off);
88
89                                 if (copy_to_user(buf, &val, 1))
90                                         return -EFAULT;
91                         }
92
93                         filled = 1;
94                 } else {
95                         /* Fill reads with -1, drop writes */
96                         filled = min(count, (size_t)(x_end - off));
97                         if (!iswrite) {
98                                 u8 val = 0xFF;
99                                 size_t i;
100
101                                 for (i = 0; i < filled; i++)
102                                         if (copy_to_user(buf + i, &val, 1))
103                                                 return -EFAULT;
104                         }
105                 }
106
107                 count -= filled;
108                 done += filled;
109                 off += filled;
110                 buf += filled;
111         }
112
113         return done;
114 }
115
116 ssize_t vfio_pci_bar_rw(struct vfio_pci_device *vdev, char __user *buf,
117                         size_t count, loff_t *ppos, bool iswrite)
118 {
119         struct pci_dev *pdev = vdev->pdev;
120         loff_t pos = *ppos & VFIO_PCI_OFFSET_MASK;
121         int bar = VFIO_PCI_OFFSET_TO_INDEX(*ppos);
122         size_t x_start = 0, x_end = 0;
123         resource_size_t end;
124         void __iomem *io;
125         ssize_t done;
126
127         if (!pci_resource_start(pdev, bar))
128                 return -EINVAL;
129
130         end = pci_resource_len(pdev, bar);
131
132         if (pos >= end)
133                 return -EINVAL;
134
135         count = min(count, (size_t)(end - pos));
136
137         if (bar == PCI_ROM_RESOURCE) {
138                 /*
139                  * The ROM can fill less space than the BAR, so we start the
140                  * excluded range at the end of the actual ROM.  This makes
141                  * filling large ROM BARs much faster.
142                  */
143                 io = pci_map_rom(pdev, &x_start);
144                 if (!io)
145                         return -ENOMEM;
146                 x_end = end;
147         } else if (!vdev->barmap[bar]) {
148                 int ret;
149
150                 ret = pci_request_selected_regions(pdev, 1 << bar, "vfio");
151                 if (ret)
152                         return ret;
153
154                 io = pci_iomap(pdev, bar, 0);
155                 if (!io) {
156                         pci_release_selected_regions(pdev, 1 << bar);
157                         return -ENOMEM;
158                 }
159
160                 vdev->barmap[bar] = io;
161         } else
162                 io = vdev->barmap[bar];
163
164         if (bar == vdev->msix_bar) {
165                 x_start = vdev->msix_offset;
166                 x_end = vdev->msix_offset + vdev->msix_size;
167         }
168
169         done = do_io_rw(io, buf, pos, count, x_start, x_end, iswrite);
170
171         if (done >= 0)
172                 *ppos += done;
173
174         if (bar == PCI_ROM_RESOURCE)
175                 pci_unmap_rom(pdev, io);
176
177         return done;
178 }
179
180 ssize_t vfio_pci_vga_rw(struct vfio_pci_device *vdev, char __user *buf,
181                                size_t count, loff_t *ppos, bool iswrite)
182 {
183         int ret;
184         loff_t off, pos = *ppos & VFIO_PCI_OFFSET_MASK;
185         void __iomem *iomem = NULL;
186         unsigned int rsrc;
187         bool is_ioport;
188         ssize_t done;
189
190         if (!vdev->has_vga)
191                 return -EINVAL;
192
193         switch (pos) {
194         case 0xa0000 ... 0xbffff:
195                 count = min(count, (size_t)(0xc0000 - pos));
196                 iomem = ioremap_nocache(0xa0000, 0xbffff - 0xa0000 + 1);
197                 off = pos - 0xa0000;
198                 rsrc = VGA_RSRC_LEGACY_MEM;
199                 is_ioport = false;
200                 break;
201         case 0x3b0 ... 0x3bb:
202                 count = min(count, (size_t)(0x3bc - pos));
203                 iomem = ioport_map(0x3b0, 0x3bb - 0x3b0 + 1);
204                 off = pos - 0x3b0;
205                 rsrc = VGA_RSRC_LEGACY_IO;
206                 is_ioport = true;
207                 break;
208         case 0x3c0 ... 0x3df:
209                 count = min(count, (size_t)(0x3e0 - pos));
210                 iomem = ioport_map(0x3c0, 0x3df - 0x3c0 + 1);
211                 off = pos - 0x3c0;
212                 rsrc = VGA_RSRC_LEGACY_IO;
213                 is_ioport = true;
214                 break;
215         default:
216                 return -EINVAL;
217         }
218
219         if (!iomem)
220                 return -ENOMEM;
221
222         ret = vga_get_interruptible(vdev->pdev, rsrc);
223         if (ret) {
224                 is_ioport ? ioport_unmap(iomem) : iounmap(iomem);
225                 return ret;
226         }
227
228         done = do_io_rw(iomem, buf, off, count, 0, 0, iswrite);
229
230         vga_put(vdev->pdev, rsrc);
231
232         is_ioport ? ioport_unmap(iomem) : iounmap(iomem);
233
234         if (done >= 0)
235                 *ppos += done;
236
237         return done;
238 }