xref: /kvmtool/disk/core.c (revision 5236b505167f076ed4b0a9c6cc680712ab38baf9)
1 #include "kvm/disk-image.h"
2 #include "kvm/qcow.h"
3 #include "kvm/virtio-blk.h"
4 
5 #include <linux/err.h>
6 #include <sys/eventfd.h>
7 #include <sys/poll.h>
8 
9 #define AIO_MAX 256
10 
11 int debug_iodelay;
12 
13 #ifdef CONFIG_HAS_AIO
14 static void *disk_image__thread(void *param)
15 {
16 	struct disk_image *disk = param;
17 	struct io_event event[AIO_MAX];
18 	struct timespec notime = {0};
19 	int nr, i;
20 	u64 dummy;
21 
22 	while (read(disk->evt, &dummy, sizeof(dummy)) > 0) {
23 		nr = io_getevents(disk->ctx, 1, ARRAY_SIZE(event), event, &notime);
24 		for (i = 0; i < nr; i++)
25 			disk->disk_req_cb(event[i].data, event[i].res);
26 	}
27 
28 	return NULL;
29 }
30 #endif
31 
32 struct disk_image *disk_image__new(int fd, u64 size,
33 				   struct disk_image_operations *ops,
34 				   int use_mmap)
35 {
36 	struct disk_image *disk;
37 	int r;
38 
39 	disk = malloc(sizeof *disk);
40 	if (!disk)
41 		return ERR_PTR(-ENOMEM);
42 
43 	*disk = (struct disk_image) {
44 		.fd	= fd,
45 		.size	= size,
46 		.ops	= ops,
47 	};
48 
49 	if (use_mmap == DISK_IMAGE_MMAP) {
50 		/*
51 		 * The write to disk image will be discarded
52 		 */
53 		disk->priv = mmap(NULL, size, PROT_RW, MAP_PRIVATE | MAP_NORESERVE, fd, 0);
54 		if (disk->priv == MAP_FAILED) {
55 			r = -errno;
56 			free(disk);
57 			return ERR_PTR(r);
58 		}
59 	}
60 
61 #ifdef CONFIG_HAS_AIO
62 	if (disk) {
63 		pthread_t thread;
64 
65 		disk->evt = eventfd(0, 0);
66 		io_setup(AIO_MAX, &disk->ctx);
67 		r = pthread_create(&thread, NULL, disk_image__thread, disk);
68 		if (r) {
69 			r = -errno;
70 			free(disk);
71 			return ERR_PTR(r);
72 		}
73 	}
74 #endif
75 	return disk;
76 }
77 
78 struct disk_image *disk_image__open(const char *filename, bool readonly, bool direct)
79 {
80 	struct disk_image *disk;
81 	struct stat st;
82 	int fd, flags;
83 
84 	if (readonly)
85 		flags = O_RDONLY;
86 	else
87 		flags = O_RDWR;
88 	if (direct)
89 		flags |= O_DIRECT;
90 
91 	if (stat(filename, &st) < 0)
92 		return ERR_PTR(-errno);
93 
94 	/* blk device ?*/
95 	disk = blkdev__probe(filename, flags, &st);
96 	if (!IS_ERR_OR_NULL(disk))
97 		return disk;
98 
99 	fd = open(filename, flags);
100 	if (fd < 0)
101 		return ERR_PTR(fd);
102 
103 	/* qcow image ?*/
104 	disk = qcow_probe(fd, true);
105 	if (!IS_ERR_OR_NULL(disk)) {
106 		pr_warning("Forcing read-only support for QCOW");
107 		return disk;
108 	}
109 
110 	/* raw image ?*/
111 	disk = raw_image__probe(fd, &st, readonly);
112 	if (!IS_ERR_OR_NULL(disk))
113 		return disk;
114 
115 	if (close(fd) < 0)
116 		pr_warning("close() failed");
117 
118 	return ERR_PTR(-ENOSYS);
119 }
120 
121 struct disk_image **disk_image__open_all(struct disk_image_params *params, int count)
122 {
123 	struct disk_image **disks;
124 	const char *filename;
125 	bool readonly;
126 	bool direct;
127 	void *err;
128 	int i;
129 
130 	if (!count)
131 		return ERR_PTR(-EINVAL);
132 	if (count > MAX_DISK_IMAGES)
133 		return ERR_PTR(-ENOSPC);
134 
135 	disks = calloc(count, sizeof(*disks));
136 	if (!disks)
137 		return ERR_PTR(-ENOMEM);
138 
139 	for (i = 0; i < count; i++) {
140 		filename = params[i].filename;
141 		readonly = params[i].readonly;
142 		direct = params[i].direct;
143 		if (!filename)
144 			continue;
145 
146 		disks[i] = disk_image__open(filename, readonly, direct);
147 		if (IS_ERR_OR_NULL(disks[i])) {
148 			pr_err("Loading disk image '%s' failed", filename);
149 			err = disks[i];
150 			goto error;
151 		}
152 	}
153 
154 	return disks;
155 error:
156 	for (i = 0; i < count; i++)
157 		if (!IS_ERR_OR_NULL(disks[i]))
158 			disk_image__close(disks[i]);
159 
160 	free(disks);
161 	return err;
162 }
163 
164 int disk_image__flush(struct disk_image *disk)
165 {
166 	if (disk->ops->flush)
167 		return disk->ops->flush(disk);
168 
169 	return fsync(disk->fd);
170 }
171 
172 int disk_image__close(struct disk_image *disk)
173 {
174 	/* If there was no disk image then there's nothing to do: */
175 	if (!disk)
176 		return 0;
177 
178 	if (disk->ops->close)
179 		return disk->ops->close(disk);
180 
181 	if (close(disk->fd) < 0)
182 		pr_warning("close() failed");
183 
184 	free(disk);
185 
186 	return 0;
187 }
188 
189 int disk_image__close_all(struct disk_image **disks, int count)
190 {
191 	while (count)
192 		disk_image__close(disks[--count]);
193 
194 	free(disks);
195 
196 	return 0;
197 }
198 
199 /*
200  * Fill iov with disk data, starting from sector 'sector'.
201  * Return amount of bytes read.
202  */
203 ssize_t disk_image__read(struct disk_image *disk, u64 sector,
204 			 const struct iovec *iov, int iovcount, void *param)
205 {
206 	ssize_t total = 0;
207 
208 	if (debug_iodelay)
209 		msleep(debug_iodelay);
210 
211 	if (disk->ops->read) {
212 		total = disk->ops->read(disk, sector, iov, iovcount, param);
213 		if (total < 0) {
214 			pr_info("disk_image__read error: total=%ld\n", (long)total);
215 			return total;
216 		}
217 	}
218 
219 	if (!disk->async && disk->disk_req_cb)
220 		disk->disk_req_cb(param, total);
221 
222 	return total;
223 }
224 
225 /*
226  * Write iov to disk, starting from sector 'sector'.
227  * Return amount of bytes written.
228  */
229 ssize_t disk_image__write(struct disk_image *disk, u64 sector,
230 			  const struct iovec *iov, int iovcount, void *param)
231 {
232 	ssize_t total = 0;
233 
234 	if (debug_iodelay)
235 		msleep(debug_iodelay);
236 
237 	if (disk->ops->write) {
238 		/*
239 		 * Try writev based operation first
240 		 */
241 
242 		total = disk->ops->write(disk, sector, iov, iovcount, param);
243 		if (total < 0) {
244 			pr_info("disk_image__write error: total=%ld\n", (long)total);
245 			return total;
246 		}
247 	} else {
248 		/* Do nothing */
249 	}
250 
251 	if (!disk->async && disk->disk_req_cb)
252 		disk->disk_req_cb(param, total);
253 
254 	return total;
255 }
256 
257 ssize_t disk_image__get_serial(struct disk_image *disk, void *buffer, ssize_t *len)
258 {
259 	struct stat st;
260 	int r;
261 
262 	r = fstat(disk->fd, &st);
263 	if (r)
264 		return r;
265 
266 	*len = snprintf(buffer, *len, "%llu%llu%llu",
267 			(u64)st.st_dev, (u64)st.st_rdev, (u64)st.st_ino);
268 	return *len;
269 }
270 
271 void disk_image__set_callback(struct disk_image *disk,
272 			      void (*disk_req_cb)(void *param, long len))
273 {
274 	disk->disk_req_cb = disk_req_cb;
275 }
276