| 1 |
#define _BSD_SOURCE // for endian.h |
| 2 |
|
| 3 |
#include <stddef.h> |
| 4 |
#include <stdlib.h> |
| 5 |
#include <stdio.h> |
| 6 |
#include <stdint.h> |
| 7 |
#include <string.h> |
| 8 |
|
| 9 |
#include <sys/types.h> |
| 10 |
#include <sys/socket.h> |
| 11 |
#include <netinet/in.h> |
| 12 |
#include <netinet/tcp.h> |
| 13 |
|
| 14 |
#include <sys/mman.h> |
| 15 |
#include <linux/nbd.h> |
| 16 |
#include <arpa/inet.h> |
| 17 |
|
| 18 |
#include <endian.h> |
| 19 |
|
| 20 |
// how stupid is that, the 1.2 header files define CL_VERSION_1_1, |
| 21 |
// but then fail to define the api functions unless you ALSO define |
| 22 |
// this. This breaks 100% of the opencl 1.1 apps, for what reason? |
| 23 |
// after all, the functions are deprecated, not removed. |
| 24 |
// in addition, you cannot test for this in any future-proof way. |
| 25 |
// each time a new opencl version comes out, you need to make a new |
| 26 |
// release. |
| 27 |
#define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
| 28 |
|
| 29 |
#ifdef __APPLE__ |
| 30 |
#define CLHDR(name) <OpenCL/name> |
| 31 |
#else |
| 32 |
#define CLHDR(name) <CL/name> |
| 33 |
#endif |
| 34 |
|
| 35 |
#include CLHDR(opencl.h) |
| 36 |
|
| 37 |
#undef NDEBUG |
| 38 |
#include <assert.h> |
| 39 |
|
| 40 |
static cl_int res; |
| 41 |
|
| 42 |
#define FAIL(name) \ |
| 43 |
do { fprintf (stderr, "cl" # name ": error %d", res); exit (1); } while (0) |
| 44 |
|
| 45 |
#define NEED_SUCCESS(name,args) \ |
| 46 |
do { \ |
| 47 |
res = cl ## name args; \ |
| 48 |
\ |
| 49 |
if (res) \ |
| 50 |
FAIL (name); \ |
| 51 |
} while (0) |
| 52 |
|
| 53 |
#define NEED_SUCCESS_ARG(retdecl, name, args) \ |
| 54 |
retdecl = cl ## name args; \ |
| 55 |
if (res) \ |
| 56 |
FAIL (name); |
| 57 |
|
| 58 |
typedef uint32_t off_type; |
| 59 |
|
| 60 |
static cl_platform_id platform; |
| 61 |
static cl_device_id device; |
| 62 |
static cl_context context; |
| 63 |
static cl_command_queue queue; |
| 64 |
|
| 65 |
#define LISTEN_PORT 3843 |
| 66 |
|
| 67 |
static off_type capacity = 768 * 1024 * 1024; |
| 68 |
static cl_mem disk; |
| 69 |
|
| 70 |
static struct nbd_request req; |
| 71 |
static struct nbd_reply reply; |
| 72 |
static char *buf; |
| 73 |
static uint32_t bufsize; |
| 74 |
|
| 75 |
static int fd; |
| 76 |
|
| 77 |
#ifndef MAP_LOCKED |
| 78 |
# define MAP_LOCKED 0 |
| 79 |
#endif |
| 80 |
|
| 81 |
static void |
| 82 |
xread (void *data, int len) |
| 83 |
{ |
| 84 |
int n = 0; |
| 85 |
|
| 86 |
while (len > n) |
| 87 |
{ |
| 88 |
int r = read (fd, n + (char *)data, len - n); |
| 89 |
assert (r > 0); |
| 90 |
n += r; |
| 91 |
} |
| 92 |
} |
| 93 |
|
| 94 |
static void |
| 95 |
xwrite (const void *data, int len) |
| 96 |
{ |
| 97 |
int n = 0; |
| 98 |
|
| 99 |
while (len > n) |
| 100 |
{ |
| 101 |
int r = write (fd, n + (char *)data, len - n); |
| 102 |
assert (r > 0); |
| 103 |
n += r; |
| 104 |
} |
| 105 |
} |
| 106 |
static buf_alloc (uint32_t size) |
| 107 |
{ |
| 108 |
if (bufsize >= size) |
| 109 |
return; |
| 110 |
|
| 111 |
if (buf) |
| 112 |
munmap (buf, bufsize); |
| 113 |
|
| 114 |
bufsize = (size + 4095) & ~4095; // too lazy to query PAGE_SIZE |
| 115 |
|
| 116 |
buf = mmap (0, bufsize, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_LOCKED, -1, 0); |
| 117 |
|
| 118 |
if (!MAP_LOCKED) |
| 119 |
mlock (buf, bufsize); |
| 120 |
} |
| 121 |
|
| 122 |
static void |
| 123 |
oclread (off_type off, uint32_t len, void *buf) |
| 124 |
{ |
| 125 |
NEED_SUCCESS (EnqueueReadBuffer, (queue, disk, 1, off, len, buf, 0, 0, 0)); |
| 126 |
} |
| 127 |
|
| 128 |
static void |
| 129 |
oclwrite (off_type off, uint32_t len, void *buf) |
| 130 |
{ |
| 131 |
NEED_SUCCESS (EnqueueWriteBuffer, (queue, disk, 0, off, len, buf, 0, 0, 0)); |
| 132 |
} |
| 133 |
|
| 134 |
static void |
| 135 |
init (void) |
| 136 |
{ |
| 137 |
NEED_SUCCESS_ARG (disk, CreateBuffer, (context, 0, capacity, 0, &res)); |
| 138 |
|
| 139 |
const int block = 64 * 1024; |
| 140 |
off_type o; |
| 141 |
|
| 142 |
buf_alloc (block); |
| 143 |
|
| 144 |
for (o = 0; o < capacity; o += block) |
| 145 |
{ |
| 146 |
oclwrite (o, block, buf); |
| 147 |
oclread (o, block, buf); |
| 148 |
} |
| 149 |
|
| 150 |
static struct sockaddr_in sa; |
| 151 |
|
| 152 |
sa.sin_family = AF_INET; |
| 153 |
sa.sin_port = htons (LISTEN_PORT); |
| 154 |
sa.sin_addr.s_addr = htonl (0x7f000001); |
| 155 |
|
| 156 |
int listener = socket (AF_INET, SOCK_STREAM, 0); |
| 157 |
bind (listener, (void *)&sa, sizeof (sa)); |
| 158 |
listen (listener, 1); |
| 159 |
|
| 160 |
printf ("waiting for connect on port %d...", LISTEN_PORT); fflush (stdout); |
| 161 |
fd = accept (listener, 0, 0); |
| 162 |
assert (fd >= 0); |
| 163 |
printf (" ok\n"); |
| 164 |
|
| 165 |
close (listener); |
| 166 |
|
| 167 |
{ |
| 168 |
xwrite ("NBDMAGIC", 8); |
| 169 |
xwrite ("\x00\x00\x42\x02\x81\x86\x12\x53", 8); |
| 170 |
|
| 171 |
uint64_t cap = htobe64 (capacity); |
| 172 |
xwrite (&cap, 8); |
| 173 |
|
| 174 |
int i; |
| 175 |
cap = 0; |
| 176 |
for (i = 0; i < 128 / 8; ++i) |
| 177 |
xwrite (&cap, 8); |
| 178 |
} |
| 179 |
} |
| 180 |
|
| 181 |
static void |
| 182 |
run (void) |
| 183 |
{ |
| 184 |
for (;;) |
| 185 |
{ |
| 186 |
xread (&req, sizeof (req)); |
| 187 |
assert (req.magic == htonl (NBD_REQUEST_MAGIC)); |
| 188 |
|
| 189 |
int type = ntohl (req.type); |
| 190 |
off_type off = be64toh (req.from); |
| 191 |
uint32_t len = ntohl (req.len); |
| 192 |
|
| 193 |
reply.magic = htonl (NBD_REPLY_MAGIC); |
| 194 |
memcpy (reply.handle, req.handle, sizeof (reply.handle)); |
| 195 |
reply.error = 0; |
| 196 |
|
| 197 |
//printf ("type %d off %d len %d\n", type, (int)off, (int)len);//D |
| 198 |
|
| 199 |
if (off + len > capacity) |
| 200 |
reply.error = htonl (1); |
| 201 |
else if (type == NBD_CMD_READ) |
| 202 |
{ |
| 203 |
buf_alloc (len); |
| 204 |
|
| 205 |
clFinish (queue); |
| 206 |
oclread (off, len, buf); |
| 207 |
|
| 208 |
xwrite (&reply, sizeof (reply)); |
| 209 |
xwrite (buf, len); |
| 210 |
} |
| 211 |
else if (type == NBD_CMD_WRITE) |
| 212 |
{ |
| 213 |
clFinish (queue); |
| 214 |
|
| 215 |
buf_alloc (len); |
| 216 |
xread (buf, len); |
| 217 |
xwrite (&reply, sizeof (reply)); |
| 218 |
|
| 219 |
oclwrite (off, len, buf); |
| 220 |
} |
| 221 |
else |
| 222 |
abort (); |
| 223 |
} |
| 224 |
} |
| 225 |
|
| 226 |
int |
| 227 |
main (int argc, char **argv) |
| 228 |
{ |
| 229 |
cl_platform_id *list; |
| 230 |
cl_uint count; |
| 231 |
int i; |
| 232 |
|
| 233 |
// we would really like to MCL_FUTURE, as being swapped out could turn |
| 234 |
// out to be fatal but we don't, as at least nvidia's opencl results |
| 235 |
// in a 1.5gb allocation (on my 1gb card) instantly. |
| 236 |
//mlockall (MCL_CURRENT | MCL_FUTURE); |
| 237 |
mlockall (MCL_CURRENT); |
| 238 |
|
| 239 |
NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); |
| 240 |
list = malloc (sizeof (*list) * count); |
| 241 |
NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
| 242 |
|
| 243 |
for (i = 0; i < count; ++i) |
| 244 |
{ |
| 245 |
platform = list [i]; |
| 246 |
|
| 247 |
cl_device_id *list; |
| 248 |
cl_uint count; |
| 249 |
int i; |
| 250 |
|
| 251 |
NEED_SUCCESS (GetDeviceIDs, (platform, CL_DEVICE_TYPE_GPU, 0, 0, &count)); |
| 252 |
list = malloc (sizeof (*list) * count); |
| 253 |
NEED_SUCCESS (GetDeviceIDs, (platform, CL_DEVICE_TYPE_GPU, count, list, 0)); |
| 254 |
|
| 255 |
for (i = 0; i < count; ++i) |
| 256 |
{ |
| 257 |
device = list [i]; |
| 258 |
NEED_SUCCESS_ARG (context, CreateContext, (0, 1, &device, 0, 0, &res)); |
| 259 |
NEED_SUCCESS_ARG (queue, CreateCommandQueue, (context, device, 0, &res)); |
| 260 |
|
| 261 |
init (); |
| 262 |
run (); |
| 263 |
exit (0); |
| 264 |
} |
| 265 |
|
| 266 |
free (list); |
| 267 |
} |
| 268 |
|
| 269 |
fprintf (stderr, "unable to create suitable opencl context\n"); |
| 270 |
return 1; |
| 271 |
} |
| 272 |
|
