1 /*
2 * libhdfs engine
3 *
4 * this engine helps perform read/write operations on hdfs cluster using
5 * libhdfs. hdfs does not support modification of data once file is created.
6 *
7 * so to mimic that create many files of small size (e.g 256k), and this
8 * engine select a file based on the offset generated by fio.
9 *
10 * thus, random reads and writes can also be achieved with this logic.
11 *
12 */
13
14 #include <math.h>
15 #include <hdfs.h>
16
17 #include "../fio.h"
18 #include "../optgroup.h"
19
20 #define CHUNCK_NAME_LENGTH_MAX 80
21 #define CHUNCK_CREATION_BUFFER_SIZE 65536
22
23 struct hdfsio_data {
24 hdfsFS fs;
25 hdfsFile fp;
26 uint64_t curr_file_id;
27 };
28
29 struct hdfsio_options {
30 void *pad; /* needed because offset can't be 0 for a option defined used offsetof */
31 char *host;
32 char *directory;
33 unsigned int port;
34 unsigned int chunck_size;
35 unsigned int single_instance;
36 unsigned int use_direct;
37 };
38
39 static struct fio_option options[] = {
40 {
41 .name = "namenode",
42 .lname = "hfds namenode",
43 .type = FIO_OPT_STR_STORE,
44 .off1 = offsetof(struct hdfsio_options, host),
45 .def = "localhost",
46 .help = "Namenode of the HDFS cluster",
47 .category = FIO_OPT_C_ENGINE,
48 .group = FIO_OPT_G_HDFS,
49 },
50 {
51 .name = "hostname",
52 .lname = "hfds namenode",
53 .type = FIO_OPT_STR_STORE,
54 .off1 = offsetof(struct hdfsio_options, host),
55 .def = "localhost",
56 .help = "Namenode of the HDFS cluster",
57 .category = FIO_OPT_C_ENGINE,
58 .group = FIO_OPT_G_HDFS,
59 },
60 {
61 .name = "port",
62 .lname = "hdfs namenode port",
63 .type = FIO_OPT_INT,
64 .off1 = offsetof(struct hdfsio_options, port),
65 .def = "9000",
66 .minval = 1,
67 .maxval = 65535,
68 .help = "Port used by the HDFS cluster namenode",
69 .category = FIO_OPT_C_ENGINE,
70 .group = FIO_OPT_G_HDFS,
71 },
72 {
73 .name = "hdfsdirectory",
74 .lname = "hfds directory",
75 .type = FIO_OPT_STR_STORE,
76 .off1 = offsetof(struct hdfsio_options, directory),
77 .def = "/",
78 .help = "The HDFS directory where fio will create chunks",
79 .category = FIO_OPT_C_ENGINE,
80 .group = FIO_OPT_G_HDFS,
81 },
82 {
83 .name = "chunk_size",
84 .alias = "chunck_size",
85 .lname = "Chunk size",
86 .type = FIO_OPT_INT,
87 .off1 = offsetof(struct hdfsio_options, chunck_size),
88 .def = "1048576",
89 .help = "Size of individual chunk",
90 .category = FIO_OPT_C_ENGINE,
91 .group = FIO_OPT_G_HDFS,
92 },
93 {
94 .name = "single_instance",
95 .lname = "Single Instance",
96 .type = FIO_OPT_BOOL,
97 .off1 = offsetof(struct hdfsio_options, single_instance),
98 .def = "1",
99 .help = "Use a single instance",
100 .category = FIO_OPT_C_ENGINE,
101 .group = FIO_OPT_G_HDFS,
102 },
103 {
104 .name = "hdfs_use_direct",
105 .lname = "HDFS Use Direct",
106 .type = FIO_OPT_BOOL,
107 .off1 = offsetof(struct hdfsio_options, use_direct),
108 .def = "0",
109 .help = "Use readDirect instead of hdfsRead",
110 .category = FIO_OPT_C_ENGINE,
111 .group = FIO_OPT_G_HDFS,
112 },
113 {
114 .name = NULL,
115 },
116 };
117
118
get_chunck_name(char * dest,char * file_name,uint64_t chunk_id)119 static int get_chunck_name(char *dest, char *file_name, uint64_t chunk_id) {
120 return snprintf(dest, CHUNCK_NAME_LENGTH_MAX, "%s_%lu", file_name, chunk_id);
121 }
122
fio_hdfsio_prep(struct thread_data * td,struct io_u * io_u)123 static int fio_hdfsio_prep(struct thread_data *td, struct io_u *io_u)
124 {
125 struct hdfsio_options *options = td->eo;
126 struct hdfsio_data *hd = td->io_ops_data;
127 unsigned long f_id;
128 char fname[CHUNCK_NAME_LENGTH_MAX];
129 int open_flags;
130
131 /* find out file id based on the offset generated by fio */
132 f_id = floor(io_u->offset / options-> chunck_size);
133
134 if (f_id == hd->curr_file_id) {
135 /* file is already open */
136 return 0;
137 }
138
139 if (hd->curr_file_id != -1) {
140 if ( hdfsCloseFile(hd->fs, hd->fp) == -1) {
141 log_err("hdfs: unable to close file: %s\n", strerror(errno));
142 return errno;
143 }
144 hd->curr_file_id = -1;
145 }
146
147 if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_SYNC) {
148 open_flags = O_RDONLY;
149 } else if (io_u->ddir == DDIR_WRITE) {
150 open_flags = O_WRONLY;
151 } else {
152 log_err("hdfs: Invalid I/O Operation\n");
153 return 0;
154 }
155
156 get_chunck_name(fname, io_u->file->file_name, f_id);
157 hd->fp = hdfsOpenFile(hd->fs, fname, open_flags, 0, 0,
158 options->chunck_size);
159 if(hd->fp == NULL) {
160 log_err("hdfs: unable to open file: %s: %d\n", fname, strerror(errno));
161 return errno;
162 }
163 hd->curr_file_id = f_id;
164
165 return 0;
166 }
167
fio_hdfsio_queue(struct thread_data * td,struct io_u * io_u)168 static enum fio_q_status fio_hdfsio_queue(struct thread_data *td,
169 struct io_u *io_u)
170 {
171 struct hdfsio_data *hd = td->io_ops_data;
172 struct hdfsio_options *options = td->eo;
173 int ret;
174 unsigned long offset;
175
176 offset = io_u->offset % options->chunck_size;
177
178 if( (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) &&
179 hdfsTell(hd->fs, hd->fp) != offset && hdfsSeek(hd->fs, hd->fp, offset) != 0 ) {
180 log_err("hdfs: seek failed: %s, are you doing random write smaller than chunk size ?\n", strerror(errno));
181 io_u->error = errno;
182 return FIO_Q_COMPLETED;
183 };
184
185 // do the IO
186 if (io_u->ddir == DDIR_READ) {
187 if (options->use_direct) {
188 ret = readDirect(hd->fs, hd->fp, io_u->xfer_buf, io_u->xfer_buflen);
189 } else {
190 ret = hdfsRead(hd->fs, hd->fp, io_u->xfer_buf, io_u->xfer_buflen);
191 }
192 } else if (io_u->ddir == DDIR_WRITE) {
193 ret = hdfsWrite(hd->fs, hd->fp, io_u->xfer_buf,
194 io_u->xfer_buflen);
195 } else if (io_u->ddir == DDIR_SYNC) {
196 ret = hdfsFlush(hd->fs, hd->fp);
197 } else {
198 log_err("hdfs: Invalid I/O Operation: %d\n", io_u->ddir);
199 ret = EINVAL;
200 }
201
202 // Check if the IO went fine, or is incomplete
203 if (ret != (int)io_u->xfer_buflen) {
204 if (ret >= 0) {
205 io_u->resid = io_u->xfer_buflen - ret;
206 io_u->error = 0;
207 return FIO_Q_COMPLETED;
208 } else {
209 io_u->error = errno;
210 }
211 }
212
213 if (io_u->error)
214 td_verror(td, io_u->error, "xfer");
215
216 return FIO_Q_COMPLETED;
217 }
218
fio_hdfsio_open_file(struct thread_data * td,struct fio_file * f)219 int fio_hdfsio_open_file(struct thread_data *td, struct fio_file *f)
220 {
221 if (td->o.odirect) {
222 td->error = EINVAL;
223 return 0;
224 }
225
226 return 0;
227 }
228
fio_hdfsio_close_file(struct thread_data * td,struct fio_file * f)229 int fio_hdfsio_close_file(struct thread_data *td, struct fio_file *f)
230 {
231 struct hdfsio_data *hd = td->io_ops_data;
232
233 if (hd->curr_file_id != -1) {
234 if ( hdfsCloseFile(hd->fs, hd->fp) == -1) {
235 log_err("hdfs: unable to close file: %s\n", strerror(errno));
236 return errno;
237 }
238 hd->curr_file_id = -1;
239 }
240 return 0;
241 }
242
fio_hdfsio_io_u_init(struct thread_data * td,struct io_u * io_u)243 static int fio_hdfsio_io_u_init(struct thread_data *td, struct io_u *io_u)
244 {
245 struct hdfsio_options *options = td->eo;
246 struct hdfsio_data *hd = td->io_ops_data;
247 struct fio_file *f;
248 uint64_t j,k;
249 int i, failure = 0;
250 uint8_t buffer[CHUNCK_CREATION_BUFFER_SIZE];
251 uint64_t bytes_left;
252 char fname[CHUNCK_NAME_LENGTH_MAX];
253 hdfsFile fp;
254 hdfsFileInfo *fi;
255 tOffset fi_size;
256
257 for_each_file(td, f, i) {
258 k = 0;
259 for(j=0; j < f->real_file_size; j += options->chunck_size) {
260 get_chunck_name(fname, f->file_name, k++);
261 fi = hdfsGetPathInfo(hd->fs, fname);
262 fi_size = fi ? fi->mSize : 0;
263 // fill exist and is big enough, nothing to do
264 if( fi && fi_size >= options->chunck_size) {
265 continue;
266 }
267 fp = hdfsOpenFile(hd->fs, fname, O_WRONLY, 0, 0,
268 options->chunck_size);
269 if(fp == NULL) {
270 failure = errno;
271 log_err("hdfs: unable to prepare file chunk %s: %s\n", fname, strerror(errno));
272 break;
273 }
274 bytes_left = options->chunck_size;
275 memset(buffer, 0, CHUNCK_CREATION_BUFFER_SIZE);
276 while( bytes_left > CHUNCK_CREATION_BUFFER_SIZE) {
277 if( hdfsWrite(hd->fs, fp, buffer, CHUNCK_CREATION_BUFFER_SIZE)
278 != CHUNCK_CREATION_BUFFER_SIZE) {
279 failure = errno;
280 log_err("hdfs: unable to prepare file chunk %s: %s\n", fname, strerror(errno));
281 break;
282 };
283 bytes_left -= CHUNCK_CREATION_BUFFER_SIZE;
284 }
285 if(bytes_left > 0) {
286 if( hdfsWrite(hd->fs, fp, buffer, bytes_left)
287 != bytes_left) {
288 failure = errno;
289 break;
290 };
291 }
292 if( hdfsCloseFile(hd->fs, fp) != 0) {
293 failure = errno;
294 log_err("hdfs: unable to prepare file chunk %s: %s\n", fname, strerror(errno));
295 break;
296 }
297 }
298 if(failure) {
299 break;
300 }
301 }
302
303 if( !failure ) {
304 fio_file_set_size_known(f);
305 }
306
307 return failure;
308 }
309
fio_hdfsio_setup(struct thread_data * td)310 static int fio_hdfsio_setup(struct thread_data *td)
311 {
312 struct hdfsio_data *hd;
313 struct fio_file *f;
314 int i;
315 uint64_t file_size, total_file_size;
316
317 if (!td->io_ops_data) {
318 hd = malloc(sizeof(*hd));
319 memset(hd, 0, sizeof(*hd));
320
321 hd->curr_file_id = -1;
322
323 td->io_ops_data = hd;
324 }
325
326 total_file_size = 0;
327 file_size = 0;
328
329 for_each_file(td, f, i) {
330 if(!td->o.file_size_low) {
331 file_size = floor(td->o.size / td->o.nr_files);
332 total_file_size += file_size;
333 }
334 else if (td->o.file_size_low == td->o.file_size_high)
335 file_size = td->o.file_size_low;
336 else {
337 file_size = get_rand_file_size(td);
338 }
339 f->real_file_size = file_size;
340 }
341 /* If the size doesn't divide nicely with the chunk size,
342 * make the last files bigger.
343 * Used only if filesize was not explicitly given
344 */
345 if (!td->o.file_size_low && total_file_size < td->o.size) {
346 f->real_file_size += (td->o.size - total_file_size);
347 }
348
349 return 0;
350 }
351
fio_hdfsio_init(struct thread_data * td)352 static int fio_hdfsio_init(struct thread_data *td)
353 {
354 struct hdfsio_data *hd = td->io_ops_data;
355 struct hdfsio_options *options = td->eo;
356 int failure;
357 struct hdfsBuilder *bld;
358
359 if (options->host == NULL || options->port == 0) {
360 log_err("hdfs: server not defined\n");
361 return EINVAL;
362 }
363
364 bld = hdfsNewBuilder();
365 if (!bld) {
366 failure = errno;
367 log_err("hdfs: unable to allocate connect builder\n");
368 return failure;
369 }
370 hdfsBuilderSetNameNode(bld, options->host);
371 hdfsBuilderSetNameNodePort(bld, options->port);
372 if(! options->single_instance) {
373 hdfsBuilderSetForceNewInstance(bld);
374 }
375 hd->fs = hdfsBuilderConnect(bld);
376
377 /* hdfsSetWorkingDirectory succeed on non-existent directory */
378 if (hdfsExists(hd->fs, options->directory) < 0 || hdfsSetWorkingDirectory(hd->fs, options->directory) < 0) {
379 failure = errno;
380 log_err("hdfs: invalid working directory %s: %s\n", options->directory, strerror(errno));
381 return failure;
382 }
383
384 return 0;
385 }
386
fio_hdfsio_io_u_free(struct thread_data * td,struct io_u * io_u)387 static void fio_hdfsio_io_u_free(struct thread_data *td, struct io_u *io_u)
388 {
389 struct hdfsio_data *hd = td->io_ops_data;
390
391 if (hd->fs && hdfsDisconnect(hd->fs) < 0) {
392 log_err("hdfs: disconnect failed: %d\n", errno);
393 }
394 }
395
396 FIO_STATIC struct ioengine_ops ioengine = {
397 .name = "libhdfs",
398 .version = FIO_IOOPS_VERSION,
399 .flags = FIO_SYNCIO | FIO_DISKLESSIO | FIO_NODISKUTIL,
400 .setup = fio_hdfsio_setup,
401 .init = fio_hdfsio_init,
402 .prep = fio_hdfsio_prep,
403 .queue = fio_hdfsio_queue,
404 .open_file = fio_hdfsio_open_file,
405 .close_file = fio_hdfsio_close_file,
406 .io_u_init = fio_hdfsio_io_u_init,
407 .io_u_free = fio_hdfsio_io_u_free,
408 .option_struct_size = sizeof(struct hdfsio_options),
409 .options = options,
410 };
411
412
fio_hdfsio_register(void)413 static void fio_init fio_hdfsio_register(void)
414 {
415 register_ioengine(&ioengine);
416 }
417
fio_hdfsio_unregister(void)418 static void fio_exit fio_hdfsio_unregister(void)
419 {
420 unregister_ioengine(&ioengine);
421 }
422