2 * UCW Library -- Fast Buffered I/O on O_DIRECT Files
4 * (c) 2006 Martin Mares <mj@ucw.cz>
6 * This software may be freely distributed and used according to the terms
7 * of the GNU Lesser General Public License.
11 * This is a fastbuf backend for fast streaming I/O using O_DIRECT and
12 * the asynchronous I/O module. It's designed for use on large files
13 * which don't fit in the disk cache.
17 * - All operations with a single fbdirect handle must be done
18 * within a single thread, unless you provide a custom I/O queue
19 * and take care of locking.
21 * FIXME: what if the OS doesn't support O_DIRECT?
22 * FIXME: unaligned seeks and partial writes?
23 * FIXME: merge with other file-oriented fastbufs
29 #include "lib/fastbuf.h"
33 #include "lib/threads.h"
39 static uns fbdir_cheat;
40 static uns fbdir_buffer_size = 65536;
41 static uns fbdir_read_ahead = 1;
42 static uns fbdir_write_back = 1;
44 static struct cf_section fbdir_cf = {
46 CF_UNS("Cheat", &fbdir_cheat),
47 CF_UNS("BufferSize", &fbdir_buffer_size),
48 CF_UNS("ReadAhead", &fbdir_read_ahead),
49 CF_UNS("WriteBack", &fbdir_write_back),
54 #define FBDIR_ALIGN 512
56 enum fbdir_mode { // Current operating mode
64 int fd; // File descriptor
65 int is_temp_file; // 0=normal file, 1=temporary file, delete on close, -1=shared FD
66 struct asio_queue *io_queue; // I/O queue to use
67 struct asio_queue *user_queue; // If io_queue was supplied by the user
68 struct asio_request *pending_read;
69 struct asio_request *done_read;
70 struct asio_request *active_buffer;
74 #define FB_DIRECT(f) ((struct fb_direct *)(f)->is_fastbuf)
76 static void CONSTRUCTOR
77 fbdir_global_init(void)
79 cf_declare_section("FBDirect", &fbdir_cf, 0);
83 fbdir_read_sync(struct fb_direct *F)
85 while (F->pending_read)
87 struct asio_request *r = asio_wait(F->io_queue);
89 struct fb_direct *G = r->user_data;
91 ASSERT(G->pending_read == r && !G->done_read);
92 G->pending_read = NULL;
98 fbdir_change_mode(struct fb_direct *F, enum fbdir_mode mode)
102 DBG("FB-DIRECT: Switching mode to %d", mode);
108 fbdir_read_sync(F); // Wait for read-ahead requests to finish
109 if (F->done_read) // Return read-ahead requests if any
111 asio_put(F->done_read);
116 asio_sync(F->io_queue); // Wait for pending writebacks
119 if (F->active_buffer)
121 asio_put(F->active_buffer);
122 F->active_buffer = NULL;
128 fbdir_submit_read(struct fb_direct *F)
130 struct asio_request *r = asio_get(F->io_queue);
133 r->len = F->io_queue->buffer_size;
140 fbdir_refill(struct fastbuf *f)
142 struct fb_direct *F = FB_DIRECT(f);
144 DBG("FB-DIRECT: Refill");
148 if (!F->pending_read)
150 fbdir_change_mode(F, M_READ);
151 fbdir_submit_read(F);
154 ASSERT(F->done_read);
157 struct asio_request *r = F->done_read;
159 if (F->active_buffer)
160 asio_put(F->active_buffer);
161 F->active_buffer = r;
165 die("Error reading %s: %s", f->name, strerror(r->returned_errno));
166 f->bptr = f->buffer = r->buffer;
167 f->bstop = f->bufend = f->buffer + r->status;
170 fbdir_submit_read(F); // Read-ahead the next block
176 fbdir_spout(struct fastbuf *f)
178 struct fb_direct *F = FB_DIRECT(f);
179 struct asio_request *r;
181 DBG("FB-DIRECT: Spout");
183 fbdir_change_mode(F, M_WRITE);
184 r = F->active_buffer;
185 if (r && f->bptr > f->bstop)
187 r->op = ASIO_WRITE_BACK;
189 r->len = f->bptr - f->bstop;
190 ASSERT(!(f->pos % FBDIR_ALIGN) || fbdir_cheat);
192 if (!fbdir_cheat && r->len % FBDIR_ALIGN) // Have to simulate incomplete writes
194 r->len = ALIGN_TO(r->len, FBDIR_ALIGN);
196 asio_sync(F->io_queue);
197 DBG("FB-DIRECT: Truncating at %Ld", (long long)f->pos);
198 if (sh_ftruncate(F->fd, f->pos) < 0)
199 die("Error truncating %s: %m", f->name);
206 r = asio_get(F->io_queue);
207 f->bstop = f->bptr = f->buffer = r->buffer;
208 f->bufend = f->buffer + F->io_queue->buffer_size;
209 F->active_buffer = r;
213 fbdir_seek(struct fastbuf *f, sh_off_t pos, int whence)
215 DBG("FB-DIRECT: Seek %Ld %d", (long long)pos, whence);
217 if (whence == SEEK_SET && pos == f->pos)
220 fbdir_change_mode(FB_DIRECT(f), M_NULL); // Wait for all async requests to finish
221 sh_off_t l = sh_seek(FB_DIRECT(f)->fd, pos, whence);
223 die("lseek on %s: %m", f->name);
227 static struct asio_queue *
228 fbdir_get_io_queue(void)
230 struct ucwlib_context *ctx = ucwlib_thread_context();
231 struct asio_queue *q = ctx->io_queue;
234 q = xmalloc_zero(sizeof(struct asio_queue));
235 q->buffer_size = fbdir_buffer_size;
236 q->max_writebacks = fbdir_write_back;
241 DBG("FB-DIRECT: Got I/O queue, uc=%d", q->use_count);
246 fbdir_put_io_queue(void)
248 struct ucwlib_context *ctx = ucwlib_thread_context();
249 struct asio_queue *q = ctx->io_queue;
251 DBG("FB-DIRECT: Put I/O queue, uc=%d", q->use_count);
254 asio_cleanup_queue(q);
256 ctx->io_queue = NULL;
261 fbdir_close(struct fastbuf *f)
263 struct fb_direct *F = FB_DIRECT(f);
265 DBG("FB-DIRECT: Close");
267 fbdir_change_mode(F, M_NULL);
269 fbdir_put_io_queue();
271 switch (F->is_temp_file)
274 if (unlink(f->name) < 0)
275 log(L_ERROR, "unlink(%s): %m", f->name);
284 fbdir_config(struct fastbuf *f, uns item, int value)
288 case BCONFIG_IS_TEMP_FILE:
289 FB_DIRECT(f)->is_temp_file = value;
296 static struct fastbuf *
297 fbdir_open_internal(byte *name, int fd, struct asio_queue *q)
299 int namelen = strlen(name) + 1;
300 struct fb_direct *F = xmalloc(sizeof(struct fb_direct) + namelen);
301 struct fastbuf *f = &F->fb;
303 DBG("FB-DIRECT: Open");
304 bzero(F, sizeof(*F));
306 memcpy(f->name, name, namelen);
309 F->io_queue = F->user_queue = q;
311 F->io_queue = fbdir_get_io_queue();
312 f->refill = fbdir_refill;
313 f->spout = fbdir_spout;
314 f->seek = fbdir_seek;
315 f->close = fbdir_close;
316 f->config = fbdir_config;
317 f->can_overwrite_buffer = 2;
322 fbdir_open_try(byte *name, uns mode, struct asio_queue *q)
326 int fd = sh_open(name, mode, 0666);
329 struct fastbuf *b = fbdir_open_internal(name, fd, q);
331 fbdir_seek(b, 0, SEEK_END);
336 fbdir_open(byte *name, uns mode, struct asio_queue *q)
338 struct fastbuf *b = fbdir_open_try(name, mode, q);
340 die("Unable to %s file %s: %m",
341 (mode & O_CREAT) ? "create" : "open", name);
346 fbdir_open_fd(int fd, struct asio_queue *q)
350 sprintf(x, "fd%d", fd);
351 if (!fbdir_cheat && fcntl(fd, F_SETFL, fcntl(fd, F_GETFL) | O_DIRECT) < 0)
352 log(L_WARN, "Cannot set O_DIRECT on fd %d: %m", fd);
353 return fbdir_open_internal(x, fd, q);
357 fbdir_open_tmp(struct asio_queue *q)
359 byte buf[TEMP_FILE_NAME_LEN];
363 f = fbdir_open(buf, O_RDWR | O_CREAT | O_TRUNC, q);
364 bconfig(f, BCONFIG_IS_TEMP_FILE, 1);
370 #include "lib/getopt.h"
372 int main(int argc, char **argv)
374 struct fastbuf *f, *t;
377 if (cf_getopt(argc, argv, CF_SHORT_OPTS, CF_NO_LONG_OPTS, NULL) >= 0)
378 die("Hey, whaddya want?");
379 f = (optind < argc) ? fbdir_open(argv[optind++], O_RDONLY, NULL) : fbdir_open_fd(0, NULL);
380 t = (optind < argc) ? fbdir_open(argv[optind++], O_RDWR | O_CREAT | O_TRUNC, NULL) : fbdir_open_fd(1, NULL);
383 ASSERT(btell(f) == btell(t));
385 #if 0 // This triggers unaligned write
392 ASSERT(btell(t) == 1);