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: doc: don't mix threads
23 * FIXME: unaligned seeks and partial writes?
24 * FIXME: merge with other file-oriented fastbufs
30 #include "lib/fastbuf.h"
34 #include "lib/threads.h"
40 static uns fbdir_cheat;
41 static uns fbdir_buffer_size = 65536;
42 static uns fbdir_read_ahead = 1;
43 static uns fbdir_write_back = 1;
45 static struct cf_section fbdir_cf = {
47 CF_UNS("Cheat", &fbdir_cheat),
48 CF_UNS("BufferSize", &fbdir_buffer_size),
49 CF_UNS("ReadAhead", &fbdir_read_ahead),
50 CF_UNS("WriteBack", &fbdir_write_back),
55 #define FBDIR_ALIGN 512
57 enum fbdir_mode { // Current operating mode
65 int fd; // File descriptor
66 int is_temp_file; // 0=normal file, 1=temporary file, delete on close, -1=shared FD
67 struct asio_queue *io_queue; // I/O queue to use
68 struct asio_queue *user_queue; // If io_queue was supplied by the user
69 struct asio_request *pending_read;
70 struct asio_request *done_read;
71 struct asio_request *active_buffer;
75 #define FB_DIRECT(f) ((struct fb_direct *)(f)->is_fastbuf)
77 static void CONSTRUCTOR
78 fbdir_global_init(void)
80 cf_declare_section("FBDirect", &fbdir_cf, 0);
84 fbdir_read_sync(struct fb_direct *F)
86 while (F->pending_read)
88 struct asio_request *r = asio_wait(F->io_queue);
90 struct fb_direct *G = r->user_data;
92 ASSERT(G->pending_read == r && !G->done_read);
93 G->pending_read = NULL;
99 fbdir_change_mode(struct fb_direct *F, enum fbdir_mode mode)
103 DBG("FB-DIRECT: Switching mode to %d", mode);
109 fbdir_read_sync(F); // Wait for read-ahead requests to finish
110 if (F->done_read) // Return read-ahead requests if any
112 asio_put(F->done_read);
117 asio_sync(F->io_queue); // Wait for pending writebacks
120 if (F->active_buffer)
122 asio_put(F->active_buffer);
123 F->active_buffer = NULL;
129 fbdir_submit_read(struct fb_direct *F)
131 struct asio_request *r = asio_get(F->io_queue);
134 r->len = F->io_queue->buffer_size;
141 fbdir_refill(struct fastbuf *f)
143 struct fb_direct *F = FB_DIRECT(f);
145 DBG("FB-DIRECT: Refill");
149 if (!F->pending_read)
151 fbdir_change_mode(F, M_READ);
152 fbdir_submit_read(F);
155 ASSERT(F->done_read);
158 struct asio_request *r = F->done_read;
160 if (F->active_buffer)
161 asio_put(F->active_buffer);
162 F->active_buffer = r;
166 die("Error reading %s: %s", f->name, strerror(r->returned_errno));
167 f->bptr = f->buffer = r->buffer;
168 f->bstop = f->bufend = f->buffer + r->status;
171 fbdir_submit_read(F); // Read-ahead the next block
177 fbdir_spout(struct fastbuf *f)
179 struct fb_direct *F = FB_DIRECT(f);
180 struct asio_request *r;
182 DBG("FB-DIRECT: Spout");
184 fbdir_change_mode(F, M_WRITE);
185 r = F->active_buffer;
186 if (r && f->bptr > f->bstop)
188 r->op = ASIO_WRITE_BACK;
190 r->len = f->bptr - f->bstop;
191 ASSERT(!(f->pos % FBDIR_ALIGN) || fbdir_cheat);
193 if (!fbdir_cheat && r->len % FBDIR_ALIGN) // Have to simulate incomplete writes
195 r->len = ALIGN_TO(r->len, FBDIR_ALIGN);
197 asio_sync(F->io_queue);
198 DBG("FB-DIRECT: Truncating at %Ld", (long long)f->pos);
199 if (sh_ftruncate(F->fd, f->pos) < 0)
200 die("Error truncating %s: %m", f->name);
207 r = asio_get(F->io_queue);
208 f->bstop = f->bptr = f->buffer = r->buffer;
209 f->bufend = f->buffer + F->io_queue->buffer_size;
210 F->active_buffer = r;
214 fbdir_seek(struct fastbuf *f, sh_off_t pos, int whence)
216 DBG("FB-DIRECT: Seek %Ld %d", (long long)pos, whence);
218 if (whence == SEEK_SET && pos == f->pos)
221 fbdir_change_mode(FB_DIRECT(f), M_NULL); // Wait for all async requests to finish
222 sh_off_t l = sh_seek(FB_DIRECT(f)->fd, pos, whence);
224 die("lseek on %s: %m", f->name);
228 static struct asio_queue *
229 fbdir_get_io_queue(void)
231 struct ucwlib_context *ctx = ucwlib_thread_context();
232 struct asio_queue *q = ctx->io_queue;
235 q = xmalloc_zero(sizeof(struct asio_queue));
236 q->buffer_size = fbdir_buffer_size;
237 q->max_writebacks = fbdir_write_back;
242 DBG("FB-DIRECT: Got I/O queue, uc=%d", q->use_count);
247 fbdir_put_io_queue(void)
249 struct ucwlib_context *ctx = ucwlib_thread_context();
250 struct asio_queue *q = ctx->io_queue;
252 DBG("FB-DIRECT: Put I/O queue, uc=%d", q->use_count);
255 asio_cleanup_queue(q);
257 ctx->io_queue = NULL;
262 fbdir_close(struct fastbuf *f)
264 struct fb_direct *F = FB_DIRECT(f);
266 DBG("FB-DIRECT: Close");
268 fbdir_change_mode(F, M_NULL);
270 fbdir_put_io_queue();
272 switch (F->is_temp_file)
275 if (unlink(f->name) < 0)
276 log(L_ERROR, "unlink(%s): %m", f->name);
285 fbdir_config(struct fastbuf *f, uns item, int value)
289 case BCONFIG_IS_TEMP_FILE:
290 FB_DIRECT(f)->is_temp_file = value;
297 static struct fastbuf *
298 fbdir_open_internal(byte *name, int fd, struct asio_queue *q)
300 int namelen = strlen(name) + 1;
301 struct fb_direct *F = xmalloc(sizeof(struct fb_direct) + namelen);
302 struct fastbuf *f = &F->fb;
304 DBG("FB-DIRECT: Open");
305 bzero(F, sizeof(*F));
307 memcpy(f->name, name, namelen);
310 F->io_queue = F->user_queue = q;
312 F->io_queue = fbdir_get_io_queue();
313 f->refill = fbdir_refill;
314 f->spout = fbdir_spout;
315 f->seek = fbdir_seek;
316 f->close = fbdir_close;
317 f->config = fbdir_config;
318 f->can_overwrite_buffer = 2;
323 fbdir_open_try(byte *name, uns mode, struct asio_queue *q)
327 int fd = sh_open(name, mode, 0666);
330 struct fastbuf *b = fbdir_open_internal(name, fd, q);
332 fbdir_seek(b, 0, SEEK_END);
337 fbdir_open(byte *name, uns mode, struct asio_queue *q)
339 struct fastbuf *b = fbdir_open_try(name, mode, q);
341 die("Unable to %s file %s: %m",
342 (mode & O_CREAT) ? "create" : "open", name);
347 fbdir_open_fd(int fd, struct asio_queue *q)
351 sprintf(x, "fd%d", fd);
352 if (!fbdir_cheat && fcntl(fd, F_SETFL, fcntl(fd, F_GETFL) | O_DIRECT) < 0)
353 log(L_WARN, "Cannot set O_DIRECT on fd %d: %m", fd);
354 return fbdir_open_internal(x, fd, q);
358 fbdir_open_tmp(struct asio_queue *q)
360 byte buf[TEMP_FILE_NAME_LEN];
364 f = fbdir_open(buf, O_RDWR | O_CREAT | O_TRUNC, q);
365 bconfig(f, BCONFIG_IS_TEMP_FILE, 1);
371 #include "lib/getopt.h"
373 int main(int argc, char **argv)
375 struct fastbuf *f, *t;
378 if (cf_getopt(argc, argv, CF_SHORT_OPTS, CF_NO_LONG_OPTS, NULL) >= 0)
379 die("Hey, whaddya want?");
380 f = (optind < argc) ? fbdir_open(argv[optind++], O_RDONLY, NULL) : fbdir_open_fd(0, NULL);
381 t = (optind < argc) ? fbdir_open(argv[optind++], O_RDWR | O_CREAT | O_TRUNC, NULL) : fbdir_open_fd(1, NULL);
384 ASSERT(btell(f) == btell(t));
386 #if 0 // This triggers unaligned write
393 ASSERT(btell(t) == 1);