arch/um/os-Linux/aio.c

   1 /*
   2  * Copyright (C) 2004 Jeff Dike (jdike@addtoit.com)
   3  * Licensed under the GPL
   4  */
   5
   6 #include <stdlib.h>
   7 #include <unistd.h>
   8 #include <signal.h>
   9 #include <errno.h>
  10 #include <sched.h>
  11 #include <sys/syscall.h>
  12 #include "os.h"
  13 #include "aio.h"
  14 #include "init.h"
  15 #include "user.h"
  16 #include "mode.h"
  17
  18 struct aio_thread_req {
  19         enum aio_type type;
  20         int io_fd;
  21         unsigned long long offset;
  22         char *buf;
  23         int len;
  24         struct aio_context *aio;
  25 };
  26
  27 static int aio_req_fd_r = -1;
  28 static int aio_req_fd_w = -1;
  29
  30 #if defined(HAVE_AIO_ABI)
  31 #include <linux/aio_abi.h>
  32
  33 /* If we have the headers, we are going to build with AIO enabled.
  34  * If we don't have aio in libc, we define the necessary stubs here.
  35  */
  36
  37 #if !defined(HAVE_AIO_LIBC)
  38
  39 static long io_setup(int n, aio_context_t *ctxp)
  40 {
  41         return syscall(__NR_io_setup, n, ctxp);
  42 }
  43
  44 static long io_submit(aio_context_t ctx, long nr, struct iocb **iocbpp)
  45 {
  46         return syscall(__NR_io_submit, ctx, nr, iocbpp);
  47 }
  48
  49 static long io_getevents(aio_context_t ctx_id, long min_nr, long nr,
  50                          struct io_event *events, struct timespec *timeout)
  51 {
  52         return syscall(__NR_io_getevents, ctx_id, min_nr, nr, events, timeout);
  53 }
  54
  55 #endif
  56
  57 /* The AIO_MMAP cases force the mmapped page into memory here
  58  * rather than in whatever place first touches the data.  I used
  59  * to do this by touching the page, but that's delicate because
  60  * gcc is prone to optimizing that away.  So, what's done here
  61  * is we read from the descriptor from which the page was
  62  * mapped.  The caller is required to pass an offset which is
  63  * inside the page that was mapped.  Thus, when the read
  64  * returns, we know that the page is in the page cache, and
  65  * that it now backs the mmapped area.
  66  */
  67
  68 static int do_aio(aio_context_t ctx, enum aio_type type, int fd, char *buf,
  69                   int len, unsigned long long offset, struct aio_context *aio)
  70 {
  71         struct iocb iocb, *iocbp = &iocb;
  72         char c;
  73         int err;
  74
  75         iocb = ((struct iocb) { .aio_data       = (unsigned long) aio,
  76                                 .aio_reqprio    = 0,
  77                                 .aio_fildes     = fd,
  78                                 .aio_buf        = (unsigned long) buf,
  79                                 .aio_nbytes     = len,
  80                                 .aio_offset     = offset,
  81                                 .aio_reserved1  = 0,
  82                                 .aio_reserved2  = 0,
  83                                 .aio_reserved3  = 0 });
  84
  85         switch(type){
  86         case AIO_READ:
  87                 iocb.aio_lio_opcode = IOCB_CMD_PREAD;
  88                 err = io_submit(ctx, 1, &iocbp);
  89                 break;
  90         case AIO_WRITE:
  91                 iocb.aio_lio_opcode = IOCB_CMD_PWRITE;
  92                 err = io_submit(ctx, 1, &iocbp);
  93                 break;
  94         case AIO_MMAP:
  95                 iocb.aio_lio_opcode = IOCB_CMD_PREAD;
  96                 iocb.aio_buf = (unsigned long) &c;
  97                 iocb.aio_nbytes = sizeof(c);
  98                 err = io_submit(ctx, 1, &iocbp);
  99                 break;
 100         default:
 101                 printk("Bogus op in do_aio - %d\n", type);
 102                 err = -EINVAL;
 103                 break;
 104         }
 105
 106         if(err > 0)
 107                 err = 0;
 108         else
 109                 err = -errno;
 110
 111         return err;
 112 }
 113
 114 static aio_context_t ctx = 0;
 115
 116 static int aio_thread(void *arg)
 117 {
 118         struct aio_thread_reply reply;
 119         struct io_event event;
 120         int err, n, reply_fd;
 121
 122         signal(SIGWINCH, SIG_IGN);
 123
 124         while(1){
 125                 n = io_getevents(ctx, 1, 1, &event, NULL);
 126                 if(n < 0){
 127                         if(errno == EINTR)
 128                                 continue;
 129                         printk("aio_thread - io_getevents failed, "
 130                                "errno = %d\n", errno);
 131                 }
 132                 else {
 133                         reply = ((struct aio_thread_reply)
 134                                 { .data = (void *) (long) event.data,
 135                                                 .err    = event.res });
 136                         reply_fd = ((struct aio_context *) reply.data)->reply_fd;
 137                         err = os_write_file(reply_fd, &reply, sizeof(reply));
 138                         if(err != sizeof(reply))
 139                                 printk("aio_thread - write failed, fd = %d, "
 140                                        "err = %d\n", aio_req_fd_r, -err);
 141                 }
 142         }
 143         return 0;
 144 }
 145
 146 #endif
 147
 148 static int do_not_aio(struct aio_thread_req *req)
 149 {
 150         char c;
 151         int err;
 152
 153         switch(req->type){
 154         case AIO_READ:
 155                 err = os_seek_file(req->io_fd, req->offset);
 156                 if(err)
 157                         goto out;
 158
 159                 err = os_read_file(req->io_fd, req->buf, req->len);
 160                 break;
 161         case AIO_WRITE:
 162                 err = os_seek_file(req->io_fd, req->offset);
 163                 if(err)
 164                         goto out;
 165
 166                 err = os_write_file(req->io_fd, req->buf, req->len);
 167                 break;
 168         case AIO_MMAP:
 169                 err = os_seek_file(req->io_fd, req->offset);
 170                 if(err)
 171                         goto out;
 172
 173                 err = os_read_file(req->io_fd, &c, sizeof(c));
 174                 break;
 175         default:
 176                 printk("do_not_aio - bad request type : %d\n", req->type);
 177                 err = -EINVAL;
 178                 break;
 179         }
 180
 181 out:
 182         return err;
 183 }
 184
 185 static int not_aio_thread(void *arg)
 186 {
 187         struct aio_thread_req req;
 188         struct aio_thread_reply reply;
 189         int err;
 190
 191         signal(SIGWINCH, SIG_IGN);
 192         while(1){
 193                 err = os_read_file(aio_req_fd_r, &req, sizeof(req));
 194                 if(err != sizeof(req)){
 195                         if(err < 0)
 196                                 printk("not_aio_thread - read failed, "
 197                                        "fd = %d, err = %d\n", aio_req_fd_r,
 198                                        -err);
 199                         else {
 200                                 printk("not_aio_thread - short read, fd = %d, "
 201                                        "length = %d\n", aio_req_fd_r, err);
 202                         }
 203                         continue;
 204                 }
 205                 err = do_not_aio(&req);
 206                 reply = ((struct aio_thread_reply) { .data      = req.aio,
 207                                          .err   = err });
 208                 err = os_write_file(req.aio->reply_fd, &reply, sizeof(reply));
 209                 if(err != sizeof(reply))
 210                         printk("not_aio_thread - write failed, fd = %d, "
 211                                "err = %d\n", aio_req_fd_r, -err);
 212         }
 213
 214         return 0;
 215 }
 216
 217 static int aio_pid = -1;
 218
 219 static int init_aio_24(void)
 220 {
 221         unsigned long stack;
 222         int fds[2], err;
 223
 224         err = os_pipe(fds, 1, 1);
 225         if(err)
 226                 goto out;
 227
 228         aio_req_fd_w = fds[0];
 229         aio_req_fd_r = fds[1];
 230         err = run_helper_thread(not_aio_thread, NULL,
 231                                 CLONE_FILES | CLONE_VM | SIGCHLD, &stack, 0);
 232         if(err < 0)
 233                 goto out_close_pipe;
 234
 235         aio_pid = err;
 236         goto out;
 237
 238 out_close_pipe:
 239         os_close_file(fds[0]);
 240         os_close_file(fds[1]);
 241         aio_req_fd_w = -1;
 242         aio_req_fd_r = -1;
 243 out:
 244 #ifndef HAVE_AIO_ABI
 245         printk("/usr/include/linux/aio_abi.h not present during build\n");
 246 #endif
 247         printk("2.6 host AIO support not used - falling back to I/O "
 248                "thread\n");
 249         return 0;
 250 }
 251
 252 #ifdef HAVE_AIO_ABI
 253 #define DEFAULT_24_AIO 0
 254 static int init_aio_26(void)
 255 {
 256         unsigned long stack;
 257         int err;
 258
 259         if(io_setup(256, &ctx)){
 260                 err = -errno;
 261                 printk("aio_thread failed to initialize context, err = %d\n",
 262                        errno);
 263                 return err;
 264         }
 265
 266         err = run_helper_thread(aio_thread, NULL,
 267                                 CLONE_FILES | CLONE_VM | SIGCHLD, &stack, 0);
 268         if(err < 0)
 269                 return err;
 270
 271         aio_pid = err;
 272
 273         printk("Using 2.6 host AIO\n");
 274         return 0;
 275 }
 276
 277 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
 278                          unsigned long long offset, struct aio_context *aio)
 279 {
 280         struct aio_thread_reply reply;
 281         int err;
 282
 283         err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
 284         if(err){
 285                 reply = ((struct aio_thread_reply) { .data = aio,
 286                                          .err  = err });
 287                 err = os_write_file(aio->reply_fd, &reply, sizeof(reply));
 288                 if(err != sizeof(reply))
 289                         printk("submit_aio_26 - write failed, "
 290                                "fd = %d, err = %d\n", aio->reply_fd, -err);
 291                 else err = 0;
 292         }
 293
 294         return err;
 295 }
 296
 297 #else
 298 #define DEFAULT_24_AIO 1
 299 static int init_aio_26(void)
 300 {
 301         return -ENOSYS;
 302 }
 303
 304 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
 305                          unsigned long long offset, struct aio_context *aio)
 306 {
 307         return -ENOSYS;
 308 }
 309 #endif
 310
 311 static int aio_24 = DEFAULT_24_AIO;
 312
 313 static int __init set_aio_24(char *name, int *add)
 314 {
 315         aio_24 = 1;
 316         return 0;
 317 }
 318
 319 __uml_setup("aio=2.4", set_aio_24,
 320 "aio=2.4\n"
 321 "    This is used to force UML to use 2.4-style AIO even when 2.6 AIO is\n"
 322 "    available.  2.4 AIO is a single thread that handles one request at a\n"
 323 "    time, synchronously.  2.6 AIO is a thread which uses the 2.6 AIO \n"
 324 "    interface to handle an arbitrary number of pending requests.  2.6 AIO \n"
 325 "    is not available in tt mode, on 2.4 hosts, or when UML is built with\n"
 326 "    /usr/include/linux/aio_abi.h not available.  Many distributions don't\n"
 327 "    include aio_abi.h, so you will need to copy it from a kernel tree to\n"
 328 "    your /usr/include/linux in order to build an AIO-capable UML\n\n"
 329 );
 330
 331 static int init_aio(void)
 332 {
 333         int err;
 334
 335         CHOOSE_MODE(({ if(!aio_24){
 336                             printk("Disabling 2.6 AIO in tt mode\n");
 337                             aio_24 = 1;
 338                     } }), (void) 0);
 339
 340         if(!aio_24){
 341                 err = init_aio_26();
 342                 if(err && (errno == ENOSYS)){
 343                         printk("2.6 AIO not supported on the host - "
 344                                "reverting to 2.4 AIO\n");
 345                         aio_24 = 1;
 346                 }
 347                 else return err;
 348         }
 349
 350         if(aio_24)
 351                 return init_aio_24();
 352
 353         return 0;
 354 }
 355
 356 /* The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
 357  * needs to be called when the kernel is running because it calls run_helper,
 358  * which needs get_free_page.  exit_aio is a __uml_exitcall because the generic
 359  * kernel does not run __exitcalls on shutdown, and can't because many of them
 360  * break when called outside of module unloading.
 361  */
 362 __initcall(init_aio);
 363
 364 static void exit_aio(void)
 365 {
 366         if(aio_pid != -1)
 367                 os_kill_process(aio_pid, 1);
 368 }
 369
 370 __uml_exitcall(exit_aio);
 371
 372 static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
 373                          unsigned long long offset, struct aio_context *aio)
 374 {
 375         struct aio_thread_req req = { .type             = type,
 376                                       .io_fd            = io_fd,
 377                                       .offset           = offset,
 378                                       .buf              = buf,
 379                                       .len              = len,
 380                                       .aio              = aio,
 381         };
 382         int err;
 383
 384         err = os_write_file(aio_req_fd_w, &req, sizeof(req));
 385         if(err == sizeof(req))
 386                 err = 0;
 387
 388         return err;
 389 }
 390
 391 int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
 392                unsigned long long offset, int reply_fd,
 393                struct aio_context *aio)
 394 {
 395         aio->reply_fd = reply_fd;
 396         if(aio_24)
 397                 return submit_aio_24(type, io_fd, buf, len, offset, aio);
 398         else {
 399                 return submit_aio_26(type, io_fd, buf, len, offset, aio);
 400         }
 401 }