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