arch/um/os-Linux/aio.c

   1 /*
   2  * Copyright (C) 2004 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
   3  * Licensed under the GPL
   4  */
   5
   6 #include <unistd.h>
   7 #include <sched.h>
   8 #include <signal.h>
   9 #include <errno.h>
  10 #include <sys/time.h>
  11 #include <asm/unistd.h>
  12 #include "aio.h"
  13 #include "init.h"
  14 #include "kern_constants.h"
  15 #include "os.h"
  16 #include "user.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 #if defined(HAVE_AIO_ABI)
  28 #include <linux/aio_abi.h>
  29
  30 /*
  31  * If we have the headers, we are going to build with AIO enabled.
  32  * If we don't have aio in libc, we define the necessary stubs here.
  33  */
  34
  35 #if !defined(HAVE_AIO_LIBC)
  36
  37 static long io_setup(int n, aio_context_t *ctxp)
  38 {
  39         return syscall(__NR_io_setup, n, ctxp);
  40 }
  41
  42 static long io_submit(aio_context_t ctx, long nr, struct iocb **iocbpp)
  43 {
  44         return syscall(__NR_io_submit, ctx, nr, iocbpp);
  45 }
  46
  47 static long io_getevents(aio_context_t ctx_id, long min_nr, long nr,
  48                          struct io_event *events, struct timespec *timeout)
  49 {
  50         return syscall(__NR_io_getevents, ctx_id, min_nr, nr, events, timeout);
  51 }
  52
  53 #endif
  54
  55 /*
  56  * The AIO_MMAP cases force the mmapped page into memory here
  57  * rather than in whatever place first touches the data.  I used
  58  * to do this by touching the page, but that's delicate because
  59  * gcc is prone to optimizing that away.  So, what's done here
  60  * is we read from the descriptor from which the page was
  61  * mapped.  The caller is required to pass an offset which is
  62  * inside the page that was mapped.  Thus, when the read
  63  * returns, we know that the page is in the page cache, and
  64  * that it now backs the mmapped area.
  65  */
  66
  67 static int do_aio(aio_context_t ctx, enum aio_type type, int fd, char *buf,
  68                   int len, unsigned long long offset, struct aio_context *aio)
  69 {
  70         struct iocb *iocbp = & ((struct iocb) {
  71                                     .aio_data       = (unsigned long) aio,
  72                                     .aio_fildes     = fd,
  73                                     .aio_buf        = (unsigned long) buf,
  74                                     .aio_nbytes     = len,
  75                                     .aio_offset     = offset
  76                              });
  77         char c;
  78
  79         switch (type) {
  80         case AIO_READ:
  81                 iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
  82                 break;
  83         case AIO_WRITE:
  84                 iocbp->aio_lio_opcode = IOCB_CMD_PWRITE;
  85                 break;
  86         case AIO_MMAP:
  87                 iocbp->aio_lio_opcode = IOCB_CMD_PREAD;
  88                 iocbp->aio_buf = (unsigned long) &c;
  89                 iocbp->aio_nbytes = sizeof(c);
  90                 break;
  91         default:
  92                 printk(UM_KERN_ERR "Bogus op in do_aio - %d\n", type);
  93                 return -EINVAL;
  94         }
  95
  96         return (io_submit(ctx, 1, &iocbp) > 0) ? 0 : -errno;
  97 }
  98
  99 /* Initialized in an initcall and unchanged thereafter */
 100 static aio_context_t ctx = 0;
 101
 102 static int aio_thread(void *arg)
 103 {
 104         struct aio_thread_reply reply;
 105         struct io_event event;
 106         int err, n, reply_fd;
 107
 108         signal(SIGWINCH, SIG_IGN);
 109
 110         while (1) {
 111                 n = io_getevents(ctx, 1, 1, &event, NULL);
 112                 if (n < 0) {
 113                         if (errno == EINTR)
 114                                 continue;
 115                         printk(UM_KERN_ERR "aio_thread - io_getevents failed, "
 116                                "errno = %d\n", errno);
 117                 }
 118                 else {
 119                         reply = ((struct aio_thread_reply)
 120                                 { .data = (void *) (long) event.data,
 121                                                 .err    = event.res });
 122                         reply_fd = ((struct aio_context *) reply.data)->reply_fd;
 123                         err = write(reply_fd, &reply, sizeof(reply));
 124                         if (err != sizeof(reply))
 125                                 printk(UM_KERN_ERR "aio_thread - write failed, "
 126                                        "fd = %d, err = %d\n", reply_fd, errno);
 127                 }
 128         }
 129         return 0;
 130 }
 131
 132 #endif
 133
 134 static int do_not_aio(struct aio_thread_req *req)
 135 {
 136         char c;
 137         unsigned long long actual;
 138         int n;
 139
 140         actual = lseek64(req->io_fd, req->offset, SEEK_SET);
 141         if (actual != req->offset)
 142                 return -errno;
 143
 144         switch(req->type) {
 145         case AIO_READ:
 146                 n = read(req->io_fd, req->buf, req->len);
 147                 break;
 148         case AIO_WRITE:
 149                 n = write(req->io_fd, req->buf, req->len);
 150                 break;
 151         case AIO_MMAP:
 152                 n = read(req->io_fd, &c, sizeof(c));
 153                 break;
 154         default:
 155                 printk(UM_KERN_ERR "do_not_aio - bad request type : %d\n",
 156                        req->type);
 157                 return -EINVAL;
 158         }
 159
 160         if (n < 0)
 161                 return -errno;
 162         return 0;
 163 }
 164
 165 /* These are initialized in initcalls and not changed */
 166 static int aio_req_fd_r = -1;
 167 static int aio_req_fd_w = -1;
 168 static int aio_pid = -1;
 169 static unsigned long aio_stack;
 170
 171 static int not_aio_thread(void *arg)
 172 {
 173         struct aio_thread_req req;
 174         struct aio_thread_reply reply;
 175         int err;
 176
 177         signal(SIGWINCH, SIG_IGN);
 178         while (1) {
 179                 err = read(aio_req_fd_r, &req, sizeof(req));
 180                 if (err != sizeof(req)) {
 181                         if (err < 0)
 182                                 printk(UM_KERN_ERR "not_aio_thread - "
 183                                        "read failed, fd = %d, err = %d\n",
 184                                        aio_req_fd_r,
 185                                        errno);
 186                         else {
 187                                 printk(UM_KERN_ERR "not_aio_thread - short "
 188                                        "read, fd = %d, length = %d\n",
 189                                        aio_req_fd_r, err);
 190                         }
 191                         continue;
 192                 }
 193                 err = do_not_aio(&req);
 194                 reply = ((struct aio_thread_reply) { .data      = req.aio,
 195                                                      .err       = err });
 196                 err = write(req.aio->reply_fd, &reply, sizeof(reply));
 197                 if (err != sizeof(reply))
 198                         printk(UM_KERN_ERR "not_aio_thread - write failed, "
 199                                "fd = %d, err = %d\n", req.aio->reply_fd, errno);
 200         }
 201
 202         return 0;
 203 }
 204
 205 static int init_aio_24(void)
 206 {
 207         int fds[2], err;
 208
 209         err = os_pipe(fds, 1, 1);
 210         if (err)
 211                 goto out;
 212
 213         aio_req_fd_w = fds[0];
 214         aio_req_fd_r = fds[1];
 215
 216         err = os_set_fd_block(aio_req_fd_w, 0);
 217         if (err)
 218                 goto out_close_pipe;
 219
 220         err = run_helper_thread(not_aio_thread, NULL,
 221                                 CLONE_FILES | CLONE_VM, &aio_stack);
 222         if (err < 0)
 223                 goto out_close_pipe;
 224
 225         aio_pid = err;
 226         goto out;
 227
 228 out_close_pipe:
 229         close(fds[0]);
 230         close(fds[1]);
 231         aio_req_fd_w = -1;
 232         aio_req_fd_r = -1;
 233 out:
 234 #ifndef HAVE_AIO_ABI
 235         printk(UM_KERN_INFO "/usr/include/linux/aio_abi.h not present during "
 236                "build\n");
 237 #endif
 238         printk(UM_KERN_INFO "2.6 host AIO support not used - falling back to "
 239                "I/O thread\n");
 240         return 0;
 241 }
 242
 243 #ifdef HAVE_AIO_ABI
 244 #define DEFAULT_24_AIO 0
 245 static int init_aio_26(void)
 246 {
 247         int err;
 248
 249         if (io_setup(256, &ctx)) {
 250                 err = -errno;
 251                 printk(UM_KERN_ERR "aio_thread failed to initialize context, "
 252                        "err = %d\n", errno);
 253                 return err;
 254         }
 255
 256         err = run_helper_thread(aio_thread, NULL,
 257                                 CLONE_FILES | CLONE_VM, &aio_stack);
 258         if (err < 0)
 259                 return err;
 260
 261         aio_pid = err;
 262
 263         printk(UM_KERN_INFO "Using 2.6 host AIO\n");
 264         return 0;
 265 }
 266
 267 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
 268                          unsigned long long offset, struct aio_context *aio)
 269 {
 270         struct aio_thread_reply reply;
 271         int err;
 272
 273         err = do_aio(ctx, type, io_fd, buf, len, offset, aio);
 274         if (err) {
 275                 reply = ((struct aio_thread_reply) { .data = aio,
 276                                          .err  = err });
 277                 err = write(aio->reply_fd, &reply, sizeof(reply));
 278                 if (err != sizeof(reply)) {
 279                         err = -errno;
 280                         printk(UM_KERN_ERR "submit_aio_26 - write failed, "
 281                                "fd = %d, err = %d\n", aio->reply_fd, -err);
 282                 }
 283                 else err = 0;
 284         }
 285
 286         return err;
 287 }
 288
 289 #else
 290 #define DEFAULT_24_AIO 1
 291 static int init_aio_26(void)
 292 {
 293         return -ENOSYS;
 294 }
 295
 296 static int submit_aio_26(enum aio_type type, int io_fd, char *buf, int len,
 297                          unsigned long long offset, struct aio_context *aio)
 298 {
 299         return -ENOSYS;
 300 }
 301 #endif
 302
 303 /* Initialized in an initcall and unchanged thereafter */
 304 static int aio_24 = DEFAULT_24_AIO;
 305
 306 static int __init set_aio_24(char *name, int *add)
 307 {
 308         aio_24 = 1;
 309         return 0;
 310 }
 311
 312 __uml_setup("aio=2.4", set_aio_24,
 313 "aio=2.4\n"
 314 "    This is used to force UML to use 2.4-style AIO even when 2.6 AIO is\n"
 315 "    available.  2.4 AIO is a single thread that handles one request at a\n"
 316 "    time, synchronously.  2.6 AIO is a thread which uses the 2.6 AIO \n"
 317 "    interface to handle an arbitrary number of pending requests.  2.6 AIO \n"
 318 "    is not available in tt mode, on 2.4 hosts, or when UML is built with\n"
 319 "    /usr/include/linux/aio_abi.h not available.  Many distributions don't\n"
 320 "    include aio_abi.h, so you will need to copy it from a kernel tree to\n"
 321 "    your /usr/include/linux in order to build an AIO-capable UML\n\n"
 322 );
 323
 324 static int init_aio(void)
 325 {
 326         int err;
 327
 328         if (!aio_24) {
 329                 err = init_aio_26();
 330                 if (err && (errno == ENOSYS)) {
 331                         printk(UM_KERN_INFO "2.6 AIO not supported on the "
 332                                "host - reverting to 2.4 AIO\n");
 333                         aio_24 = 1;
 334                 }
 335                 else return err;
 336         }
 337
 338         if (aio_24)
 339                 return init_aio_24();
 340
 341         return 0;
 342 }
 343
 344 /*
 345  * The reason for the __initcall/__uml_exitcall asymmetry is that init_aio
 346  * needs to be called when the kernel is running because it calls run_helper,
 347  * which needs get_free_page.  exit_aio is a __uml_exitcall because the generic
 348  * kernel does not run __exitcalls on shutdown, and can't because many of them
 349  * break when called outside of module unloading.
 350  */
 351 __initcall(init_aio);
 352
 353 static void exit_aio(void)
 354 {
 355         if (aio_pid != -1) {
 356                 os_kill_process(aio_pid, 1);
 357                 free_stack(aio_stack, 0);
 358         }
 359 }
 360
 361 __uml_exitcall(exit_aio);
 362
 363 static int submit_aio_24(enum aio_type type, int io_fd, char *buf, int len,
 364                          unsigned long long offset, struct aio_context *aio)
 365 {
 366         struct aio_thread_req req = { .type             = type,
 367                                       .io_fd            = io_fd,
 368                                       .offset           = offset,
 369                                       .buf              = buf,
 370                                       .len              = len,
 371                                       .aio              = aio,
 372         };
 373         int err;
 374
 375         err = write(aio_req_fd_w, &req, sizeof(req));
 376         if (err == sizeof(req))
 377                 err = 0;
 378         else err = -errno;
 379
 380         return err;
 381 }
 382
 383 int submit_aio(enum aio_type type, int io_fd, char *buf, int len,
 384                unsigned long long offset, int reply_fd,
 385                struct aio_context *aio)
 386 {
 387         aio->reply_fd = reply_fd;
 388         if (aio_24)
 389                 return submit_aio_24(type, io_fd, buf, len, offset, aio);
 390         else
 391                 return submit_aio_26(type, io_fd, buf, len, offset, aio);
 392 }