mmap engine: allow large files on 32-bit archs

[fio.git] / engines / mmap.c

diff --git a/engines/mmap.c b/engines/mmap.c
index 08bbd991463d4e8546ea40a31d14b9213ffd0d11..05a4d5163ce5dbd45a3ba351419417e642b65be1 100644 (file)
--- a/engines/mmap.c
+++ b/engines/mmap.c
@@ -1,75 +1,174 @@
 /*
 /*

- * regular read/write sync io engine
+ * mmap engine
+ *
+ * IO engine that reads/writes from files by doing memcpy to/from
+ * a memory mapped region of the file.

  *
  */
 #include <stdio.h>
 #include <stdlib.h>
 #include <unistd.h>
 #include <errno.h>
  *
  */
 #include <stdio.h>
 #include <stdlib.h>
 #include <unistd.h>
 #include <errno.h>

-#include <assert.h>

 #include <sys/mman.h>
 
 #include "../fio.h"
 #include <sys/mman.h>
 
 #include "../fio.h"

-#include "../os.h"
+
+/*
+ * Limits us to 2GB of mapped files in total
+ */
+#define MMAP_TOTAL_SZ  (2 * 1024 * 1024 * 1024UL)
+
+static unsigned long mmap_map_size;
+static unsigned long mmap_map_mask;
+
+static int fio_mmap_file(struct thread_data *td, struct fio_file *f,
+                        size_t length, off_t off)
+{
+       int flags = 0;
+       int ret = 0;
+
+       if (td_rw(td))
+               flags = PROT_READ | PROT_WRITE;
+       else if (td_write(td)) {
+               flags = PROT_WRITE;
+
+               if (td->o.verify != VERIFY_NONE)
+                       flags |= PROT_READ;
+       } else
+               flags = PROT_READ;
+
+       f->mmap_ptr = mmap(NULL, length, flags, MAP_SHARED, f->fd, off);
+       if (f->mmap_ptr == MAP_FAILED) {
+               int err = errno;
+
+               f->mmap_ptr = NULL;
+               td_verror(td, err, "mmap");
+               if (err == EINVAL && f->io_size > 2*1024*1024*1024UL)
+                       log_err("fio: mmap size likely too large\n");
+               goto err;
+       }
+
+       if (file_invalidate_cache(td, f))
+               goto err;
+
+       if (!td_random(td)) {
+               if (madvise(f->mmap_ptr, length, MADV_SEQUENTIAL) < 0) {
+                       td_verror(td, errno, "madvise");
+                       goto err;
+               }
+       } else {
+               if (madvise(f->mmap_ptr, length, MADV_RANDOM) < 0) {
+                       td_verror(td, errno, "madvise");
+                       goto err;
+               }
+       }
+
+err:
+       return ret;
+}
+
+static int fio_mmapio_prep(struct thread_data *td, struct io_u *io_u)
+{
+       struct fio_file *f = io_u->file;
+       int ret = 0;
+
+       if (io_u->buflen > mmap_map_size) {
+               log_err("fio: bs too big for mmap engine\n");
+               ret = EIO;
+               goto err;
+       }
+
+       if (io_u->offset >= f->mmap_off &&
+           io_u->offset + io_u->buflen < f->mmap_off + f->mmap_sz)
+               goto done;
+
+       if (f->mmap_ptr) {
+               if (munmap(f->mmap_ptr, f->mmap_sz) < 0) {
+                       ret = errno;
+                       goto err;
+               }
+               f->mmap_ptr = NULL;
+       }
+
+       f->mmap_sz = mmap_map_size;
+       if (f->mmap_sz  > f->io_size)
+               f->mmap_sz = f->io_size;
+
+       f->mmap_off = io_u->offset & ~mmap_map_mask;
+       if (io_u->offset + io_u->buflen >= f->mmap_off + f->mmap_sz)
+               f->mmap_off -= io_u->buflen;
+
+       ret = fio_mmap_file(td, f, f->mmap_sz, f->mmap_off);
+done:
+       if (!ret)
+               io_u->mmap_data = f->mmap_ptr + io_u->offset - f->mmap_off -
+                                       f->file_offset;
+err:
+       return ret;
+}

 
 static int fio_mmapio_queue(struct thread_data *td, struct io_u *io_u)
 {
        struct fio_file *f = io_u->file;
 
 static int fio_mmapio_queue(struct thread_data *td, struct io_u *io_u)
 {
        struct fio_file *f = io_u->file;

-       unsigned long long real_off = io_u->offset - f->file_offset;
+
+       fio_ro_check(td, io_u);

 
        if (io_u->ddir == DDIR_READ)
 
        if (io_u->ddir == DDIR_READ)

-               memcpy(io_u->xfer_buf, f->mmap + real_off, io_u->xfer_buflen);
+               memcpy(io_u->xfer_buf, io_u->mmap_data, io_u->xfer_buflen);

        else if (io_u->ddir == DDIR_WRITE)
        else if (io_u->ddir == DDIR_WRITE)

-               memcpy(f->mmap + real_off, io_u->xfer_buf, io_u->xfer_buflen);
+               memcpy(io_u->mmap_data, io_u->xfer_buf, io_u->xfer_buflen);

        else if (io_u->ddir == DDIR_SYNC) {
        else if (io_u->ddir == DDIR_SYNC) {

-               if (msync(f->mmap, f->file_size, MS_SYNC))
+               if (msync(f->mmap_ptr, f->mmap_sz, MS_SYNC)) {

                        io_u->error = errno;
                        io_u->error = errno;

+                       td_verror(td, io_u->error, "msync");
+               }

        }
 
        /*
         * not really direct, but should drop the pages from the cache
         */
        }
 
        /*
         * not really direct, but should drop the pages from the cache
         */

-       if (td->odirect && io_u->ddir != DDIR_SYNC) {
-               if (msync(f->mmap + real_off, io_u->xfer_buflen, MS_SYNC) < 0)
+       if (td->o.odirect && io_u->ddir != DDIR_SYNC) {
+               if (msync(io_u->mmap_data, io_u->xfer_buflen, MS_SYNC) < 0) {

                        io_u->error = errno;
                        io_u->error = errno;

-               if (madvise(f->mmap + real_off, io_u->xfer_buflen,  MADV_DONTNEED) < 0)
+                       td_verror(td, io_u->error, "msync");
+               }
+               if (madvise(io_u->mmap_data, io_u->xfer_buflen,  MADV_DONTNEED) < 0) {

                        io_u->error = errno;
                        io_u->error = errno;

+                       td_verror(td, io_u->error, "madvise");
+               }

        }
 
        }
 

-       if (io_u->error)
-               td_verror(td, io_u->error);
-

        return FIO_Q_COMPLETED;
 }
 
 static int fio_mmapio_init(struct thread_data *td)
 {
        return FIO_Q_COMPLETED;
 }
 
 static int fio_mmapio_init(struct thread_data *td)
 {

-       struct fio_file *f;
-       int i;
-
-       if (td->ddir == DDIR_READ && !td_rw(td))
-               return 0;
-
-       /*
-        * We need to truncate the files to the right size, if
-        * we are writing to it.
-        */
-       for_each_file(td, f, i) {
-               if (ftruncate(f->fd, f->file_size) < 0) {
-                       td_verror(td, errno);
-                       return 1;
-               }
-       }
+       unsigned long shift, mask;

 
 

+       mmap_map_size = MMAP_TOTAL_SZ / td->o.nr_files;
+       mask = mmap_map_size;
+       shift = 0;
+       do {
+               mask >>= 1;
+               if (!mask)
+                       break;
+               shift++;
+       } while (1);
+               
+       mmap_map_mask = 1UL << shift;

        return 0;
 }
 
 static struct ioengine_ops ioengine = {
        .name           = "mmap",
        .version        = FIO_IOOPS_VERSION,
        return 0;
 }
 
 static struct ioengine_ops ioengine = {
        .name           = "mmap",
        .version        = FIO_IOOPS_VERSION,

-       .queue          = fio_mmapio_queue,

        .init           = fio_mmapio_init,
        .init           = fio_mmapio_init,

-       .flags          = FIO_SYNCIO | FIO_MMAPIO,
+       .prep           = fio_mmapio_prep,
+       .queue          = fio_mmapio_queue,
+       .open_file      = generic_open_file,
+       .close_file     = generic_close_file,
+       .get_file_size  = generic_get_file_size,
+       .flags          = FIO_SYNCIO | FIO_NOEXTEND,

 };
 
 static void fio_init fio_mmapio_register(void)
 };
 
 static void fio_init fio_mmapio_register(void)