mmap engine: allow large files on 32-bit archs

Map chunks of 2GB at the time in total, remapping as we go
along.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

diff --git a/engines/mmap.c b/engines/mmap.c
index 5b8d800c4190177ff33db12681a87c2659105761..05a4d5163ce5dbd45a3ba351419417e642b65be1 100644 (file)
--- a/engines/mmap.c
+++ b/engines/mmap.c
@@ -13,59 +13,19 @@
 
 #include "../fio.h"
 
-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)
-               memcpy(io_u->xfer_buf, f->mmap + real_off, io_u->xfer_buflen);
-       else if (io_u->ddir == DDIR_WRITE)
-               memcpy(f->mmap + real_off, io_u->xfer_buf, io_u->xfer_buflen);
-       else if (io_u->ddir == DDIR_SYNC) {
-               size_t len = (f->io_size + page_size - 1) & ~page_mask;
-
-               if (msync(f->mmap, len, MS_SYNC)) {
-                       io_u->error = errno;
-                       td_verror(td, io_u->error, "msync");
-               }
-       }
-
-       /*
-        * not really direct, but should drop the pages from the cache
-        */
-       if (td->o.odirect && io_u->ddir != DDIR_SYNC) {
-               size_t len = (io_u->xfer_buflen + page_size - 1) & ~page_mask;
-               unsigned long long off = real_off & ~page_mask;
-
-               if (msync(f->mmap + off, len, MS_SYNC) < 0) {
-                       io_u->error = errno;
-                       td_verror(td, io_u->error, "msync");
-               }
-               if (madvise(f->mmap + off, len,  MADV_DONTNEED) < 0) {
-                       io_u->error = errno;
-                       td_verror(td, io_u->error, "madvise");
-               }
-       }
+/*
+ * Limits us to 2GB of mapped files in total
+ */
+#define MMAP_TOTAL_SZ  (2 * 1024 * 1024 * 1024UL)
 
-       return FIO_Q_COMPLETED;
-}
+static unsigned long mmap_map_size;
+static unsigned long mmap_map_mask;
 
-static int fio_mmapio_open(struct thread_data *td, struct fio_file *f)
+static int fio_mmap_file(struct thread_data *td, struct fio_file *f,
+                        size_t length, off_t off)
 {
-       int ret, flags;
-
-       ret = generic_open_file(td, f);
-       if (ret)
-               return ret;
-
-       /*
-        * for size checkup, don't mmap anything.
-        */
-       if (!f->io_size)
-               return 0;
+       int flags = 0;
+       int ret = 0;
 
        if (td_rw(td))
                flags = PROT_READ | PROT_WRITE;
@@ -77,11 +37,11 @@ static int fio_mmapio_open(struct thread_data *td, struct fio_file *f)
        } else
                flags = PROT_READ;
 
-       f->mmap = mmap(NULL, f->io_size, flags, MAP_SHARED, f->fd, f->file_offset);
-       if (f->mmap == MAP_FAILED) {
+       f->mmap_ptr = mmap(NULL, length, flags, MAP_SHARED, f->fd, off);
+       if (f->mmap_ptr == MAP_FAILED) {
                int err = errno;
 
-               f->mmap = NULL;
+               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");
@@ -92,49 +52,121 @@ static int fio_mmapio_open(struct thread_data *td, struct fio_file *f)
                goto err;
 
        if (!td_random(td)) {
-               if (madvise(f->mmap, f->io_size, MADV_SEQUENTIAL) < 0) {
+               if (madvise(f->mmap_ptr, length, MADV_SEQUENTIAL) < 0) {
                        td_verror(td, errno, "madvise");
                        goto err;
                }
        } else {
-               if (madvise(f->mmap, f->io_size, MADV_RANDOM) < 0) {
+               if (madvise(f->mmap_ptr, length, MADV_RANDOM) < 0) {
                        td_verror(td, errno, "madvise");
                        goto err;
                }
        }
 
-       return 0;
-
 err:
-       td->io_ops->close_file(td, f);
-       return 1;
+       return ret;
 }
 
-static int fio_mmapio_close(struct thread_data fio_unused *td,
-                           struct fio_file *f)
+static int fio_mmapio_prep(struct thread_data *td, struct io_u *io_u)
 {
-       int ret = 0, ret2;
+       struct fio_file *f = io_u->file;
+       int ret = 0;
 
-       if (f->mmap) {
-               if (munmap(f->mmap, f->io_size) < 0)
-                       ret = errno;
+       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;
 
-               f->mmap = NULL;
+       if (f->mmap_ptr) {
+               if (munmap(f->mmap_ptr, f->mmap_sz) < 0) {
+                       ret = errno;
+                       goto err;
+               }
+               f->mmap_ptr = NULL;
        }
 
-       ret2 = generic_close_file(td, f);
-       if (!ret && ret2)
-               ret = ret2;
+       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;
+
+       fio_ro_check(td, io_u);
+
+       if (io_u->ddir == DDIR_READ)
+               memcpy(io_u->xfer_buf, io_u->mmap_data, io_u->xfer_buflen);
+       else if (io_u->ddir == DDIR_WRITE)
+               memcpy(io_u->mmap_data, io_u->xfer_buf, io_u->xfer_buflen);
+       else if (io_u->ddir == DDIR_SYNC) {
+               if (msync(f->mmap_ptr, f->mmap_sz, MS_SYNC)) {
+                       io_u->error = errno;
+                       td_verror(td, io_u->error, "msync");
+               }
+       }
+
+       /*
+        * not really direct, but should drop the pages from the cache
+        */
+       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;
+                       td_verror(td, io_u->error, "msync");
+               }
+               if (madvise(io_u->mmap_data, io_u->xfer_buflen,  MADV_DONTNEED) < 0) {
+                       io_u->error = errno;
+                       td_verror(td, io_u->error, "madvise");
+               }
+       }
+
+       return FIO_Q_COMPLETED;
+}
+
+static int fio_mmapio_init(struct thread_data *td)
+{
+       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,
+       .init           = fio_mmapio_init,
+       .prep           = fio_mmapio_prep,
        .queue          = fio_mmapio_queue,
-       .open_file      = fio_mmapio_open,
-       .close_file     = fio_mmapio_close,
+       .open_file      = generic_open_file,
+       .close_file     = generic_close_file,
        .get_file_size  = generic_get_file_size,
        .flags          = FIO_SYNCIO | FIO_NOEXTEND,
 };

diff --git a/filesetup.c b/filesetup.c
index 1372177efc67a6499c4a3c48483ff6b92981d981..a3f96a15ef0e2a965406c3c050d57cfa47bbce38 100644 (file)
--- a/filesetup.c
+++ b/filesetup.c
@@ -269,8 +269,8 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
        /*
         * FIXME: add blockdev flushing too
         */
-       if (f->mmap)
-               ret = madvise(f->mmap, len, MADV_DONTNEED);
+       if (f->mmap_ptr)
+               ret = madvise(f->mmap_ptr, f->mmap_sz, MADV_DONTNEED);
        else if (f->filetype == FIO_TYPE_FILE) {
                ret = fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
        } else if (f->filetype == FIO_TYPE_BD) {

diff --git a/fio.h b/fio.h
index c9fa6a9fa8672b5af02181cced08bd17189882f0..2e5431e99a6d21d953617154e2df9538e938dcab 100644 (file)
--- a/fio.h
+++ b/fio.h
@@ -135,6 +135,7 @@ struct io_u {
 #ifdef FIO_HAVE_SOLARISAIO
                aio_result_t resultp;
 #endif
+               void *mmap_data;
        };
        struct timeval start_time;
        struct timeval issue_time;
@@ -313,9 +314,12 @@ struct fio_file {
         * filename and possible memory mapping
         */
        char *file_name;
-       void *mmap;
        unsigned int major, minor;
 
+       void *mmap_ptr;
+       size_t mmap_sz;
+       off_t mmap_off;
+
        /*
         * size of the file, offset into file, and io size from that offset
         */