#include <stdlib.h>
#include <unistd.h>
#include <errno.h>
-#include <assert.h>
#include <sys/mman.h>
#include "../fio.h"
-#include "../os.h"
+#include "../verify.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;
-
- 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->file_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->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 1GB of mapped files in total
+ */
+#define MMAP_TOTAL_SZ (1 * 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;
+ int flags = 0;
if (td_rw(td))
flags = PROT_READ | PROT_WRITE;
else if (td_write(td)) {
flags = PROT_WRITE;
- if (td->verify != VERIFY_NONE)
+ if (td->o.verify != VERIFY_NONE)
flags |= PROT_READ;
} else
flags = PROT_READ;
- f->mmap = mmap(NULL, f->file_size, flags, MAP_SHARED, f->fd, f->file_offset);
- if (f->mmap == MAP_FAILED) {
- f->mmap = NULL;
+ f->mmap_ptr = mmap(NULL, length, flags, MAP_SHARED, f->fd, off);
+ if (f->mmap_ptr == MAP_FAILED) {
+ f->mmap_ptr = NULL;
td_verror(td, errno, "mmap");
goto err;
}
- if (file_invalidate_cache(td, f))
- goto err;
-
if (!td_random(td)) {
- if (madvise(f->mmap, f->file_size, MADV_SEQUENTIAL) < 0) {
+ if (posix_madvise(f->mmap_ptr, length, POSIX_MADV_SEQUENTIAL) < 0) {
td_verror(td, errno, "madvise");
goto err;
}
} else {
- if (madvise(f->mmap, f->file_size, MADV_RANDOM) < 0) {
+ if (posix_madvise(f->mmap_ptr, length, POSIX_MADV_RANDOM) < 0) {
td_verror(td, errno, "madvise");
goto err;
}
}
+err:
+ if (td->error && f->mmap_ptr)
+ munmap(f->mmap_ptr, length);
+
+ return td->error;
+}
+
+/*
+ * Just mmap an appropriate portion, we cannot mmap the full extent
+ */
+static int fio_mmapio_prep_limited(struct thread_data *td, struct io_u *io_u)
+{
+ struct fio_file *f = io_u->file;
+
+ if (io_u->buflen > mmap_map_size) {
+ log_err("fio: bs too big for mmap engine\n");
+ return EIO;
+ }
+
+ 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;
+
+ return fio_mmap_file(td, f, f->mmap_sz, f->mmap_off);
+}
+
+/*
+ * Attempt to mmap the entire file
+ */
+static int fio_mmapio_prep_full(struct thread_data *td, struct io_u *io_u)
+{
+ struct fio_file *f = io_u->file;
+ int ret;
+
+ if (fio_file_partial_mmap(f))
+ return EINVAL;
+
+ f->mmap_sz = f->io_size;
+ f->mmap_off = 0;
+
+ ret = fio_mmap_file(td, f, f->mmap_sz, f->mmap_off);
+ if (ret)
+ fio_file_set_partial_mmap(f);
+
+ 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;
+
+ /*
+ * It fits within existing mapping, use it
+ */
+ if (io_u->offset >= f->mmap_off &&
+ io_u->offset + io_u->buflen < f->mmap_off + f->mmap_sz)
+ goto done;
+
+ /*
+ * unmap any existing mapping
+ */
+ if (f->mmap_ptr) {
+ if (munmap(f->mmap_ptr, f->mmap_sz) < 0)
+ return errno;
+ f->mmap_ptr = NULL;
+ }
+
+ if (fio_mmapio_prep_full(td, io_u)) {
+ td_clear_error(td);
+ ret = fio_mmapio_prep_limited(td, io_u);
+ if (ret)
+ return ret;
+ }
+
+done:
+ io_u->mmap_data = f->mmap_ptr + io_u->offset - f->mmap_off -
+ f->file_offset;
return 0;
+}
-err:
- td->io_ops->close_file(td, f);
- return 1;
+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 (ddir_sync(io_u->ddir)) {
+ if (msync(f->mmap_ptr, f->mmap_sz, MS_SYNC)) {
+ io_u->error = errno;
+ td_verror(td, io_u->error, "msync");
+ }
+ } else if (io_u->ddir == DDIR_TRIM) {
+ int ret = do_io_u_trim(td, io_u);
+
+ if (!ret)
+ td_verror(td, io_u->error, "trim");
+ }
+
+
+ /*
+ * not really direct, but should drop the pages from the cache
+ */
+ if (td->o.odirect && ddir_rw(io_u->ddir)) {
+ 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 (posix_madvise(io_u->mmap_data, io_u->xfer_buflen, POSIX_MADV_DONTNEED) < 0) {
+ io_u->error = errno;
+ td_verror(td, io_u->error, "madvise");
+ }
+ }
+
+ return FIO_Q_COMPLETED;
}
-static void fio_mmapio_close(struct thread_data fio_unused *td,
- struct fio_file *f)
+static int fio_mmapio_init(struct thread_data *td)
{
- if (f->mmap) {
- munmap(f->mmap, f->file_size);
- f->mmap = NULL;
+ struct thread_options *o = &td->o;
+ unsigned long shift, mask;
+
+ if ((td->o.rw_min_bs & page_mask) &&
+ (o->odirect || o->fsync_blocks || o->fdatasync_blocks)) {
+ log_err("fio: mmap options dictate a minimum block size of "
+ "%lu bytes\n", page_size);
+ return 1;
}
- generic_close_file(td, f);
+
+ 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,
};