The transparent hugepage Linux-specific memory advisory has potentially
significant implications for how the memory management behaves. Add a
new mmap specific option that enables private TLP mmap mode when set so
we can test running fio with anonymous memory (i.e. mmap /dev/zero).

Signed-off-by: Keith Busch <[email protected]>
---
 engines/mmap.c | 39 ++++++++++++++++++++++++++++++++++++---
 optgroup.h     |  2 ++
 2 files changed, 38 insertions(+), 3 deletions(-)

diff --git a/engines/mmap.c b/engines/mmap.c
index 308b4665..7bca6c20 100644
--- a/engines/mmap.c
+++ b/engines/mmap.c
@@ -11,6 +11,7 @@
 #include <sys/mman.h>
 
 #include "../fio.h"
+#include "../optgroup.h"
 #include "../verify.h"
 
 /*
@@ -26,6 +27,26 @@ struct fio_mmap_data {
        off_t mmap_off;
 };
 
+struct mmap_options {
+       void *pad;
+       unsigned int thp;
+};
+
+static struct fio_option options[] = {
+       {
+               .name   = "thp",
+               .lname  = "Transparent Huge Pages",
+               .type   = FIO_OPT_INT,
+               .off1   = offsetof(struct mmap_options, thp),
+               .help   = "Memory Advise Huge Page",
+               .category = FIO_OPT_C_ENGINE,
+               .group  = FIO_OPT_G_MMAP,
+       },
+       {
+               .name = NULL,
+       },
+};
+
 static bool fio_madvise_file(struct thread_data *td, struct fio_file *f,
                             size_t length)
 
@@ -54,7 +75,8 @@ static int fio_mmap_file(struct thread_data *td, struct 
fio_file *f,
                         size_t length, off_t off)
 {
        struct fio_mmap_data *fmd = FILE_ENG_DATA(f);
-       int flags = 0;
+       struct mmap_options *o = td->eo;
+       int flags = 0, shared = MAP_SHARED;
 
        if (td_rw(td) && !td->o.verify_only)
                flags = PROT_READ | PROT_WRITE;
@@ -66,7 +88,12 @@ static int fio_mmap_file(struct thread_data *td, struct 
fio_file *f,
        } else
                flags = PROT_READ;
 
-       fmd->mmap_ptr = mmap(NULL, length, flags, MAP_SHARED, f->fd, off);
+#if FIO_OS==os_linux
+       if (o->thp)
+               shared = MAP_PRIVATE;
+#endif
+
+       fmd->mmap_ptr = mmap(NULL, length, flags, shared, f->fd, off);
        if (fmd->mmap_ptr == MAP_FAILED) {
                fmd->mmap_ptr = NULL;
                td_verror(td, errno, "mmap");
@@ -146,6 +173,7 @@ static int fio_mmapio_prep(struct thread_data *td, struct 
io_u *io_u)
 {
        struct fio_file *f = io_u->file;
        struct fio_mmap_data *fmd = FILE_ENG_DATA(f);
+       struct mmap_options *o = td->eo;
        int ret;
 
        /*
@@ -170,10 +198,13 @@ static int fio_mmapio_prep(struct thread_data *td, struct 
io_u *io_u)
                if (ret)
                        return ret;
        }
-
 done:
        io_u->mmap_data = fmd->mmap_ptr + io_u->offset - fmd->mmap_off -
                                f->file_offset;
+#if FIO_OS==os_linux
+       if (o->thp)
+               madvise(io_u->mmap_data, fmd->mmap_sz, MADV_HUGEPAGE);
+#endif
        return 0;
 }
 
@@ -275,6 +306,8 @@ static struct ioengine_ops ioengine = {
        .close_file     = fio_mmapio_close_file,
        .get_file_size  = generic_get_file_size,
        .flags          = FIO_SYNCIO | FIO_NOEXTEND,
+       .options                = options,
+       .option_struct_size     = sizeof(struct mmap_options),
 };
 
 static void fio_init fio_mmapio_register(void)
diff --git a/optgroup.h b/optgroup.h
index adf4d09b..bf1bb036 100644
--- a/optgroup.h
+++ b/optgroup.h
@@ -61,6 +61,7 @@ enum opt_category_group {
        __FIO_OPT_G_MTD,
        __FIO_OPT_G_HDFS,
        __FIO_OPT_G_SG,
+       __FIO_OPT_G_MMAP,
        __FIO_OPT_G_NR,
 
        FIO_OPT_G_RATE          = (1ULL << __FIO_OPT_G_RATE),
@@ -97,6 +98,7 @@ enum opt_category_group {
        FIO_OPT_G_MTD           = (1ULL << __FIO_OPT_G_MTD),
        FIO_OPT_G_HDFS          = (1ULL << __FIO_OPT_G_HDFS),
        FIO_OPT_G_SG            = (1ULL << __FIO_OPT_G_SG),
+       FIO_OPT_G_MMAP          = (1ULL << __FIO_OPT_G_MMAP),
        FIO_OPT_G_INVALID       = (1ULL << __FIO_OPT_G_NR),
 };
 
-- 
2.14.4

Reply via email to