fio: add device dax engine
authorDave Jiang <dave.jiang@intel.com>
Wed, 7 Dec 2016 22:54:30 +0000 (15:54 -0700)
committerJens Axboe <axboe@fb.com>
Thu, 15 Dec 2016 20:24:23 +0000 (13:24 -0700)
Adding an ioengine that access DAX device directly. Very similar to the
mmap engine. Uses the libpmem's pmem_memcpy_persist() call for writes.

Signed-off-by: Dave Jiang <dave.jiang@intel.com>
Reviewed-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
Makefile
configure
engines/dev-dax.c [new file with mode: 0644]

index b3a12dd..262ceea 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -127,6 +127,9 @@ endif
 ifdef CONFIG_PMEMBLK
   SOURCE += engines/pmemblk.c
 endif
+ifdef CONFIG_LINUX_DEVDAX
+  SOURCE += engines/dev-dax.c
+endif
 
 ifeq ($(CONFIG_TARGET_OS), Linux)
   SOURCE += diskutil.c fifo.c blktrace.c cgroup.c trim.c engines/sg.c \
index 833b6d3..03bed3b 100755 (executable)
--- a/configure
+++ b/configure
@@ -136,6 +136,7 @@ exit_val=0
 gfio_check="no"
 libhdfs="no"
 pmemblk="no"
+devdax="no"
 disable_lex=""
 prefix=/usr/local
 
@@ -174,6 +175,8 @@ for opt do
   ;;
   --enable-pmemblk) pmemblk="yes"
   ;;
+  --enable-devdax) devdax="yes"
+  ;;
   --disable-lex) disable_lex="yes"
   ;;
   --enable-lex) disable_lex="no"
@@ -205,6 +208,7 @@ if test "$show_help" = "yes" ; then
   echo "--disable-gfapi        Disable gfapi"
   echo "--enable-libhdfs       Enable hdfs support"
   echo "--enable-pmemblk       Enable NVML libpmemblk support"
+  echo "--enable-devdax        Enable NVM Device Dax support"
   echo "--disable-lex          Disable use of lex/yacc for math"
   echo "--enable-lex           Enable use of lex/yacc for math"
   echo "--disable-shm          Disable SHM support"
@@ -1566,6 +1570,10 @@ echo "MTD                           $mtd"
 # Report whether pmemblk engine is enabled
 echo "NVML libpmemblk engine        $pmemblk"
 
+##########################################
+# Report whether dev-dax engine is enabled
+echo "NVM Device Dax engine        $devdax"
+
 # Check if we have lex/yacc available
 yacc="no"
 yacc_is_bison="no"
@@ -1917,6 +1925,9 @@ fi
 if test "$pmemblk" = "yes" ; then
   output_sym "CONFIG_PMEMBLK"
 fi
+if test "$devdax" = "yes" ; then
+  output_sym "CONFIG_LINUX_DEVDAX"
+fi
 if test "$arith" = "yes" ; then
   output_sym "CONFIG_ARITHMETIC"
   if test "$yacc_is_bison" = "yes" ; then
diff --git a/engines/dev-dax.c b/engines/dev-dax.c
new file mode 100644 (file)
index 0000000..6372576
--- /dev/null
@@ -0,0 +1,368 @@
+/*
+ * device DAX engine
+ *
+ * IO engine that reads/writes from files by doing memcpy to/from
+ * a memory mapped region of DAX enabled device.
+ *
+ * Copyright (C) 2016 Intel Corp
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License,
+ * version 2 as published by the Free Software Foundation..
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU General Public License for more details.
+ *
+ */
+
+/*
+ * device dax engine
+ * IO engine that access a DAX device directly for read and write data
+ *
+ * To use:
+ *   ioengine=dev-dax
+ *
+ *   Other relevant settings:
+ *     iodepth=1
+ *     direct=0           REQUIRED
+ *     filename=/dev/daxN.N
+ *     bs=2m
+ *
+ *     direct should be left to 0. Using dev-dax implies that memory access
+ *     is direct. However, dev-dax does not support O_DIRECT flag by design
+ *     since it is not necessary.
+ *
+ *     bs should adhere to the device dax alignment at minimally.
+ *
+ * libpmem.so
+ *   By default, the dev-dax engine will let the system find the libpmem.so
+ *   that it uses. You can use an alternative libpmem by setting the
+ *   FIO_PMEM_LIB environment variable to the full path to the desired
+ *   libpmem.so.
+ */
+
+#include <stdio.h>
+#include <limits.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <errno.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <sys/sysmacros.h>
+#include <dlfcn.h>
+#include <libgen.h>
+
+#include "../fio.h"
+#include "../verify.h"
+
+/*
+ * Limits us to 1GB of mapped files in total to model after
+ * mmap engine behavior
+ */
+#define MMAP_TOTAL_SZ  (1 * 1024 * 1024 * 1024UL)
+
+struct fio_devdax_data {
+       void *devdax_ptr;
+       size_t devdax_sz;
+       off_t devdax_off;
+};
+
+static void * (*pmem_memcpy_persist)(void *dest, const void *src, size_t len);
+
+static int fio_devdax_file(struct thread_data *td, struct fio_file *f,
+                          size_t length, off_t off)
+{
+       struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
+       int flags = 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;
+
+       fdd->devdax_ptr = mmap(NULL, length, flags, MAP_SHARED, f->fd, off);
+       if (fdd->devdax_ptr == MAP_FAILED) {
+               fdd->devdax_ptr = NULL;
+               td_verror(td, errno, "mmap");
+       }
+
+       if (td->error && fdd->devdax_ptr)
+               munmap(fdd->devdax_ptr, length);
+
+       return td->error;
+}
+
+/*
+ * Just mmap an appropriate portion, we cannot mmap the full extent
+ */
+static int fio_devdax_prep_limited(struct thread_data *td, struct io_u *io_u)
+{
+       struct fio_file *f = io_u->file;
+       struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
+
+       if (io_u->buflen > f->real_file_size) {
+               log_err("fio: bs too big for dev-dax engine\n");
+               return EIO;
+       }
+
+       fdd->devdax_sz = min(MMAP_TOTAL_SZ, f->real_file_size);
+       if (fdd->devdax_sz > f->io_size)
+               fdd->devdax_sz = f->io_size;
+
+       fdd->devdax_off = io_u->offset;
+
+       return fio_devdax_file(td, f, fdd->devdax_sz, fdd->devdax_off);
+}
+
+/*
+ * Attempt to mmap the entire file
+ */
+static int fio_devdax_prep_full(struct thread_data *td, struct io_u *io_u)
+{
+       struct fio_file *f = io_u->file;
+       struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
+       int ret;
+
+       if (fio_file_partial_mmap(f))
+               return EINVAL;
+
+       if (io_u->offset != (size_t) io_u->offset ||
+           f->io_size != (size_t) f->io_size) {
+               fio_file_set_partial_mmap(f);
+               return EINVAL;
+       }
+
+       fdd->devdax_sz = f->io_size;
+       fdd->devdax_off = 0;
+
+       ret = fio_devdax_file(td, f, fdd->devdax_sz, fdd->devdax_off);
+       if (ret)
+               fio_file_set_partial_mmap(f);
+
+       return ret;
+}
+
+static int fio_devdax_prep(struct thread_data *td, struct io_u *io_u)
+{
+       struct fio_file *f = io_u->file;
+       struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
+       int ret;
+
+       /*
+        * It fits within existing mapping, use it
+        */
+       if (io_u->offset >= fdd->devdax_off &&
+           io_u->offset + io_u->buflen < fdd->devdax_off + fdd->devdax_sz)
+               goto done;
+
+       /*
+        * unmap any existing mapping
+        */
+       if (fdd->devdax_ptr) {
+               if (munmap(fdd->devdax_ptr, fdd->devdax_sz) < 0)
+                       return errno;
+               fdd->devdax_ptr = NULL;
+       }
+
+       if (fio_devdax_prep_full(td, io_u)) {
+               td_clear_error(td);
+               ret = fio_devdax_prep_limited(td, io_u);
+               if (ret)
+                       return ret;
+       }
+
+done:
+       io_u->mmap_data = fdd->devdax_ptr + io_u->offset - fdd->devdax_off -
+                               f->file_offset;
+       return 0;
+}
+
+static int fio_devdax_queue(struct thread_data *td, struct io_u *io_u)
+{
+       fio_ro_check(td, io_u);
+       io_u->error = 0;
+
+       switch (io_u->ddir) {
+       case DDIR_READ:
+               memcpy(io_u->xfer_buf, io_u->mmap_data, io_u->xfer_buflen);
+               break;
+       case DDIR_WRITE:
+               pmem_memcpy_persist(io_u->mmap_data, io_u->xfer_buf,
+                                   io_u->xfer_buflen);
+               break;
+       case DDIR_SYNC:
+       case DDIR_DATASYNC:
+       case DDIR_SYNC_FILE_RANGE:
+               break;
+       default:
+               io_u->error = EINVAL;
+               break;
+       }
+
+       return FIO_Q_COMPLETED;
+}
+
+static int fio_devdax_init(struct thread_data *td)
+{
+       struct thread_options *o = &td->o;
+       const char *path;
+       void *dl;
+
+       if ((o->rw_min_bs & page_mask) &&
+           (o->fsync_blocks || o->fdatasync_blocks)) {
+               log_err("fio: mmap options dictate a minimum block size of "
+                       "%llu bytes\n", (unsigned long long) page_size);
+               return 1;
+       }
+
+       path = getenv("FIO_PMEM_LIB");
+       if (!path)
+               path = "libpmem.so";
+
+       dl = dlopen(path, RTLD_NOW | RTLD_NODELETE);
+       if (!dl) {
+               log_err("fio: unable to open libpmem: %s\n", dlerror());
+               return 1;
+       }
+
+       pmem_memcpy_persist = dlsym(dl, "pmem_memcpy_persist");
+       if (!pmem_memcpy_persist) {
+               log_err("fio: unable to load libpmem: %s\n", dlerror());
+               return 1;
+       }
+
+       return 0;
+}
+
+static int fio_devdax_open_file(struct thread_data *td, struct fio_file *f)
+{
+       struct fio_devdax_data *fdd;
+       int ret;
+
+       ret = generic_open_file(td, f);
+       if (ret)
+               return ret;
+
+       fdd = calloc(1, sizeof(*fdd));
+       if (!fdd) {
+               int fio_unused __ret;
+               __ret = generic_close_file(td, f);
+               return 1;
+       }
+
+       FILE_SET_ENG_DATA(f, fdd);
+
+       return 0;
+}
+
+static int fio_devdax_close_file(struct thread_data *td, struct fio_file *f)
+{
+       struct fio_devdax_data *fdd = FILE_ENG_DATA(f);
+
+       FILE_SET_ENG_DATA(f, NULL);
+       free(fdd);
+       fio_file_clear_partial_mmap(f);
+
+       return generic_close_file(td, f);
+}
+
+static int
+fio_devdax_get_file_size(struct thread_data *td, struct fio_file *f)
+{
+       char spath[PATH_MAX];
+       char npath[PATH_MAX];
+       char *rpath;
+       FILE *sfile;
+       uint64_t size;
+       struct stat st;
+       int rc;
+
+       if (fio_file_size_known(f))
+               return 0;
+
+       if (f->filetype != FIO_TYPE_CHAR)
+               return -EINVAL;
+
+       rc = stat(f->file_name, &st);
+       if (rc < 0) {
+               log_err("%s: failed to stat file %s: %d\n",
+                       td->o.name, f->file_name, errno);
+               return -errno;
+       }
+
+       snprintf(spath, PATH_MAX, "/sys/dev/char/%d:%d/subsystem",
+                major(st.st_rdev), minor(st.st_rdev));
+
+       rpath = realpath(spath, npath);
+       if (!rpath) {
+               log_err("%s: realpath on %s failed: %d\n",
+                       td->o.name, spath, errno);
+               return -errno;
+       }
+
+       /* check if DAX device */
+       if (strcmp("/sys/class/dax", rpath)) {
+               log_err("%s: %s not a DAX device!\n",
+                       td->o.name, f->file_name);
+       }
+
+       snprintf(spath, PATH_MAX, "/sys/dev/char/%d:%d/size",
+                major(st.st_rdev), minor(st.st_rdev));
+
+       sfile = fopen(spath, "r");
+       if (!sfile) {
+               log_err("%s: fopen on %s failed: %d\n",
+                       td->o.name, spath, errno);
+               return 1;
+       }
+
+       rc = fscanf(sfile, "%lu", &size);
+       if (rc < 0) {
+               log_err("%s: fscanf on %s failed: %d\n",
+                       td->o.name, spath, errno);
+               return 1;
+       }
+
+       f->real_file_size = size;
+
+       fclose(sfile);
+
+       if (f->file_offset > f->real_file_size) {
+               log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
+                                       (unsigned long long) f->file_offset,
+                                       (unsigned long long) f->real_file_size);
+               return 1;
+       }
+
+       fio_file_set_size_known(f);
+       return 0;
+}
+
+static struct ioengine_ops ioengine = {
+       .name           = "dev-dax",
+       .version        = FIO_IOOPS_VERSION,
+       .init           = fio_devdax_init,
+       .prep           = fio_devdax_prep,
+       .queue          = fio_devdax_queue,
+       .open_file      = fio_devdax_open_file,
+       .close_file     = fio_devdax_close_file,
+       .get_file_size  = fio_devdax_get_file_size,
+       .flags          = FIO_SYNCIO | FIO_DISKLESSIO | FIO_NOEXTEND | FIO_NODISKUTIL,
+};
+
+static void fio_init fio_devdax_register(void)
+{
+       register_ioengine(&ioengine);
+}
+
+static void fio_exit fio_devdax_unregister(void)
+{
+       unregister_ioengine(&ioengine);
+}