fusion atomic: support for new SDK API
authorSanthosh Koundinya <SKoundinya@fusionio.com>
Sat, 18 May 2013 06:44:01 +0000 (08:44 +0200)
committerJens Axboe <axboe@kernel.dk>
Sat, 18 May 2013 06:44:01 +0000 (08:44 +0200)
vsl_vectored_write() is dead, long live nvm_atomic_write().

Signed-off-by: Jens Axboe <axboe@kernel.dk>
configure
engines/fusion-aw.c

index 95765cc..98af5ff 100755 (executable)
--- a/configure
+++ b/configure
@@ -809,15 +809,18 @@ echo "GUASI                         $guasi"
 # fusion-aw probe
 fusion_aw="no"
 cat > $TMPC << EOF
-#include <nvm/vectored_write.h>
+#include <nvm/nvm_primitives.h>
 int main(int argc, char **argv)
 {
-  struct vsl_iovec iov;
-  return vsl_vectored_write(0, &iov, 0, O_ATOMIC);
+  nvm_version_t ver_info;
+  nvm_handle_t handle;
+
+  handle = nvm_get_handle(0, &ver_info);
+  return nvm_atomic_write(handle, 0, 0, 0);
 }
 EOF
-if compile_prog "" "-L/usr/lib/fio -lnvm-primitives" "fusion-aw"; then
-  LIBS="-L/usr/lib/fio -lnvm-primitives $LIBS"
+if compile_prog "" "-L/usr/lib/fio -L/usr/lib/nvm -lnvm-primitives -lvsl -ldl" "fusion-aw"; then
+  LIBS="-L/usr/lib/fio -L/usr/lib/nvm -lnvm-primitives -lvsl -ldl $LIBS"
   fusion_aw="yes"
 fi
 echo "Fusion-io atomic engine       $fusion_aw"
index 0bc5c0b..9092eca 100644 (file)
@@ -1,8 +1,8 @@
 /*
  * Custom fio(1) engine that submits synchronous atomic writes to file.
  *
- * Copyright (C) 2012 Fusion-io, Inc.
- * Author: Santhosh Kumar Koundinya.
+ * Copyright (C) 2013 Fusion-io, Inc.
+ * Author: Santhosh Kumar Koundinya (skoundinya@fusionio.com).
  *
  * This program is free software; you can redistribute it and/or modify it
  * under the terms of the GNU General Public License as published by the Free
 
 #include "../fio.h"
 
-#include <nvm/vectored_write.h>
+#include <nvm/nvm_primitives.h>
 
-/* Fix sector size to 512 bytes independent of actual sector size, just like
- * the linux kernel. */
-#define SECTOR_SHIFT    9
-#define SECTOR_SIZE    (1U<<SECTOR_SHIFT)
-
-struct acs_engine_data {
-       struct vsl_iovec iov[IO_VECTOR_LIMIT];
+struct fas_data {
+       nvm_handle_t nvm_handle;
+       size_t xfer_buf_align;
+       size_t xfer_buflen_align;
+       size_t xfer_buflen_max;
+       size_t sector_size;
 };
 
 static int queue(struct thread_data *td, struct io_u *io_u)
 {
        int rc;
-       int iov_index;
-       off_t offset;
-       char *xfer_buf;
-       size_t xfer_buflen;
-       struct acs_engine_data *d =
-               (struct acs_engine_data *) io_u->file->engine_data;
+       struct fas_data *d = (struct fas_data *) io_u->file->engine_data;
 
        if (io_u->ddir != DDIR_WRITE) {
-               td_vmsg(td, -EIO, "only writes supported", "io_u->ddir");
-               rc = -EIO;
+               td_vmsg(td, EINVAL, "only writes supported", "io_u->ddir");
+               rc = -EINVAL;
                goto out;
        }
-       if (io_u->xfer_buflen > IO_SIZE_MAX) {
-               td_vmsg(td, -EIO, "data too big", "io_u->xfer_buflen");
-               rc = -EIO;
+
+       if ((size_t) io_u->xfer_buf % d->xfer_buf_align) {
+               td_vmsg(td, EINVAL, "unaligned data buffer", "io_u->xfer_buf");
+               rc = -EINVAL;
                goto out;
        }
-       if (io_u->xfer_buflen & (SECTOR_SIZE - 1)) {
-               td_vmsg(td, -EIO, "unaligned data size", "io_u->xfer_buflen");
-               rc = -EIO;
+
+       if (io_u->xfer_buflen % d->xfer_buflen_align) {
+               td_vmsg(td, EINVAL, "unaligned data size", "io_u->xfer_buflen");
+               rc = -EINVAL;
                goto out;
        }
 
-       /* Chop up the write into minimal number of iovec's necessary */
-       iov_index = 0;
-       offset = io_u->offset;
-       xfer_buf = io_u->xfer_buf;
-       xfer_buflen = io_u->xfer_buflen;
-       while (xfer_buflen) {
-               struct vsl_iovec *iov = &d->iov[iov_index++];
-
-               iov->iov_len = xfer_buflen > IO_VECTOR_MAX_SIZE ?
-                   IO_VECTOR_MAX_SIZE : xfer_buflen;
-               iov->iov_base = (uint64_t) xfer_buf;
-               iov->sector = offset >> SECTOR_SHIFT;
-               iov->iov_flag = VSL_IOV_WRITE;
-
-               offset += iov->iov_len;
-               xfer_buf += iov->iov_len;
-               xfer_buflen -= iov->iov_len;
+       if (io_u->xfer_buflen > d->xfer_buflen_max) {
+               td_vmsg(td, EINVAL, "data too big", "io_u->xfer_buflen");
+               rc = -EINVAL;
+               goto out;
        }
-       assert(xfer_buflen == 0);
-       assert(iov_index <= IO_VECTOR_LIMIT);
 
-       rc = vsl_vectored_write(io_u->file->fd, d->iov, iov_index, O_ATOMIC);
+       rc = nvm_atomic_write(d->nvm_handle, (uint64_t) io_u->xfer_buf,
+               io_u->xfer_buflen, io_u->offset / d->sector_size);
        if (rc == -1) {
-               td_verror(td, -errno, "vsl_vectored_write");
-               rc = -EIO;
+               td_verror(td, errno, "nvm_atomic_write");
+               rc = -errno;
                goto out;
-       } else {
-               io_u->error = 0;
-               rc = FIO_Q_COMPLETED;
        }
-
+       rc = FIO_Q_COMPLETED;
 out:
        if (rc < 0)
-               io_u->error = rc;
+               io_u->error = -rc;
 
        return rc;
 }
@@ -100,34 +79,79 @@ out:
 static int open_file(struct thread_data *td, struct fio_file *f)
 {
        int rc;
-       struct acs_engine_data *d = NULL;
+       int fio_unused close_file_rc;
+       struct fas_data *d;
+       nvm_version_t nvm_version;
+       nvm_capability_t nvm_capability[4];
+
 
        d = malloc(sizeof(*d));
        if (!d) {
-               td_verror(td, -ENOMEM, "malloc");
-               rc = -ENOMEM;
+               td_verror(td, ENOMEM, "malloc");
+               rc = ENOMEM;
                goto error;
        }
+       d->nvm_handle = -1;
        f->engine_data = (uintptr_t) d;
 
        rc = generic_open_file(td, f);
 
+       if (rc)
+               goto free_engine_data;
+
+       /* Set the version of the library as seen when engine is compiled */
+       nvm_version.major = NVM_PRIMITIVES_API_MAJOR;
+       nvm_version.minor = NVM_PRIMITIVES_API_MINOR;
+       nvm_version.micro = NVM_PRIMITIVES_API_MICRO;
+
+       d->nvm_handle = nvm_get_handle(f->fd, &nvm_version);
+       if (d->nvm_handle == -1) {
+               td_vmsg(td, errno, "nvm_get_handle failed", "nvm_get_handle");
+               rc = errno;
+               goto close_file;
+       }
+
+       nvm_capability[0].cap_id = NVM_CAP_ATOMIC_WRITE_START_ALIGN_ID;
+       nvm_capability[1].cap_id = NVM_CAP_ATOMIC_WRITE_MULTIPLICITY_ID;
+       nvm_capability[2].cap_id = NVM_CAP_ATOMIC_WRITE_MAX_VECTOR_SIZE_ID;
+       nvm_capability[3].cap_id = NVM_CAP_SECTOR_SIZE_ID;
+       rc = nvm_get_capabilities(d->nvm_handle, nvm_capability, 4, 0);
+       if (rc == -1) {
+               td_vmsg(td, errno, "error in getting atomic write capabilities", "nvm_get_capabilities");
+               rc = errno;
+               goto close_file;
+       } else if (rc < 4) {
+               td_vmsg(td, EINVAL, "couldn't get all the atomic write capabilities" , "nvm_get_capabilities");
+               rc = ECANCELED;
+               goto close_file;
+       }
+       /* Reset rc to 0 because we got all capabilities we needed */
+       rc = 0;
+       d->xfer_buf_align = nvm_capability[0].cap_value;
+       d->xfer_buflen_align = nvm_capability[1].cap_value;
+       d->xfer_buflen_max = d->xfer_buflen_align * nvm_capability[2].cap_value;
+       d->sector_size = nvm_capability[3].cap_value;
+
 out:
        return rc;
-
+close_file:
+       close_file_rc = generic_close_file(td, f);
+free_engine_data:
+       free(d);
 error:
        f->fd = -1;
        f->engine_data = 0;
-       if (d)
-               free(d);
-
        goto out;
 }
 
 static int close_file(struct thread_data *td, struct fio_file *f)
 {
-       if (f->engine_data) {
-               free((void *) f->engine_data);
+       struct fas_data *d = (struct fas_data *) f->engine_data;
+
+       if (d) {
+               if (d->nvm_handle != -1)
+                       nvm_release_handle(d->nvm_handle);
+               free(d);
                f->engine_data = 0;
        }