Recent changes (master)

index 3684fe70..0b39b01a 100644
@@ -2559,6 +2559,22 @@ with the caveat that when used on the command line, they must come after the
 	map and release for each IO. This is more efficient, and reduces the
 	IO latency as well.
 
+.. option:: dmabuf=str : [io_uring]
+
+	Select the DMA-BUF exporter for the IO buffers. Allowed values are:
+
+		**none**
+			Do not use DMA-BUF-backed buffers. This is the default.
+
+		**udmabuf**
+			Allocate the IO buffers from a udmabuf (``/dev/udmabuf``)
+			region and register them with io_uring as DMA-BUF-backed
+			fixed buffers, so that the device DMAs directly to/from the
+			DMA-BUF. Requires :option:`direct` =1; buffered IO is not
+			supported. Requires a kernel that supports the extended
+			buffer registration update. Not supported io_uring_cmd mode
+			as kernel lacks support for passthrough.
+
 .. option:: nonvectored=int : [io_uring] [io_uring_cmd]
 
 	With this option, fio will use non-vectored read/write commands, where

index 2ab83371..60d0ef98 100755
@@ -2709,6 +2709,38 @@ else
   nvme_uring_cmd="no"
 fi
 print_config "NVMe uring command support" "$nvme_uring_cmd"
+
+##########################################
+# Check udmabuf allocation support
+cat > $TMPC << EOF
+#define _GNU_SOURCE
+#include <fcntl.h>
+#include <sys/mman.h>
+#include <sys/ioctl.h>
+#include <linux/udmabuf.h>
+int main(void)
+{
+  struct udmabuf_create create = {
+    .memfd = 0,
+    .flags = 0,
+    .offset = 0,
+    .size = 4096,
+  };
+  int (*create_memfd)(const char *, unsigned int) = memfd_create;
+  int fd = create_memfd("fio-udmabuf", MFD_ALLOW_SEALING);
+
+  if (fcntl(fd, F_ADD_SEALS, F_SEAL_SHRINK) < 0)
+    return 1;
+  return ioctl(fd, UDMABUF_CREATE, &create);
+}
+EOF
+if compile_prog "" "" "udmabuf"; then
+  output_sym "CONFIG_UDMABUF"
+  udmabuf="yes"
+else
+  udmabuf="no"
+fi
+print_config "udmabuf allocation support" "$udmabuf"
 fi
 
 ##########################################

new file mode 100644
index 00000000..49037bf1
@@ -0,0 +1,14 @@
+[global]
+filename=/dev/nvme1n1
+rw=write
+bs=4096
+direct=1
+runtime=30
+time_based=1
+group_reporting=1
+
+[file1]
+iodepth=128
+fixedbufs=0
+dmabuf=udmabuf
+ioengine=io_uring
\ No newline at end of file

index 7dc5c353..08153d30 100644
@@ -2342,6 +2342,22 @@ release them when IO is done. If this option is set, the pages are pre-mapped
 before IO is started. This eliminates the need to map and release for each IO.
 This is more efficient, and reduces the IO latency as well.
 .TP
+.BI (io_uring)dmabuf \fR=\fPstr
+Select the DMA-BUF exporter for the IO buffers. Allowed values are:
+.RS
+.TP
+.B none
+Do not use DMA-BUF-backed buffers. This is the default.
+.TP
+.B udmabuf
+Allocate the IO buffers from a udmabuf (\fB/dev/udmabuf\fR) region and
+register them with io_uring as DMA-BUF-backed fixed buffers, so that the
+device DMAs directly to/from the DMA-BUF. Requires \fBdirect\fR=1;
+buffered IO is not supported. Requires a kernel that supports
+the extended buffer registration update. Not supported io_uring_cmd mode
+as kernel lacks support for passthrough.
+.RE
+.TP
 .BI (io_uring,io_uring_cmd)nonvectored \fR=\fPint
 With this option, fio will use non-vectored read/write commands, where address
 must contain the address directly. Default is -1.