<?xml version='1.0'?> <!--*-nxml-*-->
<!DOCTYPE refentry PUBLIC "-//OASIS//DTD DocBook XML V4.2//EN"
"http://www.oasis-open.org/docbook/xml/4.2/docbookx.dtd">
<!--
Written 2012 by David Herrmann <dh.herrmann@googlemail.com>
Dedicated to the Public Domain
-->
<refentry id="drm-memory">
<refentryinfo>
<title>Direct Rendering Manager</title>
<productname>libdrm</productname>
<date>September 2012</date>
<authorgroup>
<author>
<contrib>Developer</contrib>
<firstname>David</firstname>
<surname>Herrmann</surname>
<email>dh.herrmann@googlemail.com</email>
</author>
</authorgroup>
</refentryinfo>
<refmeta>
<refentrytitle>drm-memory</refentrytitle>
<manvolnum>7</manvolnum>
</refmeta>
<refnamediv>
<refname>drm-memory</refname>
<refname>drm-mm</refname>
<refname>drm-gem</refname>
<refname>drm-ttm</refname>
<refpurpose>DRM Memory Management</refpurpose>
</refnamediv>
<refsynopsisdiv>
<funcsynopsis>
<funcsynopsisinfo>#include <xf86drm.h></funcsynopsisinfo>
</funcsynopsis>
</refsynopsisdiv>
<refsect1>
<title>Description</title>
<para>Many modern high-end GPUs come with their own memory managers. They
even include several different caches that need to be synchronized
during access. Textures, framebuffers, command buffers and more need
to be stored in memory that can be accessed quickly by the GPU.
Therefore, memory management on GPUs is highly driver- and
hardware-dependent.</para>
<para>However, there are several frameworks in the kernel that are used by
more than one driver. These can be used for trivial mode-setting
without requiring driver-dependent code. But for
hardware-accelerated rendering you need to read the manual pages for
the driver you want to work with.</para>
<refsect2>
<title>Dumb-Buffers</title>
<para>Almost all in-kernel DRM hardware drivers support an API called
<emphasis>Dumb-Buffers</emphasis>. This API allows to create buffers
of arbitrary size that can be used for scanout. These buffers can be
memory mapped via
<citerefentry><refentrytitle>mmap</refentrytitle><manvolnum>2</manvolnum></citerefentry>
so you can render into them on the CPU. However, GPU access to these
buffers is often not possible. Therefore, they are fine for simple
tasks but not suitable for complex compositions and
renderings.</para>
<para>The <constant>DRM_IOCTL_MODE_CREATE_DUMB</constant> ioctl can be
used to create a dumb buffer. The kernel will return a 32bit handle
that can be used to manage the buffer with the DRM API. You can
create framebuffers with
<citerefentry><refentrytitle>drmModeAddFB</refentrytitle><manvolnum>3</manvolnum></citerefentry>
and use it for mode-setting and scanout. To access the buffer, you
first need to retrieve the offset of the buffer. The
<constant>DRM_IOCTL_MODE_MAP_DUMB</constant> ioctl requests the DRM
subsystem to prepare the buffer for memory-mapping and returns a
fake-offset that can be used with
<citerefentry><refentrytitle>mmap</refentrytitle><manvolnum>2</manvolnum></citerefentry>.</para>
<para>The <constant>DRM_IOCTL_MODE_CREATE_DUMB</constant> ioctl takes as
argument a structure of type
<structname>struct drm_mode_create_dumb</structname>:
<programlisting>
struct drm_mode_create_dumb {
__u32 height;
__u32 width;
__u32 bpp;
__u32 flags;
__u32 handle;
__u32 pitch;
__u64 size;
};
</programlisting>
The fields <structfield>height</structfield>,
<structfield>width</structfield>, <structfield>bpp</structfield> and
<structfield>flags</structfield> have to be provided by the caller.
The other fields are filled by the kernel with the return values.
<structfield>height</structfield> and
<structfield>width</structfield> are the dimensions of the
rectangular buffer that is created. <structfield>bpp</structfield>
is the number of bits-per-pixel and must be a multiple of
<literal>8</literal>. You most commonly want to pass
<literal>32</literal> here. The <structfield>flags</structfield>
field is currently unused and must be zeroed. Different flags to
modify the behavior may be added in the future. After calling the
ioctl, the <structfield>handle</structfield>,
<structfield>pitch</structfield> and <structfield>size</structfield>
fields are filled by the kernel. <structfield>handle</structfield>
is a 32bit gem handle that identifies the buffer. This is used by
several other calls that take a gem-handle or memory-buffer as
argument. The <structfield>pitch</structfield> field is the
pitch (or stride) of the new buffer. Most drivers use 32bit or 64bit
aligned stride-values. The <structfield>size</structfield> field
contains the absolute size in bytes of the buffer. This can normally
also be computed with
<emphasis>(height * pitch + width) * bpp / 4</emphasis>.</para>
<para>To prepare the buffer for
<citerefentry><refentrytitle>mmap</refentrytitle><manvolnum>2</manvolnum></citerefentry>
you need to use the <constant>DRM_IOCTL_MODE_MAP_DUMB</constant>
ioctl. It takes as argument a structure of type
<structname>struct drm_mode_map_dumb</structname>:
<programlisting>
struct drm_mode_map_dumb {
__u32 handle;
__u32 pad;
__u64 offset;
};
</programlisting>
You need to put the gem-handle that was previously retrieved via
<constant>DRM_IOCTL_MODE_CREATE_DUMB</constant> into the
<structfield>handle</structfield> field. The
<structfield>pad</structfield> field is unused padding and must be
zeroed. After completion, the <structfield>offset</structfield>
field will contain an offset that can be used with
<citerefentry><refentrytitle>mmap</refentrytitle><manvolnum>2</manvolnum></citerefentry>
on the DRM file-descriptor.</para>
<para>If you don't need your dumb-buffer, anymore, you have to destroy it
with <constant>DRM_IOCTL_MODE_DESTROY_DUMB</constant>. If you close
the DRM file-descriptor, all open dumb-buffers are automatically
destroyed. This ioctl takes as argument a structure of type
<structname>struct drm_mode_destroy_dumb</structname>:
<programlisting>
struct drm_mode_destroy_dumb {
__u32 handle;
};
</programlisting>
You only need to put your handle into the
<structfield>handle</structfield> field. After this call, the handle
is invalid and may be reused for new buffers by the dumb-API.</para>
</refsect2>
<refsect2>
<title>TTM</title>
<para><emphasis>TTM</emphasis> stands for
<emphasis>Translation Table Manager</emphasis> and is a generic
memory-manager provided by the kernel. It does not provide a common
user-space API so you need to look at each driver interface if you
want to use it. See for instance the radeon manpages for more
information on memory-management with radeon and TTM.</para>
</refsect2>
<refsect2>
<title>GEM</title>
<para><emphasis>GEM</emphasis> stands for
<emphasis>Graphics Execution Manager</emphasis> and is a generic DRM
memory-management framework in the kernel, that is used by many
different drivers. Gem is designed to manage graphics memory,
control access to the graphics device execution context and handle
essentially NUMA environment unique to modern graphics hardware. Gem
allows multiple applications to share graphics device resources
without the need to constantly reload the entire graphics card. Data
may be shared between multiple applications with gem ensuring that
the correct memory synchronization occurs.</para>
<para>Gem provides simple mechanisms to manage graphics data and control
execution flow within the linux DRM subsystem. However, gem is not a
complete framework that is fully driver independent. Instead, if
provides many functions that are shared between many drivers, but
each driver has to implement most of memory-management with
driver-dependent ioctls. This manpage tries to describe the
semantics (and if it applies, the syntax) that is shared between all
drivers that use gem.</para>
<para>All GEM APIs are defined as
<citerefentry><refentrytitle>ioctl</refentrytitle><manvolnum>2</manvolnum></citerefentry>
on the DRM file descriptor. An application must be authorized via
<citerefentry><refentrytitle>drmAuthMagic</refentrytitle><manvolnum>3</manvolnum></citerefentry>
to the current DRM-Master to access the GEM subsystem. A driver that
does not support gem will return <constant>ENODEV</constant> for all
these ioctls. Invalid object handles return
<constant>EINVAL</constant> and invalid object names return
<constant>ENOENT</constant>.</para>
<para>Gem provides explicit memory management primitives. System pages are
allocated when the object is created, either as the fundamental
storage for hardware where system memory is used by the graphics
processor directly, or as backing store for graphics-processor
resident memory.</para>
<para>Objects are referenced from user-space using handles. These are, for
all intents and purposes, equivalent to file descriptors but avoid
the overhead. Newer kernel drivers also support the
<citerefentry><refentrytitle>drm-prime</refentrytitle><manvolnum>7</manvolnum></citerefentry>
infrastructure which can return real file-descriptor for gem-handles
using the linux dma-buf API. Objects may be published with a name so
that other applications and processes can access them. The name
remains valid as long as the object exists. Gem-objects are
reference counted in the kernel. The object is only destroyed when
all handles from user-space were closed.</para>
<para>Gem-buffers cannot be created with a generic API. Each driver
provides its own API to create gem-buffers. See for example
<constant>DRM_I915_GEM_CREATE</constant>,
<constant>DRM_NOUVEAU_GEM_NEW</constant> or
<constant>DRM_RADEON_GEM_CREATE</constant>. Each of these ioctls
returns a gem-handle that can be passed to different generic ioctls.
The <emphasis>libgbm</emphasis> library from the
<emphasis>mesa3D</emphasis> distribution tries to provide a
driver-independent API to create gbm buffers and retrieve a
gbm-handle to them. It allows to create buffers for different
use-cases including scanout, rendering, cursors and CPU-access. See
the libgbm library for more information or look at the
driver-dependent man-pages (for example
<citerefentry><refentrytitle>drm-intel</refentrytitle><manvolnum>7</manvolnum></citerefentry>
or
<citerefentry><refentrytitle>drm-radeon</refentrytitle><manvolnum>7</manvolnum></citerefentry>).</para>
<para>Gem-buffers can be closed with the
<constant>DRM_IOCTL_GEM_CLOSE</constant> ioctl. It takes as argument
a structure of type <structname>struct drm_gem_close</structname>:
<programlisting>
struct drm_gem_close {
__u32 handle;
__u32 pad;
};
</programlisting>
The <structfield>handle</structfield> field is the gem-handle to be
closed. The <structfield>pad</structfield> field is unused padding.
It must be zeroed. After this call the gem handle cannot be used by
this process anymore and may be reused for new gem objects by the
gem API.</para>
<para>If you want to share gem-objects between different processes, you
can create a name for them and pass this name to other processes
which can then open this gem-object. Names are currently 32bit
integer IDs and have no special protection. That is, if you put a
name on your gem-object, every other client that has access to the
DRM device and is authenticated via
<citerefentry><refentrytitle>drmAuthMagic</refentrytitle><manvolnum>3</manvolnum></citerefentry>
to the current DRM-Master, can <emphasis>guess</emphasis> the name
and open or access the gem-object. If you want more fine-grained
access control, you can use the new
<citerefentry><refentrytitle>drm-prime</refentrytitle><manvolnum>7</manvolnum></citerefentry>
API to retrieve file-descriptors for gem-handles. To create a name
for a gem-handle, you use the
<constant>DRM_IOCTL_GEM_FLINK</constant> ioctl. It takes as argument
a structure of type <structname>struct drm_gem_flink</structname>:
<programlisting>
struct drm_gem_flink {
__u32 handle;
__u32 name;
};
</programlisting>
You have to put your handle into the
<structfield>handle</structfield> field. After completion, the
kernel has put the new unique name into the
<structfield>name</structfield> field. You can now pass this name to
other processes which can then import the name with the
<constant>DRM_IOCTL_GEM_OPEN</constant> ioctl. It takes as argument
a structure of type <structname>struct drm_gem_open</structname>:
<programlisting>
struct drm_gem_open {
__u32 name;
__u32 handle;
__u32 size;
};
</programlisting>
You have to fill in the <structfield>name</structfield> field with
the name of the gem-object that you want to open. The kernel will
fill in the <structfield>handle</structfield> and
<structfield>size</structfield> fields with the new handle and size
of the gem-object. You can now access the gem-object via the handle
as if you created it with the gem API.</para>
<para>Besides generic buffer management, the GEM API does not provide any
generic access. Each driver implements its own functionality on top
of this API. This includes execution-buffers, GTT management,
context creation, CPU access, GPU I/O and more. The next
higher-level API is <emphasis>OpenGL</emphasis>. So if you want to
use more GPU features, you should use the
<emphasis>mesa3D</emphasis> library to create OpenGL contexts on DRM
devices. This does <emphasis>not</emphasis> require any
windowing-system like X11, but can also be done on raw DRM devices.
However, this is beyond the scope of this man-page. You may have a
look at other mesa3D manpages, including libgbm and libEGL. 2D
software-rendering (rendering with the CPU) can be achieved with the
dumb-buffer-API in a driver-independent fashion, however, for
hardware-accelerated 2D or 3D rendering you must use OpenGL. Any
other API that tries to abstract the driver-internals to access
GEM-execution-buffers and other GPU internals, would simply reinvent
OpenGL so it is not provided. But if you need more detailed
information for a specific driver, you may have a look into the
driver-manpages, including
<citerefentry><refentrytitle>drm-intel</refentrytitle><manvolnum>7</manvolnum></citerefentry>,
<citerefentry><refentrytitle>drm-radeon</refentrytitle><manvolnum>7</manvolnum></citerefentry>
and
<citerefentry><refentrytitle>drm-nouveau</refentrytitle><manvolnum>7</manvolnum></citerefentry>.
However, the
<citerefentry><refentrytitle>drm-prime</refentrytitle><manvolnum>7</manvolnum></citerefentry>
infrastructure and the generic gem API as described here allow
display-managers to handle graphics-buffers and render-clients
without any deeper knowledge of the GPU that is used. Moreover, it
allows to move objects between GPUs and implement complex
display-servers that don't do any rendering on their own. See its
man-page for more information.</para>
</refsect2>
</refsect1>
<refsect1>
<title>Examples</title>
<para>This section includes examples for basic memory-management
tasks.</para>
<refsect2>
<title>Dumb-Buffers</title>
<para>This examples shows how to create a dumb-buffer via the generic
DRM API. This is driver-independent (as long as the driver
supports dumb-buffers) and provides memory-mapped buffers that can
be used for scanout. This example creates a full-HD 1920x1080
buffer with 32 bits-per-pixel and a color-depth of 24 bits. The
buffer is then bound to a framebuffer which can be used for
scanout with the KMS API (see
<citerefentry><refentrytitle>drm-kms</refentrytitle><manvolnum>7</manvolnum></citerefentry>).</para>
<programlisting>
struct drm_mode_create_dumb creq;
struct drm_mode_destroy_dumb dreq;
struct drm_mode_map_dumb mreq;
uint32_t fb;
int ret;
void *map;
/* create dumb buffer */
memset(&creq, 0, sizeof(creq));
creq.width = 1920;
creq.height = 1080;
creq.bpp = 32;
ret = drmIoctl(fd, DRM_IOCTL_MODE_CREATE_DUMB, &creq);
if (ret < 0) {
/* buffer creation failed; see "errno" for more error codes */
...
}
/* creq.pitch, creq.handle and creq.size are filled by this ioctl with
* the requested values and can be used now. */
/* create framebuffer object for the dumb-buffer */
ret = drmModeAddFB(fd, 1920, 1080, 24, 32, creq.pitch, creq.handle, &fb);
if (ret) {
/* frame buffer creation failed; see "errno" */
...
}
/* the framebuffer "fb" can now used for scanout with KMS */
/* prepare buffer for memory mapping */
memset(&mreq, 0, sizeof(mreq));
mreq.handle = creq.handle;
ret = drmIoctl(fd, DRM_IOCTL_MODE_MAP_DUMB, &mreq);
if (ret) {
/* DRM buffer preparation failed; see "errno" */
...
}
/* mreq.offset now contains the new offset that can be used with mmap() */
/* perform actual memory mapping */
map = mmap(0, creq.size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, mreq.offset);
if (map == MAP_FAILED) {
/* memory-mapping failed; see "errno" */
...
}
/* clear the framebuffer to 0 */
memset(map, 0, creq.size);
</programlisting>
</refsect2>
</refsect1>
<refsect1>
<title>Reporting Bugs</title>
<para>Bugs in this manual should be reported to
http://bugs.freedesktop.org under the "Mesa" product, with "Other" or
"libdrm" as the component.</para>
</refsect1>
<refsect1>
<title>See Also</title>
<para>
<citerefentry><refentrytitle>drm</refentrytitle><manvolnum>7</manvolnum></citerefentry>,
<citerefentry><refentrytitle>drm-kms</refentrytitle><manvolnum>7</manvolnum></citerefentry>,
<citerefentry><refentrytitle>drm-prime</refentrytitle><manvolnum>7</manvolnum></citerefentry>,
<citerefentry><refentrytitle>drmAvailable</refentrytitle><manvolnum>3</manvolnum></citerefentry>,
<citerefentry><refentrytitle>drmOpen</refentrytitle><manvolnum>3</manvolnum></citerefentry>,
<citerefentry><refentrytitle>drm-intel</refentrytitle><manvolnum>7</manvolnum></citerefentry>,
<citerefentry><refentrytitle>drm-radeon</refentrytitle><manvolnum>7</manvolnum></citerefentry>,
<citerefentry><refentrytitle>drm-nouveau</refentrytitle><manvolnum>7</manvolnum></citerefentry>
</para>
</refsect1>
</refentry>
