Hi,
Here are my own notes from the Linaro memory management mini-summit in
Budapest. I've written them from my own point of view, which is mostly
V4L2 in embedded devices and camera related use cases. I attempted to
summarise the discussion mostly concentrating into parts which I've
considered important and ignored the rest.
So please do not consider this as the generic notes of the mini-summit.
:-) I still felt like sharing this since it might be found useful by
those who are working with similar systems with similar problems.
Memory buffer management --- the future
=======================================
The memory buffer management can be split to following sub-problems
which may have dependencies, both in implementation and possibly in
the APIs as well:
- Fulfilling buffer allocation requirements
- API to allocate buffers
- Sharing buffers among kernel subsystems (e.g. V4L2, DRM, FB)
- Sharing buffers between processes
- Cache coherency
What has been agreed that we need kernel to recognise a DMA buffer
which may be passed between user processes and different kernel subsystems.
Fulfilling buffer allocation requirements
-----------------------------------------
APIs, as well as devices, have different requirements on the buffers.
It is difficult to come up with generic requirements for buffer
allocation, and to keep the solution future-proof is challenging as
well. In principle the user is interested in being able to share
buffers between subsystems without knowing the exact requirements of
the devices, which makes it possible to keep the requirement handling
internal to the kernel. Whether this is the way to go or not, will be
seen in the future. The buffer allocation remains a problem to be
resolved in the future.
Majority of the devices' requirements could be filled using a few
allocators; one for physically continugous memory and the other for
physically non-contiguous memory of single page allocations. Being
able to allocate large pages would also be beneficial in many cases.
API to allocate buffers
-----------------------
It was agreed there was a need to have a generic interface for buffer
object creation. This could be either a new system call which would be
supported by all devices supporting such buffers in subsystem APIs
(such as V4L2), or a new dedicated character device.
Different subsystems have different ways of describing the properties
of the buffers, such as how the data in the buffer should be
interpreted. The V4L2 has width, height, bytesperline and pixel
format, for example. The generic buffers should not recognise such
properties since this is very subsystem specific information. Instead,
the user which is aware of the different subsystems must come with
matching set of buffer properties using the subsystem specific
interfaces.
Sharing buffers among kernel subsystems
---------------------------------------
There was discussion on how to refer to generic DMA buffers, and the
audience was first mostly split between using buffer IDs to refer to
the buffers and using file handles for the purpose. Using file handles
have pros and cons compared to the numeric IDs:
+ Easy life cycle management. Deallocation of buffers no longer in use
is trivial.
+ Access control for files exists already. Passing file descriptors
between processes is possible throught Unix sockets.
- Allocating extremely large number of buffers would require as many
file descriptors. This is not likely to be an important issue.
Before the day ended, it was felt that the file handles are the right
way to go.
The generic DMA buffers further need to be associated to the subsystem
buffers. This is up to the subsystem APIs. In V4L2, this would most
likely mean that there will be a new buffer type for the generic DMA
buffers.
Sharing buffers between processes
---------------------------------
Numeric IDs can be easily shared between processes while sharing file
handles is more difficult. However, it can be done using the Unix
sockets between any two processes. This also gives automatically
the same access control mechanism as every other file. Access control
mechanisms are mandatory when making the buffer shareable between
processes.
Cache coherency
---------------
Cache coherency is seen largely orthogonal to any other sub-problems
in memory buffer management. In few cases this might have something in
common with buffer allocation. Some architectures, ARM in particular, do
not have coherent caches, meaning that the operating system must know
when to invalidate or clean various parts of the cache. There are two
ways to approach the issue, independently of the cache implementation:
1. Allocate non-cacheable memory, or
2. invalidate or clean (or flush) the cache when necessary.
Allocating non-cacheable memory is a valid solution to cache coherency
handling in some situations, but mostly only when the buffer is only
partially accessed by the CPU or at least not multiple times. In other
cases, invalidating or cleaning the cache is the way to go.
The exact circumstances in which using non-cacheable memory gives a
performance benefit over invalidating or cleaning the cache when
necessary are very system and use case dependent. This should be
selectable from the user space.
The cache invalidation or cleaning can be either on the whole (data)
cache or a particular memory area. Performing the operation on a
particular memory area may be difficult since it should be done to all
mappings of the memory in the system. Also, there is a limit beyond
which performing invalidation or clean for an area is always more
expensive than a full cache flush: on many machines the cache line
size is 64 bytes, and the invalidate/clean must be performed for the
whole buffer, which in cameras could be tens of megabytes in size, per
every cache line.
Mapping buffers to application memory is not always necessary --- the
buffers may only be used by the devices, in which case a scatterlist
of the pages in the buffer is necessary to map the buffer to the IOMMU.
More (impartial :-)) information can be found here:
<URL:http://summit.ubuntu.com/uds-o/meeting/linaro-graphics-memory-managemen…>
<URL:http://summit.ubuntu.com/uds-o/meeting/linaro-graphics-memory-managemen…>
<URL:http://summit.ubuntu.com/uds-o/meeting/linaro-graphics-memory-managemen…>
Regards,
--
Sakari Ailus
sakari.ailus(a)maxwell.research.nokia.com
Hi all,
During the Budapest meetings it was mentioned that you can pass a fd between
processes. How does that work? Does someone have a code example or a link to
code that does that? Just to satisfy my curiosity.
Regards,
Hans
Thanks Jesse for initiating the mailing list.
We need to address the requirements of Graphics and Multimedia Accelerators
(IPs).
What we really need is a permanent solution (at upstream) which accommodates
the following requirements and conforms to Graphics and Multimedia use
cases.
1.Mechanism to map/unmap the memory. Some of the IPs’ have the ability to
address virtual memory and some can address only physically contiguous
address space. We need to address both these cases.
2.Mechanism to allocate and release memory.
3.Method to share the memory (ZERO copy is a MUST for better performance)
between different device drivers (example output of camera to multimedia
encoder).
4.Method to share the memory with different processes in userspace. The
sharing mechanism should include built-in security features.
Are there any special requirements from V4L or DRM perspectives?
Thanks,
Sree
(Disclaimer: I come from a graphics background, so sorry if I use graphicsy
terminology; please let me know if any of this isn't clear. I tried.)
There is an wide range of hardware capabilities that require different
programming approaches in order to perform optimally. We need to define an
interface that is flexible enough to handle each of them, or else it won't be
used and we'll be right back where we are today: with vendors rolling their own
support for the things they need.
I'm going to try to enumerate some of the more unique usage patterns as
I see them here.
- Many or all engines may sit behind asynchronous command stream interfaces.
Programming is done through "batch buffers"; a set of commands operating on a
set of in-memory buffers is prepared and then submitted to the kernel to be
queued. The kernel will first make sure all of the buffers are resident
(which may require paging or mapping into an IOMMU/GART, a.k.a. "pinning"),
then queue the batch of commands. The hardware will process the commands at
its earliest convenience, and then interrupt the CPU to notify it that it's
done with the buffers (i.e. it can now be "unpinned").
Those familiar with graphics may recognize this programming model as a
classic GPU command stream. But it doesn't need to be used exclusively with
GPUs; any number of devices may have such an on-demand paging mechanism.
- In contrast, some engines may also stream to or from memory continuously
(e.g., video capture or scanout); such buffers need to be pinned for an
extended period of time, not tied to the command streams described above.
- There can be multiple different command streams working at the same time on
the same buffers. (There may be hardware synchronization primitives between
the multiple command streams so the CPU doesn't have to babysit too much, for
both performance and power reasons.)
- In some systems, IOMMU/GART may be much smaller than physical memory; older
GPUs and SoCs have this. To support these, we need to be able to map and
unmap pages into the IOMMU on demand in our host command stream flow. This
model also requires patching up pending batch buffers before queueing them to
the hardware, to update them to point to the newly-mapped location in the
IOMMU.
- In other systems, IOMMU/GART may be much larger than physical memory; more
modern GPUs and SoCs have this. With these, we can reserve virtual (IOMMU)
address space for each buffer up front. To userspace, the buffers always
appear "mapped". This is similar in concept to how the CPU virtual space in
userspace sticks around even when the underlying memory is paged out to disk.
In this case, pinning is performed at the same time as the small-IOMMU case
above, but in the normal/fast case, the pages are never paged out of the
IOMMU, and the pin step just increments a refcount to prevent the pages from
being evicted.
It is desirable to keep the same IOMMU address for:
a) implementing features such as
http://www.opengl.org/registry/specs/NV/shader_buffer_load.txt
(OpenGL client applications and shaders manipulate GPU vaddr pointers
directly; a GPU virtual address is assumed to be valid forever).
b) performance: scanning through the command buffers to patch up pointers can
be very expensive.
One other important note: buffer format properties may be necessary to set up
mappings (both CPU and iommu mappings). For example, both types of mappings
may need to know tiling properties of the buffer. This may be a property of
the mapping itself (consider it baked into the page table entries), not
necessarily something a different driver or userspace can program later
independently.
Some of the discussion I heard this morning tended towards being overly
simplistic and didn't seem to cover each of these cases well. Hopefully this
will help get everyone on the same page.
Thanks,
Robert
Hi all,
A bit later than what I've hoped for, but here we go [Jesse and Dave,
please correct/clarify/extend where you see fit]:
The core idea of GEM is to identify graphic buffer objects with 32bit ids.
The reason being "X runs out of open fds" (KDE easily reaches a few
thousand).
The core design principle behind GEM is that the kernel is in full control
of the allocation of these buffer objects and is free to move the around
in any way it sees fit. This is to make concurrent rendering by multiple
processes possible while userspace can still assume that it is in sole
possession of the gpu - GEM means "graphics execution manager".
Below some more details on what GEM is and does, what it does
_not_ do and how it relates to other graphic subsystems.
GEM does ...
------------
- lifecycle management. Userspace references are associated with the drm
fd and get reaped on close (in case userspace forgets about them).
- per-device global names to exchange buffers between processes (eg dri2).
These names are again 32bit ids. These global ids do not count as
userspace references and don't prevent a buffer from being reaped.
- it implements very few generic ioctls:
* flink for creating a global name for a buffer object
* open for getting a per-fd handle to a buffer object with a global name
* close for dropping a per-fd handle.
- a little bit of kernel-internal helpers to facilitate mmap (by blending
multiple buffer objects into the single drm device address space) and a
few other things.
That's it, i.e. GEM is very much meant to be as simple as possible.
Driver-specific GEM ioctls
--------------------------
The generic GEM stuff is obviously not very useful. So drivers implement
quite a bit driver-specific ioctls, like:
- buffer creation. In recent kernels there is some support to create dumb
scanout objects for KMS. But they're only really useful for boot-splashs
and unaccelerated dumb KMS drivers. Creating buffers usable for
rendering is only possible with driver specific ioctls.
- command submission. An important part is mapping abstract buffer ids to
actual gpu address (and rewriting batchbuffers with these). In the
future, with support for virtual gpu address spaces this might change.
- tiling management. The kernel needs to know this to correctly
tile/detile buffers when moving them around (e.g. evicting from vram).
- command completion signalling and gpu/cpu synchronization.
There are currently two approaches for implementing a GEM driver:
- roll-your-own, used by drm/i915 (and sometimes getting flaked for NIH).
- ttm-base: radeon & nouveau.
GEM does not ...
----------------
This still leaves out a few things that I've seen mentioned as
ideas/requirements here and elsewhere:
- cross-device buffer sharing and namespaces (see below) and
- buffer format handling and mediation between different users (except
tiling as mentioned above). The reason here is that gpus are a mess
and one of the worst parts is format handling. Better keep that out
of the kernel ...
KMS (kernel mode setting)
-------------------------
KMS is essentially just a port of the xrandr api to the kernel as an ioctl
interface:
- crtcs feed (possible multiple) outputs and get their data from a
framebuffer object. A major part of KMS is also the support for
vsynced-pageflipping of framebuffers.
- Internally there's some support infrastructure to simplify drivers (all
the drm_*_helper.c code).
- framebuffers are created from a opaque driver-specific 32bit id and a
format description. For GEM drivers these ids name GEM objects, but that
need not be: The recently merged qemu kms driver does not implement gem
and has one unique buffer object with id 0.
- as mentioned above there newly is a generic ioctl to create an object
suitable as a dumb scanout (plus some support to mmap it).
- currently KMS has no generic support for overlays (there are
driver-specific ioctls in i915 and vmgfx, though). Jesse Barnes has
posted an RFC to remedy this:
http://www.mail-archive.com/dri-devel@lists.freedesktop.org/msg10415.html
GEM and PRIME
-------------
PRIME is a proof-of-concept implementation from Dave Airlie for sharing
GEM objects between drivers/devices: Buffer sharing is done with a list of
struct page pointers. While being shared, buffers can't be moved anymore.
No further buffer description is passed along in the kernel, format/layout
mediation is to be handled in userspace.
Blog-post describing the initial design for sharing buffers between an
integrated Intel igd and a discrete ATI gpu:
http://airlied.livejournal.com/71734.html
Other code using the same framework to render on an Intel igd and display
the framebuffer on an usb-connected displayport:
http://git.kernel.org/?p=linux/kernel/git/airlied/drm-testing.git;a=shortlo…
GEM/KMS and fbdev
-----------------
There's some minimal support to emulate an fbdev with a gem/kms driver.
Resolution can't be changed and it's unaccelerated. There's been some
muttering once in a while to better integrate this with either a kms
kernel console driver or by routing fbdev resolution changes to kms.
But the main use case is to display a kernel oops, which works. For
everything else there's X (or an EGL client that understands kms).
-Daniel
--
Daniel Vetter
Mail: daniel(a)ffwll.ch
Mobile: +41 (0)79 365 57 48
Hi!
I just want to clarify some buffer object operations and terminology that seems confusing to people and that are used by most modern GPU drivers.
I think it's useful to be aware of this, going forward in the memory manager discussions.
Terminology:
Scanout buffer: Buffer that is used for continous access by a device. Needs to be permanently pinned.
Pinned buffer: A pinned buffer may not move and may not change backing pages. Allows it to be mapped to a device.
Synchronization object: An object that is either in a signaled or non-signaled state. Signaled means that the device is done with the buffer, and has flushed its caches. A synchronization object has a device-specific part that may, for example, contain flushing state.
Basic device use of a buffer:
Scanout buffers (and perhaps also capture buffers?) are typically pinned.
Other buffers that are temporarily used by a GPU and, for example, a video decoding engine or image processor are typically *not* pinned. The usage pattern for submitting any commands that affect the buffer is as follows:
1) Take a mutex that stops the buffer from being moved. This mutex could be global (stops all buffers from being moved) or per-buffer.
2) Wait on any previous synchronization objects attached to the buffer, if those sync objects would not be implicitly signaled when the device executes its work. This is where it becomes bad to have a global mutex under 1).
3) Validate the buffer. This means setting up any missing (contigous) device mappings or move to VRAM, flush cpu caches if necessary.
4) Patch up the device commands to reflect any movement of the buffer in 3). New offsets, SG-lists etc.
5) Submit the device commands.
6) Create a new synchronization object and attach it to the buffer.
7) Release the mutex taken i 1).
The buffer will not be moved until the synchronization object has signaled, and mappings set up under 3) will not be torn down until the memory manager receives a request to free up mapping resources.
I'd call this "Generation 2" device buffer management. (Intel (uses busy lists, no sync objects), Radeon, Nouveau, vmwgfx, New VIA)
"Generation 1" was using a global memory manager for pinned buffers (SiS, old VIA DRM drivers)
Generation 3 would be page based device MMUs with programmable apertures to access VRAM.
What we were discussing today is basically creating a unified gen 1 manager, with a new user-space interface.
/Thomas
DRM support for platform devices dropped last year and was drastically
improved earlier this year. Qualcomm uses it for a really weak DRM driver
that handles memory for X but does GPU and display through a different
interface. Feel free to flame me for that.. :).
https://www.codeaurora.org/gitweb/quic/la/?p=kernel/msm.git;a=blob;f=driver…
And I believe OMAP also has a solution somewhere (sorry, I couldn't find a URL).
Jordan
Hi!
Just wanted to share some thoughts about CMA, TTM and the problems with
conflicting mappings.
1) CMA seems to be a nice way to handle the problem with contigous
pages, although it seems Arnd has some concerns. It would be nice to
here about those. Some thoughts:
a) It seems fairly straightforward to interface CMA with TTM. The
benefit would be that CMA could ask TTM at any time (using a shrinker
callback) to release its contigous pages, and TTM would do so once the
GPU is finished with them (unless of course they are used with a pinned
buffer object, like a scanout buffer). CMA would need to be extended
with a small API to create / free a contigous range and to actually
populate that range with pages.
b) DRM, TTM and it seems CMA all need a range allocator. There is a
reasonable implementation in drm_mm.c, which since the original
implementation has seen a fair bit of improvement. Should we try to move
that to linux/lib ?
c) Could the CMA technique be used also to keep a pool of pages that are
unmapped from the linear kernel map? Essentially a range of HIGHMEM
pages? The benefit compared to just having a pool of HIGHMEM pages by
itself would be that the graphics subsystem would have priority over
normal system use (moving out movable contents), and could use these
pages with nonstandard caching attribute maps if needed. If this is
considered a good idea, we could perhaps consider placing the default
CMA region in HIGHMEM.
/Thomas
Hi all,
I've updated the mini-summit wiki with a couple more details:
https://wiki.linaro.org/Events/2011-05-MM
one of which is a sample use case description from Samsung. I would
encourage everyone to look at that and see if there are other use
cases they think would make more sense, or if there is clarification
or amendment of the current proposal. The discussion around this is
slated for Tuesday, so we have some time before it comes up in the
summit. As we proceed, I'll be moving sections of the agenda over
onto the discussion blueprints on launchpad (as that's how Linaro
tracks stuff), but everything will also be available on the wiki as
well as this list for those that can't or don't want to use launchpad.
Also, there are still a few components without representatives in the
summit, and while it would be nice to be able to have those in the
early parts of the sessions, I would rather flex the agenda than omit
them. Even if there is no presentation on a component at the summit,
it would still be good to have written overviews of those, so I'll ask
again for volunteers.
cheers,
Jesse
