From: Thomas Zimmermann <tzimmermann@suse.de>
To: "Christian König" <christian.koenig@amd.com>,
"Daniel Vetter" <daniel@ffwll.ch>
Cc: airlied@redhat.com, dri-devel@lists.freedesktop.org, hdegoede@redhat.com
Subject: Re: [PATCH 1/8] drm/gem: Write down some rules for vmap usage
Date: Tue, 1 Dec 2020 10:30:20 +0100 [thread overview]
Message-ID: <cdf02ecf-6b64-2bf5-0579-26e0dbb806fe@suse.de> (raw)
In-Reply-To: <70e396bf-e8a3-4871-7ad1-4a90298d8461@amd.com>
[-- Attachment #1.1.1: Type: text/plain, Size: 5958 bytes --]
Hi
Am 01.12.20 um 10:13 schrieb Christian König:
> Am 01.12.20 um 09:32 schrieb Thomas Zimmermann:
>> Hi
>>
>> Am 30.11.20 um 16:33 schrieb Christian König:
>>> Am 30.11.20 um 16:30 schrieb Daniel Vetter:
>>>> On Mon, Nov 30, 2020 at 01:04:26PM +0100, Thomas Zimmermann wrote:
>>>>> Mapping a GEM object's buffer into kernel address space prevents the
>>>>> buffer from being evicted from VRAM, which in turn may result in
>>>>> out-of-memory errors. It's therefore required to only vmap GEM BOs for
>>>>> short periods of time; unless the GEM implementation provides
>>>>> additional
>>>>> guarantees.
>>>>>
>>>>> Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
>>>>> ---
>>>>> drivers/gpu/drm/drm_prime.c | 6 ++++++
>>>>> include/drm/drm_gem.h | 16 ++++++++++++++++
>>>>> 2 files changed, 22 insertions(+)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/drm_prime.c b/drivers/gpu/drm/drm_prime.c
>>>>> index 7db55fce35d8..9c9ece9833e0 100644
>>>>> --- a/drivers/gpu/drm/drm_prime.c
>>>>> +++ b/drivers/gpu/drm/drm_prime.c
>>>>> @@ -669,6 +669,12 @@ EXPORT_SYMBOL(drm_gem_unmap_dma_buf);
>>>>> * callback. Calls into &drm_gem_object_funcs.vmap for device
>>>>> specific handling.
>>>>> * The kernel virtual address is returned in map.
>>>>> *
>>>>> + * To prevent the GEM object from being relocated, callers must
>>>>> hold the GEM
>>>>> + * object's reservation lock from when calling this function until
>>>>> releasing the
>>>>> + * mapping. Holding onto a mapping and the associated reservation
>>>>> lock for an
>>>>> + * unbound time may result in out-of-memory errors. Calls to
>>>>> drm_gem_dmabuf_vmap()
>>>>> + * should therefore be accompanied by a call to
>>>>> drm_gem_dmabuf_vunmap().
>>>>> + *
>>>>> * Returns 0 on success or a negative errno code otherwise.
>>>> This is a dma-buf hook, which means just documenting the rules you'd
>>>> like
>>>> to have here isn't enough. We need to roll this out at the dma-buf
>>>> level,
>>>> and enforce it.
>>>>
>>>> Enforce it = assert_lock_held
>>>>
>>>> Roll out = review everyone. Because this goes through dma-buf it'll
>>>> come
>>>> back through shmem helpers (and other helpers and other subsystems)
>>>> back
>>>> to any driver using vmap for gpu buffers. This includes the media
>>>> subsystem, and the media subsystem definitely doesn't cope with just
>>>> temporarily mapping buffers. So there we need to pin them, which I
>>>> think
>>>> means we'll need 2 version of dma_buf_vmap - one that's temporary and
>>>> requires we hold dma_resv lock, the other requires that the buffer is
>>>> pinned.
>>>
>>> OR start to proper use the dma_buf_pin/dma_buf_unpin functions which
>>> I added to cover this use case as well.
>>
>> While I generally agree, here are some thoughts:
>>
>> I found all generic pin functions useless, because they don't allow
>> for specifying where to pin. With fbdev emulation, this means that
>> console buffers might never make it to VRAM for scanout. If anything,
>> the policy should be that pin always pins in HW-accessible memory.
>
> Yes, completely agree. The major missing part here are some kind of
> usage flags what we want to do with the buffer.
Not sure, but wasn't that not wanted by someone? I don't really remember.
Given Daniel's reply about pin, it really feels like we have
contradicting policies for this interface.
>
>>
>> Pin has quite a bit of overhead (more locking, buffer movement), so it
>> should be the second choice after regular vmap. To make both work
>> together, pin probably relies on holding the reservation lock internally.
>
> There is a dma_resv_assert_held() at the beginning of those two functions.
>
>>
>> Therefore I think we still would want some additional helpers, such as:
>>
>> pin_unlocked(), which acquires the resv lock, calls regular pin and
>> then drops the resv lock. Same for unpin_unlocked()
>>
>> vmap_pinned(), which enforces that the buffer has been pinned and
>> then calls regalar vmap. Same for vunmap_pinned()
>
> I would rather open code that in each driver, hiding the locking logic
> in some midlayer is usually not a good idea.
These helpers are less about hiding logic and more about making drivers
do the right thing. The idea behind pin_unlocked() is that it drops the
reservation lock immediately before returning. I suspect that some
driver might not do that with an open-coded version. And vmap_pinned()
does check for the BO to be pinned. Regular vmap would assert for the
reservation lock instead.
Best regards
Thomas
>
> Regards,
> Christian.
>
>>
>> A typical pattern with these functions would look like this.
>>
>> drm_gem_object bo;
>> dma_buf_map map;
>>
>> init() {
>> pin_unlocked(bo);
>> vmap_pinned(bo, map);
>> }
>>
>> worker() {
>> begin_cpu_access()
>> // access bo via map
>> end_cpu_access()
>> }
>>
>> fini() {
>> vunmap_pinned(bo, map);
>> unpin_unlocked(bo);
>> }
>>
>> init()
>> while (...) {
>> worker()
>> }
>> fini()
>>
>> Is that reasonable for media drivers?
>>
>> Best regards
>> Thomas
>>
>>
>>>
>>> Cheers,
>>> Christian.
>>>
>>>>
>>>> That's what I meant with that this approach here is very sprawling :-/
>>>> -Daniel
>
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/dri-devel
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Felix Imendörffer
[-- Attachment #1.2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 840 bytes --]
[-- Attachment #2: Type: text/plain, Size: 160 bytes --]
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel
next prev parent reply other threads:[~2020-12-01 9:30 UTC|newest]
Thread overview: 31+ messages / expand[flat|nested] mbox.gz Atom feed top
2020-11-30 12:04 [PATCH 0/8] drm/vram-helper: Lock GEM BOs while they are mapped Thomas Zimmermann
2020-11-30 12:04 ` [PATCH 1/8] drm/gem: Write down some rules for vmap usage Thomas Zimmermann
2020-11-30 15:30 ` Daniel Vetter
2020-11-30 15:33 ` Christian König
2020-12-01 8:32 ` Thomas Zimmermann
2020-12-01 9:10 ` Daniel Vetter
2020-12-01 9:40 ` Thomas Zimmermann
2020-12-01 10:00 ` Daniel Vetter
2020-12-01 10:27 ` Thomas Zimmermann
2020-12-01 10:34 ` Christian König
2020-12-01 11:30 ` Thomas Zimmermann
2020-12-01 12:14 ` Christian König
2020-12-01 12:33 ` Thomas Zimmermann
2020-12-01 12:38 ` Christian König
2020-12-01 12:51 ` Thomas Zimmermann
2020-12-01 12:53 ` Thomas Zimmermann
2020-12-01 13:05 ` Christian König
2020-12-01 16:54 ` Daniel Vetter
2020-12-01 12:05 ` Thomas Zimmermann
2020-12-01 9:13 ` Christian König
2020-12-01 9:30 ` Thomas Zimmermann [this message]
2020-12-01 8:15 ` Thomas Zimmermann
2020-11-30 12:04 ` [PATCH 2/8] drm/ast: Only map cursor BOs during updates Thomas Zimmermann
2020-11-30 12:04 ` [PATCH 3/8] drm/vram-helper: Provide drm_gem_vram_vmap_unlocked() Thomas Zimmermann
2020-11-30 12:04 ` [PATCH 4/8] drm/ast: Use drm_gem_vram_vmap_unlocked() in ast_cursor_show() Thomas Zimmermann
2020-11-30 12:04 ` [PATCH 5/8] drm/vboxvideo: Use drm_gem_vram_vmap_unlocked() in cursor update Thomas Zimmermann
2020-11-30 12:04 ` [PATCH 6/8] drm/vram-helper: Remove pinning and locking from drm_gem_vram_vmap() Thomas Zimmermann
2020-11-30 12:04 ` [PATCH 7/8] drm/vram-helper: Remove vmap reference counting Thomas Zimmermann
2020-11-30 12:04 ` [PATCH 8/8] drm/vram-helper: Simplify vmap implementation Thomas Zimmermann
2020-11-30 12:27 ` [PATCH 0/8] drm/vram-helper: Lock GEM BOs while they are mapped Christian König
2020-11-30 12:45 ` Thomas Zimmermann
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=cdf02ecf-6b64-2bf5-0579-26e0dbb806fe@suse.de \
--to=tzimmermann@suse.de \
--cc=airlied@redhat.com \
--cc=christian.koenig@amd.com \
--cc=daniel@ffwll.ch \
--cc=dri-devel@lists.freedesktop.org \
--cc=hdegoede@redhat.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).