[PATCH] dmr/amdgpu: Fix wrongly unref of BO
Xie, AlexBin
AlexBin.Xie at amd.com
Thu Apr 20 21:56:00 UTC 2017
Hi Christian,
I read amdgpu_ttm_io_mem_reserve() and amdgpu_ttm_io_mem_free() and relevant codes from amdgpu_vram_scratch_init and amdgpu_vram_scratch_fini.
No. For the current source code, I think the premap and no-op is not working.
I add printk to prove. You can see bo_kmap_type is an invalid value. ioremap_wc is really called to map the IO range into vmalloc space.
...
Apr 20 16:31:18 axie-System-Product-Name kernel: [ 106.759623] entering amdgpu_vram_scratch_init.
Apr 20 16:31:18 axie-System-Product-Name kernel: [ 106.759631] scratch ioremap_wc
Apr 20 16:31:18 axie-System-Product-Name kernel: [ 106.759631] bo_kmap_type = 129
Apr 20 16:31:18 axie-System-Product-Name kernel: [ 106.759632] bus address = (null)
Apr 20 16:31:18 axie-System-Product-Name kernel: [ 106.759632] is_iomem = 1
Apr 20 16:31:18 axie-System-Product-Name kernel: [ 106.759633] leaving amdgpu_vram_scratch_init.
...
I don't have log on rmmod AMDGPU yet. There are quite some settings to make that happen in my computer.
I think they are symemtric. Both mapping and unmapping are not no-op.
Here is the printk patch for your reference.
From 25f95239c23225008e4b59f30b9b5363fb321f94 Mon Sep 17 00:00:00 2001
From: Alex Xie <AlexBin.Xie at amd.com>
Date: Thu, 20 Apr 2017 17:48:49 -0400
Subject: [PATCH] A hack to trace premap and noop.
Change-Id: I61fbbdbd82f62433e229b2e7e97be7a780ea8afa
Signed-off-by: Alex Xie <AlexBin.Xie at amd.com>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 22 ++++++++++++++++++++++
drivers/gpu/drm/ttm/ttm_bo.c | 1 +
drivers/gpu/drm/ttm/ttm_bo_util.c | 29 ++++++++++++++++++++++++++---
include/drm/ttm/ttm_bo_api.h | 1 +
4 files changed, 50 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
index fbb4afb..a537ea1 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
@@ -313,6 +313,7 @@ static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
static int amdgpu_vram_scratch_init(struct amdgpu_device *adev)
{
int r;
+ printk("entering amdgpu_vram_scratch_init.");
if (adev->vram_scratch.robj == NULL) {
r = amdgpu_bo_create(adev, AMDGPU_GPU_PAGE_SIZE,
@@ -340,16 +341,36 @@ static int amdgpu_vram_scratch_init(struct amdgpu_device *adev)
amdgpu_bo_unpin(adev->vram_scratch.robj);
amdgpu_bo_unreserve(adev->vram_scratch.robj);
+ /* Would like a printk to see if map / unmap is noop*/
+ adev->vram_scratch.robj->tbo.mem.bus.printk = true;
+
+ if (adev->vram_scratch.robj->kmap.bo_kmap_type == ttm_bo_map_premapped)
+ printk("amdgpu_vram_scratch premapped!");
+
+ printk("bo_kmap_type = %d", adev->vram_scratch.robj->kmap.bo_kmap_type);
+ printk("bus address = %p", adev->vram_scratch.robj->tbo.mem.bus.addr);
+ printk("is_iomem = %d", adev->vram_scratch.robj->tbo.mem.bus.is_iomem);
+ printk("leaving amdgpu_vram_scratch_init.");
+
return r;
}
static void amdgpu_vram_scratch_fini(struct amdgpu_device *adev)
{
int r;
+ printk("entering amdgpu_vram_scratch_fini.");
if (adev->vram_scratch.robj == NULL) {
return;
}
+
+ if (adev->vram_scratch.robj->kmap.bo_kmap_type == ttm_bo_map_premapped)
+ printk("amdgpu_vram_scratch premapped!");
+
+ printk("bo_kmap_type = %d", adev->vram_scratch.robj->kmap.bo_kmap_type);
+ printk("bus address = %p", adev->vram_scratch.robj->tbo.mem.bus.addr);
+ printk("is_iomem = %d", adev->vram_scratch.robj->tbo.mem.bus.is_iomem);
+
r = amdgpu_bo_reserve(adev->vram_scratch.robj, false);
if (likely(r == 0)) {
amdgpu_bo_kunmap(adev->vram_scratch.robj);
@@ -357,6 +378,7 @@ static void amdgpu_vram_scratch_fini(struct amdgpu_device *adev)
amdgpu_bo_unreserve(adev->vram_scratch.robj);
}
amdgpu_bo_unref(&adev->vram_scratch.robj);
+ printk("leaving amdgpu_vram_scratch_fini.");
}
/**
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index 989b98b..9b05d54 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -1178,6 +1178,7 @@ int ttm_bo_init_reserved(struct ttm_bo_device *bdev,
bo->mem.page_alignment = page_alignment;
bo->mem.bus.io_reserved_vm = false;
bo->mem.bus.io_reserved_count = 0;
+ bo->mem.bus.printk = false;
bo->moving = NULL;
bo->mem.placement = (TTM_PL_FLAG_SYSTEM | TTM_PL_FLAG_CACHED);
bo->persistent_swap_storage = persistent_swap_storage;
diff --git a/drivers/gpu/drm/ttm/ttm_bo_util.c b/drivers/gpu/drm/ttm/ttm_bo_util.c
index bf6e216..9d06952 100644
--- a/drivers/gpu/drm/ttm/ttm_bo_util.c
+++ b/drivers/gpu/drm/ttm/ttm_bo_util.c
@@ -526,14 +526,24 @@ static int ttm_bo_ioremap(struct ttm_buffer_object *bo,
if (bo->mem.bus.addr) {
map->bo_kmap_type = ttm_bo_map_premapped;
map->virtual = (void *)(((u8 *)bo->mem.bus.addr) + offset);
+ if (bo->mem.bus.printk)
+ printk ("scratch premapping");
+
} else {
map->bo_kmap_type = ttm_bo_map_iomap;
- if (mem->placement & TTM_PL_FLAG_WC)
+ if (mem->placement & TTM_PL_FLAG_WC) {
map->virtual = ioremap_wc(bo->mem.bus.base + bo->mem.bus.offset + offset,
size);
- else
+ if (bo->mem.bus.printk)
+ printk ("scratch ioremap_wc");
+
+ }
+ else {
map->virtual = ioremap_nocache(bo->mem.bus.base + bo->mem.bus.offset + offset,
size);
+ if (bo->mem.bus.printk)
+ printk ("scratch ioremap_nocache");
+ }
}
return (!map->virtual) ? -ENOMEM : 0;
}
@@ -618,21 +628,34 @@ void ttm_bo_kunmap(struct ttm_bo_kmap_obj *map)
struct ttm_mem_type_manager *man =
&bo->bdev->man[bo->mem.mem_type];
- if (!map->virtual)
+ if (!map->virtual) {
+ if (bo->mem.bus.printk)
+ printk ("scratch unmap return earlier");
return;
+ }
switch (map->bo_kmap_type) {
case ttm_bo_map_iomap:
+ if (bo->mem.bus.printk)
+ printk ("scratch iounmap");
iounmap(map->virtual);
break;
case ttm_bo_map_vmap:
+ if (bo->mem.bus.printk)
+ printk ("scratch vunmap");
vunmap(map->virtual);
break;
case ttm_bo_map_kmap:
+ if (bo->mem.bus.printk)
+ printk ("scratch kunmap");
kunmap(map->page);
break;
case ttm_bo_map_premapped:
+ if (bo->mem.bus.printk)
+ printk ("scratch unmap ttm_bo_map_premapped");
break;
default:
+ if (bo->mem.bus.printk)
+ printk ("scratch unmap bug");
BUG();
}
(void) ttm_mem_io_lock(man, false);
diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h
index 2d0f63e..f74a554 100644
--- a/include/drm/ttm/ttm_bo_api.h
+++ b/include/drm/ttm/ttm_bo_api.h
@@ -70,6 +70,7 @@ struct ttm_bus_placement {
bool is_iomem;
bool io_reserved_vm;
uint64_t io_reserved_count;
+ bool printk;
};
--
2.7.4
Thanks,
Alex Bin
________________________________
From: Christian König <deathsimple at vodafone.de>
Sent: Thursday, April 20, 2017 4:43 AM
To: Xie, AlexBin; Zhou, David(ChunMing); amd-gfx at lists.freedesktop.org
Subject: Re: [PATCH] dmr/amdgpu: Fix wrongly unref of BO
Hi AlexBin,
Missing kunmap mapping in vmalloc will make kernel master page table incorrect.
That's what I tried to explain yesterday, but unfortunately didn't had time to do so. There is not corruption of the kernel master page table in this case!
The call of ttm_bo_kunmap is completely optional, take a look at amdgpu_ttm_io_mem_reserve() and amdgpu_ttm_io_mem_free().
The aperture is kept mapped into the page tables for the whole time the driver is loaded. So this is a complete no-op and only done for consistency.
It is good that you agree that there is no real world bad example caused by my patch. I will not discuss whether it is an improvement or not now to save time for both of us.
Great at least we can now agree to completely drop this patch.
Thanks,
Christian.
Am 19.04.2017 um 21:30 schrieb Xie, AlexBin:
Hi Christian,
Missing kunmap mapping in vmalloc will make kernel master page table incorrect. I would not call such issue as completely harmless. Please note that AMD graphic driver can run in 32 bit system. In 32 bit system, vmalloc address space is much smaller than size of most GPU VRAM.
amdgpu_bo_free_kernel has same issue as amdgpu_vram_scratch_fini. 1. It calls amdgpu_bo_reserve interruptible too. 2. It misses kunmap when amdgpu_bo_reserve returns error too. As result, kernel master page table can become incorrect, or as you call it "completely harmless vmalloc space leaking".
Because amdgpu_bo_free_kernel is used in more places, such as psp command submission, there will be bigger chance to have other usage where signal is not blocked. This will become a real bug.
I am thinking that we may fix the issue completely when TTM releases BO. But that is a bigger change.
It is good that you agree that there is no real world bad example caused by my patch. I will not discuss whether it is an improvement or not now to save time for both of us.
Thanks,
Alex Bin Xie
________________________________
From: Christian König <deathsimple at vodafone.de><mailto:deathsimple at vodafone.de>
Sent: Wednesday, April 19, 2017 7:50 AM
To: Xie, AlexBin; Zhou, David(ChunMing); amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org>
Subject: Re: [PATCH] dmr/amdgpu: Fix wrongly unref of BO
Without correctly kunmap, page table is corrupted. Page entries point to wrong memory locations. You might call it completely harmless. But I think this is a severe bug. Leaking memory is better than a corrupted page table. Think security.
We are talking about the page tables for the vmalloc area in the kernel here, so no security problem. Leaking memory is much more problematic.
Would you provide any document and reference by saying" It is impossible to receive a signal during module load/unload"? For example, if the unload stuck in a lock, can CTRL+C stop the unload?
No, CTRL+C doesn't abort module load/unload. There have been patches to changes this a while ago, but IIRC it broke a whole bunch of driver relying on this.
What about there is some other return error? What about in future somebody improve amdgpu_bo_reserve to return other errors, then function amdgpu_vram_scratch_fini becomes buggy?
Yes, that is indeed an issue. For example -EDEADLK is possible as well. That's why I said we should use amdgpu_bo_free_kernel() instead.
While I am thinking whether there is a better way for the current situation, would you give a real world example that my patch really not working? Then we can address it.
I don't think there is because the driver can't receive a signal during load/unload, but the problem is rather that the patch doesn't improve the situation at all.
Regards,
Christian.
Am 19.04.2017 um 13:37 schrieb Xie, AlexBin:
Hi Christian,
Without correctly kunmap, page table is corrupted. Page entries point to wrong memory locations. You might call it completely harmless. But I think this is a severe bug. Leaking memory is better than a corrupted page table. Think security.
Would you provide any document and reference by saying" It is impossible to receive a signal during module load/unload"? For example, if the unload stuck in a lock, can CTRL+C stop the unload?
If "It is impossible to receive a signal during module load/unload", interruptible waiting is fine too, because function amdgpu_bo_reserve will return successfully.
What about there is some other return error? What about in future somebody improve amdgpu_bo_reserve to return other errors, then function amdgpu_vram_scratch_fini becomes buggy?
While I am thinking whether there is a better way for the current situation, would you give a real world example that my patch really not working? Then we can address it.
Thanks,
Alex Bin
________________________________
From: Christian König <deathsimple at vodafone.de><mailto:deathsimple at vodafone.de>
Sent: Wednesday, April 19, 2017 2:35 AM
To: Xie, AlexBin; Zhou, David(ChunMing); amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org>
Subject: Re: [PATCH] dmr/amdgpu: Fix wrongly unref of BO
Hi AlexBin,
the answer is ttm_bo_kunmap isn't called at all and yes in the case of an iomap we leak the address space reserved.
But that is completely harmless on a 64bit system compared to leaking the memory backing the address space.
Using amdgpu_bo_free_kernel() instead of openly coding it here is probably a good idea.
Additional to that it's probably a good idea to set the no_intr flag when reserving kernel BOs. It is impossible to receive a signal during module load/unload, but it's probably better to document that in the code as well.
Regards,
Christian.
Am 18.04.2017 um 20:54 schrieb Xie, AlexBin:
Hi Christian,
Have you found how/where/when? When you said "mapping will just be released a bit later on", you must know the answer.
It is difficult to prove something does not exist. Anyway, I will give it a try to prove such "later on" does not exist.
Function ttm_bo_kunmap is the only function to unmap. To prove this, search ttm_bo_map_iomap, only ttm_bo_kunmap use this enum to correctly kunmap.
Function ttm_bo_kunmap is not called by ttm itself. This is a hint that all TTM delay delete mechanism or unref mechanism will NOT kunmap BO later on.
Function ttm_bo_kunmap is called by AMDGPU function amdgpu_bo_kunmap and amdgpu_gem_prime_vunmap.
Search AMDGPU for amdgpu_bo_kunmap. All matches do not kunmap for scratch VRAM BO. amdgpu_bo_free_kernel is a suspect but the answer is still NO.
So all possibilities are searched. Did I miss anything?
Thanks,
Alex Bin Xie
________________________________
From: Xie, AlexBin
Sent: Tuesday, April 18, 2017 2:04:33 PM
To: Christian König; Zhou, David(ChunMing); amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org>
Subject: Re: [PATCH] dmr/amdgpu: Fix wrongly unref of BO
Hi Christian,
Would you point out where/when will kunmap happen for this BO when release? It must be somewhere in some function calls.
I checked before I asked for review. But I did not see such obvious kunmap function call.
If so, there should be a comment in function amdgpu_vram_scratch_fini to avoid future confusion.
Thanks,
Alex Bin Xie
________________________________
From: Christian König <deathsimple at vodafone.de><mailto:deathsimple at vodafone.de>
Sent: Tuesday, April 18, 2017 1:46 PM
To: Xie, AlexBin; Zhou, David(ChunMing); amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org>
Subject: Re: [PATCH] dmr/amdgpu: Fix wrongly unref of BO
Hi AlexBin,
No, David is right. This is a very common coding pattern in the kernel module.
Freeing up a BO while there still exists a kernel mapping is perfectly ok, the mapping will just be released a bit later on.
So this code is actually perfectly ok and just an optimization, but your patch breaks it and creates a memory leak.
Regards,
Christian.
Am 18.04.2017 um 17:17 schrieb Xie, AlexBin:
Hi David,
When amdgpu_bo_reserve return errors, we cannot release the BO. This is not a memory leak. General speaking, memory leak is unnoticed and unintentional.
The caller of function amdgpu_vram_scratch_fini ignores the return error value...
The "memory leak" is not caused by my patch. It is caused because reserving BO fails.
This patch only aim to make function amdgpu_vram_scratch_fini behave correctly.
To follow up, we can add a warning message when amdgpu_bo_reserve error happens in a different patch.
If function call amdgpu_bo_reserve is changed to uninterruptible, this changes driver behaviour. Without a substantial issue, I would be cautious for such a change.
Thanks,
Alex Bin Xie
________________________________
From: Zhou, David(ChunMing)
Sent: Monday, April 17, 2017 10:38 PM
To: Xie, AlexBin; amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org>
Subject: Re: [PATCH] dmr/amdgpu: Fix wrongly unref of BO
On 2017年04月17日 22:54, Xie, AlexBin wrote:
Hi David,
Thanks for the comments. However, please have look at amdgpu_bo_reserve definition.
static inline int amdgpu_bo_reserve(struct amdgpu_bo *bo, bool no_intr)
Ah, this is a wired wrapper for ttm_bo_reserve.
When we call this function like the following:
r = amdgpu_bo_reserve(adev->vram_scratch.robj, false);
The false means interruptible.
On the other hand, when amdgpu_bo_reserve function return error, why do we unref BO without kunmap and unpin the BO? This is wrong implementation when amdgpu_bo_reserve return any error.
Yeah, I see your mean, it's in driver un-loading, How about changing to no interruptible? Your patch will make a memleak if bo_reserve fails, but it seems not matter. I have no strong preference.
Regards,
David Zhou
Thanks,
Alex Bin Xie
________________________________
From: Zhou, David(ChunMing)
Sent: Friday, April 14, 2017 12:00 AM
To: Xie, AlexBin; amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org>
Subject: Re: [PATCH] dmr/amdgpu: Fix wrongly unref of BO
On 2017年04月14日 05:34, Alex Xie wrote:
> According to comment of amdgpu_bo_reserve, amdgpu_bo_reserve
> can return with -ERESTARTSYS. When this function was interrupted
> by a signal, BO should not be unref. Otherwise the BO might be
> released while is kmapped and pinned, or BO MIGHT be deref
> multiple times, etc.
r = amdgpu_bo_reserve(adev->vram_scratch.robj, false);
we have specified interruptible to false, so -ERESTARTSYS isn't possible
here.
Thanks,
David Zhou
>
> Change-Id: If76071a768950a0d3ad9d5da7fcae04881807621
> Signed-off-by: Alex Xie <AlexBin.Xie at amd.com><mailto:AlexBin.Xie at amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 53996e3..1dcc2d1 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -355,8 +355,8 @@ static void amdgpu_vram_scratch_fini(struct amdgpu_device *adev)
> amdgpu_bo_kunmap(adev->vram_scratch.robj);
> amdgpu_bo_unpin(adev->vram_scratch.robj);
> amdgpu_bo_unreserve(adev->vram_scratch.robj);
> + amdgpu_bo_unref(&adev->vram_scratch.robj);
> }
> - amdgpu_bo_unref(&adev->vram_scratch.robj);
> }
>
> /**
_______________________________________________
amd-gfx mailing list
amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org>
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org>
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx at lists.freedesktop.org<mailto:amd-gfx at lists.freedesktop.org>
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20170420/6bb9cf81/attachment-0001.html>
More information about the amd-gfx
mailing list