Message-ID: <8a331165-4435-4c2d-70e0-20655019dc51@nbd.name>
Date:   Thu, 26 Jan 2023 10:14:57 +0100
From:   Felix Fietkau <nbd@....name>
To:     Alexander H Duyck <alexander.duyck@...il.com>,
        Ilias Apalodimas <ilias.apalodimas@...aro.org>
Cc:     netdev@...r.kernel.org, Jesper Dangaard Brouer <hawk@...nel.org>,
        "David S. Miller" <davem@...emloft.net>,
        Eric Dumazet <edumazet@...gle.com>,
        Jakub Kicinski <kuba@...nel.org>,
        Paolo Abeni <pabeni@...hat.com>,
        Lorenzo Bianconi <lorenzo@...nel.org>,
        linux-kernel@...r.kernel.org, Yunsheng Lin <linyunsheng@...wei.com>
Subject: Re: [PATCH] net: page_pool: fix refcounting issues with fragmented
 allocation

On 26.01.23 07:12, Felix Fietkau wrote:
> On 25.01.23 23:14, Alexander H Duyck wrote:
>> On Wed, 2023-01-25 at 20:40 +0100, Felix Fietkau wrote:
>>> On 25.01.23 20:10, Felix Fietkau wrote:
>>> > On 25.01.23 20:02, Alexander H Duyck wrote:
>>> > > On Wed, 2023-01-25 at 19:42 +0100, Felix Fietkau wrote:
>>> > > > On 25.01.23 19:26, Alexander H Duyck wrote:
>>> > > > > On Wed, 2023-01-25 at 18:32 +0100, Felix Fietkau wrote:
>>> > > > > > On 25.01.23 18:11, Alexander H Duyck wrote:
>>> > > > > > > On Tue, 2023-01-24 at 22:30 +0100, Felix Fietkau wrote:
>>> > > > > > > > On 24.01.23 22:10, Alexander H Duyck wrote:
>>> > > > > > > > > On Tue, 2023-01-24 at 18:22 +0100, Felix Fietkau wrote:
>>> > > > > > > > > > On 24.01.23 15:11, Ilias Apalodimas wrote:
>>> > > > > > > > > > > Hi Felix,
>>> > > > > > > > > > > 
>>> > > > > > > > > > > ++cc Alexander and Yunsheng.
>>> > > > > > > > > > > 
>>> > > > > > > > > > > Thanks for the report
>>> > > > > > > > > > > 
>>> > > > > > > > > > > On Tue, 24 Jan 2023 at 14:43, Felix Fietkau <nbd@....name> wrote:
>>> > > > > > > > > > > > 
>>> > > > > > > > > > > > While testing fragmented page_pool allocation in the mt76 driver, I was able
>>> > > > > > > > > > > > to reliably trigger page refcount underflow issues, which did not occur with
>>> > > > > > > > > > > > full-page page_pool allocation.
>>> > > > > > > > > > > > It appears to me, that handling refcounting in two separate counters
>>> > > > > > > > > > > > (page->pp_frag_count and page refcount) is racy when page refcount gets
>>> > > > > > > > > > > > incremented by code dealing with skb fragments directly, and
>>> > > > > > > > > > > > page_pool_return_skb_page is called multiple times for the same fragment.
>>> > > > > > > > > > > > 
>>> > > > > > > > > > > > Dropping page->pp_frag_count and relying entirely on the page refcount makes
>>> > > > > > > > > > > > these underflow issues and crashes go away.
>>> > > > > > > > > > > > 
>>> > > > > > > > > > > 
>>> > > > > > > > > > > This has been discussed here [1].  TL;DR changing this to page
>>> > > > > > > > > > > refcount might blow up in other colorful ways.  Can we look closer and
>>> > > > > > > > > > > figure out why the underflow happens?
>>> > > > > > > > > > I don't see how the approch taken in my patch would blow up. From what I 
>>> > > > > > > > > > can tell, it should be fairly close to how refcount is handled in 
>>> > > > > > > > > > page_frag_alloc. The main improvement it adds is to prevent it from 
>>> > > > > > > > > > blowing up if pool-allocated fragments get shared across multiple skbs 
>>> > > > > > > > > > with corresponding get_page and page_pool_return_skb_page calls.
>>> > > > > > > > > > 
>>> > > > > > > > > > - Felix
>>> > > > > > > > > > 
>>> > > > > > > > > 
>>> > > > > > > > > Do you have the patch available to review as an RFC? From what I am
>>> > > > > > > > > seeing it looks like you are underrunning on the pp_frag_count itself.
>>> > > > > > > > > I would suspect the issue to be something like starting with a bad
>>> > > > > > > > > count in terms of the total number of references, or deducing the wrong
>>> > > > > > > > > amount when you finally free the page assuming you are tracking your
>>> > > > > > > > > frag count using a non-atomic value in the driver.
>>> > > > > > > > The driver patches for page pool are here:
>>> > > > > > > > https://patchwork.kernel.org/project/linux-wireless/patch/64abb23f4867c075c19d704beaae5a0a2f8e8821.1673963374.git.lorenzo@kernel.org/
>>> > > > > > > > https://patchwork.kernel.org/project/linux-wireless/patch/68081e02cbe2afa2d35c8aa93194f0adddbd0f05.1673963374.git.lorenzo@kernel.org/
>>> > > > > > > > 
>>> > > > > > > > They are also applied in my mt76 tree at:
>>> > > > > > > > https://github.com/nbd168/wireless
>>> > > > > > > > 
>>> > > > > > > > - Felix
>>> > > > > > > 
>>> > > > > > > So one thing I am thinking is that we may be seeing an issue where we
>>> > > > > > > are somehow getting a mix of frag and non-frag based page pool pages.
>>> > > > > > > That is the only case I can think of where we might be underflowing
>>> > > > > > > negative. If you could add some additional debug info on the underflow
>>> > > > > > > WARN_ON case in page_pool_defrag_page that might be useful.
>>> > > > > > > Specifically I would be curious what the actual return value is. I'm
>>> > > > > > > assuming we are only hitting negative 1, but I would want to verify we
>>> > > > > > > aren't seeing something else.
>>> > > > > > I'll try to run some more tests soon. However, I think I found the piece 
>>> > > > > > of code that is incompatible with using pp_frag_count.
>>> > > > > > When receiving an A-MSDU packet (multiple MSDUs within a single 802.11 
>>> > > > > > packet), and it is not split by the hardware, a cfg80211 function 
>>> > > > > > extracts the individual MSDUs into separate skbs. In that case, a 
>>> > > > > > fragment can be shared across multiple skbs, and get_page is used to 
>>> > > > > > increase the refcount.
>>> > > > > > You can find this in net/wireless/util.c: ieee80211_amsdu_to_8023s (and 
>>> > > > > > its helper functions).
>>> > > > > 
>>> > > > > I'm not sure if it is problematic or not. Basically it is trading off
>>> > > > > by copying over the frags, calling get_page on each frag, and then
>>> > > > > using dev_kfree_skb to disassemble and release the pp_frag references.
>>> > > > > There should be other paths in the kernel that are doing something
>>> > > > > similar.
>>> > > > > 
>>> > > > > > This code also has a bug where it doesn't set pp_recycle on the newly 
>>> > > > > > allocated skb if the previous one has it, but that's a separate matter 
>>> > > > > > and fixing it doesn't make the crash go away.
>>> > > > > 
>>> > > > > Adding the recycle would cause this bug. So one thing we might be
>>> > > > > seeing is something like that triggering this error. Specifically if
>>> > > > > the page is taken via get_page when assembling the new skb then we
>>> > > > > cannot set the recycle flag in the new skb otherwise it will result in
>>> > > > > the reference undercount we are seeing. What we are doing is shifting
>>> > > > > the references away from the pp_frag_count to the page reference count
>>> > > > > in this case. If we set the pp_recycle flag then it would cause us to
>>> > > > > decrement pp_frag_count instead of the page reference count resulting
>>> > > > > in the underrun.
>>> > > > Couldn't leaving out the pp_recycle flag potentially lead to a case 
>>> > > > where the last user of the page drops it via page_frag_free instead of 
>>> > > > page_pool_return_skb_page? Is that valid?
>>> > > 
>>> > > No. What will happen is that when the pp_frag_count is exhausted the
>>> > > page will be unmapped and evicted from the page pool. When the page is
>>> > > then finally freed it will end up going back to the page allocator
>>> > > instead of page pool.
>>> > > 
>>> > > Basically the idea is that until pp_frag_count reaches 0 there will be
>>> > > at least 1 page reference held.
>>> > > 
>>> > > > > > Is there any way I can make that part of the code work with the current 
>>> > > > > > page pool frag implementation?
>>> > > > > 
>>> > > > > The current code should work. Basically as long as the references are
>>> > > > > taken w/ get_page and skb->pp_recycle is not set then we shouldn't run
>>> > > > > into this issue because the pp_frag_count will be dropped when the
>>> > > > > original skb is freed and the page reference count will be decremented
>>> > > > > when the new one is freed.
>>> > > > > 
>>> > > > > For page pool page fragments the main thing to keep in mind is that if
>>> > > > > pp_recycle is set it will update the pp_frag_count and if it is not
>>> > > > > then it will just decrement the page reference count.
>>> > > > What takes care of DMA unmap and other cleanup if the last reference to 
>>> > > > the page is dropped via page_frag_free?
>>> > > > 
>>> > > > - Felix
>>> > > 
>>> > > When the page is freed on the skb w/ pp_recycle set it will unmap the
>>> > > page and evict it from the page pool. Basically in these cases the page
>>> > > goes from the page pool back to the page allocator.
>>> > > 
>>> > > The general idea with this is that if we are using fragments that there
>>> > > will be enough of them floating around that if one or two frags have a
>>> > > temporeary detour through a non-recycling path that hopefully by the
>>> > > time the last fragment is freed the other instances holding the
>>> > > additional page reference will have let them go. If not then the page
>>> > > will go back to the page allocator and it will have to be replaced in
>>> > > the page pool.
>>> > Thanks for the explanation, it makes sense to me now. Unfortunately it
>>> > also means that I have no idea what could cause this issue. I will
>>> > finish my mt76 patch rework which gets rid of the pp vs non-pp
>>> > allocation mix and re-run my tests to provide updated traces.
>>> Here's the updated mt76 page pool support commit:
>>> https://github.com/nbd168/wireless/commit/923cdab6d4c92a0acb3536b3b0cc4af9fee7c808
>> 
>> Yeah, so I don't see anything wrong with the patch in terms of page
>> pool.
>> 
>>> And here is the trace that I'm getting with 6.1:
>>> https://nbd.name/p/a16957f2
>>> 
>>> If you have any debug patch you'd like me to test, please let me know.
>>> 
>>> - Felix
>> 
>> So looking at the traces I am assuming what we are seeing is the
>> deferred freeing from the TCP Rx path since I don't see a driver
>> anywhere between net_rx_action and napi_consume skb. So it seems like
>> the packets are likely making it all the way up the network stack.
>> 
>> Is this the first wireless driver to add support for page pool? I'm
>> thinking we must be seeing something in the wireless path that is
>> causing an issue such as the function you called out earlier but I
>> can't see anything obvious.
> Yes, it's the first driver with page pool support.
> 
>> One thing we need to be on the lookout for is cloned skbs. When an skb
>> is cloned the pp_recycle gets copied over. In that case the reference
>> is moved over to the skb dataref count. What comes to mind is something
>> like commit 1effe8ca4e34c ("skbuff: fix coalescing for page_pool
>> fragment recycling").
> I suspect that the crash might be related to a bad interaction between
> the page reuse in A-MSDU rx + skb coalescing on TCP rx.
> If I change the A-MSDU code to copy data instead of reusing fragments,
> it doesn't crash anymore.
> I believe the issue must be specific to that codepath, since most
> received and processed packets are either not A-MSDU or A-MSDU decap has
> already been performed by the hardware.
> If I change my test to use 3 client mode interfaces instead of 4, the
> hardware is able to offload all A-MSDU rx processing and I don't see any
> crashes anymore.
> 
> Could you please take another look at ieee80211_amsdu_to_8023s to see if
> there's anything in there that could cause these issues?
Here are clues from a few more tests that I ran:
- preventing the reuse of the last skb in ieee80211_amsdu_to_8023s does 
not prevent the crashes, so the issue is indeed related to taking page 
references and putting the pages in skb fragments.
- if I return false in skb_try_coalesce, it still crashes:
https://nbd.name/p/18cac078

- Felix

Powered by blists - more mailing lists