Re: dma-api debugfs directory is not created since debugfs is not initialized

From: Greg KH <gregkh@linuxfoundation.org>
To: Robin Murphy <robin.murphy@arm.com>
Cc: Corentin Labbe <clabbe.montjoie@gmail.com>,
	rafael@kernel.org, hch@lst.de, m.szyprowski@samsung.com,
	iommu@lists.linux-foundation.org, linux-kernel@vger.kernel.org
Subject: Re: dma-api debugfs directory is not created since debugfs is not initialized
Date: Tue, 27 Apr 2021 14:41:15 +0200	[thread overview]
Message-ID: <YIgGa6yF/adg8OSN@kroah.com> (raw)
In-Reply-To: <22f268bb-9b96-1208-0292-68c9b50422a1@arm.com>

On Tue, Apr 27, 2021 at 01:32:50PM +0100, Robin Murphy wrote:
> On 2021-04-27 12:39, Greg KH wrote:
> > On Tue, Apr 27, 2021 at 01:34:27PM +0200, Corentin Labbe wrote:
> > > Hello
> > > 
> > > I try to debug some DMA problem on next-20210427, and so I have enabled CONFIG_DMA_API_DEBUG=y.
> > > But the dma-api directory does show up in debugfs, but lot of other directory exists in it.
> > 
> > Does it show up properly in 5.12?
> > 
> > > After debugging it seems due to commit 56348560d495 ("debugfs: do not attempt to create a new file before the filesystem is initalized")
> > > Reverting the commit permit to "dma-api" debugfs to be found. (but seems not the right way to fix it).
> > 
> > We have had some odd start-up ordering issues that the above commit has
> > caused to show.  Given that this commit is now in stable kernels, with
> > no report of this issue so far, I'm worried that maybe this is a dma
> > subsystem ordering issue?
> 
> Both debugfs_init() and dma_debug_init() do run at core_initcall level, and
> disassembling the vmlinux from my current working tree does indeed suggest
> that they somehow end up in the wrong relative order:
> 
> [...]
> ffff80001160d0c8 <__initcall__kmod_debug__325_918_dma_debug_init1>:
> ffff80001160d0c8:       feb0d528        .word   0xfeb0d528
> 
> [...]
> 
> ffff80001160d108 <__initcall__kmod_debugfs__357_848_debugfs_init1>:
> ffff80001160d108:       fff4326c        .word   0xfff4326c
> [...]
> 
> 
> I always had the impression that initcall ordering tended to work out
> roughly alphabetical, such that entries from fs/* might come before
> kernel/*, but I guess it's at the whims of the linker in the end :/

init call ordering happens from link ordering.

> Perhaps the easiest thing to do is split out dma_debug_fs_init() and run
> that at a later level? We do want the dma-debug infrastructure itself to be
> up as early as possible, but I think the debugfs view of its internals can
> happily wait until closer to the time that there's actually a userspace to
> be able to look at it.

That seems like a better idea here, there's no need for "special
treatment" of debugfs.

thanks,

greg k-h