Re: [PATCH v4 4/5] kdump: wait for DMA to finish when using CMA

[Jiri Bohac]

On Tue, Jun 03, 2025 at 03:15:03PM +0200, David Hildenbrand wrote:
> On 30.05.25 22:29, Jiri Bohac wrote:
> > When re-using the CMA area for kdump there is a risk of pending DMA into
> > pinned user pages in the CMA area.
> > 
> > Pages that are pinned long-term are migrated away from CMA, so these are
> > not a concern. Pages pinned without FOLL_LONGTERM remain in the CMA and may
> > possibly be the source or destination of a pending DMA transfer.
> 
> I'll note that we right now do have an upstream BUG where that is sometimes
> not the case. I mentioned it previously that such bugs will be a problem :(
> 
> https://lkml.kernel.org/r/20250523023709epcms1p236d4f55b79adb9366ec1cf6d5792b06b@epcms1p2

I'll just reitarate the whole purpose of this patchset, as
added to Documentation:

+       This option increases the risk of a kdump failure: DMA transfers
+       configured by the first kernel may end up corrupting the second
+       kernel's memory.
+
+       This reservation method is intended for systems that can't afford to
+       sacrifice enough memory for standard crashkernel reservation and where
+       less reliable and possibly incomplete kdump is preferable to no kdump at
+       all.

It is expected that kdump may be less reliable when ,cma is used.
You mentioned a bug that augments this unreliability and that is surely going to get fixed.
I think this is fine.

The whole point is getting a completely optional best-effort kdump when
otherwise we would have no kdump.

> > +static void crash_cma_clear_pending_dma(void)
> > +{
> > +	unsigned int s = cma_dma_timeout_sec;
> > +
> > +	if (!crashk_cma_cnt)
> > +		return;
> > +
> > +	while (s--)
> > +		mdelay(1000);
> 
> Any reason we cannot do it in a single mdelay() invocation?
> 
> mdelay() already is a loop around udelay on larger values IIUC.

No good reasons ;)
I just wanted to prevent a totally theoretical overflow (if cma_dma_timeout_sec was made configurable;
I also anticipated someone might want to add some progress printks into the cycle (without verifying if
that's even possible in this context).

If you want, I have no problem changing this to:
+	mdelay(cma_dma_timeout_sec * 1000); 

-- 
Jiri Bohac <jbohac@xxxxxxx>
SUSE Labs, Prague, Czechia