[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20230127070705.GA4180@lst.de>
Date: Fri, 27 Jan 2023 08:07:05 +0100
From: Christoph Hellwig <hch@....de>
To: "H. Peter Anvin" <hpa@...or.com>
Cc: Christoph Hellwig <hch@....de>,
Dave Hansen <dave.hansen@...el.com>,
Georgi Djakov <quic_c_gdjako@...cinc.com>,
catalin.marinas@....com, will@...nel.org,
dave.hansen@...ux.intel.com, luto@...nel.org, peterz@...radead.org,
tglx@...utronix.de, mingo@...hat.com, bp@...en8.de,
m.szyprowski@...sung.com, robin.murphy@....com,
linux-arm-kernel@...ts.infradead.org, iommu@...ts.linux.dev,
linux-kernel@...r.kernel.org, djakov@...nel.org
Subject: Re: [RFC] mm: Allow ZONE_DMA32 to be disabled via kernel command
line
On Thu, Jan 26, 2023 at 10:52:43PM -0800, H. Peter Anvin wrote:
> >If people want ZONE_DMA32 to go away we need something to replace
> >it first, like a large enough CMA region in the 32-bit addressable
> >range.
>
> Not to mention all kinds of odd masks like 30, 31, 39, 40, 46, ... bits.
Yes. Out of those all >= 32 are falling straight into ZONE_DM32,
the lower ones we do a first try in ZONE_DMA32 and then fall back to
ZONE_DMA. <= 29 mask OTOH are really rate in modern systems for
actual devices. So with a CMA region for what is currently ZONE_DMA
and one for the first 1G we'd probably cover most of what's actually
needed for x86_64. Of course on 32-bit architetures things become
a lot more complicated due to highmem.
Powered by blists - more mailing lists