Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20220711122459.13773-1-me@linux.beauty> <20220711122459.13773-5-me@linux.beauty>
 <CAK8P3a2Mr0ZMXGDx6htYEbBBtm4mubk-meSASJjPRK1j1O-hEA@mail.gmail.com>
 <181efcca6ae.de84203d522625.7740936811073442334@linux.beauty>
 <CAK8P3a30o1RLifV1TMqDJ26vLhVdOzz3wP6yPrayLV2GPxUtwQ@mail.gmail.com> <18267b7a61f.12b26bd91245310.4476663913461696630@linux.beauty>
In-Reply-To: <18267b7a61f.12b26bd91245310.4476663913461696630@linux.beauty>
From:   Arnd Bergmann <arnd@arndb.de>
Date:   Thu, 4 Aug 2022 10:24:20 +0200
Message-ID: <CAK8P3a0zSGqj3YEi+i9yfSLk8-aJtyiY6Bj069cxCdErk81+cw@mail.gmail.com>
Subject: Re: [PATCH 4/4] sample/reserved_mem: Introduce a sample of struct
 page and dio support to no-map rmem
To:     Li Chen <me@linux.beauty>
Cc:     Arnd Bergmann <arnd@arndb.de>,
        Catalin Marinas <catalin.marinas@arm.com>,
        Will Deacon <will@kernel.org>,
        Rob Herring <robh+dt@kernel.org>,
        Frank Rowand <frowand.list@gmail.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Li Chen <lchen@ambarella.com>,
        Linux ARM <linux-arm-kernel@lists.infradead.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        DTML <devicetree@vger.kernel.org>, Linux-MM <linux-mm@kvack.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Thu, Aug 4, 2022 at 9:17 AM Li Chen <me@linux.beauty> wrote:
> ---- On Tue, 12 Jul 2022 16:50:46 +0900  Arnd Bergmann  wrote ---
>  > Does your hardware require a fixed address for the buffer? If it can be
>  > anywhere in memory (or at least within a certain range) but just has to
>  > be physically contiguous, the normal way would be to use a CMA area
>  > to allocate from, which gives you 'struct page' backed pages.
>
> CMA does support Direct I/O, but it has its own issue:
> It does not guarantee that the memory previously borrowed by the OS will be returned to the device.
>
> We've been plagued by examples like this in the past:
> Many other kernel modules/subsystems have already allocated much memory from both non-CMA and CMA memory,
> When our DSP driver got probed then, cma_alloc will fail in that non-CMA system memory is not enough
> for CMA memory to migrate.

This part should at least be possible to solve by declaring the amount
and location of
CMA areas in the right way. It's not great to fine-tune the DT for a
particular kernel's
use, but if you know which other drivers require CMA type allocations
you can find a lower
bound that should suffice.

Most coherent allocations tend to be long-lived and only for very
small memory regions.
If you have another driver that uses large or periodic
dma_alloc_coherent() type allocations,
you can consider either giving that device its own CMA area, or fixing
it to use streaming
mappings.

        Arnd