Received-SPF: pass (google.com: domain of linux-kernel+bounces-212285-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.80.249 as permitted sender) client-ip=147.75.80.249;
Precedence: bulk
MIME-Version: 1.0
References: <20240612081216.1319089-1-zhai.he@nxp.com> <20240612114748.bf5983b50634f23d674bc749@linux-foundation.org>
In-Reply-To: <20240612114748.bf5983b50634f23d674bc749@linux-foundation.org>
From: Barry Song <baohua@kernel.org>
Date: Thu, 13 Jun 2024 09:37:15 +1200
Message-ID: <CAGsJ_4wsAh8C08PXutYZx9xV3rLRwLG-E6Mq-JgoSO5LA1ns=A@mail.gmail.com>
Subject: Re: [PATCH v2] Supports to use the default CMA when the
 device-specified CMA memory is not enough.
To: Andrew Morton <akpm@linux-foundation.org>
Cc: "zhai.he" <zhai.he@nxp.com>, sboyd@kernel.org, linux-mm@kvack.org, 
	linux-kernel@vger.kernel.org, stable@vger.kernel.org, zhipeng.wang_1@nxp.com, 
	jindong.yue@nxp.com, Christoph Hellwig <hch@lst.de>
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable

On Thu, Jun 13, 2024 at 6:47=E2=80=AFAM Andrew Morton <akpm@linux-foundatio=
n.org> wrote:
>
> On Wed, 12 Jun 2024 16:12:16 +0800 "zhai.he" <zhai.he@nxp.com> wrote:
>
> > From: He Zhai <zhai.he@nxp.com>
>
> (cc Barry & Christoph)
>
> What was your reason for adding cc:stable to the email headers?  Does
> this address some serious problem?  If so, please fully describe that
> problem.
>
> > In the current code logic, if the device-specified CMA memory
> > allocation fails, memory will not be allocated from the default CMA are=
a.
> > This patch will use the default cma region when the device's
> > specified CMA is not enough.
> >
> > In addition, the log level of allocation failure is changed to debug.
> > Because these logs will be printed when memory allocation from the
> > device specified CMA fails, but if the allocation fails, it will be
> > allocated from the default cma area. It can easily mislead developers'
> > judgment.

I am not convinced that this patch is correct. If device-specific CMA
is too small,
why not increase it in the device tree? Conversely, if the default CMA
size is too
large, why not reduce it via the cmdline?  CMA offers all kinds of
flexible configuration
options based on users=E2=80=99 needs.

One significant benefit of device-specific CMA is that it helps
decrease fragmentation
in the common CMA pool. While many devices allocate memory from the same po=
ol,
they have different memory requirements in terms of sizes and
alignments. Occasions
of memory allocation and release can lead to situations where the CMA
pool has enough
free space, yet someone fails to obtain contiguous memory from it.

This patch entirely negates the advantage we gain from device-specific CMA.
My point is that instead of modifying the core code, please consider correc=
ting
your device tree or cmdline configurations.

> >
> > ...
> >
> > --- a/kernel/dma/contiguous.c
> > +++ b/kernel/dma/contiguous.c
> > @@ -357,8 +357,13 @@ struct page *dma_alloc_contiguous(struct device *d=
ev, size_t size, gfp_t gfp)
> >       /* CMA can be used only in the context which permits sleeping */
> >       if (!gfpflags_allow_blocking(gfp))
> >               return NULL;
> > -     if (dev->cma_area)
> > -             return cma_alloc_aligned(dev->cma_area, size, gfp);
> > +     if (dev->cma_area) {
> > +             struct page *page =3D NULL;
> > +
> > +             page =3D cma_alloc_aligned(dev->cma_area, size, gfp);
> > +             if (page)
> > +                     return page;
> > +     }
> >       if (size <=3D PAGE_SIZE)
> >               return NULL;
>
> The dma_alloc_contiguous() kerneldoc should be updated for this.
>
> The patch prompts the question "why does the device-specified CMA area
> exist?".  Why not always allocate from the global pool?  If the
> device-specified area exists to prevent one device from going crazy and
> consuming too much contiguous memory, this patch violates that intent?
>
> > @@ -406,6 +411,8 @@ void dma_free_contiguous(struct device *dev, struct=
 page *page, size_t size)
> >       if (dev->cma_area) {
> >               if (cma_release(dev->cma_area, page, count))
> >                       return;
> > +             if (cma_release(dma_contiguous_default_area, page, count)=
)
> > +                     return;
> >       } else {
> >               /*
> >                * otherwise, page is from either per-numa cma or default=
 cma
> > diff --git a/mm/cma.c b/mm/cma.c
> > index 3e9724716bad..6e12faf1bea7 100644
> > --- a/mm/cma.c
> > +++ b/mm/cma.c
> > @@ -495,8 +495,8 @@ struct page *cma_alloc(struct cma *cma, unsigned lo=
ng count,
> >       }
> >
> >       if (ret && !no_warn) {
> > -             pr_err_ratelimited("%s: %s: alloc failed, req-size: %lu p=
ages, ret: %d\n",
> > -                                __func__, cma->name, count, ret);
> > +             pr_debug("%s: alloc failed, req-size: %lu pages, ret: %d,=
 try to use default cma\n",
> > +                         cma->name, count, ret);
> >               cma_debug_show_areas(cma);
> >       }
>

Thanks
Barry