Message-ID: <ZbpPm_KCkONR7R3U@raptor>
Date: Wed, 31 Jan 2024 13:48:11 +0000
From: Alexandru Elisei <alexandru.elisei@....com>
To: Anshuman Khandual <anshuman.khandual@....com>
Cc: catalin.marinas@....com, will@...nel.org, oliver.upton@...ux.dev,
	maz@...nel.org, james.morse@....com, suzuki.poulose@....com,
	yuzenghui@...wei.com, arnd@...db.de, akpm@...ux-foundation.org,
	mingo@...hat.com, peterz@...radead.org, juri.lelli@...hat.com,
	vincent.guittot@...aro.org, dietmar.eggemann@....com,
	rostedt@...dmis.org, bsegall@...gle.com, mgorman@...e.de,
	bristot@...hat.com, vschneid@...hat.com, mhiramat@...nel.org,
	rppt@...nel.org, hughd@...gle.com, pcc@...gle.com,
	steven.price@....com, vincenzo.frascino@....com, david@...hat.com,
	eugenis@...gle.com, kcc@...gle.com, hyesoo.yu@...sung.com,
	linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org,
	kvmarm@...ts.linux.dev, linux-fsdevel@...r.kernel.org,
	linux-arch@...r.kernel.org, linux-mm@...ck.org,
	linux-trace-kernel@...r.kernel.org
Subject: Re: [PATCH RFC v3 09/35] mm: cma: Introduce cma_remove_mem()

Hi,

On Wed, Jan 31, 2024 at 06:49:34PM +0530, Anshuman Khandual wrote:
> On 1/30/24 17:03, Alexandru Elisei wrote:
> > Hi,
> > 
> > I really appreciate the feedback you have given me so far. I believe the
> > commit message isn't clear enough and there has been a confusion.
> > 
> > A CMA user adds a CMA area to the cma_areas array with
> > cma_declare_contiguous_nid() or cma_init_reserved_mem().
> > init_cma_reserved_pageblock() then iterates over the array and activates
> > all cma areas.
> 
> Agreed.
> 
> > 
> > The function cma_remove_mem() is intended to be used to remove a cma area
> > from the cma_areas array **before** the area has been activated.
> 
> Understood.
> 
> > 
> > Usecase: a driver (in this case, the arm64 dynamic tag storage code)
> > manages several cma areas. The driver successfully adds the first area to
> > the cma_areas array. When the driver tries to adds the second area, the
> > function fails. Without cma_remove_mem(), the driver has no way to prevent
> > the first area from being freed to the page allocator. cma_remove_mem() is
> > about providing a means to do cleanup in case of error.
> > 
> > Does that make more sense now?
> 
> How to ensure that cma_remove_mem() should get called by the driver before
> core_initcall()---> cma_init_reserved_areas()---> cma_activate_area() chain
> happens. Else cma_remove_mem() will miss out to clear cma->count and given
> area will proceed to get activated like always.

The same way drivers today call cma_declare_contiguous_nid() and
cma_init_reserved_mem() before cma_init_reserved_areas(). For an example,
have a look at kernel/dma/contiguous.c:: rmem_cma_setup().

As for how the series uses cma_remove_mem(), have a look at patch #20
("arm64: mte: Add tag storage memory to CMA") [1], specifically this bit:

        for (i = 0; i < num_tag_regions; i++) {
                region = &tag_regions[i];

		// code removed for clarity

                ret = cma_init_reserved_mem(PFN_PHYS(region->tag_range.start),
                                PFN_PHYS(range_len(&region->tag_range)),
                                order, NULL, &region->cma);
                if (ret) {
                        for (j = 0; j < i; j++)
                                cma_remove_mem(&region->cma);
                        goto out_disabled;
                }
	}

	// code removed for clarity

out_disabled:
        num_tag_regions = 0;
        pr_info("MTE tag storage region management disabled");

I'll try to walk you through it. The driver manages 2 cma regions.

cma_init_reserved_mem() succeeds for the first region.

cma_init_reserved_mem() fails for the second region.

As a result, the first region will be activated (pages will be placed on
the MIGRATE_CMA list), but the second region will not be activated.

The driver can function only when **all** cma regions have been
successfully activated.

Driver removes first region from CMA, so now no regions will be activated,
and probing fails.

In a more general sense, cma_remove_mem() is **not** about removing a
region that failed initialization or activation, it's about removing a cma
area that was added to cma_areas successfully, but the driver doesn't want
to activate anymore for whatever reason (it can be because of a probing
error totally unrelated to CMA).

Does it make more sense now? I hope that this example also answers the rest
of your questions.

[1] https://lore.kernel.org/linux-arm-kernel/20240125164256.4147-21-alexandru.elisei@arm.com/

Thanks,
Alex

> 
> > 
> > Ok Tue, Jan 30, 2024 at 11:20:56AM +0530, Anshuman Khandual wrote:
> >>
> >>
> >> On 1/25/24 22:12, Alexandru Elisei wrote:
> >>> Memory is added to CMA with cma_declare_contiguous_nid() and
> >>> cma_init_reserved_mem(). This memory is then put on the MIGRATE_CMA list in
> >>> cma_init_reserved_areas(), where the page allocator can make use of it.
> >>
> >> cma_declare_contiguous_nid() reserves memory in memblock and marks the
> > 
> > You forgot about about cma_init_reserved_mem() which does the same thing,
> > but yes, you are right.
> 
> Agreed, missed that. There are some direct cma_init_reserved_mem() calls as well.
> 
> > 
> >> for subsequent CMA usage, where as cma_init_reserved_areas() activates
> >> these memory areas through init_cma_reserved_pageblock(). Standard page
> >> allocator only receives these memory via free_reserved_page() - only if
> > 
> > I don't think that's correct. init_cma_reserved_pageblock() clears the
> > PG_reserved page flag, sets the migratetype to MIGRATE_CMA and then frees
> > the page. After that, the page is available to the standard page allocator
> > to use for allocation. Otherwise, what would be the point of the
> > MIGRATE_CMA migratetype?
> 
> Understood and agreed.
> 
> > 
> >> the page block activation fails.
> > 
> > For the sake of having a complete picture, I'll add that that only happens
> > if cma->reserve_pages_on_error is false. If the CMA user sets the field to
> > 'true' (with cma_reserve_pages_on_error()), then the pages in the CMA
> > region are kept PG_reserved if activation fails.
> 
> Why cannot you use cma_reserve_pages_on_error() ?
> 
> > 
> >>
> >>>
> >>> If a device manages multiple CMA areas, and there's an error when one of
> >>> the areas is added to CMA, there is no mechanism for the device to prevent
> >>
> >> What kind of error ? init_cma_reserved_pageblock() fails ? But that will
> >> not happen until cma_init_reserved_areas().
> > 
> > I think I haven't been clear enough. When I say that "an area is added
> > to CMA", I mean that the memory region is added to cma_areas array, via
> > cma_declare_contiguous_nid() or cma_init_reserved_mem(). There are several
> > ways in which either function can fail.
> 
> Okay.
> 
> > 
> >>
> >>> the rest of the areas, which were added before the error occured, from
> >>> being later added to the MIGRATE_CMA list.
> >>
> >> Why is this mechanism required ? cma_init_reserved_areas() scans over all
> >> CMA areas and try and activate each of them sequentially. Why is not this
> >> sufficient ?
> > 
> > This patch is about removing a struct cma from the cma_areas array after it
> > has been added to the array, with cma_declare_contiguous_nid() or
> > cma_init_reserved_mem(), to prevent the area from being activated in
> > cma_init_reserved_areas(). Sorry for the confusion.
> > 
> > I'll add a check in cma_remove_mem() to fail if the cma area has been
> > activated, and a comment to the function to explain its usage.
> 
> That will be a good check.
> 
> > 
> >>
> >>>
> >>> Add cma_remove_mem() which allows a previously reserved CMA area to be
> >>> removed and thus it cannot be used by the page allocator.
> >>
> >> Successfully activated CMA areas do not get used by the buddy allocator.
> > 
> > I don't believe that is correct, see above.
> Apologies, it's my bad.
> 
> > 
> >>
> >>>
> >>> Signed-off-by: Alexandru Elisei <alexandru.elisei@....com>
> >>> ---
> >>>
> >>> Changes since rfc v2:
> >>>
> >>> * New patch.
> >>>
> >>>  include/linux/cma.h |  1 +
> >>>  mm/cma.c            | 30 +++++++++++++++++++++++++++++-
> >>>  2 files changed, 30 insertions(+), 1 deletion(-)
> >>>
> >>> diff --git a/include/linux/cma.h b/include/linux/cma.h
> >>> index e32559da6942..787cbec1702e 100644
> >>> --- a/include/linux/cma.h
> >>> +++ b/include/linux/cma.h
> >>> @@ -48,6 +48,7 @@ extern int cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
> >>>  					unsigned int order_per_bit,
> >>>  					const char *name,
> >>>  					struct cma **res_cma);
> >>> +extern void cma_remove_mem(struct cma **res_cma);
> >>>  extern struct page *cma_alloc(struct cma *cma, unsigned long count, unsigned int align,
> >>>  			      bool no_warn);
> >>>  extern int cma_alloc_range(struct cma *cma, unsigned long start, unsigned long count,
> >>> diff --git a/mm/cma.c b/mm/cma.c
> >>> index 4a0f68b9443b..2881bab12b01 100644
> >>> --- a/mm/cma.c
> >>> +++ b/mm/cma.c
> >>> @@ -147,8 +147,12 @@ static int __init cma_init_reserved_areas(void)
> >>>  {
> >>>  	int i;
> >>>  
> >>> -	for (i = 0; i < cma_area_count; i++)
> >>> +	for (i = 0; i < cma_area_count; i++) {
> >>> +		/* Region was removed. */
> >>> +		if (!cma_areas[i].count)
> >>> +			continue;
> >>
> >> Skip previously added CMA area (now zeroed out) ?
> > 
> > Yes, that's what I meant with the comment "Region was removed". Do you
> > think I should reword the comment?
> > 
> >>
> >>>  		cma_activate_area(&cma_areas[i]);
> >>> +	}
> >>>  
> >>>  	return 0;
> >>>  }
> >>
> >> cma_init_reserved_areas() gets called via core_initcall(). Some how
> >> platform/device needs to call cma_remove_mem() before core_initcall()
> >> gets called ? This might be time sensitive.
> > 
> > I don't understand your point.
> > 
> >>
> >>> @@ -216,6 +220,30 @@ int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
> >>>  	return 0;
> >>>  }
> >>>  
> >>> +/**
> >>> + * cma_remove_mem() - remove cma area
> >>> + * @res_cma: Pointer to the cma region.
> >>> + *
> >>> + * This function removes a cma region created with cma_init_reserved_mem(). The
> >>> + * ->count is set to 0.
> >>> + */
> >>> +void __init cma_remove_mem(struct cma **res_cma)
> >>> +{
> >>> +	struct cma *cma;
> >>> +
> >>> +	if (WARN_ON_ONCE(!res_cma || !(*res_cma)))
> >>> +		return;
> >>> +
> >>> +	cma = *res_cma;
> >>> +	if (WARN_ON_ONCE(!cma->count))
> >>> +		return;
> >>> +
> >>> +	totalcma_pages -= cma->count;
> >>> +	cma->count = 0;
> >>> +
> >>> +	*res_cma = NULL;
> >>> +}
> >>> +
> >>>  /**
> >>>   * cma_declare_contiguous_nid() - reserve custom contiguous area
> >>>   * @base: Base address of the reserved area optional, use 0 for any
> >>
> >> But first please do explain what are the errors device or platform might
> > 
> > cma_declare_contiguous_nid() and cma_init_reserved_mem() can fail in a
> > number of ways, the code should be self documenting.
> 
> But when they do fail - would not cma->count be left uninitialized as 0 ?
> Hence the proposed check (!cma->count) in cma_init_reserved_areas() should
> just do the trick without requiring an explicit cma_remove_mem() call.
> 
> > 
> >> see on a previously marked CMA area so that removing them on way becomes
> >> necessary preventing their activation via cma_init_reserved_areas().
> > 
> > I've described how the function is supposed to be used at the top of my
> > reply.
> > 
> > Thanks,
> > Alex
> 

Powered by blists - more mailing lists