[PATCH RFC v3 06/35] mm: cma: Make CMA_ALLOC_SUCCESS/FAIL count the number of pages

[Anshuman Khandual]

On 1/29/24 17:21, Alexandru Elisei wrote:
> Hi,
> 
> On Mon, Jan 29, 2024 at 02:54:20PM +0530, Anshuman Khandual wrote:
>>
>>
>> On 1/25/24 22:12, Alexandru Elisei wrote:
>>> The CMA_ALLOC_SUCCESS, respectively CMA_ALLOC_FAIL, are increased by one
>>> after each cma_alloc() function call. This is done even though cma_alloc()
>>> can allocate an arbitrary number of CMA pages. When looking at
>>> /proc/vmstat, the number of successful (or failed) cma_alloc() calls
>>> doesn't tell much with regards to how many CMA pages were allocated via
>>> cma_alloc() versus via the page allocator (regular allocation request or
>>> PCP lists refill).
>>>
>>> This can also be rather confusing to a user who isn't familiar with the
>>> code, since the unit of measurement for nr_free_cma is the number of pages,
>>> but cma_alloc_success and cma_alloc_fail count the number of cma_alloc()
>>> function calls.
>>>
>>> Let's make this consistent, and arguably more useful, by having
>>> CMA_ALLOC_SUCCESS count the number of successfully allocated CMA pages, and
>>> CMA_ALLOC_FAIL count the number of pages the cma_alloc() failed to
>>> allocate.
>>>
>>> For users that wish to track the number of cma_alloc() calls, there are
>>> tracepoints for that already implemented.
>>>
>>> Signed-off-by: Alexandru Elisei <alexandru.elisei at arm.com>
>>> ---
>>>  mm/cma.c | 4 ++--
>>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/mm/cma.c b/mm/cma.c
>>> index f49c95f8ee37..dbf7fe8cb1bd 100644
>>> --- a/mm/cma.c
>>> +++ b/mm/cma.c
>>> @@ -517,10 +517,10 @@ struct page *cma_alloc(struct cma *cma, unsigned long count,
>>>  	pr_debug("%s(): returned %p\n", __func__, page);
>>>  out:
>>>  	if (page) {
>>> -		count_vm_event(CMA_ALLOC_SUCCESS);
>>> +		count_vm_events(CMA_ALLOC_SUCCESS, count);
>>>  		cma_sysfs_account_success_pages(cma, count);
>>>  	} else {
>>> -		count_vm_event(CMA_ALLOC_FAIL);
>>> +		count_vm_events(CMA_ALLOC_FAIL, count);
>>>  		if (cma)
>>>  			cma_sysfs_account_fail_pages(cma, count);
>>>  	}
>>
>> Without getting into the merits of this patch - which is actually trying to do
>> semantics change to /proc/vmstat, wondering how is this even related to this
>> particular series ? If required this could be debated on it's on separately.
> 
> Having the number of CMA pages allocated and the number of CMA pages freed
> allows someone to infer how many tagged pages are in use at a given time:

That should not be done in CMA which is a generic multi purpose allocator.

> (allocated CMA pages - CMA pages allocated by drivers* - CMA pages
> released) * 32. That is valuable information for software and hardware
> designers.
> 
> Besides that, for every iteration of the series, this has proven invaluable
> for discovering bugs with freeing and/or reserving tag storage pages.

I am afraid that might not be enough justification for getting something
merged mainline.

> 
> *that would require userspace reading cma_alloc_success and
> cma_release_success before any tagged allocations are performed.

While assuming that no other non-memory-tagged CMA based allocation amd free
call happens in the meantime ? That would be on real thin ice.

I suppose arm64 tagged memory specific allocation or free related counters
need to be created on the caller side, including arch_free_pages_prepare().