Re: [RFC PATCH 2/2] filemap: use high-order folios in filemap sync RA

[Stepanov Anatoly <stepanov.anatoly@huawei.com>]

On 4/15/2026 3:06 PM, Pedro Falcato wrote:
> On Thu, Apr 16, 2026 at 03:28:53AM +0800, Anatoly Stepanov wrote:
>> [Idea]
>>
>> If a mmap'ed file being accessed such that async RA never
>> kicks in, we might end up with only 0-order folios in the page cache.
>>
>> if fault_around_bytes is larger than 1 single page, then
>> it's beneficial to use high-order folios, which brings significant
>> filemap_map_pages() speedup.
>> So, let's just use fault_around_bytes as a starting point here.
> 
> Well, this heuristic looks arbitrary. I don't like to mix different concepts.
> 
> With this, in practice most file folios will be 64K. Why? Why is it related
> to faultaround when faultaround is a separate mechanism that isn't particularly
> relevant here?
> 
>>
>> if an arch supports PTE-coalescing we can get more of those for free.
>> (see arm64 example below)
>>
>> We don't save the new order to "ra->order", so if async RA will happen
>> it would normally start from order-0.
>>
>> [Things to be discussed]
>>
>> But at the same time, i can see drawback for 16K, 64K pages, in this case fault_around will still be 64K by default.
>> In this case, it seems makes sense to make the fault_around_bytes be like order-N of PAGE_SIZE, not fixed bytes number.
>>
>> Another issue is - when fault_around=0, but we'd like to use high-order folios for sync_RA, for cont-PTE for example,
>> For this we can use kind of "max(fault_around_order, cont_pte_order)".
>>
>> Or introduce some dedicated tunable like "sync_mmap_order".
>>
>> [Benchmark]
>>
>> Simple benchmark below reading 100M file in 4M (RA size) chunks
>> such that async RA doesn't kick in and the page cache ends up being
>> filled up with 0-order folios.

> 
> Well, the problem is that you are _never_ getting RA to kick in. Folio
> size is the least of your concern, you are effectively not doing much
> readahead since the kernel thinks you're doing random accesses.
No, that's not true, "sync mmap readahead" actually works in the case
the problem  is that "async RA" doesn't kick in. 

>>
>> The patched kernel gives ~3 times increase in throughput,
>> considering the page cache is filled up at the moment.
>>
>> The main speedup comes from filemap_map_pages() due to high-order
>> folios usage.
>>
>> As a bonus, we get better cont_pte bit coverage for Arm64.
>>
>> Example:
>> // Open 100M file and read every 4M chunk, given max_ra=4M
>> // Perform 10 runs, measure the throughput.
>> ...
>>  char *map = mmap(NULL, filesize, PROT_READ, MAP_PRIVATE, fd, 0);
>>     if (map == MAP_FAILED) {
>>         perror("Error mapping file");
>>         close(fd);
>>         return 1;
>>     }
>>
>>     struct timespec start, end;
>>     clock_gettime(CLOCK_MONOTONIC, &start);
>>
>>     unsigned int size_4M = 4*1024*1024;
>>     unsigned int num_reads = filesize / size_4M;
>>     volatile char val;
>>     for (int i = 0; i < num_reads; i++) {
>>         off_t offset = (off_t)i * size_4M;
>>         val = map[offset];
>>     }
> 
> This doesn't seem like a real issue. And if it is, you can always issue
> readahead manually. But the whole pattern of "every perfectly-sized RA
> window, access 4 bytes and advance" is completely bizarre. And _if_ this
> is your workload, then having order-0 folios at the read site is much better
> than filling your page cache with data you are not accessing.
> 
> Do you have an actual use case for this? Where have you observed these
> problems?
> 


-- 
Anatoly Stepanov, Huawei