Re: [PATCH] fuse: enable larger read buffers for readdir [v2].

[Bernd Schubert <bernd.schubert@fastmail.fm>]

On 7/27/23 17:35, Miklos Szeredi wrote:
> On Thu, 27 Jul 2023 at 10:13, Jaco Kroon <jaco@uls.co.za> wrote:
>>
>> This patch does not mess with the caching infrastructure like the
>> previous one, which we believe caused excessive CPU and broke directory
>> listings in some cases.
>>
>> This version only affects the uncached read, which then during parse adds an
>> entry at a time to the cached structures by way of copying, and as such,
>> we believe this should be sufficient.
>>
>> We're still seeing cases where getdents64 takes ~10s (this was the case
>> in any case without this patch, the difference now that we get ~500
>> entries for that time rather than the 14-18 previously).  We believe
>> that that latency is introduced on glusterfs side and is under separate
>> discussion with the glusterfs developers.
>>
>> This is still a compile-time option, but a working one compared to
>> previous patch.  For now this works, but it's not recommended for merge
>> (as per email discussion).
>>
>> This still uses alloc_pages rather than kvmalloc/kvfree.
>>
>> Signed-off-by: Jaco Kroon <jaco@uls.co.za>
>> ---
>>   fs/fuse/Kconfig   | 16 ++++++++++++++++
>>   fs/fuse/readdir.c | 18 ++++++++++++------
>>   2 files changed, 28 insertions(+), 6 deletions(-)
>>
>> diff --git a/fs/fuse/Kconfig b/fs/fuse/Kconfig
>> index 038ed0b9aaa5..0783f9ee5cd3 100644
>> --- a/fs/fuse/Kconfig
>> +++ b/fs/fuse/Kconfig
>> @@ -18,6 +18,22 @@ config FUSE_FS
>>            If you want to develop a userspace FS, or if you want to use
>>            a filesystem based on FUSE, answer Y or M.
>>
>> +config FUSE_READDIR_ORDER
>> +       int
>> +       range 0 5
>> +       default 5
>> +       help
>> +               readdir performance varies greatly depending on the size of the read.
>> +               Larger buffers results in larger reads, thus fewer reads and higher
>> +               performance in return.
>> +
>> +               You may want to reduce this value on seriously constrained memory
>> +               systems where 128KiB (assuming 4KiB pages) cache pages is not ideal.
>> +
>> +               This value reprents the order of the number of pages to allocate (ie,
>> +               the shift value).  A value of 0 is thus 1 page (4KiB) where 5 is 32
>> +               pages (128KiB).
>> +
>>   config CUSE
>>          tristate "Character device in Userspace support"
>>          depends on FUSE_FS
>> diff --git a/fs/fuse/readdir.c b/fs/fuse/readdir.c
>> index dc603479b30e..47cea4d91228 100644
>> --- a/fs/fuse/readdir.c
>> +++ b/fs/fuse/readdir.c
>> @@ -13,6 +13,12 @@
>>   #include <linux/pagemap.h>
>>   #include <linux/highmem.h>
>>
>> +#define READDIR_PAGES_ORDER            CONFIG_FUSE_READDIR_ORDER
>> +#define READDIR_PAGES                  (1 << READDIR_PAGES_ORDER)
>> +#define READDIR_PAGES_SIZE             (PAGE_SIZE << READDIR_PAGES_ORDER)
>> +#define READDIR_PAGES_MASK             (READDIR_PAGES_SIZE - 1)
>> +#define READDIR_PAGES_SHIFT            (PAGE_SHIFT + READDIR_PAGES_ORDER)
>> +
>>   static bool fuse_use_readdirplus(struct inode *dir, struct dir_context *ctx)
>>   {
>>          struct fuse_conn *fc = get_fuse_conn(dir);
>> @@ -328,25 +334,25 @@ static int fuse_readdir_uncached(struct file *file, struct dir_context *ctx)
>>          struct fuse_mount *fm = get_fuse_mount(inode);
>>          struct fuse_io_args ia = {};
>>          struct fuse_args_pages *ap = &ia.ap;
>> -       struct fuse_page_desc desc = { .length = PAGE_SIZE };
>> +       struct fuse_page_desc desc = { .length = READDIR_PAGES_SIZE };
> 
> Does this really work?  I would've thought we are relying on single
> page lengths somewhere.
> 
>>          u64 attr_version = 0;
>>          bool locked;
>>
>> -       page = alloc_page(GFP_KERNEL);
>> +       page = alloc_pages(GFP_KERNEL, READDIR_PAGES_ORDER);
>>          if (!page)
>>                  return -ENOMEM;
>>
>>          plus = fuse_use_readdirplus(inode, ctx);
>>          ap->args.out_pages = true;
>> -       ap->num_pages = 1;
>> +       ap->num_pages = READDIR_PAGES;
> 
> No.  This is the array lenght, which is 1.  This is the hack I guess,
> which makes the above trick work.

Hmm, ap->num_pages / ap->pages[] is used in fuse_copy_pages, but so is 
ap->descs[] - shouldn't the patch caused an out-of-bound access?
Out of interest, would you mind to explain how the hack worked?


Thanks,
Bernd