Hi, Jan
On 2025/7/14 17:33, Jan Kara wrote:
On Fri 11-07-25 13:55:09, Youling Tang wrote:
From: Youling Tang <tangyouling@xxxxxxxxxx>
On XFS systems with pagesize=4K, blocksize=16K, and CONFIG_TRANSPARENT_HUGEPAGE
enabled, We observed the following readahead behaviors:
# echo 3 > /proc/sys/vm/drop_caches
# dd if=test of=/dev/null bs=64k count=1
# ./tools/mm/page-types -r -L -f /mnt/xfs/test
foffset offset flags
0 136d4c __RU_l_________H______t_________________F_1
1 136d4d __RU_l__________T_____t_________________F_1
2 136d4e __RU_l__________T_____t_________________F_1
3 136d4f __RU_l__________T_____t_________________F_1
...
c 136bb8 __RU_l_________H______t_________________F_1
d 136bb9 __RU_l__________T_____t_________________F_1
e 136bba __RU_l__________T_____t_________________F_1
f 136bbb __RU_l__________T_____t_________________F_1 <-- first read
10 13c2cc ___U_l_________H______t______________I__F_1 <-- readahead flag
11 13c2cd ___U_l__________T_____t______________I__F_1
12 13c2ce ___U_l__________T_____t______________I__F_1
13 13c2cf ___U_l__________T_____t______________I__F_1
...
1c 1405d4 ___U_l_________H______t_________________F_1
1d 1405d5 ___U_l__________T_____t_________________F_1
1e 1405d6 ___U_l__________T_____t_________________F_1
1f 1405d7 ___U_l__________T_____t_________________F_1
[ra_size = 32, req_count = 16, async_size = 16]
# echo 3 > /proc/sys/vm/drop_caches
# dd if=test of=/dev/null bs=60k count=1
# ./page-types -r -L -f /mnt/xfs/test
foffset offset flags
0 136048 __RU_l_________H______t_________________F_1
...
c 110a40 __RU_l_________H______t_________________F_1
d 110a41 __RU_l__________T_____t_________________F_1
e 110a42 __RU_l__________T_____t_________________F_1 <-- first read
f 110a43 __RU_l__________T_____t_________________F_1 <-- first readahead flag
10 13e7a8 ___U_l_________H______t_________________F_1
...
20 137a00 ___U_l_________H______t_______P______I__F_1 <-- second readahead flag (20 - 2f)
21 137a01 ___U_l__________T_____t_______P______I__F_1
...
3f 10d4af ___U_l__________T_____t_______P_________F_1
[first readahead: ra_size = 32, req_count = 15, async_size = 17]
When reading 64k data (same for 61-63k range, where last_index is page-aligned
in filemap_get_pages()), 128k readahead is triggered via page_cache_sync_ra()
and the PG_readahead flag is set on the next folio (the one containing 0x10 page).
When reading 60k data, 128k readahead is also triggered via page_cache_sync_ra().
However, in this case the readahead flag is set on the 0xf page. Although the
requested read size (req_count) is 60k, the actual read will be aligned to
folio size (64k), which triggers the readahead flag and initiates asynchronous
readahead via page_cache_async_ra(). This results in two readahead operations
totaling 256k.
The root cause is that when the requested size is smaller than the actual read
size (due to folio alignment), it triggers asynchronous readahead. By changing
last_index alignment from page size to folio size, we ensure the requested size
matches the actual read size, preventing the case where a single read operation
triggers two readahead operations.
After applying the patch:
# echo 3 > /proc/sys/vm/drop_caches
# dd if=test of=/dev/null bs=60k count=1
# ./page-types -r -L -f /mnt/xfs/test
foffset offset flags
0 136d4c __RU_l_________H______t_________________F_1
1 136d4d __RU_l__________T_____t_________________F_1
2 136d4e __RU_l__________T_____t_________________F_1
3 136d4f __RU_l__________T_____t_________________F_1
...
c 136bb8 __RU_l_________H______t_________________F_1
d 136bb9 __RU_l__________T_____t_________________F_1
e 136bba __RU_l__________T_____t_________________F_1 <-- first read
f 136bbb __RU_l__________T_____t_________________F_1
10 13c2cc ___U_l_________H______t______________I__F_1 <-- readahead flag
11 13c2cd ___U_l__________T_____t______________I__F_1
12 13c2ce ___U_l__________T_____t______________I__F_1
13 13c2cf ___U_l__________T_____t______________I__F_1
...
1c 1405d4 ___U_l_________H______t_________________F_1
1d 1405d5 ___U_l__________T_____t_________________F_1
1e 1405d6 ___U_l__________T_____t_________________F_1
1f 1405d7 ___U_l__________T_____t_________________F_1
[ra_size = 32, req_count = 16, async_size = 16]
The same phenomenon will occur when reading from 49k to 64k. Set the readahead
flag to the next folio.
Because the minimum order of folio in address_space equals the block size (at
least in xfs and bcachefs that already support bs > ps), having request_count
aligned to block size will not cause overread.
Co-developed-by: Chi Zhiling <chizhiling@xxxxxxxxxx>
Signed-off-by: Chi Zhiling <chizhiling@xxxxxxxxxx>
Signed-off-by: Youling Tang <tangyouling@xxxxxxxxxx>
I agree with analysis of the problem but not quite with the solution. See
below.
diff --git a/mm/filemap.c b/mm/filemap.c
index 765dc5ef6d5a..56a8656b6f86 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -2584,8 +2584,9 @@ static int filemap_get_pages(struct kiocb *iocb, size_t count,
unsigned int flags;
int err = 0;
- /* "last_index" is the index of the page beyond the end of the read */
- last_index = DIV_ROUND_UP(iocb->ki_pos + count, PAGE_SIZE);
+ /* "last_index" is the index of the folio beyond the end of the read */
+ last_index = round_up(iocb->ki_pos + count, mapping_min_folio_nrbytes(mapping));
+ last_index >>= PAGE_SHIFT;
I think that filemap_get_pages() shouldn't be really trying to guess what
readahead code needs and round last_index based on min folio order. After
all the situation isn't special for LBS filesystems. It can also happen
that the readahead mark ends up in the middle of large folio for other
reasons. In fact, we already do have code in page_cache_ra_order() ->
ra_alloc_folio() that handles rounding of index where mark should be placed
so your changes essentially try to outsmart that code which is not good. I
think the solution should really be placed in page_cache_ra_order() +
ra_alloc_folio() instead.
In fact the problem you are trying to solve was kind of introduced (or at
least made more visible) by my commit ab4443fe3ca62 ("readahead: avoid
multiple marked readahead pages"). There I've changed the code to round the
index down because I've convinced myself it doesn't matter and rounding
down is easier to handle in that place. But your example shows there are
cases where rounding down has weird consequences and rounding up would have
been better. So I think we need to come up with a method how to round up
the index of marked folio to fix your case without reintroducing problems
mentioned in commit ab4443fe3ca62.
Yes, I simply replaced round_up() in ra_alloc_folio() with round_down()
to avoid this phenomenon before submitting this patch.
But at present, I haven't found a suitable way to solve both of these
problems
simultaneously. Do you have a better solution on your side?
Thanks,
Youling.
Honza
