Reading some old email... Maybe the reply is still useful.
On Wed 12-10-22 17:40:11, lijiazi wrote:
> I recently encountered a CMA page migration failure issue.
> This page has private, and private data is buffer_head struct pointer.
> buffer_head->b_count is not zero, so drop_buffers failed.
>
> This leads to the failure of both directly reclaim and migration attempts for
> this page.
> Finally, CMA memory alloc failed.
>
> This buffer_head detail info are as follows:
>
> crash> struct buffer_head 0xffffffec9f0200d0 -x
> struct buffer_head {
> b_state = 0x29, //has Uptodate, Req, Mapped flags
> b_this_page = 0xffffffec9f0200d0,
> b_page = 0xffffffbfb4bb0080,
> b_blocknr = 0x801b,
> b_size = 0x1000,
> b_data = 0xffffffed2ec02000 "\244\201",
> b_bdev = 0xffffffed169b2580,
> b_end_io = 0xffffff91006c44e4 <end_buffer_read_sync>,
> b_private = 0x0,
> b_assoc_buffers = {
> next = 0xffffffec9f020118,
> prev = 0xffffffec9f020118
> },
> b_assoc_map = 0x0,
> b_count = {
> counter = 0x1
> }
> }
>
> The b_count is 1, just because it's in cpu6 bh_lru:
> crash> p bh_lrus:a | grep 0xffffffec9f0200d0 -B 1
> per_cpu(bh_lrus, 6) = $7 = {
> bhs = {0xffffffed146867b8, 0xffffffec9f020548, 0xffffffed0f7e3138, 0xffffffed0f7e30d0,
> 0xffffffed0f6f7340, 0xffffffed0b8c59c0, 0xffffffeb7bdb7888, 0xffffffed0b8c5548,
> 0xffffffed0f7b7270, 0xffffffed0f7b7208, 0xffffffed0f7b7138, 0xffffffec9f0200d0,//this entry
> 0xffffffed0f7b7068, 0xffffffed0f7b7000, 0xffffffed0f7b7bc8, 0xffffffec9f020068}
>
> On my device using kernel-4.19, inactive bh may be in bh_lrus for a long
> time, and cause the corresponding page migration failure.
>
> In function buffer_busy, can we check if b_count is greater than zero
> just because it's in bh_lrus?
> If yes, can we evict some inactive bhs to improve the success rate of
> migration?
Not sure about your codebase but at least upstream __buffer_migrate_folio()
(which used to be buffer_migrate_page() in your kernel) does
invalidate_bh_lrus() call if it finds any buffer is busy, exactly to avoid
this problem. Maybe you're missing some backport?
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
