On Fri 03-06-22 15:40:13, Matthew Wilcox wrote: > On Thu, May 26, 2022 at 10:43:51PM -0700, Christoph Hellwig wrote: > > > static ssize_t jfs_quota_read(struct super_block *sb, int type, char *data, > > > + size_t len, loff_t pos) > > > > And this whole helper is generic now. It might be worth to move it > > into fs/quota/dquot.c as generic_quota_read. > > I've been working on that this week. Unfortunately, you have to convert > both quota_read and quota_write at the same time, it turns out. At > least ext4_quota_write() uses the bdev's inode's page cache to back > the buffer_heads, so quota_read() and quota_write() are incoherent > with each other: > > 00017 gqr: mapping:00000000ee19acfb index:0x1 pos:0x1470 len:0x30 > 00017 4qw: mapping:000000007f9a811e index:0x18405 pos:0x1440 len:0x30 Yes, reads and writes have to use the same cache. Otherwise bad things happen... > I don't know if there's a way around this. Can't really use > read_mapping_folio() on the bdev's inode in generic_quota_read() -- the > blocks for a given page might be fragmented on disk. I don't know > if there's a way to tell ext4_bread() to use the inode's page cache > instead of the bdev's. There's no way for ext4_bread() to read from inode's page cache. And that is deliberate - ext4_bread() is used for filesystem metadata (and quota is treated as filesystem metadata) and we use bdev page cache for all the metadata. > And if we did that, would it even work as being part of a transaction? In principle it could work because we would then treat quota as journalled data and jbd2 supports that. But honestly, special-casing quota as journalled data IMHO brings more hassle on the write side than it can save by sharing some code on the read side. Honza -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
