Skip to contentSkip to conversation

synchronous readpage for buffer_heads

RFC In review Matthew Wilcox 2 participants 3 messages Oct 22, 2020 – Oct 23, 2020

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

RFC 1/1 synchronous readpage for buffer_heads

I'm working on making readpage synchronous so that it can actually return errors instead of futilely setting PageError. Something that's common between most of the block based filesystems is the need to submit N I/Os and wait for them to all complete (some filesystems don't support sub-page block size, so they don't have this problem).

I ended up coming up with a fairly nice data structure which I've called the blk_completion. It waits for 'n' events to happen, then wakes the task that cares, unless the task has got bored and wandered off to do something else.

block_read_full_page() then uses this data structure to submit 'n' buffer heads and wait for them to all complete. The fscrypt code doesn't work in this scheme, so I bailed on that for now. I have ideas for fixing it, but they can wait.

Obviously this all needs documentation, but I'd like feedback on the idea before I do that. I have given it some light testing, but there aren't too many filesystems left that use block_read_full_page() so I haven't done a proper xfstests run.

include/linux/bio.h +11 -0
index f254bc79bb3a..0bde05f5548c 100644
@@ -814,4 +814,15 @@ static inline void bio_set_polled(struct bio *bio, struct kiocb *kiocb)
 		bio->bi_opf |= REQ_NOWAIT;
 }
 
+struct blk_completion {
+	struct task_struct *cmpl_task;
+	spinlock_t cmpl_lock;
+	int cmpl_count;
+	blk_status_t cmpl_status;
+};
+
+void blk_completion_init(struct blk_completion *, int n);
+int blk_completion_sub(struct blk_completion *, blk_status_t status, int n);
+int blk_completion_wait_killable(struct blk_completion *);
+
 #endif /* __LINUX_BIO_H */
block/blk-core.c +61 -0

67 lines not on this page · raw message

fs/buffer.c +100 -14

162 lines not on this page · raw message

#1 lore raw
Quoted 16 lines · Matthew Wilcox

I'm working on making readpage synchronous so that it can actually return errors instead of futilely setting PageError. Something that's common between most of the block based filesystems is the need to submit N I/Os and wait for them to all complete (some filesystems don't support sub-page block size, so they don't have this problem).

I ended up coming up with a fairly nice data structure which I've called the blk_completion. It waits for 'n' events to happen, then wakes the task that cares, unless the task has got bored and wandered off to do something else.

block_read_full_page() then uses this data structure to submit 'n' buffer heads and wait for them to all complete. The fscrypt code doesn't work in this scheme, so I bailed on that for now. I have ideas for fixing it, but they can wait.

Quoted 5 lines

Obviously this all needs documentation, but I'd like feedback on the idea before I do that. I have given it some light testing, but there aren't too many filesystems left that use block_read_full_page() so I haven't done a proper xfstests run.

Hmm. You are aware, of course, that hch et al are working on replacing bhs with iomap, right? So wouldn't it be more useful to concentrate on the iomap code, and ensure that _that_ is working correctly?

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer
#2 lore raw
Quoted 3 lines · Hannes Reinecke

On 10/22/20 5:22 PM, Matthew Wilcox wrote:
Hmm. You are aware, of course, that hch et al are working on replacing bhs with iomap, right?

$ git shortlog --author=Wilcox origin/master -- fs/iomap |head -1
Matthew Wilcox (Oracle) (17):

But actually, I don't see anyone working on a mass migration of filesystems from either using BHs directly or using the mpage code to using iomap. I have a summer student for next summer who I'm going to let loose on this problem, but I fear buffer_heads will be with us for a long time to come.

I mean, who's going to convert reiserfs to iomap?

$ git log --no-merges --since=2015-01-01 origin/master fs/reiserfs |grep -c ^comm
130

Not exactly a thriving development community. It doesn't even support fallocate.

Quoted 2 lines

So wouldn't it be more useful to concentrate on the iomap code, and ensure that _that_ is working correctly?

Did that one first, then did mpage_readpage(), now I've moved on to block_read_full_page(). Now I'm going to go back and redo iomap with everything I learned doing block_read_full_page(). It's going to use blk_completion.