iomap: Do not use GFP_NORETRY to allocate BIOs
Subsystems (2)
PATCH 1/1 iomap: Do not use GFP_NORETRY to allocate BIOs
From: "Matthew Wilcox (Oracle)" <[email protected]>
If we use GFP_NORETRY, we have to be able to handle failures, and it's tricky to handle failure here. Other implementations of ->readpages do not attempt to handle BIO allocation failures, so this is no worse.
--- fs/iomap/buffered-io.c | 2 -- 1 file changed, 2 deletions(-)
fs/iomap/buffered-io.c +0 -2
@@ -307,8 +307,6 @@ iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,if (ctx->bio)submit_bio (ctx->bio); -if (ctx->is_readahead)/* same as readahead_gfp_mask */ - gfp |= __GFP_NORETRY | __GFP_NOWARN; ctx->bio =bio_alloc (gfp,min (BIO_MAX_PAGES , nr_vecs)); ctx->bio->bi_opf =REQ_OP_READ ;if (ctx->is_readahead)
Signature · 1 line
2.25.1
Quoted 5 lines · Matthew Wilcox
From: "Matthew Wilcox (Oracle)" <[email protected]>
If we use
GFP_NORETRY, we have to be able to handle failures, and it's tricky to handle failure here. Other implementations of ->readpages do not attempt to handle BIO allocation failures, so this is no worse.
do_mpage_readpage tries to use it, I guess that is wher I copied it from..
But I don't think it is a bad idea, so the patch itself looks good:
It probably wants a fixes tag, though:
Quoted 9 lines · Christoph Hellwig
Quoted 5 lines · Matthew Wilcox
From: "Matthew Wilcox (Oracle)" <[email protected]>
If we use
GFP_NORETRY, we have to be able to handle failures, and it's tricky to handle failure here. Other implementations of ->readpages do not attempt to handle BIO allocation failures, so this is no worse.
do_mpage_readpagetries to use it, I guess that is wher I copied it from..
Oh, I see that now. It uses readahead_gfp_mask(), and I was grepping for GFP_NORETRY so I didn't spot it. It falls back to block_read_full_page() which we can't do. That will allocate smaller BIOs, so there's an argument that we should do the same. How about this:
+++ b/fs/iomap/buffered-io.c
@@ -302,6 +302,7 @@ iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
if (!ctx->bio || !is_contig || bio_full(ctx->bio, plen)) {
gfp_t gfp = mapping_gfp_constraint(page->mapping, GFP_KERNEL);
+ gfp_t orig_gfp = gfp;
int nr_vecs = (length + PAGE_SIZE - 1) >> PAGE_SHIFT;
if (ctx->bio)
@@ -310,6 +311,8 @@ iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,
if (ctx->is_readahead) /* same as readahead_gfp_mask */
gfp |= __GFP_NORETRY | __GFP_NOWARN;
ctx->bio = bio_alloc(gfp, min(BIO_MAX_PAGES, nr_vecs));
+ if (!ctx->bio)
+ ctx->bio = bio_alloc(orig_gfp, 1);
ctx->bio->bi_opf = REQ_OP_READ;
if (ctx->is_readahead)
ctx->bio->bi_opf |= REQ_RAHEAD;Quoted 4 lines · Matthew Wilcox
Oh, I see that now. It uses
readahead_gfp_mask(), and I was grepping forGFP_NORETRYso I didn't spot it. It falls back toblock_read_full_page()which we can't do. That will allocate smaller BIOs, so there's an argument that we should do the same. How about this:
That looks silly to me. This just means we'll keep iterating over small bios for readahead.. Either we just ignore the different gfp mask, or we need to go all the way and handle errors, although that doesn't really look nice.
Quoted 10 lines · Christoph Hellwig
Quoted 4 lines · Matthew Wilcox
Oh, I see that now. It uses
readahead_gfp_mask(), and I was grepping forGFP_NORETRYso I didn't spot it. It falls back toblock_read_full_page()which we can't do. That will allocate smaller BIOs, so there's an argument that we should do the same. How about this:That looks silly to me. This just means we'll keep iterating over small bios for readahead.. Either we just ignore the different gfp mask, or we need to go all the way and handle errors, although that doesn't really look nice.
I'm not sure it's silly, although I'd love to see bio_alloc() support nr_iovecs == 0 meaning "allocate me any size biovec and tell me what size I got in ->bi_max_vecs". By allocating a small biovec this time, we do one allocation rather than two, and maybe by the time we come to allocate the next readahead bio, kswapd will have succeeded in freeing up more memory for us.
Quoted 6 lines · Matthew Wilcox
Quoted 4 lines
That looks silly to me. This just means we'll keep iterating over small bios for readahead.. Either we just ignore the different gfp mask, or we need to go all the way and handle errors, although that doesn't really look nice.
I'm not sure it's silly,
Oh well, I'm not going to be in the way of fixing a bug I added. So feel free to go ahead with this and mention it matches mpage_readpages.
Quoted 6 lines
although I'd love to see
bio_alloc()supportnr_iovecs== 0 meaning "allocate me any size biovec and tell me what size I got in ->bi_max_vecs". By allocating a small biovec this time, we do one allocation rather than two, and maybe by the time we come to allocate the next readahead bio, kswapd will have succeeded in freeing up more memory for us.
Sounds easy enough - especially as callers don't need to look at bi_max_vecs anyway, that is the job of bio_add_page and friends. That being said an upper bound still sounds useful - no need to allocate a a gigantic bio if we know we only need a few pages.
Quoted 10 lines · Christoph Hellwig
Quoted 6 lines · Matthew Wilcox
Quoted 4 lines
That looks silly to me. This just means we'll keep iterating over small bios for readahead.. Either we just ignore the different gfp mask, or we need to go all the way and handle errors, although that doesn't really look nice.
I'm not sure it's silly,
Oh well, I'm not going to be in the way of fixing a bug I added. So feel free to go ahead with this and mention it matches
mpage_readpages.
Are you going to submit this as a formal patch?
Quoted 7 lines · Matthew Wilcox
From: "Matthew Wilcox (Oracle)" <[email protected]>
If we use
GFP_NORETRY, we have to be able to handle failures, and it's tricky to handle failure here. Other implementations of ->readpages do not attempt to handle BIO allocation failures, so this is no worse.Signed-off-by Matthew Wilcox (Oracle) <[email protected]>
This fixes the regression I saw, though I see what looks like a new patch forming further down in this thread, so please let me know if you'd prefer I take this change, the other one, or both.
--D
Quoted 20 lines
--- fs/iomap/buffered-io.c | 2 -- 1 file changed, 2 deletions(-)fs/iomap/buffered-io.c +0
-2@@ -307,8 +307,6 @@ iomap_readpage_actor(struct inode *inode, loff_t pos, loff_t length, void *data,if (ctx->bio)submit_bio (ctx->bio); -if (ctx->is_readahead)/* same as readahead_gfp_mask */ - gfp |= __GFP_NORETRY | __GFP_NOWARN; ctx->bio =bio_alloc (gfp,min (BIO_MAX_PAGES , nr_vecs)); ctx->bio->bi_opf =REQ_OP_READ ;if (ctx->is_readahead)Signature · 1 line
2.25.1