bcachefs: use kvzalloc() for journal bios

We can write quite large journal entries.

It's possible that JOURNAL_ENTRY_SIZE_MAX is far larger than it needs to
be and should be reduced to something more reasonable (1-2MB), but
performance can be sensitive to anything that affects journal
pipelining - we've had issues there in the past, so we shouldn't change
that without first doing performance testing.

For now, apply the simple fix to deal with the following, which caused a
mount to fail:

[43579.348135] mount.bcachefs: page allocation failure: order:5, mode:0x40dc0(GFP_KERNEL|__GFP_ZERO|__GFP_COMP), nodemask=(null),cpuset=openrc.sshd,mems_allowed=0
[43579.349792] CPU: 3 UID: 0 PID: 14702 Comm: mount.bcachefs Tainted: G     U              6.16.0-rc4 #18 VOLUNTARY
[43579.349798] Tainted: [U]=USER
[43579.349799] Hardware name: MSI MS-7982/B150M PRO-VDH (MS-7982), BIOS 3.H0 07/10/2018
[43579.349800] Call Trace:
[43579.349803]  <TASK>
[43579.349806] dump_stack_lvl (lib/dump_stack.c:122)
[43579.349814] warn_alloc (mm/page_alloc.c:3744)
[43579.349821] ? __alloc_pages_direct_compact (./arch/x86/include/asm/jump_label.h:36 ./include/linux/delayacct.h:211 mm/page_alloc.c:3882)
[43579.349827] __alloc_pages_slowpath.constprop.0 (mm/page_alloc.c:4699)
[43579.349833] __alloc_frozen_pages_noprof (mm/page_alloc.c:4972)
[43579.349838] __alloc_pages_noprof (mm/page_alloc.c:4994)
[43579.349843] ___kmalloc_large_node (./include/linux/gfp.h:284 ./include/linux/gfp.h:311 mm/slub.c:4272)
[43579.349848] __kmalloc_large_noprof (./arch/x86/include/asm/bitops.h:414 ./include/asm-generic/getorder.h:46 mm/slub.c:4292)
[43579.349852] bch2_dev_journal_init (fs/bcachefs/journal.c:1629 (discriminator 4))
[43579.349857] __bch2_dev_attach_bdev (fs/bcachefs/super.c:1588)
[43579.349861] ? kernfs_create_link (fs/kernfs/symlink.c:48)
[43579.349865] bch2_dev_attach_bdev (fs/bcachefs/super.c:1630)
[43579.349868] bch2_fs_open (fs/bcachefs/super.c:2440)
[43579.349874] bch2_fs_get_tree (fs/bcachefs/fs.c:2474)
[43579.349880] vfs_get_tree (fs/super.c:1805)

Reported-by: Marcin Mirosław <[email protected]>
Signed-off-by: Kent Overstreet <[email protected]>

Author: Kent Overstreet

fs/bcachefs/journal.c

@@ -1593,7 +1593,7 @@ void bch2_dev_journal_exit(struct bch_dev *ca)
 	struct journal_device *ja = &ca->journal;
 
 	for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
-		kfree(ja->bio[i]);
+		kvfree(ja->bio[i]);
 		ja->bio[i] = NULL;
 	}
 
@@ -1630,7 +1630,16 @@ int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
 	unsigned nr_bvecs = DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE);
 
 	for (unsigned i = 0; i < ARRAY_SIZE(ja->bio); i++) {
-		ja->bio[i] = kzalloc(struct_size(ja->bio[i], bio.bi_inline_vecs,
+		/*
+		 * kvzalloc() is not what we want to be using here:
+		 * JOURNAL_ENTRY_SIZE_MAX is probably quite a bit bigger than it
+		 * needs to be.
+		 *
+		 * But changing that will require performance testing -
+		 * performance can be sensitive to anything that affects journal
+		 * pipelining.
+		 */
+		ja->bio[i] = kvzalloc(struct_size(ja->bio[i], bio.bi_inline_vecs,
 				     nr_bvecs), GFP_KERNEL);
 		if (!ja->bio[i])
 			return bch_err_throw(c, ENOMEM_dev_journal_init);