On Fri, May 10, 2019 at 8:33 AM Adam Borowski <kilobyte(a)angband.pl> wrote:
Even if allocation is done synchronously, data would be lost except on
actual pmem. Explicit msync()s don't need MAP_SYNC, and don't require
a sync per page.
Signed-off-by: Adam Borowski <kilobyte(a)angband.pl>
---
MAP_SYNC can't be allowed unconditionally, as cacheline flushes don't help
guarantee persistency in page cache. This fixes an error in my earlier
patch "btrfs: allow MAP_SYNC mmap" -- you'd probably want to amend that.
fs/btrfs/file.c | 7 +++++++
1 file changed, 7 insertions(+)
diff --git a/fs/btrfs/file.c b/fs/btrfs/file.c
index 362a9cf9dcb2..0bc5428037ba 100644
--- a/fs/btrfs/file.c
+++ b/fs/btrfs/file.c
@@ -2233,6 +2233,13 @@ static int btrfs_file_mmap(struct file *filp, struct
vm_area_struct *vma)
if (!IS_DAX(inode) && !mapping->a_ops->readpage)
return -ENOEXEC;
+ /*
+ * Normal operation of btrfs is pretty much an antithesis of MAP_SYNC;
+ * supporting it outside DAX is pointless.
+ */
+ if (!IS_DAX(inode) && (vma->vm_flags & VM_SYNC))
+ return -EOPNOTSUPP;
+
If the virtio-pmem patch set goes upstream prior to btrfs-dax support
this will need to switch over to the new daxdev_mapping_supported()
helper.
https://lore.kernel.org/lkml/20190426050039.17460-5-pagupta@redhat.com/