Document special rule when using RWF_ATOMIC flag in conjunction with O_DIRECT (which is the only time which it can be used). Some block devices have a virtual boundary, which is a boundary at which write iovec vectors start and end addresses need to align to - see virt_boundary_mask in linux kernel sysfs-block Doc. To avoid splitting writes with torn-write protection in the kernel for when vectors cross this boundary, special iovec boundary rules are added such that we don't cross this boundary. Even though a virt_boundary_mask may be larger than the CPU page size, we just impose that as a reasonable limit, which matches what the linux kernel NVMe driver uses today - NVMe is of special interest for supporting RWF_ATOMIC. Signed-off-by: John Garry <john.g.garry@xxxxxxxxxx> --- man2/open.2 | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/man2/open.2 b/man2/open.2 index 4c921723c95c..e5b9d107859a 100644 --- a/man2/open.2 +++ b/man2/open.2 @@ -1782,6 +1782,19 @@ blockdev \-\-getss .EE .in .PP +When +.B O_DIRECT +is used in conjunction with +.BR pwritev2 () +and +.BR RWF_ATOMIC +flag, in addition to any alignment rules already imposed by the filesystem and +underlying block device, it must be ensured that +.I iovec +vectors have no gaps such that the end alignment of a vector must have the same +start alignment of any subsequent vector and that alignment must be least +at a page boundary. +.PP .B O_DIRECT I/Os should never be run concurrently with the .BR fork (2) -- 2.31.1
