Date: Tue, 7 Feb 2023 05:47:30 +0200 From: Konstantin Belousov <kostikbel@gmail.com> To: Alan Somers <asomers@freebsd.org> Cc: Rick Macklem <rick.macklem@gmail.com>, FreeBSD CURRENT <freebsd-current@freebsd.org> Subject: Re: RFC: Should fspacectl() commit changes to stable storage? Message-ID: <Y%2BHJ0jENs/KiuFqz@kib.kiev.ua> In-Reply-To: <CAOtMX2hCUp3e9qBMqkSGS6mpwMig173OFQ90XNboE%2BK52e_F0Q@mail.gmail.com> References: <CAM5tNy6Dkv==TceNF5n02wY20-rEAgn_8S%2BtKNwv6iia1qqLYA@mail.gmail.com> <CAOtMX2hCUp3e9qBMqkSGS6mpwMig173OFQ90XNboE%2BK52e_F0Q@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, Feb 06, 2023 at 06:59:59PM -0700, Alan Somers wrote: > On Mon, Feb 6, 2023 at 6:23 PM Rick Macklem <rick.macklem@gmail.com> wrote: > > > > PR#269328 reports an issue related to fspacectl() being > > mixed with mmap'd I/O. > > > > When working on a fix for this for the NFS client, I realized that > > "man fspacectl" does not clarify if the deallocation should commit > > changes to stable storage before returning success. > > vop_stddeallocate() currently does not do this so, if a machine > > crashes immediately after fspacectl() returns success, the hole > > may not be there when the machine reboots. > > > > For POSIX writes, it is clear that recently written data may be > > lost when a machine crashes/reboots and fsync(2) is used to > > ensure the data is on stable storage and won't be lost when a > > crash/reboot occurs. > > > > The question is "Should fsync(2) be required to ensure a hole > > punched by fspacectl(2) is not lost or should it be guaranteed > > to not be lost when fspacectl(2) returns? > > Since fspacectl(2) is FreeBSD specific and there is no standard, > > it could be defined either way. > > > > So, what do you think? rick > > It think it should be just like write. An fsync should be required to > ensure that the effects will be visible after a reboot. I agree. It is easy for an application to do fsync() after fspacectl() to get synchronous behavior. But if we make fspacectl() synchronous, application cannot revert the behavior to the speedy one. As Alan said, it should be same as write(). This does not preclude some filesystems to choose stronger guarantees, of course.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Y%2BHJ0jENs/KiuFqz>