On Mon, Sep 27, 2021 at 2:58 PM Jane Chu <jane.chu@xxxxxxxxxx> wrote: > > On 9/27/2021 2:07 PM, Dave Chinner wrote: > > On Thu, Sep 23, 2021 at 06:35:16PM -0700, Darrick J. Wong wrote: > >> On Thu, Sep 23, 2021 at 06:21:19PM -0700, Jane Chu wrote: > >>> > >>> On 9/23/2021 6:18 PM, Dan Williams wrote: > >>>> On Thu, Sep 23, 2021 at 3:54 PM Dave Chinner <david@xxxxxxxxxxxxx> wrote: > >>>>> > >>>>> On Wed, Sep 22, 2021 at 10:42:11PM -0700, Dan Williams wrote: > >>>>>> On Wed, Sep 22, 2021 at 7:43 PM Dan Williams <dan.j.williams@xxxxxxxxx> wrote: > >>>>>>> > >>>>>>> On Wed, Sep 22, 2021 at 6:42 PM Dave Chinner <david@xxxxxxxxxxxxx> wrote: > >>>>>>> [..] > >>>>>>>> Hence this discussion leads me to conclude that fallocate() simply > >>>>>>>> isn't the right interface to clear storage hardware poison state and > >>>>>>>> it's much simpler for everyone - kernel and userspace - to provide a > >>>>>>>> pwritev2(RWF_CLEAR_HWERROR) flag to directly instruct the IO path to > >>>>>>>> clear hardware error state before issuing this user write to the > >>>>>>>> hardware. > >>>>>>> > >>>>>>> That flag would slot in nicely in dax_iomap_iter() as the gate for > >>>>>>> whether dax_direct_access() should allow mapping over error ranges, > >>>>>>> and then as a flag to dax_copy_from_iter() to indicate that it should > >>>>>>> compare the incoming write to known poison and clear it before > >>>>>>> proceeding. > >>>>>>> > >>>>>>> I like the distinction, because there's a chance the application did > >>>>>>> not know that the page had experienced data loss and might want the > >>>>>>> error behavior. The other service the driver could offer with this > >>>>>>> flag is to do a precise check of the incoming write to make sure it > >>>>>>> overlaps known poison and then repair the entire page. Repairing whole > >>>>>>> pages makes for a cleaner implementation of the code that tries to > >>>>>>> keep poison out of the CPU speculation path, {set,clear}_mce_nospec(). > >>>>>> > >>>>>> This flag could also be useful for preadv2() as there is currently no > >>>>>> way to read the good data in a PMEM page with poison via DAX. So the > >>>>>> flag would tell dax_direct_access() to again proceed in the face of > >>>>>> errors, but then the driver's dax_copy_to_iter() operation could > >>>>>> either read up to the precise byte offset of the error in the page, or > >>>>>> autoreplace error data with zero's to try to maximize data recovery. > >>>>> > >>>>> Yes, it could. I like the idea - say RWF_IGNORE_HWERROR - to read > >>>>> everything that can be read from the bad range because it's the > >>>>> other half of the problem RWF_RESET_HWERROR is trying to address. > >>>>> That is, the operation we want to perform on a range with an error > >>>>> state is -data recovery-, not "reinitialisation". Data recovery > >>>>> requires two steps: > >>>>> > >>>>> - "try to recover the data from the bad storage"; and > >>>>> - "reinitialise the data and clear the error state" > >>>>> > >>>>> These naturally map to read() and write() operations, not > >>>>> fallocate(). With RWF flags they become explicit data recovery > >>>>> operations, unlike fallocate() which needs to imply that "writing > >>>>> zeroes" == "reset hardware error state". While that reset method > >>>>> may be true for a specific pmem hardware implementation it is not a > >>>>> requirement for all storage hardware. It's most definitely not a > >>>>> requirement for future storage hardware, either. > >>>>> > >>>>> It also means that applications have no choice in what data they can > >>>>> use to reinitialise the damaged range with because fallocate() only > >>>>> supports writing zeroes. If we've recovered data via a read() as you > >>>>> suggest we could, then we can rebuild the data from other redundant > >>>>> information and immediately write that back to the storage, hence > >>>>> repairing the fault. > >>>>> > >>>>> That, in turn, allows the filesystem to turn the RWF_RESET_HWERROR > >>>>> write into an exclusive operation and hence allow the > >>>>> reinitialisation with the recovered/repaired state to run atomically > >>>>> w.r.t. all other filesystem operations. i.e. the reset write > >>>>> completes the recovery operation instead of requiring separate > >>>>> "reset" and "write recovered data into zeroed range" steps that > >>>>> cannot be executed atomically by userspace... > >>>> > >>>> /me nods > >>>> > >>>> Jane, want to take a run at patches for this ^^^? > >>>> > >>> > >>> Sure, I'll give it a try. > >>> > >>> Thank you all for the discussions! > >> > >> Cool, thank you! > > > > I'd like to propose a slight modification to the API: a single RWF > > flag called RWF_RECOVER_DATA. On read, this means the storage tries > > to read all the data it can from the range, and for the parts it > > can't read data from (cachelines, sectors, whatever) it returns as > > zeroes. > > > > On write, this means the errors over the range get cleared and the > > user data provided gets written over the top of whatever was there. > > Filesystems should perform this as an exclusive operation to that > > range of the file. > > > > That way we only need one IOCB_RECOVERY flag, and for communicating > > with lower storage layers (e.g. dm/md raid and/or hardware) only one > > REQ_RECOVERY flag is needed in the bio. > > > > Thoughts? > > Sounds cleaner. Dan, your thoughts? I like it. I was struggling with a way to unify the flag names, and "recovery" is a good term for not surprising the caller with zeros. I.e. don't use this flow to avoid errors, use this flow to maximize data recovery.