On Sun, Mar 3, 2024 at 4:28 PM Rick Macklem <rick.mack...@gmail.com> wrote: > > On Sun, Mar 3, 2024 at 3:27 PM Rick Macklem <rick.mack...@gmail.com> wrote: > > > > On Sun, Mar 3, 2024 at 1:17 PM Rick Macklem <rick.macklem@gmailcom> wrote: > > > > > > On Sat, Mar 2, 2024 at 8:28 PM Garrett Wollman <woll...@bimajority.org> > > > wrote: > > > > > > > > > > > > I wrote previously: > > > > > PID TID COMM TDNAME KSTACK > > > > > 997 108481 nfsd nfsd: master mi_switch > > > > > sleepq_timedwait _sleep nfsv4_lock nfsrvd_dorpc nfssvc_program > > > > > svc_run_internal svc_run nfsrvd_nfsd nfssvc_nfsd sys_nfssvc > > > > > amd64_syscall fast_syscall_common > > > > > 997 960918 nfsd nfsd: service mi_switch > > > > > sleepq_timedwait _sleep nfsv4_lock nfsrv_setclient nfsrvd_exchangeid > > > > > nfsrvd_dorpc nfssvc_program svc_run_internal svc_thread_start > > > > > fork_exit fork_trampoline > > > > > 997 962232 nfsd nfsd: service mi_switch _cv_wait > > > > > txg_wait_synced_impl txg_wait_synced dmu_offset_next zfs_holey > > > > > zfs_freebsd_ioctl vn_generic_copy_file_range vop_stdcopy_file_range > > > > > VOP_COPY_FILE_RANGE vn_copy_file_range nfsrvd_copy_file_range > > > > > nfsrvd_dorpc nfssvc_program svc_run_internal svc_thread_start > > > > > fork_exit fork_trampoline > > > > > > > > I spent some time this evening looking at this last stack trace, and > > > > stumbled across the following comment in > > > > sys/contrib/openzfs/module/zfs/dmu.c: > > > > > > > > | /* > > > > | * Enable/disable forcing txg sync when dirty checking for holes with > > > > lseek(). > > > > | * By default this is enabled to ensure accurate hole reporting, it > > > > can result > > > > | * in a significant performance penalty for lseek(SEEK_HOLE) heavy > > > > workloads. > > > > | * Disabling this option will result in holes never being reported in > > > > dirty > > > > | * files which is always safe. > > > > | */ > > > > | int zfs_dmu_offset_next_sync = 1; > > > > > > > > I believe this explains why vn_copy_file_range sometimes takes much > > > > longer than a second: our servers often have lots of data waiting to > > > > be written to disk, and if the file being copied was recently modified > > > > (and so is dirty), this might take several seconds. I've set > > > > vfs.zfs.dmu_offset_next_sync=0 on the server that was hurting the most > > > > and am watching to see if we have more freezes. > > > > > > > > If this does the trick, then I can delay deploying a new kernel until > > > > April, after my upcoming vacation. > > > Interesting. Please let us know how it goes. > > Btw, I just tried this for my trivial test and it worked very well. > > A 1Gbyte file was cpied in two Copy RPCs of 1sec and slightly less than > > 1sec. > Oops, I spoke too soon. > The Copy RPCs worked fine (as above) but the Commit RPCs took > a long time, so it still looks like you may need the patches. And I should mention that my test is done on a laptop without a ZIL, so maybe a ZIL on a separate device might generate different results.
rick > > rick > > > > > So, your vacation may be looking better, rick > > > > > > > > And enjoy your vacation, rick > > > > > > > > > > > -GAWollman > > > >
