On Sat, Sep 05, 2026 at 09:43:05AM -0500, Timothy Pearson wrote: > After upgrading a POWER9 pSeries guest on a PowerNV host from Linux 6.1.0-18 > to 6.12.94, rebuilding a degraded RAID10 array behind an LSI SAS3008 > controller passed through via VFIO proceeds at ~1MB/s with 100% CPU load on > one core. Booting the same guest with the old 6.1 series kernel and resuming > the rebuild of the same array proceeds at over 200MB/s with minimal CPU load. > For completeness, the default 6.12 readahead settings were adjusted back > down to the same values as the 6.1 kernel, with no improvement observed. > > No variation in the regression was found between host kernel 6.12 and host > kernel 6.18. Since there are no other reports of similar issues from amd64 > users, it is presumed this is a POWER-specific regression. >
Hi Timothy, We suspect this degradation is caused by the spurious interrupt problem you reported. I had posted a fix for that problem [1], would it be possible for you to see if this problem persists with my patch applied? [1]: https://lore.kernel.org/linuxppc-dev/[email protected]/ Thanks, Gautam
