aboutsummaryrefslogtreecommitdiffstats
path: root/package
diff options
context:
space:
mode:
authorDaniel Golle <daniel@makrotopia.org>2020-08-10 23:33:00 +0100
committerDaniel Golle <daniel@makrotopia.org>2020-08-29 01:17:26 +0100
commit5a1e4a7fdb2d732a485afb7cdcfc359e15ac259e (patch)
tree948c68284b340f3bdf7e7140148aabe7d14cf134 /package
parent19b8696dd7e6e1c4c2a66bbfe32df08b24ed0ca6 (diff)
downloadupstream-5a1e4a7fdb2d732a485afb7cdcfc359e15ac259e.tar.gz
upstream-5a1e4a7fdb2d732a485afb7cdcfc359e15ac259e.tar.bz2
upstream-5a1e4a7fdb2d732a485afb7cdcfc359e15ac259e.zip
oxnas: reduce size of ATA DMA descriptor space
After years of trying to find the reason for random kernel crashes while both CPU and SATA are under load it has been found. Some odd commented-out #defines in kref's single-port driver [1] which were copied from the vendor driver made me develop a theory: The IO-mapped memory area for DMA descriptors apparetly got some holes just before the alignment boundaries. This feels like an off-by-one bug in the hardware or maybe those fields are used internally by the SATA controller's firmware. Whatever the cause is: they cannot be used and trying to use them results in reading back unexpected stuff and ends up with oopsing Unable to handle kernel paging request at virtual address d085c004 Work around the issue by reducing the area used for bmdma descriptors. This reduces SATA performance (iops) quite a bit, but finally makes things work reliably. Possibly one could optimize this much more by really just skipping the holes in that memory area -- however, that seems to be non-trivial with the driver and libata in it's current form (suggestions are welcome). The 'proper' way to have good SATA performance would be to make use of the hardware RAID features (one can use the JBOD mode to access even just a single disc transparently through the RAID controller integrated in the SATA host instead of accessing the SATA ports 'raw' as we do now). [1]: https://github.com/kref/linux-oxnas/blob/master/drivers/ata/sata_oxnas.c#L25 Signed-off-by: Daniel Golle <daniel@makrotopia.org> (cherry picked from commit 5793112f751ee3d9f841af4846d68e6b1ff1bff4, including fixup commit d75e75306301852a848824cf268d8b58eda28a8a)
Diffstat (limited to 'package')
0 files changed, 0 insertions, 0 deletions