| Message ID | 20210429173315.1252465-1-jxgao@google.com |
|---|---|
| Headers | show |
| Series | preserve DMA offsets when using swiotlb | expand |
On Thu, Apr 29, 2021 at 10:34 AM Jianxiong Gao <jxgao@google.com> wrote: > > We observed several NVMe failures when running with SWIOTLB. The root > cause of the issue is that when data is mapped via SWIOTLB, the address > offset is not preserved. Several device drivers including the NVMe > driver relies on this offset to function correctly. > > Even though we discovered the error when running using AMD SEV, we have > reproduced the same error in Rhel 8 without SEV. By adding swiotlb=force > option to the boot command line parameter, NVMe funcionality is > impacted. For example formatting a disk into xfs format returns an > error. Christoph, are you OK with backporting this patch set to LTS, based on the rationale in the cover letter above? Thanks, Marc
On Tue, May 04, 2021 at 01:22:14PM -0700, Marc Orr wrote: > Christoph, are you OK with backporting this patch set to LTS, based on > the rationale in the cover letter above? I don't object, but I really do not have time to review it in detail.
On Thu, Apr 29, 2021 at 05:33:06PM +0000, Jianxiong Gao wrote: > We observed several NVMe failures when running with SWIOTLB. The root > cause of the issue is that when data is mapped via SWIOTLB, the address > offset is not preserved. Several device drivers including the NVMe > driver relies on this offset to function correctly. > > Even though we discovered the error when running using AMD SEV, we have > reproduced the same error in Rhel 8 without SEV. By adding swiotlb=force > option to the boot command line parameter, NVMe funcionality is > impacted. For example formatting a disk into xfs format returns an > error. > > > ---- > Changes in v2: > Rebased patches to 5.10.33 It looks like if I were to take these now, we need to also have a version for 5.11.y because you can not upgrade from an older kernel and have a "regression" like this, right? 5.11.y will still be alive for at least a week or so, let me see if your backports work there or not... thanks, greg k-h
On Wed, May 05, 2021 at 09:23:28AM +0200, Greg KH wrote: > On Thu, Apr 29, 2021 at 05:33:06PM +0000, Jianxiong Gao wrote: > > We observed several NVMe failures when running with SWIOTLB. The root > > cause of the issue is that when data is mapped via SWIOTLB, the address > > offset is not preserved. Several device drivers including the NVMe > > driver relies on this offset to function correctly. > > > > Even though we discovered the error when running using AMD SEV, we have > > reproduced the same error in Rhel 8 without SEV. By adding swiotlb=force > > option to the boot command line parameter, NVMe funcionality is > > impacted. For example formatting a disk into xfs format returns an > > error. > > > > > > ---- > > Changes in v2: > > Rebased patches to 5.10.33 > > It looks like if I were to take these now, we need to also have a > version for 5.11.y because you can not upgrade from an older kernel and > have a "regression" like this, right? > > 5.11.y will still be alive for at least a week or so, let me see if your > backports work there or not... Ok, looks like it worked, now queued up. greg k-h
We observed several NVMe failures when running with SWIOTLB. The root cause of the issue is that when data is mapped via SWIOTLB, the address offset is not preserved. Several device drivers including the NVMe driver relies on this offset to function correctly. Even though we discovered the error when running using AMD SEV, we have reproduced the same error in Rhel 8 without SEV. By adding swiotlb=force option to the boot command line parameter, NVMe funcionality is impacted. For example formatting a disk into xfs format returns an error. ---- Changes in v2: Rebased patches to 5.10.33 Updated patch description to correct format. Jianxiong Gao (9): driver core: add a min_align_mask field to struct device_dma_parameters swiotlb: add a IO_TLB_SIZE define swiotlb: factor out an io_tlb_offset helper swiotlb: factor out a nr_slots helper swiotlb: clean up swiotlb_tbl_unmap_single swiotlb: refactor swiotlb_tbl_map_single swiotlb: don't modify orig_addr in swiotlb_tbl_sync_single swiotlb: respect min_align_mask nvme-pci: set min_align_mask drivers/nvme/host/pci.c | 1 + include/linux/device.h | 1 + include/linux/dma-mapping.h | 16 +++ include/linux/swiotlb.h | 1 + kernel/dma/swiotlb.c | 259 ++++++++++++++++++++---------------- 5 files changed, 164 insertions(+), 114 deletions(-)