Uploaded image for project: 'Lustre'
  1. Lustre
  2. LU-16138

remove need for T10-PI block integrity patches from server kernel

    XMLWordPrintable

Details

    • Improvement
    • Resolution: Unresolved
    • Major
    • Lustre 2.17.0
    • None
    • 9223372036854775807

    Description

      Kernels since v4.3-rc4-30-g0f8087ecdeac separate the block integrity verify and generate functions into a separate struct blk_integrity_profile instead of copying the pointers into struct blk_integrity. This might be used to facilitate replacing the blk_integrity_profile pointer on the block device for the duration that osd-ldiskfs is mounting it, and potentially avoid the need to patch the kernel for T10-PI at all.

      As a starting point, we would need to call blk_integrity_register() using a template holding the integrity methods provided by osd-ldiskfs - osd_dif_type<1,3>generate<crc,ip>() and osd_dif_type<1,3>verify<crc,ip>(). It looks like this function is just copying the individual fields from the provided template into the disk queue's blk_integrity fields, so we should save the original blk_integrity to be restored when osdldiskfs unmounts the target, otherwise the kernel would crash trying to access the old function pointers if there is any block device IO after osd-ldiskfs is removed from the kernel.

      It seems prudent to call blk_integrity_unregister() before unloading the module, as that would force all block device IO using the osd-ldiskfs.ko methods to be finished, and then blk_integrity_register() the saved values so that T10-PI is still functional on the device.

      The osd_dif_generate() and osd_dif_verify() functions need to handle being called on block read/write calls that are unrelated to osd-ldiskfs BRW IOs, since that will now happen for every block in the filesystem whether or not it came from osd-ldiskfs. One way to separate OSD vs. non-OSD IO might be by putting a magic at the end of struct osd_bio_private that can be used to reliably detect if this is a BRW IO or if it is IO on some other ldiskfs block. Non-BRW blocks would return NULL from find_lnb() and not dereference bio_private if NULL, and it looks like the rest of the code will "just work".

      Attachments

        Issue Links

          Activity

            People

              dongyang Dongyang Li
              adilger Andreas Dilger
              Votes:
              0 Vote for this issue
              Watchers:
              7 Start watching this issue

              Dates

                Created:
                Updated: