On Jun 29, 2018, at 2:46 AM, Lukas Czerner <lczerner@xxxxxxxxxx> wrote: > > On Thu, Jun 28, 2018 at 09:05:04PM -0600, Andreas Dilger wrote: >> >>> On Jun 28, 2018, at 7:45 PM, Theodore Y. Ts'o <tytso@xxxxxxx> wrote: >>> >>> On Thu, Jun 28, 2018 at 06:07:47PM -0600, Andreas Dilger wrote: >>>>> But does an RDMA operation actually do a block allocation? Really? >>>>> And if it is willing to do a block allocation, why is it not willing >>>>> to bump i_size? >>>> >>>> It's not that the RDMA does block allocation, but rather that the RDMA >>>> always transfers and writes the full PAGE_SIZE of data, even if i_size >>>> is less than the end of the last block. This simplifies the RDMA code >>>> so that it can always write the data instead of having to stop at i_size. > > Every time we write past i_size we need to extend the i_size, why is > this case different ? > >>> >>> Right. So there are two choices: >>> >>> 1) Keep the blocks beyond i_size marked as uninitialized. You >>> transfer and write the full PAGE_SIZE of data, but it simply will >>> never be available to the user. > > Yes, that's for extent mapped files. > >>> >>> 2) Zero the page, write it out to the file, and then extend i_size and >>> mark the extents as uninitialized. > > Except at that point you do not really need to mark the extent as > unitialized, the blocks are allocated and written to and i_size is > extended. That's how it needs to be done for indirect block mapped > files. > >> >> The end of the page would already be zeroed before write. >> >>> Why is it that Lustre is choosing to keep i_size where it is, but to >>> mark the blocks beyond it as initialized? >> >> This isn't about initialized vs. uninitialized extents. It is only about >> allocated vs. unallocated blocks, possibly with block-mapped files. There >> is no way to have uninitialized blocks with a block-mapped file. >> >> The code is checking whether there are any blocks allocated beyond i_size, >> and if there are, without the patch it considers i_size broken and extends >> it to the end of the last allocated block. The patch allows a small number >> of blocks to be allocated beyond i_size without triggering this heuristic. > > I do not think that ext4 has any capacity to allocate initialized blocks > beyond i_size without actually writing to it and if we write beyong i_size > then we need to extend i_size so if we see anything like that it's a problem > that needs fixing. I think it's as simple as that. > > Unless you can show me where in upstream ext4 we can get to this > situation I am strongly for fixing e2fsck. But maybe I am missing > something. Like I said previously, this is done with Lustre, which has a different IO submission path than stock ext4. I don't think there is any requirement that this only be in upstream ext4, since e2fsprogs also has code to support running on BSD, Windows, even Hurd. >> The only difference vs. the previous code is that it correctly calculates >> what the PAGE_SIZE aligned block number is (the old code assumed that lblock >> was the base-1 block number instead of the base-0 block number that it is). > > Right, but the test is broken when PAGE_SIZE > 4k that's how I found > out in the first place. See me previous email about f_eofblocks and > f_pgsize_gt_blksize failures. I already have a patch for the f_pgsize_gt_blksize failure that should work with both 4KB and 64KB (and other) PAGE_SIZE systems, since it is generating the filesystem itself. I probably need to make a script for f_eofblocks so that it skips the test on PAGE_SIZE=64KB or uses a different expect.1 file. Cheers, Andreas
Attachment:
signature.asc
Description: Message signed with OpenPGP
