[dedicated box testing] ABD2 + Illumos #4950 + Illumos #2605 (master January 16th, 2015) #4236
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
zfsolinux currently uses vmalloc backed slab for ARC buffers. There are some major problems with this approach. One is that 32-bit system have only a handful of vmalloc space. Another is that the fragmentation in slab will easily deplete memory in busy system. With ABD, we use scatterlist to allocate data buffers. In this approach we can allocate in HIGHMEM, which alleviates vmalloc space pressure on 32-bit. Also, we don't have to rely on slab, so there's no fragmentation issue. But for metadata buffers, we still uses linear buffer from slab. The reason for this is that there are a lot of *_phys pointers directly point to metadata buffers. Thus, it would be much more complicated to use scatter buffer for metadata. Currently, ABD is not enabled and its API will treat them as normal buffers. We will enable it once all relevant code is modified to use the API. Signed-off-by: Chunwei Chen <[email protected]>
Modify/Add incremental fletcher function prototype to match abd_iterate_rfunc callback type. Also, reduce duplicated code a bit in zfs_fletcher.c. Signed-off-by: Chunwei Chen <[email protected]>
1. Use abd_t in arc_buf_t->b_data, dmu_buf_t->db_data, zio_t->io_data and zio_transform_t->zt_orig_data 2. zio_* function take abd_t for data Signed-off-by: Chunwei Chen <[email protected]>
1. Add checksum function for abd_t 2. Use abd_t version checksum function in zio_checksum_table 3. Make zio_checksum_compute and zio_checksum_error handle abd_t Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
…d zil.c Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
Use ABD API on related pointers and functions.(b_data, db_data, zio_*(), etc.) Suggested-by: DHE <[email protected]> Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
Currently, abd_uiomove repeatedly calls abd_copy_*_off. The problem is that it will need to do abd_miter_advance repeatedly over the parts that were skipped before. We split out the miter part of the abd_copy_*_off into abd_miter_copy_*. These new function will take miter directly and they will automatically advance it after finish. We initialize an miter in uiomove and use the iterator copy functions to solve the stated problem. Signed-off-by: Chunwei Chen <[email protected]>
The check is needed to make sure the user buffer is indeed in user space. Also
change copy_{to,from}_user to __copy_{to,from}_user so that we don't
repeatedly call access_ok.
Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
When we aren't allocating in HIGHMEM, we can try to allocate contiguous pages, we can also use sg_alloc_table_from_pages to merge adjacent pages for us. This will allow more efficient cache prefetch and also reduce sg iterator overhead. And this has been tested to greatly improve performance. Signed-off-by: Jinshan Xiong <[email protected]> Signed-off-by: Chunwei Chen <[email protected]>
This patch adds non-highmem scatter ABD and also adds a new set of functions, namely abd_buf_segment and friends, for accessing those ABD. This is a preparation for metadata scatter ABD. Signed-off-by: Chunwei Chen <[email protected]>
Some metadata types are not required or not easily to use scatter ABD. So to allow ARC to accommodate both types of metadata, we add a new flag ARC_FLAG_META_SCATTER to indicate which type the buffer belongs. We also introduce a new type arc_buf_alloc_t, which is basically arc_buf_contents_t with a flag indicating scatter metadata. Users can pass it to arc_buf_alloc to decide which buffer type to allocate. Signed-off-by: Chunwei Chen <[email protected]>
We add a new member ot_scatter to dmu_ot to determine the ABD type for each DMU type. We temporary set them all to FALSE, until some types are ready to handle scatter abd. BP_GET_BUFA_TYPE and DBUF_GET_BUFA_TYPE will returns with the scatter flag set up accordingly, so they can be passed to arc_buf_alloc, and the ARC subsystem will returns the correct ABD type. Signed-off-by: Chunwei Chen <[email protected]>
Use non-highmem scatter ABD for indirect block, i.e. level > 0. abd_array should be used to access the blkptr. Signed-off-by: Chunwei Chen <[email protected]>
Use non-highmem scatter ABD for dnode array blocks, abd_array should be used to access the dnodes. Signed-off-by: Chunwei Chen <[email protected]>
abd_alloc_scatter always allocate nr*PAGE_SIZE memory. So in order to save
memory, we allow it to fallback to do linear allocation when size is less than
PAGE_SIZE.
Note that orignally, we requires that zio->io_data remains the same type
before it reaches lower level, vdev_{queue,cache,disk}. After this patch,
however, transformation like compression may change a scatter io_data to
linear. This is fine, because every scatter ABD aware API can also take linear
ABD, but not vice versa. So we relax the type check in push transform to check
linear only.
Signed-off-by: Chunwei Chen <[email protected]>
In dsl_scan_scrub_cb and spa_load_verify_cb, we originally always allocated linear ABD. Now we try to allocate scatter ABD according to the BUFA type of the blkptr to reduce unnecessary spl slab allocation. Also in zio_ddt_read_start, we match the parent zio->io_data ABD type for the same reason. Signed-off-by: Chunwei Chen <[email protected]>
SPA history is access purely through dmu_read/dmu_write. It doesn't require any modification except for setting ot_scatter. Signed-off-by: Chunwei Chen <[email protected]>
Add abd version byteswap function with the name "abd_<old bswap func name>". Note that abd_byteswap_uint*_array and abd_dnode_buf_byteswap can handle scatter buffer, so now we don't need extra borrow/copy. Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Chunwei Chen <[email protected]>
Update arc_c under a mutex from openzfs#4197 ABD2 openzfs#3441 ABD: linear/scatter dual typed buffer for ARC (ver 2) 4950 openzfs#4207 Illumos openzfs#4950 files sometimes can't be removed from a full filesystem 2605 openzfs#4213 Illumos 2605 want to resume interrupted zfs send
kernelOfTruth
changed the title
|
Closing as stale. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
big blob patch with
Update arc_c under a mutex from #4197
ABD2 #3441 ABD: linear/scatter dual typed buffer for ARC (ver 2)
4950 #4207 Illumos #4950 files sometimes can't be removed from a full filesystem
2605 #4213 Illumos 2605 want to resume interrupted zfs send
edit:
ABD2 is of course below the blob and NOT included - but not changing the commit message, it's already TEST'ing 👍