Messages in this thread | | | Subject | Re: [PATCH v3 6/6] btrfs: Use larger zlib buffer for s390 hardware compression | From | Josef Bacik <> | Date | Mon, 6 Jan 2020 11:40:58 -0500 |
| |
On 1/3/20 5:33 PM, Mikhail Zaslonko wrote: > In order to benefit from s390 zlib hardware compression support, > increase the btrfs zlib workspace buffer size from 1 to 4 pages (if > s390 zlib hardware support is enabled on the machine). This brings up > to 60% better performance in hardware on s390 compared to the PAGE_SIZE > buffer and much more compared to the software zlib processing in btrfs. > In case of memory pressure fall back to a single page buffer during > workspace allocation. > > Signed-off-by: Mikhail Zaslonko <zaslonko@linux.ibm.com> > --- > fs/btrfs/compression.c | 2 +- > fs/btrfs/zlib.c | 128 ++++++++++++++++++++++++++++++----------- > 2 files changed, 94 insertions(+), 36 deletions(-) > > diff --git a/fs/btrfs/compression.c b/fs/btrfs/compression.c > index ee834ef7beb4..6bd0e75a822c 100644 > --- a/fs/btrfs/compression.c > +++ b/fs/btrfs/compression.c > @@ -1285,7 +1285,7 @@ int btrfs_decompress_buf2page(const char *buf, unsigned long buf_start, > /* copy bytes from the working buffer into the pages */ > while (working_bytes > 0) { > bytes = min_t(unsigned long, bvec.bv_len, > - PAGE_SIZE - buf_offset); > + PAGE_SIZE - (buf_offset % PAGE_SIZE)); > bytes = min(bytes, working_bytes); > > kaddr = kmap_atomic(bvec.bv_page); > diff --git a/fs/btrfs/zlib.c b/fs/btrfs/zlib.c > index a6c90a003c12..159486801631 100644 > --- a/fs/btrfs/zlib.c > +++ b/fs/btrfs/zlib.c > @@ -20,9 +20,12 @@ > #include <linux/refcount.h> > #include "compression.h" > > +#define ZLIB_DFLTCC_BUF_SIZE (4 * PAGE_SIZE) > + > struct workspace { > z_stream strm; > char *buf; > + unsigned long buf_size; > struct list_head list; > int level; > }; > @@ -61,7 +64,17 @@ struct list_head *zlib_alloc_workspace(unsigned int level) > zlib_inflate_workspacesize()); > workspace->strm.workspace = kvmalloc(workspacesize, GFP_KERNEL); > workspace->level = level; > - workspace->buf = kmalloc(PAGE_SIZE, GFP_KERNEL); > + workspace->buf = NULL; > + if (zlib_deflate_dfltcc_enabled()) { > + workspace->buf = kmalloc(ZLIB_DFLTCC_BUF_SIZE, > + __GFP_NOMEMALLOC | __GFP_NORETRY | > + __GFP_NOWARN | GFP_NOIO); > + workspace->buf_size = ZLIB_DFLTCC_BUF_SIZE; > + } > + if (!workspace->buf) { > + workspace->buf = kmalloc(PAGE_SIZE, GFP_KERNEL); > + workspace->buf_size = PAGE_SIZE; > + } > if (!workspace->strm.workspace || !workspace->buf) > goto fail; > > @@ -78,6 +91,7 @@ int zlib_compress_pages(struct list_head *ws, struct address_space *mapping, > unsigned long *total_in, unsigned long *total_out) > { > struct workspace *workspace = list_entry(ws, struct workspace, list); > + int i; > int ret; > char *data_in; > char *cpage_out; > @@ -85,6 +99,7 @@ int zlib_compress_pages(struct list_head *ws, struct address_space *mapping, > struct page *in_page = NULL; > struct page *out_page = NULL; > unsigned long bytes_left; > + unsigned long in_buf_pages; > unsigned long len = *total_out; > unsigned long nr_dest_pages = *out_pages; > const unsigned long max_out = nr_dest_pages * PAGE_SIZE; > @@ -102,9 +117,6 @@ int zlib_compress_pages(struct list_head *ws, struct address_space *mapping, > workspace->strm.total_in = 0; > workspace->strm.total_out = 0; > > - in_page = find_get_page(mapping, start >> PAGE_SHIFT); > - data_in = kmap(in_page); > - > out_page = alloc_page(GFP_NOFS | __GFP_HIGHMEM); > if (out_page == NULL) { > ret = -ENOMEM; > @@ -114,12 +126,48 @@ int zlib_compress_pages(struct list_head *ws, struct address_space *mapping, > pages[0] = out_page; > nr_pages = 1; > > - workspace->strm.next_in = data_in; > + workspace->strm.next_in = workspace->buf; > + workspace->strm.avail_in = 0; > workspace->strm.next_out = cpage_out; > workspace->strm.avail_out = PAGE_SIZE; > - workspace->strm.avail_in = min(len, PAGE_SIZE); > > while (workspace->strm.total_in < len) { > + /* get next input pages and copy the contents to > + * the workspace buffer if required > + */ > + if (workspace->strm.avail_in == 0) { > + bytes_left = len - workspace->strm.total_in; > + in_buf_pages = min(DIV_ROUND_UP(bytes_left, PAGE_SIZE), > + workspace->buf_size / PAGE_SIZE); > + if (in_buf_pages > 1) { > + for (i = 0; i < in_buf_pages; i++) { > + if (in_page) { > + kunmap(in_page); > + put_page(in_page); > + } > + in_page = find_get_page(mapping, > + start >> PAGE_SHIFT); > + data_in = kmap(in_page); > + memcpy(workspace->buf + i*PAGE_SIZE, > + data_in, PAGE_SIZE); > + start += PAGE_SIZE; > + }
Is there a reason to leave the last in_page mapped here? I realize we'll clean it up further down, but since we're copying everything into the buf anyway why not just map->copy->unmap for everything?
> + workspace->strm.next_in = workspace->buf; > + } else { > + if (in_page) { > + kunmap(in_page); > + put_page(in_page); > + } > + in_page = find_get_page(mapping, > + start >> PAGE_SHIFT); > + data_in = kmap(in_page); > + start += PAGE_SIZE; > + workspace->strm.next_in = data_in; > + } > + workspace->strm.avail_in = min(bytes_left, > + workspace->buf_size); > + } > + > ret = zlib_deflate(&workspace->strm, Z_SYNC_FLUSH); > if (ret != Z_OK) { > pr_debug("BTRFS: deflate in loop returned %d\n", > @@ -136,6 +184,7 @@ int zlib_compress_pages(struct list_head *ws, struct address_space *mapping, > ret = -E2BIG; > goto out; > } > +
Extra newline. Thanks,
Josef
| |