lkml.org 
[lkml]   [2022]   [Jul]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [RFC PATCH] fs/isofs: Replace kmap() with kmap_local_page()
On Mon 25-07-22 18:23:31, Fabio M. De Francesco wrote:
> The use of kmap() is being deprecated in favor of kmap_local_page().
>
> There are two main problems with kmap(): (1) It comes with an overhead as
> mapping space is restricted and protected by a global lock for
> synchronization and (2) it also requires global TLB invalidation when the
> kmap’s pool wraps and it might block when the mapping space is fully
> utilized until a slot becomes available.
>
> With kmap_local_page() the mappings are per thread, CPU local, can take
> page faults, and can be called from any context (including interrupts).
> Tasks can be preempted and, when scheduled to run again, the kernel
> virtual addresses are restored and still valid. It is faster than kmap()
> in kernels with HIGHMEM enabled.
>
> Since kmap_local_page() can be safely used in compress.c, it should be
> called everywhere instead of kmap().
>
> Therefore, replace kmap() with kmap_local_page() in compress.c. Where it
> is needed, use memzero_page() instead of open coding kmap_local_page()
> plus memset() to fill the pages with zeros. Delete the redundant
> flush_dcache_page() in the two call sites of memzero_page().
>
> This is an RFC because these changes have not been tested (tests are
> welcome!), therefore I'm not entirely sure whether these conversions work
> properly. I'd like to hear comments from more experienced developers
> before sending a real patch. Suggestions about how to run tests would
> also be much appreciated.
>
> Suggested-by: Ira Weiny <ira.weiny@intel.com>
> Signed-off-by: Fabio M. De Francesco <fmdefrancesco@gmail.com>

What you propose makes sense to me. But the lack of testing is troublesome.
You can always at least test your patch without highmem on isofs image with
compression (mkisofs seems to support creating such filesystems). Even that
would detect a bug in your patch ;) - see below.

> diff --git a/fs/isofs/compress.c b/fs/isofs/compress.c
> index 95a19f25d61c..a1562124bb91 100644
> --- a/fs/isofs/compress.c
> +++ b/fs/isofs/compress.c
> @@ -120,8 +119,7 @@ static loff_t zisofs_uncompress_block(struct inode *inode, loff_t block_start,
> zerr != Z_STREAM_END) {
> if (!stream.avail_out) {
> if (pages[curpage]) {
> - stream.next_out = page_address(pages[curpage])
> - + poffset;
> + stream.next_out = kmap_local_page(pages[curpage] + poffset);

This is wrong. Most importantly because you need to add 'poffset' to the
final address provided by kmap_local_page(), not to the struct page pointer.

Secondly please wrap the line to fit into 80 chars.

> stream.avail_out = PAGE_SIZE - poffset;
> poffset = 0;
> } else {
> @@ -170,6 +168,12 @@ static loff_t zisofs_uncompress_block(struct inode *inode, loff_t block_start,
> }
> }
>
> + if (stream.next_out)
> + if (stream.next_out != (char *)zisofs_sink_page) {
> + kunmap_local(stream.next_out);
> + stream.next_out = NULL;
> + }
> +

This looks buggy as well. If we mapped page above, we'll unmap it here even
if stream.avail_out > 0 and we want to still write to it. I think you
should unmap the page here only if stream.avail_out == 0 and we are going
to switch to the next page...

> if (!stream.avail_out) {
> /* This page completed */
> if (pages[curpage]) {
> @@ -183,6 +187,9 @@ static loff_t zisofs_uncompress_block(struct inode *inode, loff_t block_start,
> }
> inflate_out:
> zlib_inflateEnd(&stream);
> + if (stream.next_out)
> + if (stream.next_out != (char *)zisofs_sink_page)
> + kunmap_local(stream.next_out);

This is correct but I'd simplify it to:

if (stream.next_out && stream.next_out != (char *)zisofs_sink_page)
kunmap_local(stream.next_out);

Honza
--
Jan Kara <jack@suse.com>
SUSE Labs, CR

\
 
 \ /
  Last update: 2022-07-26 16:51    [W:0.066 / U:0.576 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site