notroj opened a new pull request, #76:
URL: https://github.com/apache/apr/pull/76

   Two changes for the memory growth reported in httpd [PR 
70167](https://bz.apache.org/bugzilla/show_bug.cgi?id=70167), where mod_ssl 
splits a FILE bucket at an offset which is not page-aligned, and the httpd 
child's memory then grows in proportion to the size of the file sent.
   
   Trunk already page-aligns the offset in `apr_mmap_create()`, so the growth 
as reported does not reproduce here; it needs APR 1.7.x with APR-util 1.6.x. 
Both changes are written to be backported to those branches.
   
   ### Avoid allocating from the pool when `apr_mmap_create()` fails
   
   `apr_mmap_create()` allocated the `apr_mmap_t` before attempting the 
mapping, so every failed call left memory in the caller's pool. A FILE bucket 
retries the mmap for each read, so a persistent failure consumed pool memory 
for each read of the file. The structure is now allocated only once the file 
has been mapped, for both the Unix and Windows implementations.
   
   The Windows change has not been compiled or tested.
   
   ### Read a FILE bucket up to a page boundary before using mmap
   
   A FILE bucket which does not start at a page-aligned offset is now read only 
as far as the next page boundary, so the remainder is mapped from a 
page-aligned offset. This applies only if:
   
   - mmap is enabled for the bucket;
   - the bucket is larger than the read buffer size, so could not be read in a 
single read anyway;
   - the remainder beyond the page boundary is at least `APR_MMAP_THRESHOLD`.
   
   The page size comes from `sysconf(_SC_PAGESIZE)`; where that is not 
available the behaviour is unchanged.
   
   One cost on trunk: a bucket between the read buffer size and 
`APR_MMAP_LIMIT` at an unaligned offset was previously a single mmap, and is 
now one short read followed by one mmap.
   
   ### Tests
   
   - `test_mmap_create_fail` (testmmap): a failed `apr_mmap_create()` does not 
allocate from the pool. The pool check uses `apr_pool_num_bytes()`, so it is 
only effective in a pool-debug build.
   - `test_mmap_unaligned` (testbuckets): a FILE bucket at an unaligned offset, 
larger than `APR_MMAP_LIMIT`, is read up to the page boundary and then as MMAP 
buckets.
   - `test_mmap_fail` (testbuckets): a file which cannot be mapped is read 
correctly via the fallback path.
   
   The first two fail without the corresponding source change (the first in a 
pool-debug build). The full test suite passes on Linux x86_64.
   
   ### Not covered
   
   - Each successfully mapped chunk still leaves one `apr_mmap_t` in the pool, 
about 14 bytes per MB read.
   - Neither change has been tested on the 1.x branches or end-to-end with 
httpd.
   
   🤖 Generated with [Claude Code](https://claude.com/claude-code)
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to