On 9/22/20 5:12 AM, Matthew Wilcox wrote: > This is a fun little race. > > Dramatis personae: Pages P0, P1 are consecutive and aligned. > Threads A, B, C. > > Page P0 is allocated to the page cache. > Page P1 is free. > > Thread A calls find_get_entry() > P0 is returned from xas_load() > > Thread B removes the page from the page cache (eg truncate, invalidatepage). > P0 is buddy-merged with P1. > > Thread C calls alloc_pages, order 1, does not specify GFP_COMP. P0 now > has refcount 1. > > Thread A calls page_cache_get_speculative(). P0 has refcount 2. > > Thread C calls __free_page(P0, 1) > put_page_testzero is _false_. Do not call free_the_page(). > > Thread A calls put_page(P0) > We free P0 and nobody knows to free P1. > > > Weird solution: In __free_page(), if put_page_testzero() fails and page > is not PageHead, convert it to a compound page. Then the put_page() > by Thread A will free P1. I can imagine doing the conversion in a manner that deals with races properly will be rather tricky... > Better ideas? IMHO, alloc_pages() with order > 0 and without __GFP_COMP is a weird beast. In that case it would be probably best to refcount each base page separately. I don't know how many assumptions this would break :/