[parisc-linux] Bug in flush_dcache_page?

[Matthew Wilcox]

On Wed, Oct 31, 2001 at 05:39:05PM +0000, Matthew Wilcox wrote:
> static inline void flush_dcache_page(struct page *page)
> {
> 	if (!page->mapping || (page->mapping && !page->mapping->i_mmap &&
> 			!page->mapping->i_mmap_shared)) {
> 		set_bit(PG_dcache_dirty, &page->flags);
> 	} else {
> 		struct vm_area_struct *vma;
> 		for (vma = page->mapping->i_mmap; vma; vma = vma->vm_next) {
> 			flush_user_page(vma->vm_mm, vma->vm_start + page->index);
> 		}
> 		if (page->mapping->i_mmap_shared) {
> 			vma = page->mapping->i_mmap_shared;
> 			flush_user_page(vma->vm_mm, vma->vm_start + page->index);
> 		}
> 		flush_kernel_dcache_page(page_address(page));
> 	}
> }

Aha.  A couple of things from talking to various VM/VFS people...
1. page->mapping is guaranteed to exist for pages in the page cache --
   this is how they're located.
2. If a page is simultaneously mapped in both user and kernel space, and
   user space touches it, either it leaves the page cache or it was mapped
   shared.

flush_dcache_page is only called on pages in the page cache, so we can
assume that page->mapping exists, and we don't need to flush any user
mappings of this page.

	if (!page->mapping->i_mmap & !page->mapping->i_mmap_shared) {
		set_bit(PG_dcache_dirty, &page->flags);
	} else {
		struct vm_area_struct *vma = page->mapping->i_mmap_shared;
		if (vma) {
			flush_user_page(vma->vm_mm, vma->vm_start + page->index);
		}
		flush_kernel_dcache_page(page_address(page));
	}

-- 
Revolutions do not require corporate support.