On Mon, Sep 09, 2002 at 09:09:36PM -0400, Haijin Yan wrote: > > > > > > IIRC, it points to the list of buffer heads describing the disk > > > location of the data resident in the page. Buffer heads are the > > > structs used to schedule traditional (read()/write()) disk > > > I/O, whereas page->mapping is used for mmap() I/O. So a > > > "buffer" is just a logical block of data within a page > > > that correponds to a physical disk block somewhere; and > > > the buffer head is a description of that relationship. In the > > > older 2.4 kernels, they were complimentary: when a mmap()ed page > > > had to be written out, the VM created buffers describing the page > > > and then handed those buffers off to the disk I/O subsystem. I'm > > > not sure if things work the same now. (Probably so.) > > > > Each page in page cache belongs to an inode or is anonymous. But > > page_launder() writes the pages out independently of what they belong > > to. So it one cache. Mappings are there so page_launder knows where to > > write the page (it calls down to writepage method in mapping). > > > > With buffers, IIRC it's still the same way. When a page has to be > > written to disk, it's turned into buffers and these are queued for disk > > write. > > > > Are you guys saying that pages for traditional I/O(read/write) do not > have page->mapping while clean mmap()ed pages do not have page->buffer? No. I was trying to say exactly otherwise. Every page that's from a file _on disk_ or anonymous and swap in on disk will have buffers when it's dirty. No matter if it became dirty via write or mmap. > As far as I know, a page(not anonymous) with its mapping points to > mapping of inode of the file in page_cache_read() function like this: > > inode = file->f_dentry->d_inode; > struct address_space *mapping = inode->i_mapping; > page = page_cache_alloc(mapping); > page->mapping = mapping; > > when page got swapped, it's mapping is set to > page->mapping = &swapper_space; > > so when is the condition (!page->mapping) true and what it represents for? When _anonymous_ page get swapped, it get page->mapping = &swapper_space. When page is from a file, it has it's mapping being in memory or not. I looked in the source and it seems, that setting swapper_space is "lazy" in some circumstances, ie. the page has NULL mapping until we actually want to swap it the first time, but I am not sure. > If page->buffers is true, Does that ony mean either page is brought in by > traditional IO or a disk block is allocated for this mmap()ed page for > written(because page is dirty)? Well, it means there is some block device IO to be done on the page. mmap()ing a page does not matter for this purpose. mmap()ing just creates PTE for the page, but does not affect how IO works on it. When a page is written using write, buffers are created. When it's written using mmap, buffers are created when kernel walks PTEs and notices the PTE is dirty. ------------------------------------------------------------------------------- Jan 'Bulb' Hudec <bulb@ucw.cz> -- Kernelnewbies: Help each other learn about the Linux kernel. Archive: http://mail.nl.linux.org/kernelnewbies/ FAQ: http://kernelnewbies.org/faq/
