Hi folks, After I proposed that we use active references to the perag to be able to gate shrink removing AGs and hence perags safely, it was obvious that we've got a fair bit of work to do actually use perags in all the places we need to. There's a lot of code right now that iterates ag numbers and then looks up perags from that, often multiple times for the same perag in the one operation. IF we want to use reference counted perags for access control, then we nee dto convert all these uses to perag iterators, not agno iterators. This patchset does not include any of the active/passive reference counting needed for shrink gating - we have to get perags in use in all the palces we need first before that will work effectively, and that's what this patchset starts to address. It's also been clear as I've been doing these conversions that having a perag available in places that are doing AG specific work allows for significant cleanups and optimisations to be made. One such example is fleshed out in this patch (inode allocation), but there are many more if we do things like start moving AG geometry information into the perag. This means we no longer need to run a calculation to determine what the size of the AG is, which is important because the verify functions consume a large amount of CPU doing exactly this sort of check on block and inode numbers throughout the code. It also leads to repeated patterns where we have a perag in hand before we have to read an AGI or AGF buffer to lock the AG for the operation we are about to perform. There are many optimisations on both the buffer caching and AG locking strategies that we can build on from this. e.g. moving AGI/AGF locking into the pag rather than using the buffer lock, doing pag+agbno based buffer cache lookups instead of daddr based lookups that then have to look up the pag, etc. IOWs, this turns a lot of the code we have on it's head and there's significant potential for code simplification and algorithmic optimisations to be made as a result. A lot of this sort of thing will be medium term work rather than done up front - shrink is the initial priority, so widespread conversion comes first. [Patches 1-4] The first step of this is consolidating all the perag management - init, free, get, put, etc into a common location. THis is spread all over the place right now, so move it all into libxfs/xfs_ag.[ch]. This does expose kernel only bits of the perag to libxfs and hence userspace, so the structures and code is rearranged to minimise the number of ifdefs that need to be added to the userspace codebase. The perag iterator in xfs_icache.c is promoted to a first class API and expanded to the needs of the code as required. [Patches 5-10] These are the first basic perag iterator conversions and changes to pass the perag down the stack from those iterators where appropriate. A lot of this is obvious, simple changes, though in some places we stop passing the perag down the stack because the code enters into an as yet unconverted subsystem that still uses raw AGs. [Patches 11-16] These replace the agno passed in the btree cursor for per-ag btree operations with a perag that is passed to the cursor init function. The cursor takes it's own reference to the perag, and the reference is dropped when the cursor is deleted. Hence we get reference coverage for the entire time the cursor is active, even if the code that initialised the cursor drops it's reference before the cursor or any of it's children (duplicates) have been deleted. The first patch adds the perag infrastructure for the cursor, the next four patches convert a btree cursor at a time, and the last removes the agno from the cursor once it is unused. [Patches 17-21] These patches are a demonstration of the simplifications and cleanups that come from plumbing the perag through interfaces that select and then operate on a specific AG. In this case the inode allocation algorithm does up to three walks across all AGs before it either allocates an inode or fails. Two of these walks are purely just to select the AG, and even then it doesn't guarantee inode allocation success so there's a third walk if the selected AG allocation fails. These patches collapse the selection and allocation into a single loop, simplifies the error handling because xfs_dir_ialloc() always returns ENOSPC if no AG was selected for inode allocation or we fail to allocate an inode in any AG, gets rid of xfs_dir_ialloc() wrapper, converts inode allocation to run entirely from a single perag instance, and then factors xfs_dialloc() into a much, much simpler loop which is easy to understand. Hence we end up with the same inode allocation logic, but it only needs two complete iterations at worst, makes AG selection and allocation atomic w.r.t. shrink and chops out out over 100 lines of code from this hot code path. [Patch 22] Converts the unlink path to pass perags through it. There's more conversion work to be done, but this patchset gets through a large chunk of it in one hit. Most of the iterators are converted, so once this is solidified we can move on to converting these to active references for being able to free perags while the fs is still active. Indeed, this allows more than just shrink - if we can safely detect a perag is unreferenced and take it out of service, we have the infrastructure we need to be able to implement a memory shrinker for perags. That is a big step towards supporting extremely large numbers of AGs in the filesystem - we can't really support millions of AGs in a filesystem if they must all be loading into memory at all times. We can already do demand based initialisation of perags, but we cannot do memory pressure based reclaim. Reference counting for shrink gives us the necessary capability for demand based reclaim of perags.... This approach solves more than one problem we really need to solve, and hence I think it's worth making this scope of changes now to support shrink operations.... Thoughts, comments, welcome.. Cheers, Dave.
