| Message ID | 20210519190900.320044-9-hch@lst.de (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [01/11] xfs: cleanup error handling in xfs_buf_get_map | expand |
On Wed, May 19, 2021 at 09:08:57PM +0200, Christoph Hellwig wrote: > Factor out two helpers that do everything needed for allocating and > freeing pages that back a buffer, and remove the duplication between > the different interfaces. > > Signed-off-by: Christoph Hellwig <hch@lst.de> This seems really confused. Up until this point in the patch set you are pulling code out of xfs_buf_alloc_pages() into helpers. Now you are getting rid of the helpers and putting the slightly modified code back into xfs_buf_alloc_pages(). This doesn't make any sense at all. The freeing helper now requires the buffer state to be manipulated on allocation failure so that the free function doesn't run off the end of the bp->b_pages array. That's a bit of a landmine, and it doesn't really clean anything up much at all. And on the allocation side there is new "fail fast" behaviour because you've lifted the readahead out of xfs_buf_alloc_pages. You also lifted the zeroing checks, which I note that you immediately put back inside xfs_buf_alloc_pages() in the next patch. The stuff up to this point in the series makes sense. From this patch onwards it seems to me that you're just undoing the factoring and cleanups from the first few patches... I mean, like the factoring of xfs_buf_alloc_slab(), you could have just factored out xfs_buf_alloc_pages(bp, page_count) from xfs_buf_allocate_memory() and used that directly in xfs_buf_get_uncached() and avoided a bunch of this factoring, make a slight logic modification and recombine churn. And it would be trivial to do on top of the bulk allocation patch which already converts both of these functions to use bulk allocation.... Cheers, Dave.
On Thu, May 20, 2021 at 09:22:45AM +1000, Dave Chinner wrote: > Up until this point in the patch set you are pulling code out > of xfs_buf_alloc_pages() into helpers. Now you are getting rid of > the helpers and putting the slightly modified code back into > xfs_buf_alloc_pages(). This doesn't make any sense at all. It makes a whole lot of sense, but it seems you don't like the structure :) As stated in the commit log we now have one helper that sets a _XBF_PAGES backing with pages and the map, and one helper to tear it down. I think it makes a whole lot of sense this way. > The freeing helper now requires the buffer state to be > manipulated on allocation failure so that the free function doesn't > run off the end of the bp->b_pages array. That's a bit of a > landmine, and it doesn't really clean anything up much at all. It is something we also do elsewhere in the kernel. Another alternative would be to do a NULL check on the page, or to just pointlessly duplicate the freeing loop. > And on the allocation side there is new "fail fast" behaviour > because you've lifted the readahead out of xfs_buf_alloc_pages. You > also lifted the zeroing checks, which I note that you immediately > put back inside xfs_buf_alloc_pages() in the next patch. This is to clearly split code consolidatation from behavior changes. I could move both earlier at the downside of adding a lot of new code first that later gets removed. > I mean, like the factoring of xfs_buf_alloc_slab(), you could have > just factored out xfs_buf_alloc_pages(bp, page_count) from > xfs_buf_allocate_memory() and used that directly in > xfs_buf_get_uncached() and avoided a bunch of this factoring, make a > slight logic modification and recombine churn. And it would be > trivial to do on top of the bulk allocation patch which already > converts both of these functions to use bulk allocation.... As mentioned in the cover letter: the bulk allocation review is what trigger this as it tripped me following various lose ends. And as usual I'd rather have that kind of change at the end where the surrounding code makes sense, so the rebased version is now is patch 11 of this series.
On Thu, May 20, 2021 at 07:35:04AM +0200, Christoph Hellwig wrote: > On Thu, May 20, 2021 at 09:22:45AM +1000, Dave Chinner wrote: > > Up until this point in the patch set you are pulling code out > > of xfs_buf_alloc_pages() into helpers. Now you are getting rid of > > the helpers and putting the slightly modified code back into > > xfs_buf_alloc_pages(). This doesn't make any sense at all. > > It makes a whole lot of sense, but it seems you don't like the > structure :) > > As stated in the commit log we now have one helper that sets a > _XBF_PAGES backing with pages and the map, and one helper to > tear it down. I think it makes a whole lot of sense this way. I don't like the way the patchset is built. It creates temporary infrastructure, then tears it down again to return the code to almost exactly the same structure that it originally had. In doing this, you change the semantics of functions and helpers multiple times yet, eventually, we end up with the same semantics as we started with. It's much more obvious to factor out the end helpers first, with the exact semantics that the current have and will end up with, and then just convert and clean up the code in and around those helpers. It's much easier to follow and very correct if the function call semnatics and behaviour don't keep changing... > > The freeing helper now requires the buffer state to be > > manipulated on allocation failure so that the free function doesn't > > run off the end of the bp->b_pages array. That's a bit of a > > landmine, and it doesn't really clean anything up much at all. > > It is something we also do elsewhere in the kernel. Another > alternative would be to do a NULL check on the page, or to just > pointlessly duplicate the freeing loop. A null check in the freeing code is much simpler to understand at a glance. It's easy to miss that the error handling only works because callers have a single extra line of code that makes error handling work correctly. This is a bad pattern because it's easy for new code to get it wrong and have nobody notice that it's wrong. > > And on the allocation side there is new "fail fast" behaviour > > because you've lifted the readahead out of xfs_buf_alloc_pages. You > > also lifted the zeroing checks, which I note that you immediately > > put back inside xfs_buf_alloc_pages() in the next patch. > > This is to clearly split code consolidatation from behavior changes. > I could move both earlier at the downside of adding a lot of new > code first that later gets removed. Ah, what new code? factoring out the _alloc_pages() code at the same time as the alloc_kmem() code is the only "new" code that is necessary. Everything else is then consolidation, and this doesn't require repeatedly changing behaviour and moving code out and back into helpers.... > > I mean, like the factoring of xfs_buf_alloc_slab(), you could have > > just factored out xfs_buf_alloc_pages(bp, page_count) from > > xfs_buf_allocate_memory() and used that directly in > > xfs_buf_get_uncached() and avoided a bunch of this factoring, make a > > slight logic modification and recombine churn. And it would be > > trivial to do on top of the bulk allocation patch which already > > converts both of these functions to use bulk allocation.... > > As mentioned in the cover letter: the bulk allocation review is what > trigger this as it tripped me following various lose ends. And as > usual I'd rather have that kind of change at the end where the > surrounding code makes sense, so the rebased version is now is patch 11 > of this series. I've re-written my patch based on this cleanup series. It largely does all the same things, and ends up largely in the same place, but does things in an order that doesn't keep changing behaviour and repeatedly moving the same code around. I'll post it once I've QA'd it. Cheers, Dave.
diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c index 76a107e3cb2a22..31aff8323605cd 100644 --- a/fs/xfs/xfs_buf.c +++ b/fs/xfs/xfs_buf.c @@ -273,35 +273,17 @@ _xfs_buf_alloc( } /* - * Allocate a page array capable of holding a specified number - * of pages, and point the page buf at it. + * Free all pages allocated to the buffer including the page map. */ -STATIC int -_xfs_buf_get_pages( - struct xfs_buf *bp) +static void +xfs_buf_free_pages( + struct xfs_buf *bp) { - ASSERT(bp->b_pages == NULL); - - bp->b_page_count = DIV_ROUND_UP(BBTOB(bp->b_length), PAGE_SIZE); - if (bp->b_page_count > XB_PAGES) { - bp->b_pages = kmem_zalloc(sizeof(struct page *) * - bp->b_page_count, KM_NOFS); - if (!bp->b_pages) - return -ENOMEM; - } else { - bp->b_pages = bp->b_page_array; - } + unsigned int i; - return 0; -} + for (i = 0; i < bp->b_page_count; i++) + __free_page(bp->b_pages[i]); -/* - * Frees b_pages if it was allocated. - */ -STATIC void -_xfs_buf_free_pages( - struct xfs_buf *bp) -{ if (bp->b_pages != bp->b_page_array) { kmem_free(bp->b_pages); bp->b_pages = NULL; @@ -324,22 +306,14 @@ xfs_buf_free( ASSERT(list_empty(&bp->b_lru)); if (bp->b_flags & _XBF_PAGES) { - uint i; - if (xfs_buf_is_vmapped(bp)) vm_unmap_ram(bp->b_addr, bp->b_page_count); - - for (i = 0; i < bp->b_page_count; i++) { - struct page *page = bp->b_pages[i]; - - __free_page(page); - } + xfs_buf_free_pages(bp); if (current->reclaim_state) current->reclaim_state->reclaimed_slab += bp->b_page_count; } else if (bp->b_flags & _XBF_KMEM) kmem_free(bp->b_addr); - _xfs_buf_free_pages(bp); xfs_buf_free_maps(bp); kmem_cache_free(xfs_buf_zone, bp); } @@ -380,34 +354,33 @@ xfs_buf_alloc_slab( static int xfs_buf_alloc_pages( struct xfs_buf *bp, - uint flags) + gfp_t gfp_mask, + bool fail_fast) { - gfp_t gfp_mask = xb_to_gfp(flags); - unsigned short i; - int error; - - /* - * assure zeroed buffer for non-read cases. - */ - if (!(flags & XBF_READ)) - gfp_mask |= __GFP_ZERO; + int i; - error = _xfs_buf_get_pages(bp); - if (unlikely(error)) - return error; + ASSERT(bp->b_pages == NULL); - bp->b_flags |= _XBF_PAGES; + bp->b_page_count = DIV_ROUND_UP(BBTOB(bp->b_length), PAGE_SIZE); + if (bp->b_page_count > XB_PAGES) { + bp->b_pages = kmem_zalloc(sizeof(struct page *) * + bp->b_page_count, KM_NOFS); + if (!bp->b_pages) + return -ENOMEM; + } else { + bp->b_pages = bp->b_page_array; + } for (i = 0; i < bp->b_page_count; i++) { struct page *page; uint retries = 0; retry: page = alloc_page(gfp_mask); - if (unlikely(page == NULL)) { - if (flags & XBF_READ_AHEAD) { + if (unlikely(!page)) { + if (fail_fast) { bp->b_page_count = i; - error = -ENOMEM; - goto out_free_pages; + xfs_buf_free_pages(bp); + return -ENOMEM; } /* @@ -429,13 +402,9 @@ xfs_buf_alloc_pages( bp->b_pages[i] = page; } - return 0; -out_free_pages: - for (i = 0; i < bp->b_page_count; i++) - __free_page(bp->b_pages[i]); - bp->b_flags &= ~_XBF_PAGES; - return error; + bp->b_flags |= _XBF_PAGES; + return 0; } /* @@ -706,7 +675,13 @@ xfs_buf_get_map( */ if (BBTOB(new_bp->b_length) >= PAGE_SIZE || xfs_buf_alloc_slab(new_bp, flags) < 0) { - error = xfs_buf_alloc_pages(new_bp, flags); + gfp_t gfp_mask = xb_to_gfp(flags); + + /* assure a zeroed buffer for non-read cases */ + if (!(flags & XBF_READ)) + gfp_mask |= __GFP_ZERO; + error = xfs_buf_alloc_pages(new_bp, gfp_mask, + flags & XBF_READ_AHEAD); if (error) goto out_free_buf; } @@ -936,7 +911,7 @@ xfs_buf_get_uncached( int flags, struct xfs_buf **bpp) { - int error, i; + int error; struct xfs_buf *bp; DEFINE_SINGLE_BUF_MAP(map, XFS_BUF_DADDR_NULL, numblks); @@ -947,19 +922,10 @@ xfs_buf_get_uncached( if (error) goto fail; - error = _xfs_buf_get_pages(bp); + error = xfs_buf_alloc_pages(bp, xb_to_gfp(flags), true); if (error) goto fail_free_buf; - for (i = 0; i < bp->b_page_count; i++) { - bp->b_pages[i] = alloc_page(xb_to_gfp(flags)); - if (!bp->b_pages[i]) { - error = -ENOMEM; - goto fail_free_mem; - } - } - bp->b_flags |= _XBF_PAGES; - error = _xfs_buf_map_pages(bp, 0); if (unlikely(error)) { xfs_warn(target->bt_mount, @@ -972,9 +938,7 @@ xfs_buf_get_uncached( return 0; fail_free_mem: - while (--i >= 0) - __free_page(bp->b_pages[i]); - _xfs_buf_free_pages(bp); + xfs_buf_free_pages(bp); fail_free_buf: xfs_buf_free_maps(bp); kmem_cache_free(xfs_buf_zone, bp);
Factor out two helpers that do everything needed for allocating and freeing pages that back a buffer, and remove the duplication between the different interfaces. Signed-off-by: Christoph Hellwig <hch@lst.de> --- fs/xfs/xfs_buf.c | 110 ++++++++++++++++------------------------------- 1 file changed, 37 insertions(+), 73 deletions(-)