XFS uses kmalloc() to allocate sector sized IO buffer. Turns out buffer allocated via kmalloc(sector sized) can't be guaranteed to be 512 byte aligned, and actually slab only provides ARCH_KMALLOC_MINALIGN alignment, even though it is observed that the sector size allocation is often 512 byte aligned. When KASAN or other memory debug options are enabled, the allocated buffer becomes not aliged with 512 byte any more. This unalgined IO buffer causes at least two issues: 1) some storage controller requires IO buffer to be 512 byte aligned, and data corruption is observed 2) loop/dio requires the IO buffer to be logical block size aligned, and loop's default logcial block size is 512 byte, then one xfs image can't be mounted via loop/dio any more. Use page_frag_alloc() to allocate the sector sized buffer, then the above issue can be fixed because offset_in_page of allocated buffer is always sector aligned. Not see any regression with this patch on xfstests. Cc: Jens Axboe <axboe@xxxxxxxxx> Cc: Vitaly Kuznetsov <vkuznets@xxxxxxxxxx> Cc: Dave Chinner <dchinner@xxxxxxxxxx> Cc: Darrick J. Wong <darrick.wong@xxxxxxxxxx> Cc: Dave Chinner <dchinner@xxxxxxxxxx> Cc: Christoph Hellwig <hch@xxxxxx> Cc: Alexander Duyck <alexander.h.duyck@xxxxxxxxxxxxxxx> Cc: Aaron Lu <aaron.lu@xxxxxxxxx> Cc: Christopher Lameter <cl@xxxxxxxxx> Cc: Linux FS Devel <linux-fsdevel@xxxxxxxxxxxxxxx> Cc: linux-mm@xxxxxxxxx Cc: linux-block@xxxxxxxxxxxxxxx Link: https://marc.info/?t=153734857500004&r=1&w=2 Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx> --- fs/xfs/xfs_buf.c | 21 ++++++++++++++++++--- 1 file changed, 18 insertions(+), 3 deletions(-) diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c index 4f5f2ff3f70f..92b8cdf5e51c 100644 --- a/fs/xfs/xfs_buf.c +++ b/fs/xfs/xfs_buf.c @@ -340,12 +340,27 @@ xfs_buf_free( __free_page(page); } } else if (bp->b_flags & _XBF_KMEM) - kmem_free(bp->b_addr); + page_frag_free(bp->b_addr); _xfs_buf_free_pages(bp); xfs_buf_free_maps(bp); kmem_zone_free(xfs_buf_zone, bp); } +static DEFINE_PER_CPU(struct page_frag_cache, xfs_frag_cache); + +static void *xfs_alloc_frag(int size) +{ + struct page_frag_cache *nc; + void *data; + + preempt_disable(); + nc = this_cpu_ptr(&xfs_frag_cache); + data = page_frag_alloc(nc, size, GFP_ATOMIC); + preempt_enable(); + + return data; +} + /* * Allocates all the pages for buffer in question and builds it's page list. */ @@ -368,7 +383,7 @@ xfs_buf_allocate_memory( */ size = BBTOB(bp->b_length); if (size < PAGE_SIZE) { - bp->b_addr = kmem_alloc(size, KM_NOFS); + bp->b_addr = xfs_alloc_frag(size); if (!bp->b_addr) { /* low memory - use alloc_page loop instead */ goto use_alloc_page; @@ -377,7 +392,7 @@ xfs_buf_allocate_memory( if (((unsigned long)(bp->b_addr + size - 1) & PAGE_MASK) != ((unsigned long)bp->b_addr & PAGE_MASK)) { /* b_addr spans two pages - use alloc_page instead */ - kmem_free(bp->b_addr); + page_frag_free(bp->b_addr); bp->b_addr = NULL; goto use_alloc_page; } -- 2.9.5
