Christoph, On 2019/06/26 15:10, Christoph Hellwig wrote: > On Wed, Jun 26, 2019 at 10:47:57AM +0900, Damien Le Moal wrote: >> @@ -1501,9 +1502,14 @@ struct bio *bio_map_kern(struct request_queue *q, void *data, unsigned int len, >> unsigned long end = (kaddr + len + PAGE_SIZE - 1) >> PAGE_SHIFT; >> unsigned long start = kaddr >> PAGE_SHIFT; >> const int nr_pages = end - start; >> + bool is_vmalloc = is_vmalloc_addr(data); >> + struct page *page; >> int offset, i; >> struct bio *bio; >> >> + if (is_vmalloc) >> + invalidate_kernel_vmap_range(data, len); > > That is not correct. > > The submission path needs an unconditional flush_kernel_vmap_range call, > and the read completion path will additionally need the > invalidate_kernel_vmap_range call. > I mimicked what XFS and DM do with vmalloc-ed buffers. I guess I missed something. So in this case, the allocation is in sd_zbc.c, where the completion is too. So I think it may be better to have flush_kernel_vmap_range() right after the allocation before scsi_execute_req() is called and do the invalidate_kernel_vmap_range() before scanning the report zones output for transformation into struct blk_zone ? And do not do anything in bio_map_kern beside the change from virt_to_page() to vmalloc_to_page() ? -- Damien Le Moal Western Digital Research
