Hi Dan,
On Thu, Aug 29, 2019 at 11:51:27PM +0800, Gao Xiang wrote:
> Hi Dan,
>
> On Thu, Aug 29, 2019 at 06:43:46PM +0300, Dan Carpenter wrote:
> > > p.s. There are 2947 (un)likely places in fs/ directory.
> >
> > I was complaining about you adding new pointless ones, not existing
> > ones. The likely/unlikely annotations are supposed to be functional and
> > not decorative. I explained this very clearly.
>
> I don't think that is mostly pointless. I think it has functional use
> because all error handling paths are rarely happened, or can you remove
> the unlikely in IS_ERR as well?
>
> >
> > Probably most of the annotations in fs/ are wrong but they are also
> > harmless except for the slight messiness. However there are definitely
> > some which are important so removing them all isn't a good idea.
> >
> > > If you like, I will delete them all.
> >
> > But for erofs, I don't think that any of the likely/unlikely calls have
> > been thought about so I'm fine with removing all of them in one go.
>
> Maybe some misuse but rare, I will show you case by case. Wait a minute.
Anyway, I'm fine to delete them all if you like, but I think majority of these
are meaningful.
data.c- /* page is already locked */
data.c- DBG_BUGON(PageUptodate(page));
data.c-
data.c: if (unlikely(err))
data.c- SetPageError(page);
data.c- else
data.c- SetPageUptodate(page);
--
IO error
data.c-
data.c-repeat:
data.c- page = find_or_create_page(mapping, blkaddr, gfp);
data.c: if (unlikely(!page)) {
data.c- DBG_BUGON(nofail);
data.c- return ERR_PTR(-ENOMEM);
data.c- }
--
NO MEM
data.c- }
data.c-
data.c- err = bio_add_page(bio, page, PAGE_SIZE, 0);
data.c: if (unlikely(err != PAGE_SIZE)) {
data.c- err = -EFAULT;
data.c- goto err_out;
Internal error (since it is a single bio with one page).
data.c- }
---
data.c- lock_page(page);
data.c-
data.c- /* this page has been truncated by others */
data.c: if (unlikely(page->mapping != mapping)) {
data.c-unlock_repeat:
data.c- unlock_page(page);
data.c- put_page(page);
data.c- goto repeat;
data.c- }
truncated
data.c-
data.c: /* more likely a read error */
data.c: if (unlikely(!PageUptodate(page))) {
data.c- if (io_retries) {
data.c- --io_retries;
data.c- goto unlock_repeat;
IO err
--
data.c- nblocks = DIV_ROUND_UP(inode->i_size, PAGE_SIZE);
data.c- lastblk = nblocks - is_inode_flat_inline(inode);
data.c-
data.c: if (unlikely(offset >= inode->i_size)) {
data.c- /* leave out-of-bound access unmapped */
data.c- map->m_flags = 0;
data.c- map->m_plen = 0;
--
FS corrupted
data.c-int erofs_map_blocks(struct inode *inode,
data.c- struct erofs_map_blocks *map, int flags)
data.c-{
data.c: if (unlikely(is_inode_layout_compression(inode))) {
for compressed files, should call z_erofs_map_blocks_iter since it
behaves in iterative way. but I think that can be deleted.
data.c- int err = z_erofs_map_blocks_iter(inode, map, flags);
data.c-
data.c- if (map->mpage) {
--
data.c- unsigned int blkoff;
data.c-
data.c- err = erofs_map_blocks(inode, &map, EROFS_GET_BLOCKS_RAW);
data.c: if (unlikely(err))
data.c- goto err_out;
Error
data.c-
data.c- /* zero out the holed page */
data.c: if (unlikely(!(map.m_flags & EROFS_MAP_MAPPED))) {
no hole in erofs.
data.c- zero_user_segment(page, 0, PAGE_SIZE);
data.c- SetPageUptodate(page);
data.c-
--
data.c-submit_bio_out:
data.c- __submit_bio(bio, REQ_OP_READ, 0);
data.c-
data.c: return unlikely(err) ? ERR_PTR(err) : NULL;
err
data.c-}
data.c-
data.c-/*
--
data.c- DBG_BUGON(!list_empty(pages));
data.c-
data.c- /* the rare case (end in gaps) */
data.c: if (unlikely(bio))
data.c- __submit_bio(bio, REQ_OP_READ, 0);
data.c- return 0;
data.c-}
decompressor.c- get_page(victim);
decompressor.c- } else {
decompressor.c- victim = erofs_allocpage(pagepool, GFP_KERNEL,
false);
decompressor.c: if (unlikely(!victim))
decompressor.c- return -ENOMEM;
nomem
decompressor.c- victim->mapping = Z_EROFS_MAPPING_STAGING;
decompressor.c- }
dir.c- de_namelen = le16_to_cpu(de[1].nameoff) - nameoff;
dir.c-
dir.c- /* a corrupted entry is found */
dir.c: if (unlikely(nameoff + de_namelen > maxsize ||
dir.c- de_namelen > EROFS_NAME_LEN)) {
corrupted
dir.c- errln("bogus dirent @ nid %llu", EROFS_V(dir)->nid);
dir.c- DBG_BUGON(1);
--
dir.c-
dir.c- nameoff = le16_to_cpu(de->nameoff);
dir.c-
dir.c: if (unlikely(nameoff < sizeof(struct erofs_dirent) ||
dir.c- nameoff >= PAGE_SIZE)) {
corrupted
dir.c- errln("%s, invalid de[0].nameoff %u @ nid %llu",
dir.c- __func__, nameoff, EROFS_V(dir)->nid);
--
dir.c- dirsize - ctx->pos + ofs, PAGE_SIZE);
dir.c-
dir.c- /* search dirents at the arbitrary position */
dir.c: if (unlikely(initial)) {
as comments said
dir.c- initial = false;
dir.c-
dir.c- ofs = roundup(ofs, sizeof(struct erofs_dirent));
dir.c: if (unlikely(ofs >= nameoff))
dir.c- goto skip_this;
dir.c- }
dir.c-
--
dir.c-
dir.c- ctx->pos = blknr_to_addr(i) + ofs;
dir.c-
dir.c: if (unlikely(err))
err
dir.c- break;
dir.c- ++i;
dir.c- ofs = 0;
inode.c-
inode.c- vi->datamode = __inode_data_mapping(advise);
inode.c-
inode.c: if (unlikely(vi->datamode >= EROFS_INODE_LAYOUT_MAX)) {
inode.c- errln("unsupported data mapping %u of nid %llu",
inode.c- vi->datamode, vi->nid);
inode.c- DBG_BUGON(1);
err
--
inode.c- if (S_ISLNK(inode->i_mode) && inode->i_size < PAGE_SIZE) {
inode.c- char *lnk = erofs_kmalloc(sbi, inode->i_size + 1,
GFP_KERNEL);
inode.c-
inode.c: if (unlikely(!lnk))
inode.c- return -ENOMEM;
inode.c-
inode.c- m_pofs += vi->inode_isize + vi->xattr_isize;
inode.c-
inode.c- /* inline symlink data shouldn't across page boundary
as well */
inode.c: if (unlikely(m_pofs + inode->i_size > PAGE_SIZE)) {
inode.c- kfree(lnk);
inode.c- errln("inline data cross block boundary @ nid
%llu",
inode.c- vi->nid);
err
--
inode.c-{
inode.c- struct inode *inode = erofs_iget_locked(sb, nid);
inode.c-
inode.c: if (unlikely(!inode))
inode.c- return ERR_PTR(-ENOMEM);
err
inode.c-
inode.c- if (inode->i_state & I_NEW) {
--
inode.c- vi->nid = nid;
inode.c-
inode.c- err = fill_inode(inode, isdir);
inode.c: if (likely(!err))
inode.c- unlock_new_inode(inode);
inode.c- else {
err
inode.c- iget_failed(inode);
I will stop here.
Thanks,
Gao Xiang
>
> Thanks,
> Gao Xiang
>
> >
> > regards,
> > dan carpenter
> >
