Messages in this thread | | | Date | Mon, 10 Jul 2023 11:37:06 +0800 | Subject | Re: [PATCH] erofs: fix two loop issues when read page beyond EOF | From | Gao Xiang <> |
| |
On 2023/7/10 11:32, Chunhai Guo wrote: > Hi Xiang, > > On 2023/7/8 17:00, Gao Xiang wrote: >> Hi Chunhai, >> >> On 2023/7/8 14:24, Chunhai Guo wrote: >>> When z_erofs_read_folio() reads a page with an offset far beyond EOF, two >>> issues may occur: >>> - z_erofs_pcluster_readmore() may take a long time to loop when the offset >>> is big enough, which is unnecessary. >>> - For example, it will loop 4691368 times and take about 27 seconds >>> with following case. >>> - offset = 19217289215 >>> - inode_size = 1442672 >>> - z_erofs_do_read_page() may loop infinitely due to the inappropriate >>> truncation in the below statement. Since the offset is 64 bits and >>> min_t() truncates the result to 32 bits. The solution is to replace >>> unsigned int with another 64-bit type, such as erofs_off_t. >>> cur = end - min_t(unsigned int, offset + end - map->m_la, end); >>> - For example: >>> - offset = 0x400160000 >>> - end = 0x370 >>> - map->m_la = 0x160370 >>> - offset + end - map->m_la = 0x400000000 >>> - offset + end - map->m_la = 0x00000000 (truncated as unsigned int) >> >> Thanks for the catch! >> >> Could you split these two into two patches? >> >> how about using: >> cur = end - min_t(erofs_off_t, offend + end - map->m_la, end) >> for this? >> >> since cur and end are all [0, PAGE_SIZE - 1] for now, and >> folio_size() later. > > OK. I will split the patch. > > Sorry that I can not understand what is 'offend' refer to and what do you mean. Could you please describe it more clearly?
Sorry, there is a typo here, I meant 'offset'.
`cur` and `end` both are not exceed 4096 if your page_size is 4096.
Does cur = end - min_t(erofs_off_t, offset + end - map->m_la, end)
fix your issue?
> >>> - Expected result: >>> - cur = 0 >>> - Actual result: >>> - cur = 0x370 >>> >>> Signed-off-by: Chunhai Guo <guochunhai@vivo.com> >>> --- >>> fs/erofs/zdata.c | 13 ++++++++++--- >>> 1 file changed, 10 insertions(+), 3 deletions(-) >>> >>> diff --git a/fs/erofs/zdata.c b/fs/erofs/zdata.c >>> index 5f1890e309c6..6abbd4510076 100644 >>> --- a/fs/erofs/zdata.c >>> +++ b/fs/erofs/zdata.c >>> @@ -972,7 +972,8 @@ static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe, >>> struct erofs_map_blocks *const map = &fe->map; >>> const loff_t offset = page_offset(page); >>> bool tight = true, exclusive; >>> - unsigned int cur, end, spiltted; >>> + erofs_off_t cur, end; >>> + unsigned int spiltted; >>> int err = 0; >>> /* register locked file pages as online pages in pack */ >>> @@ -1035,7 +1036,7 @@ static int z_erofs_do_read_page(struct z_erofs_decompress_frontend *fe, >>> */ >>> tight &= (fe->mode > Z_EROFS_PCLUSTER_FOLLOWED_NOINPLACE); >>> - cur = end - min_t(unsigned int, offset + end - map->m_la, end); >>> + cur = end - min_t(erofs_off_t, offset + end - map->m_la, end); >>> if (!(map->m_flags & EROFS_MAP_MAPPED)) { >>> zero_user_segment(page, cur, end); >>> goto next_part; >>> @@ -1841,7 +1842,7 @@ static void z_erofs_pcluster_readmore(struct z_erofs_decompress_frontend *f, >>> } >>> cur = map->m_la + map->m_llen - 1; >>> - while (cur >= end) { >>> + while ((cur >= end) && (cur < i_size_read(inode))) { >>> pgoff_t index = cur >> PAGE_SHIFT; >>> struct page *page; >>> @@ -1876,6 +1877,12 @@ static int z_erofs_read_folio(struct file *file, struct folio *folio) >>> trace_erofs_readpage(page, false); >>> f.headoffset = (erofs_off_t)page->index << PAGE_SHIFT; >>> + /* when trying to read beyond EOF, return zero page directly */ >>> + if (f.headoffset >= i_size_read(inode)) { >>> + zero_user_segment(page, 0, PAGE_SIZE); >>> + return 0; >>> + } >> Do we really need to optimize this rare case? >> I guess the follow readmore fix is enough, thoughts? > >> Thanks, >> Gao Xiang > > Since the code is constantly being updated and someone may encounter this bug again, I think we had better fix it if possible.
z_erofs_do_read_page() already covers this case so I'm not sure why we need to add another logic here.
Thanks, Gao Xiang
> > Thanks. > Guo Chunhai
| |