From mboxrd@z Thu Jan 1 00:00:00 1970 From: Peilin Ye Date: Sat, 26 Sep 2020 19:39:57 +0000 Subject: Re: KASAN: use-after-free Read in bit_putcs Message-Id: <20200926193957.GA1033221@PWN> List-Id: References: <000000000000226d3f05b02dd607@google.com> In-Reply-To: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Tetsuo Handa Cc: syzbot , linux-fbdev@vger.kernel.org, b.zolnierkie@samsung.com, daniel.vetter@ffwll.ch, deller@gmx.de, syzkaller-bugs@googlegroups.com, linux-kernel@vger.kernel.org, dri-devel@lists.freedesktop.org, gregkh@linuxfoundation.org, jirislaby@kernel.org, yepeilin.cs@gmail.com On Sun, Sep 27, 2020 at 01:25:17AM +0900, Tetsuo Handa wrote: > A simplified reproducer and debug printk() patch shown below reported that > vc_font.height is increased to 9 via ioctl(VT_RESIZEX) after it was once > decreased from 16 to 2 via ioctl(PIO_FONT). > > > > Since vc_resize() with v.v_rows = 0 preserves current vc->vc_rows value, > this reproducer is bypassing > > if (v.v_clin) { > int rows = v.v_vlin / v.v_clin; > if (v.v_rows != rows) { > if (v.v_rows) /* Parameters don't add up */ > return -EINVAL; > v.v_rows = rows; > } > } > > check by setting v.v_vlin = 1 and v.v_clin = 9. > > If v.v_vcol > 0 and v.v_vcol != vc->vc_cols (though this reproducer is passing > v.v_vcol = 0), tty_do_resize() from vc_do_resize() from vc_resize() can make > "struct tty_struct"->winsize.ws_ypixel = 1 despite > "struct tty_struct"->winsize.vc->vc_rows = vc->vc_rows (which is usually larger > than 1). Does such winsize (a row has 1 / vc->vc_rows pixel) make sense? > > > > Since I don't know the meaning of "struct vt_consize"->v_clin (which is commented > with "/* number of pixel rows per character */" but does it mean font size ?), > I don't know why we can assign that value to vcp->vc_font.height via > > if (v.v_clin) > vcp->vc_font.height = v.v_clin; > > in vt_resizex(). While ioctl(PIO_FONT) needs to pass vc->vc_sw->con_font_set() > check in con_font_set(), ioctl(VT_RESIZEX) does not pass it in vt_resizex()... > > Since this problem does not happen if I remove > > if (v.v_clin) > vcp->vc_font.height = v.v_clin; Hi Tetsuo! > from vt_resizex(), I guess that some variables are getting confused by change > of vc->vc_font.height ... Yes, see bit_putcs(): (drivers/video/fbdev/core/bitblit.c) static void bit_putcs(struct vc_data *vc, struct fb_info *info, const unsigned short *s, int count, int yy, int xx, int fg, int bg) { struct fb_image image; u32 width = DIV_ROUND_UP(vc->vc_font.width, 8); u32 cellsize = width * vc->vc_font.height; ^^^^^^^^ ^^^^^^^^^^^^^^ `cellsize` is now too large. Later, in bit_putcs_aligned(): while (cnt--) { src = vc->vc_font.data + (scr_readw(s++)& charmask)*cellsize; ^^^^^^^^ `src` goes out of bounds of the data buffer. At first glance I guess this is an out-of-bound read reported as a use-after-free read? The crashlog says: [ 149.732103][ T6693] Allocated by task 6667: [ 149.732115][ T6693] kasan_save_stack (mm/kasan/common.c:48) [ 149.732121][ T6693] __kasan_kmalloc.constprop.0 (mm/kasan/common.c:56 mm/kasan/common.c:461) [ 149.732126][ T6693] __kmalloc (mm/slab.c:3656 mm/slab.c:3664) [ 149.732133][ T6693] alloc_pipe_info (fs/pipe.c:810) [ 149.732139][ T6693] create_pipe_files (fs/pipe.c:883 fs/pipe.c:914) [ 149.732145][ T6693] do_pipe2 (fs/pipe.c:965 fs/pipe.c:1012) I'm not sure, but I don't think a buffer allocated in fs/pipe.c is related here. Maybe they just live near each other on the heap? To resolve this out-of-bound issue for now, I think the easiest way is to add a range check in bit_putcs(), or bit_putcs_aligned(). ...but yeah, that `VT_RESIZEX` ioctl looks really buggy, and is already causing more issues: KASAN: global-out-of-bounds Read in fbcon_get_font Link: https://syzkaller.appspot.com/bug?idb8be45afea11888776f897895aef9ad1c3ecfd This was also caused by `VT_RESIZEX`... Thank you, Peilin Ye