On Tue, 26 Jul 2022 23:40:48 +0200, Dipanjan Das wrote: > > On Sat, Jul 23, 2022 at 3:17 AM Takashi Iwai <tiwai@xxxxxxx> wrote: > > > > On Sat, 23 Jul 2022 09:00:08 +0200, > > Greg KH wrote: > > > > > > Wondeful, do you have a fix for this that solves the reported problem > > > that you have tested with the reproducer? > > > > ... or at least more detailed information. > > Here is our analysis of the bug in the kernel v5.10.131. > > During allocation, the `size` of the DMA buffer is not page-aligned: > https://elixir.bootlin.com/linux/v5.10.131/source/sound/core/memalloc.c#L149. > However, in sound/core/pcm_native.c:798 > (https://elixir.bootlin.com/linux/v5.10.131/source/sound/core/pcm_native.c#L798), > the `size` variable is page-aligned before memset-ing the `dma_area`. > >From the other BUG_ON assertions in other parts of the code, it looks > like the DMA area is not supposed to be equal to or greater than > 0x200000 bytes. However, due to page-alignment, the `size` can indeed > get rounded up to 0x200000 which causes the out of bound access. > > > Last but not least, you should check whether it's specific to your > > 5.10.x kernel or it's also seen with the latest upstream, too. > > The bug is not reproducible on the latest mainline, because in > sound/core/memalloc.c:66 > (https://github.com/torvalds/linux/blob/5de64d44968e4ae66ebdb0a2d08b443f189d3651/sound/core/memalloc.c#L66) > the allocation function `snd_dma_alloc_dir_pages()` now page-aligns > the `size` right before allocating the DMA buffer. Therefore, any > subsequent page-alignment, like the one in `snd_pcm_hw_params()` does > not cause an out of bound access. Thanks for the analysis. A good news is that, at least for the vmalloc() case, it's a kind of false-positive; vmalloc() always takes the full pages, so practically seen, the size is page-aligned. It's fooling the memory checker, though. But the similar problem could be seen with genalloc calls, and this was fixed by the upstream commit 5c1733e33c888a3cb7f576564d8ad543d5ad4a9e ALSA: memalloc: Align buffer allocations in page size I suppose you can simply backport this commit to 5.10.y. Could you confirm that this fixes your problem? thanks, Takashi
