From mboxrd@z Thu Jan 1 00:00:00 1970 From: Mikulas Patocka Subject: Re: [PATCH] kvmalloc: always use vmalloc if CONFIG_DEBUG_VM Date: Fri, 20 Apr 2018 08:20:23 -0400 (EDT) Message-ID: References: <3e65977e-53cd-bf09-bc4b-0ce40e9091fe@gmail.com> <20180418.134651.2225112489265654270.davem@davemloft.net> <20180420114712.GB10788@bombadil.infradead.org> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20180420114712.GB10788@bombadil.infradead.org> List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: virtualization-bounces@lists.linux-foundation.org Errors-To: virtualization-bounces@lists.linux-foundation.org To: Matthew Wilcox Cc: dm-devel@redhat.com, eric.dumazet@gmail.com, mst@redhat.com, netdev@vger.kernel.org, linux-kernel@vger.kernel.org, virtualization@lists.linux-foundation.org, linux-mm@kvack.org, edumazet@google.com, Andrew Morton , David Miller , Vlastimil Babka List-Id: virtualization@lists.linuxfoundation.org On Fri, 20 Apr 2018, Matthew Wilcox wrote: > On Thu, Apr 19, 2018 at 12:12:38PM -0400, Mikulas Patocka wrote: > > Unfortunatelly, some kernel code has bugs - it uses kvmalloc and then > > uses DMA-API on the returned memory or frees it with kfree. Such bugs were > > found in the virtio-net driver, dm-integrity or RHEL7 powerpc-specific > > code. > > Maybe it's time to have the SG code handle vmalloced pages? This is > becoming more and more common with vmapped stacks (and some of our > workarounds are hideous -- allocate 4 bytes with kmalloc because we can't > DMA onto the stack any more?). We already have a few places which do > handle sgs of vmalloced addresses, such as the nx crypto driver: > > if (is_vmalloc_addr(start_addr)) > sg_addr = page_to_phys(vmalloc_to_page(start_addr)) > + offset_in_page(sg_addr); > else > sg_addr = __pa(sg_addr); > > and videobuf: > > pg = vmalloc_to_page(virt); > if (NULL == pg) > goto err; > BUG_ON(page_to_pfn(pg) >= (1 << (32 - PAGE_SHIFT))); > sg_set_page(&sglist[i], pg, PAGE_SIZE, 0); > > Yes, there's the potential that we have to produce two SG entries for a > virtually contiguous region if it crosses a page boundary, and our APIs > aren't set up right to make it happen. But this is something we should > consider fixing ... otherwise we'll end up with dozens of driver hacks. > The videobuf implementation was already copy-and-pasted into the saa7146 > driver, for example. What if the device requires physically contiguous area and the vmalloc area crosses a page? Will you use a bounce buffer? Where do you allocate the bounce buffer from? What if you run out of bounce buffers? Mikulkas