Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S969671AbdIZOxO (ORCPT ); Tue, 26 Sep 2017 10:53:14 -0400 Received: from foss.arm.com ([217.140.101.70]:35160 "EHLO foss.arm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S967152AbdIZOxN (ORCPT ); Tue, 26 Sep 2017 10:53:13 -0400 Subject: Re: [PATCH] dma-debug: fix incorrect pfn calculation To: miles.chen@mediatek.com, Andrew Morton Cc: linux-kernel@vger.kernel.org, linux-mm@kvack.org, wsd_upstream@mediatek.com, linux-mediatek@lists.infradead.org, iommu@lists.linux-foundation.org References: <1506432287-7214-1-git-send-email-miles.chen@mediatek.com> From: Robin Murphy Message-ID: <03870968-0060-d6db-d109-f2c299c35bf1@arm.com> Date: Tue, 26 Sep 2017 15:53:10 +0100 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.3.0 MIME-Version: 1.0 In-Reply-To: <1506432287-7214-1-git-send-email-miles.chen@mediatek.com> Content-Type: text/plain; charset=utf-8 Content-Language: en-US Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 5025 Lines: 134 On 26/09/17 14:24, miles.chen@mediatek.com wrote: > From: Miles Chen > > dma-debug report the following warning: > > [name:panic&]WARNING: CPU: 3 PID: 298 at kernel-4.4/lib/dma-debug.c:604 > debug _dma_assert_idle+0x1a8/0x230() > DMA-API: cpu touching an active dma mapped cacheline [cln=0x00000882300] > CPU: 3 PID: 298 Comm: vold Tainted: G W O 4.4.22+ #1 > Hardware name: MT6739 (DT) > Call trace: > [] dump_backtrace+0x0/0x1d4 > [] show_stack+0x14/0x1c > [] dump_stack+0xa8/0xe0 > [] warn_slowpath_common+0xf4/0x11c > [] warn_slowpath_fmt+0x60/0x80 > [] debug_dma_assert_idle+0x1a8/0x230 > [] wp_page_copy.isra.96+0x118/0x520 > [] do_wp_page+0x4fc/0x534 > [] handle_mm_fault+0xd4c/0x1310 > [] do_page_fault+0x1c8/0x394 > [] do_mem_abort+0x50/0xec > > I found that debug_dma_alloc_coherent() and debug_dma_free_coherent() > always use type "dma_debug_coherent" and assume that dma_alloc_coherent() > always returns a linear address. > > However if a device returns false on is_device_dma_coherent(), > dma_alloc_coherent() will create another non-cacheable mapping > (also non linear). In this case, page_to_pfn(virt_to_page(virt)) will > return an incorrect pfn. If the pfn is valid and mapped as a COW page, > we will hit the warning when doing wp_page_copy(). > > Fix this by calculating correct pfn if is_device_dma_coherent() > returns false. As the inevitable storm of kbuild robot reports will tell you soon, you can't do that: is_device_dma_coherent() is a private helper between arm{,64} arch code and xen, and should not be used anywhere else. > Signed-off-by: Miles Chen > --- > lib/dma-debug.c | 20 ++++++++++++++------ > 1 file changed, 14 insertions(+), 6 deletions(-) > > diff --git a/lib/dma-debug.c b/lib/dma-debug.c > index ea4cc3d..b17e56e 100644 > --- a/lib/dma-debug.c > +++ b/lib/dma-debug.c > @@ -47,6 +47,8 @@ enum { > dma_debug_sg, > dma_debug_coherent, > dma_debug_resource, > + dma_debug_noncoherent, > + nr_dma_debug_types, > }; > > enum map_err_types { > @@ -154,9 +156,9 @@ static inline bool dma_debug_disabled(void) > [MAP_ERR_CHECKED] = "dma map error checked", > }; > > -static const char *type2name[5] = { "single", "page", > +static const char *type2name[nr_dma_debug_types] = { "single", "page", > "scather-gather", "coherent", > - "resource" }; > + "resource", "noncoherent" }; > > static const char *dir2name[4] = { "DMA_BIDIRECTIONAL", "DMA_TO_DEVICE", > "DMA_FROM_DEVICE", "DMA_NONE" }; > @@ -1484,6 +1486,7 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size, > dma_addr_t dma_addr, void *virt) > { > struct dma_debug_entry *entry; > + bool coherent = is_device_dma_coherent(dev); > > if (unlikely(dma_debug_disabled())) > return; > @@ -1495,9 +1498,11 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size, > if (!entry) > return; > > - entry->type = dma_debug_coherent; > + entry->type = coherent ? dma_debug_coherent : > + dma_debug_noncoherent; > entry->dev = dev; > - entry->pfn = page_to_pfn(virt_to_page(virt)); There are more architectures where the virtual address returned by dma_alloc_coherent is not a linear map address - some just have a static offset between cacheable and non-cacheable aliases - so there may be other cases where this is the wrong calculation, but at least those probably don't trigger the problematic false-positive. That said, the cases where coherent allocations *are* dynamically remapped should be easy enough to handle properly without having to resort to dodgy hacks: if (is_vmalloc_addr(virt)) pfn = vmalloc_to_pfn(virt); else pfn = page_to_pfn(virt_to_page(virt)); Simple. > + entry->pfn = coherent ? page_to_pfn(virt_to_page(virt)) : > + dma_addr >> PAGE_SHIFT; And in particular, this is just as likely to just give *different* false positives, since there's no guarantee whatsoever that dma_addr has any relationship to the appropriate pfn. Robin. > entry->offset = offset_in_page(virt); > entry->size = size; > entry->dev_addr = dma_addr; > @@ -1510,10 +1515,13 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size, > void debug_dma_free_coherent(struct device *dev, size_t size, > void *virt, dma_addr_t addr) > { > + bool coherent = is_device_dma_coherent(dev); > struct dma_debug_entry ref = { > - .type = dma_debug_coherent, > + .type = coherent ? dma_debug_coherent : > + dma_debug_noncoherent, > .dev = dev, > - .pfn = page_to_pfn(virt_to_page(virt)), > + .pfn = coherent ? page_to_pfn(virt_to_page(virt)) : > + addr >> PAGE_SHIFT, > .offset = offset_in_page(virt), > .dev_addr = addr, > .size = size, >