| Message ID | 0405c6131def5aa179ff4ba5d4201ebde89cede3.1443178314.git.robin.murphy@arm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, Sep 25, 2015 at 01:15:46PM +0100, Robin Murphy wrote: > Since some dma_alloc_coherent implementations return a zeroed buffer > regardless of whether __GFP_ZERO is passed, there exist drivers which > are implicitly dependent on this and pass otherwise uninitialised > buffers to hardware. This can lead to subtle and awkward-to-debug issues > using those drivers on different platforms, where nonzero uninitialised > junk may for instance occasionally look like a valid command which > causes the hardware to start misbehaving. To help with debugging such > issues, add the option to make uninitialised buffers much more obvious. The reason people started to do this is to stop a security leak in the ALSA code: ALSA allocates the ring buffer with dma_alloc_coherent() which used to grab pages and return them uninitialised. These pages could contain anything - including the contents of /etc/shadow, or your bank details. ALSA then lets userspace mmap() that memory, which means any user process which has access to the sound devices can read data leaked from kernel memory. I think I did bring it up at the time I found it, and decided that the safest thing to do was to always return an initialised buffer - short of constantly auditing every dma_alloc_coherent() user which also mmap()s the buffer into userspace, I couldn't convince myself that it was safe to avoid initialising the buffer. I don't know whether the original problem still exists in ALSA or not, but I do know that there are dma_alloc_coherent() implementations out there which do not initialise prior to returning memory. > diff --git a/lib/dma-debug.c b/lib/dma-debug.c > index 908fb35..40514ed 100644 > --- a/lib/dma-debug.c > +++ b/lib/dma-debug.c > @@ -30,6 +30,7 @@ > #include <linux/sched.h> > #include <linux/ctype.h> > #include <linux/list.h> > +#include <linux/poison.h> > #include <linux/slab.h> > > #include <asm/sections.h> > @@ -1447,7 +1448,7 @@ void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist, > EXPORT_SYMBOL(debug_dma_unmap_sg); > > void debug_dma_alloc_coherent(struct device *dev, size_t size, > - dma_addr_t dma_addr, void *virt) > + dma_addr_t dma_addr, void *virt, gfp_t flags) > { > struct dma_debug_entry *entry; > > @@ -1457,6 +1458,9 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size, > if (unlikely(virt == NULL)) > return; > > + if (IS_ENABLED(CONFIG_DMA_API_DEBUG_POISON) && !(flags & __GFP_ZERO)) > + memset(virt, DMA_ALLOC_POISON, size); > + This is likely to be slow in the case of non-cached memory and large allocations. The config option should come with a warning.
Hi Russell, On 25/09/15 13:44, Russell King - ARM Linux wrote: > On Fri, Sep 25, 2015 at 01:15:46PM +0100, Robin Murphy wrote: >> Since some dma_alloc_coherent implementations return a zeroed buffer >> regardless of whether __GFP_ZERO is passed, there exist drivers which >> are implicitly dependent on this and pass otherwise uninitialised >> buffers to hardware. This can lead to subtle and awkward-to-debug issues >> using those drivers on different platforms, where nonzero uninitialised >> junk may for instance occasionally look like a valid command which >> causes the hardware to start misbehaving. To help with debugging such >> issues, add the option to make uninitialised buffers much more obvious. > > The reason people started to do this is to stop a security leak in the > ALSA code: ALSA allocates the ring buffer with dma_alloc_coherent() > which used to grab pages and return them uninitialised. These pages > could contain anything - including the contents of /etc/shadow, or > your bank details. > > ALSA then lets userspace mmap() that memory, which means any user process > which has access to the sound devices can read data leaked from kernel > memory. > > I think I did bring it up at the time I found it, and decided that the > safest thing to do was to always return an initialised buffer - short of > constantly auditing every dma_alloc_coherent() user which also mmap()s > the buffer into userspace, I couldn't convince myself that it was safe > to avoid initialising the buffer. > > I don't know whether the original problem still exists in ALSA or not, > but I do know that there are dma_alloc_coherent() implementations out > there which do not initialise prior to returning memory. Indeed, I think we've discussed this before, and I don't imagine we'll be changing the actual behaviour of the existing allocators any time soon. [ I still don't see that as an excuse for callers not to be fixed, though - anyone allocating something that may be exposed to userspace has a responsibility to initialise it appropriately. After all, the DMA API is just one source, what do we do if such a careless subsystem got some uninitialised pages of leftover sensitive data from, say, alloc_pages() instead? ] That's a bit of a separate issue though. If a driver itself _needs_ a zeroed buffer but doesn't specifically request one, or doesn't get one even if it did, then that's just a regular bug, and it's what this patch is intended to help weed out. We've no need for a special poison value for data protection in the general case; zero is just fine for that. >> diff --git a/lib/dma-debug.c b/lib/dma-debug.c >> index 908fb35..40514ed 100644 >> --- a/lib/dma-debug.c >> +++ b/lib/dma-debug.c >> @@ -30,6 +30,7 @@ >> #include <linux/sched.h> >> #include <linux/ctype.h> >> #include <linux/list.h> >> +#include <linux/poison.h> >> #include <linux/slab.h> >> >> #include <asm/sections.h> >> @@ -1447,7 +1448,7 @@ void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist, >> EXPORT_SYMBOL(debug_dma_unmap_sg); >> >> void debug_dma_alloc_coherent(struct device *dev, size_t size, >> - dma_addr_t dma_addr, void *virt) >> + dma_addr_t dma_addr, void *virt, gfp_t flags) >> { >> struct dma_debug_entry *entry; >> >> @@ -1457,6 +1458,9 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size, >> if (unlikely(virt == NULL)) >> return; >> >> + if (IS_ENABLED(CONFIG_DMA_API_DEBUG_POISON) && !(flags & __GFP_ZERO)) >> + memset(virt, DMA_ALLOC_POISON, size); >> + > > This is likely to be slow in the case of non-cached memory and large > allocations. The config option should come with a warning. It depends on DMA_API_DEBUG, which already has a stern performance warning, is additionally hidden behind EXPERT, and carries a slightly flippant yet largely truthful warning that actually using it could break pretty much every driver in your system; is that not enough? If I was feeling particularly antagonistic, I'd also point out that as discussed above you've already taken the hit of a memset(0) and cache flush that you _didn't_ ask for, and there was no warning on that ;) The intent is a specific troubleshooting tool - realistically it's probably only usable at all when restricting DMA debug to a per-driver basis. My hunch is that nobody's too fussed about the performance of a driver that doesn't work properly, especially once they've reached the point of dumping buffers in an attempt to figure out why, when seeing the presence or not of uniform poison values could be helpful. (Of course, sometimes you end up debugging the allocator itself - see commit 7132813c3845 - which was one of the motivating factors for this patch). Robin.
On Fri, 25 Sep 2015 18:35:39 +0100 Robin Murphy <robin.murphy@arm.com> wrote: > Hi Russell, > > On 25/09/15 13:44, Russell King - ARM Linux wrote: > > On Fri, Sep 25, 2015 at 01:15:46PM +0100, Robin Murphy wrote: > >> Since some dma_alloc_coherent implementations return a zeroed buffer > >> regardless of whether __GFP_ZERO is passed, there exist drivers which > >> are implicitly dependent on this and pass otherwise uninitialised > >> buffers to hardware. This can lead to subtle and awkward-to-debug issues > >> using those drivers on different platforms, where nonzero uninitialised > >> junk may for instance occasionally look like a valid command which > >> causes the hardware to start misbehaving. To help with debugging such > >> issues, add the option to make uninitialised buffers much more obvious. > > > > The reason people started to do this is to stop a security leak in the > > ALSA code: ALSA allocates the ring buffer with dma_alloc_coherent() > > which used to grab pages and return them uninitialised. These pages > > could contain anything - including the contents of /etc/shadow, or > > your bank details. > > > > ALSA then lets userspace mmap() that memory, which means any user process > > which has access to the sound devices can read data leaked from kernel > > memory. > > > > I think I did bring it up at the time I found it, and decided that the > > safest thing to do was to always return an initialised buffer - short of > > constantly auditing every dma_alloc_coherent() user which also mmap()s > > the buffer into userspace, I couldn't convince myself that it was safe > > to avoid initialising the buffer. > > > > I don't know whether the original problem still exists in ALSA or not, > > but I do know that there are dma_alloc_coherent() implementations out > > there which do not initialise prior to returning memory. > > Indeed, I think we've discussed this before, and I don't imagine we'll > be changing the actual behaviour of the existing allocators any time soon. If I'm understanding things correctly, some allocators zero the memory by default and others do not. And we have an unknown number of drivers which are assuming that the memory is zeroed. Correct? If so, our options are a) audit all callers, find the ones which expect zeroed memory but aren't passing __GFP_ZERO and fix them. b) convert all allocators to zero the memory by default. Obviously, a) is better. How big a job is it? This patch will help the process, if people use it. > >> + if (IS_ENABLED(CONFIG_DMA_API_DEBUG_POISON) && !(flags & __GFP_ZERO)) > >> + memset(virt, DMA_ALLOC_POISON, size); > >> + > > > > This is likely to be slow in the case of non-cached memory and large > > allocations. The config option should come with a warning. > > It depends on DMA_API_DEBUG, which already has a stern performance > warning, is additionally hidden behind EXPERT, and carries a slightly > flippant yet largely truthful warning that actually using it could break > pretty much every driver in your system; is that not enough? It might be helpful to provide a runtime knob as well - having to rebuild&reinstall just to enable/disable this feature is a bit painful.
On 29/09/15 22:27, Andrew Morton wrote: [...] > If I'm understanding things correctly, some allocators zero the memory > by default and others do not. And we have an unknown number of drivers > which are assuming that the memory is zeroed. > > Correct? That's precisely the motivation here, yes. > If so, our options are > > a) audit all callers, find the ones which expect zeroed memory but > aren't passing __GFP_ZERO and fix them. > > b) convert all allocators to zero the memory by default. > > Obviously, a) is better. How big a job is it? This I'm not so sure of, hence the very tentative first step. For a very crude guess at an an upper bound: $ git grep -E '(dma|pci)_alloc_co(her|nsist)ent' drivers/ | wc -l 1148 vs. $ git grep -E '(dma|pci)_zalloc_co(her|nsist)ent' drivers/ | wc -l 234 noting that the vast majority of the former are still probably benign, but picking out those which aren't from the code alone without knowledge of and/or access to the hardware might be non-trivial. > This patch will help the process, if people use it. > >>>> + if (IS_ENABLED(CONFIG_DMA_API_DEBUG_POISON) && !(flags & __GFP_ZERO)) >>>> + memset(virt, DMA_ALLOC_POISON, size); >>>> + >>> >>> This is likely to be slow in the case of non-cached memory and large >>> allocations. The config option should come with a warning. >> >> It depends on DMA_API_DEBUG, which already has a stern performance >> warning, is additionally hidden behind EXPERT, and carries a slightly >> flippant yet largely truthful warning that actually using it could break >> pretty much every driver in your system; is that not enough? > > It might be helpful to provide a runtime knob as well - having to > rebuild&reinstall just to enable/disable this feature is a bit painful. Good point - there's always the global DMA debug disable knob, but this particular feature probably does warrant finer-grained control to be really practical. Having thought about it some more, it's also probably wrong that this doesn't respect the dma_debug_driver filter, given that it is actually invasive; in fixing that, how about if it also *only* applied when a specific driver is filtered? Then there would be no problematic "break anything and everything" mode, and the existing debugfs controls should suffice. Robin.
On Wed, 7 Oct 2015 20:17:03 +0100 Robin Murphy <robin.murphy@arm.com> wrote: > > It might be helpful to provide a runtime knob as well - having to > > rebuild&reinstall just to enable/disable this feature is a bit painful. > > Good point - there's always the global DMA debug disable knob, but this > particular feature probably does warrant finer-grained control to be > really practical. Having thought about it some more, it's also probably > wrong that this doesn't respect the dma_debug_driver filter, given that > it is actually invasive; in fixing that, how about if it also *only* > applied when a specific driver is filtered? Then there would be no > problematic "break anything and everything" mode, and the existing > debugfs controls should suffice. Yes, this should respect the driver filtering. On reflection... The patch poisons dma buffers if CONFIG_DMA_API_DEBUG and if __GFP_ZERO wasn't explicitly used. I'm rather surprised that the dma-debug code didn't do this from day one. I'd be inclined to enable this buffer-poisoning by default. Do you have a feeling for how much overhead that will add? Presumably not much, if __GFP_ZERO is acceptable. Also, how about we remove CONFIG_DMA_API_DEBUG_POISON and switch to a debugfs knob? btw, the documentation could do with a bit of a tune-up. The comments in dma-debug.c regarding driver filtering are non-existent. Documentation/kernel-parameters.txt says "The filter can be disabled or changed to another driver later using sysfs" but Documentation/DMA-API.txt talks about debugfs.
diff --git a/include/asm-generic/dma-mapping-common.h b/include/asm-generic/dma-mapping-common.h index b1bc954..0f3e16b 100644 --- a/include/asm-generic/dma-mapping-common.h +++ b/include/asm-generic/dma-mapping-common.h @@ -260,7 +260,7 @@ static inline void *dma_alloc_attrs(struct device *dev, size_t size, return NULL; cpu_addr = ops->alloc(dev, size, dma_handle, flag, attrs); - debug_dma_alloc_coherent(dev, size, *dma_handle, cpu_addr); + debug_dma_alloc_coherent(dev, size, *dma_handle, cpu_addr, flag); return cpu_addr; } diff --git a/include/linux/dma-debug.h b/include/linux/dma-debug.h index fe8cb61..e5f539d 100644 --- a/include/linux/dma-debug.h +++ b/include/linux/dma-debug.h @@ -51,7 +51,8 @@ extern void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist, int nelems, int dir); extern void debug_dma_alloc_coherent(struct device *dev, size_t size, - dma_addr_t dma_addr, void *virt); + dma_addr_t dma_addr, void *virt, + gfp_t flags); extern void debug_dma_free_coherent(struct device *dev, size_t size, void *virt, dma_addr_t addr); @@ -132,7 +133,8 @@ static inline void debug_dma_unmap_sg(struct device *dev, } static inline void debug_dma_alloc_coherent(struct device *dev, size_t size, - dma_addr_t dma_addr, void *virt) + dma_addr_t dma_addr, void *virt, + gfp_t flags) { } diff --git a/include/linux/poison.h b/include/linux/poison.h index 317e16d..174104e 100644 --- a/include/linux/poison.h +++ b/include/linux/poison.h @@ -73,6 +73,9 @@ #define MUTEX_DEBUG_INIT 0x11 #define MUTEX_DEBUG_FREE 0x22 +/********** lib/dma_debug.c **********/ +#define DMA_ALLOC_POISON 0xee + /********** lib/flex_array.c **********/ #define FLEX_ARRAY_FREE 0x6c /* for use-after-free poisoning */ diff --git a/lib/Kconfig.debug b/lib/Kconfig.debug index ab76b99..f2da7a1 100644 --- a/lib/Kconfig.debug +++ b/lib/Kconfig.debug @@ -1752,6 +1752,16 @@ config DMA_API_DEBUG If unsure, say N. +config DMA_API_DEBUG_POISON + bool "Poison coherent DMA buffers" + depends on DMA_API_DEBUG && EXPERT + help + Poison DMA buffers returned by dma_alloc_coherent unless __GFP_ZERO + is explicitly specified, to catch drivers depending on zeroed buffers + without passing the correct flags. + + Only say Y if you're prepared for almost everything to break. + config TEST_LKM tristate "Test module loading with 'hello world' module" default n diff --git a/lib/dma-debug.c b/lib/dma-debug.c index 908fb35..40514ed 100644 --- a/lib/dma-debug.c +++ b/lib/dma-debug.c @@ -30,6 +30,7 @@ #include <linux/sched.h> #include <linux/ctype.h> #include <linux/list.h> +#include <linux/poison.h> #include <linux/slab.h> #include <asm/sections.h> @@ -1447,7 +1448,7 @@ void debug_dma_unmap_sg(struct device *dev, struct scatterlist *sglist, EXPORT_SYMBOL(debug_dma_unmap_sg); void debug_dma_alloc_coherent(struct device *dev, size_t size, - dma_addr_t dma_addr, void *virt) + dma_addr_t dma_addr, void *virt, gfp_t flags) { struct dma_debug_entry *entry; @@ -1457,6 +1458,9 @@ void debug_dma_alloc_coherent(struct device *dev, size_t size, if (unlikely(virt == NULL)) return; + if (IS_ENABLED(CONFIG_DMA_API_DEBUG_POISON) && !(flags & __GFP_ZERO)) + memset(virt, DMA_ALLOC_POISON, size); + entry = dma_entry_alloc(); if (!entry) return;
Since some dma_alloc_coherent implementations return a zeroed buffer regardless of whether __GFP_ZERO is passed, there exist drivers which are implicitly dependent on this and pass otherwise uninitialised buffers to hardware. This can lead to subtle and awkward-to-debug issues using those drivers on different platforms, where nonzero uninitialised junk may for instance occasionally look like a valid command which causes the hardware to start misbehaving. To help with debugging such issues, add the option to make uninitialised buffers much more obvious. Signed-off-by: Robin Murphy <robin.murphy@arm.com> --- include/asm-generic/dma-mapping-common.h | 2 +- include/linux/dma-debug.h | 6 ++++-- include/linux/poison.h | 3 +++ lib/Kconfig.debug | 10 ++++++++++ lib/dma-debug.c | 6 +++++- 5 files changed, 23 insertions(+), 4 deletions(-)