| Message ID | 1392339850-18686-3-git-send-email-keescook@chromium.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: > Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though > the behavior is different: it depends on STRICT_KERNMEM_PERMS, which > sets rodata read-only (but executable), where as this option additionally > splits rodata from the kernel text (resulting in potentially more memory > lost to padding) and sets it non-executable as well. The end result is > that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be > marked purely read-only. This triggers an Oops in kexec, because we have a block of code in .text which is a template for generating baremetal code to relocate the new kernel, and some literal words are written into it before copying. Possibly this should be in .rodata, not .text. There may be a few other instances of this kind of thing. Are you aware of similar situations on other arches? Cheers ---Dave
On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin@arm.com> wrote: > On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: >> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though >> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which >> sets rodata read-only (but executable), where as this option additionally >> splits rodata from the kernel text (resulting in potentially more memory >> lost to padding) and sets it non-executable as well. The end result is >> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be >> marked purely read-only. > > This triggers an Oops in kexec, because we have a block of code in .text > which is a template for generating baremetal code to relocate the new > kernel, and some literal words are written into it before copying. You're writing into the text area? I would imagine that CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the right place to be building code -- shouldn't the module area be used for that? > Possibly this should be in .rodata, not .text. Well, rodata should be neither writable nor executable. > There may be a few other instances of this kind of thing. This config will certainly find them! :) But, that's why it's behind a config. > Are you aware of similar situations on other arches? I think there were some problems a long time ago on x86 for rodata too. -Kees
On Fri, Feb 14, 2014 at 11:11:07AM -0800, Kees Cook wrote: > On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin@arm.com> wrote: > > On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: > >> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though > >> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which > >> sets rodata read-only (but executable), where as this option additionally > >> splits rodata from the kernel text (resulting in potentially more memory > >> lost to padding) and sets it non-executable as well. The end result is > >> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be > >> marked purely read-only. > > > > This triggers an Oops in kexec, because we have a block of code in .text > > which is a template for generating baremetal code to relocate the new > > kernel, and some literal words are written into it before copying. > > You're writing into the text area? I would imagine that > CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the > right place to be building code -- shouldn't the module area be used > for that? > > > Possibly this should be in .rodata, not .text. > > Well, rodata should be neither writable nor executable. We're not writing into code exactly. This code is never executed in-place in vmlinux. It gets copied, and only copies are ever executed. Some pointers and offsets get poked into the code to configure it. I think it would be better simply to put the code in .rodata, and poke paramaters into the copy, not the original -- but that's a bit more awkward to code up, since the values can't be poked simply by writing global variables. > > > There may be a few other instances of this kind of thing. > > This config will certainly find them! :) But, that's why it's behind a config. I haven't tested exhaustively, but it think this is sufficient for a Tested-by. The patch does seem to be doing what it is intended to do, and doesn't seem to be triggering false positives all over the place. > > > Are you aware of similar situations on other arches? > > I think there were some problems a long time ago on x86 for rodata too. It would be good to get this kexec case fixed -- I'll try to hack up a separate patch. Cheers ---Dave
On Mon, Feb 17, 2014 at 4:34 AM, Dave Martin <Dave.Martin@arm.com> wrote: > On Fri, Feb 14, 2014 at 11:11:07AM -0800, Kees Cook wrote: >> On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin@arm.com> wrote: >> > On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: >> >> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though >> >> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which >> >> sets rodata read-only (but executable), where as this option additionally >> >> splits rodata from the kernel text (resulting in potentially more memory >> >> lost to padding) and sets it non-executable as well. The end result is >> >> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be >> >> marked purely read-only. >> > >> > This triggers an Oops in kexec, because we have a block of code in .text >> > which is a template for generating baremetal code to relocate the new >> > kernel, and some literal words are written into it before copying. >> >> You're writing into the text area? I would imagine that >> CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the >> right place to be building code -- shouldn't the module area be used >> for that? >> >> > Possibly this should be in .rodata, not .text. >> >> Well, rodata should be neither writable nor executable. > > We're not writing into code exactly. > > This code is never executed in-place in vmlinux. It gets copied, and > only copies are ever executed. > > Some pointers and offsets get poked into the code to configure it. > > I think it would be better simply to put the code in .rodata, and > poke paramaters into the copy, not the original -- but that's a bit > more awkward to code up, since the values can't be poked simply by > writing global variables. Okay, interesting. I'll be curious to see what the patch for this looks like. >> > There may be a few other instances of this kind of thing. >> >> This config will certainly find them! :) But, that's why it's behind a config. > > I haven't tested exhaustively, but it think this is sufficient for a > Tested-by. The patch does seem to be doing what it is intended to > do, and doesn't seem to be triggering false positives all over the > place. Great, thanks for taking the time to check on it! Should I send this to the patch tracker, or wait for more feedback? >> > Are you aware of similar situations on other arches? >> >> I think there were some problems a long time ago on x86 for rodata too. > > It would be good to get this kexec case fixed -- I'll try to hack up > a separate patch. Thanks! -Kees
On Tue, Feb 18, 2014 at 10:10:03AM -0800, Kees Cook wrote: > On Mon, Feb 17, 2014 at 4:34 AM, Dave Martin <Dave.Martin@arm.com> wrote: > > On Fri, Feb 14, 2014 at 11:11:07AM -0800, Kees Cook wrote: > >> On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin@arm.com> wrote: > >> > On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: > >> >> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though > >> >> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which > >> >> sets rodata read-only (but executable), where as this option additionally > >> >> splits rodata from the kernel text (resulting in potentially more memory > >> >> lost to padding) and sets it non-executable as well. The end result is > >> >> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be > >> >> marked purely read-only. > >> > > >> > This triggers an Oops in kexec, because we have a block of code in .text > >> > which is a template for generating baremetal code to relocate the new > >> > kernel, and some literal words are written into it before copying. > >> > >> You're writing into the text area? I would imagine that > >> CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the > >> right place to be building code -- shouldn't the module area be used > >> for that? > >> > >> > Possibly this should be in .rodata, not .text. > >> > >> Well, rodata should be neither writable nor executable. > > > > We're not writing into code exactly. > > > > This code is never executed in-place in vmlinux. It gets copied, and > > only copies are ever executed. > > > > Some pointers and offsets get poked into the code to configure it. > > > > I think it would be better simply to put the code in .rodata, and > > poke paramaters into the copy, not the original -- but that's a bit > > more awkward to code up, since the values can't be poked simply by > > writing global variables. > > Okay, interesting. I'll be curious to see what the patch for this looks like. > > >> > There may be a few other instances of this kind of thing. > >> > >> This config will certainly find them! :) But, that's why it's behind a config. > > > > I haven't tested exhaustively, but it think this is sufficient for a > > Tested-by. The patch does seem to be doing what it is intended to > > do, and doesn't seem to be triggering false positives all over the > > place. > > Great, thanks for taking the time to check on it! > > Should I send this to the patch tracker, or wait for more feedback? It would be good if someone who's more familiar with the parms and vmlinux.lds stuff could take a look at it, though I don't see any obvious problem yet. If you don't receive further comments, you could try reposting once to alert people to the fact that you're still waiting. Cheers ---Dave
On Fri, Feb 21, 2014 at 12:37:04PM +0000, Dave Martin wrote: > It would be good if someone who's more familiar with the parms and > vmlinux.lds stuff could take a look at it, though I don't see any > obvious problem yet. The biggest issue with it is that we end up with: - the .text section rounded up to 1MB - the .rodata section rounded up to 1MB That means we can end up wasting up to 1MB of memory for each (in the worst case where we encroach into the next 1MB aligned region by a few bytes) and this memory can't be re-used. The alternative is to adjust the maps such that we end up mapping the .text / .rodata overlap 1MB using 4K pages, taking the additional TLB hit by doing so. The .text is aligned to 1MB, so the majority of the first 0x8000 to 0x100000 is unused. The end of the .text section is aligned to 1MB, and the start of the .data section is also aligned to 1MB. So, the minimum kernel size is: 0x100000 + MB_ALIGN(sizeof(.text)) + MB_ALIGN(sizeof(.rodata)) + MB_ALIGN(sizeof(init sections)) + sizeof(.data) - 0x8000 So, looking at this kernel I've recently built: Idx Name Size VMA LMA File off Algn 0 .head.text 00000204 c0008000 c0008000 00008000 2**2 CONTENTS, ALLOC, LOAD, READONLY, CODE --- .text this gets set to 0xc0100000, we lose 0xc0008240 to 0xc0100000 1 .text 006c4530 c0008240 c0008240 00008240 2**6 CONTENTS, ALLOC, LOAD, READONLY, CODE 2 .text.head 0000004c c06cc770 c06cc770 006cc770 2**2 CONTENTS, ALLOC, LOAD, READONLY, CODE --- sizeof(.text) + sizeof(.text.head) becomes 0x700000 --- .rodata starts at 0xc0800000 instead of 0xc06cd000 3 .rodata 0022f568 c06cd000 c06cd000 006cd000 2**6 CONTENTS, ALLOC, LOAD, READONLY, DATA 4 __bug_table 0000873c c08fc568 c08fc568 008fc568 2**0 CONTENTS, ALLOC, LOAD, READONLY, DATA 5 .pci_fixup 00000030 c0904ca4 c0904ca4 00904ca4 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA 6 __ksymtab 00008158 c0904cd4 c0904cd4 00904cd4 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA 7 __ksymtab_gpl 00006858 c090ce2c c090ce2c 0090ce2c 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA 8 __kcrctab 000040ac c0913684 c0913684 00913684 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA 9 __kcrctab_gpl 0000342c c0917730 c0917730 00917730 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA 10 __ksymtab_strings 00022a08 c091ab5c c091ab5c 0091ab5c 2**0 CONTENTS, ALLOC, LOAD, READONLY, DATA 11 __param 00000c70 c093d564 c093d564 0093d564 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA 12 __modver 00000e2c c093e1d4 c093e1d4 0093e1d4 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA 13 __ex_table 00000f18 c093f000 c093f000 0093f000 2**3 CONTENTS, ALLOC, LOAD, READONLY, DATA 14 .notes 00000024 c093ff18 c093ff18 0093ff18 2**2 CONTENTS, ALLOC, LOAD, READONLY, CODE 15 .vectors 00000020 00000000 c0940000 00940000 2**2 CONTENTS, ALLOC, LOAD, READONLY, CODE 16 .stubs 00000240 00001000 c0940020 00941000 2**5 CONTENTS, ALLOC, LOAD, READONLY, CODE 17 .init.text 00051760 c0940260 c0940260 00948260 2**5 CONTENTS, ALLOC, LOAD, READONLY, CODE 18 .exit.text 00002130 c09919c0 c09919c0 009999c0 2**2 CONTENTS, ALLOC, LOAD, READONLY, CODE 19 .init.arch.info 00000108 c0993af0 c0993af0 0099baf0 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA 20 .init.tagtable 00000048 c0993bf8 c0993bf8 0099bbf8 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA 21 .init.smpalt 000032f8 c0993c40 c0993c40 0099bc40 2**2 CONTENTS, ALLOC, LOAD, READONLY, DATA 22 .init.pv_table 00000314 c0996f38 c0996f38 0099ef38 2**0 CONTENTS, ALLOC, LOAD, READONLY, DATA 23 .init.data 0000c19c c0997250 c0997250 0099f250 2**3 CONTENTS, ALLOC, LOAD, DATA 24 .data..percpu 000035c0 c09a4000 c09a4000 009ac000 2**6 CONTENTS, ALLOC, LOAD, DATA --- sizeof previous sections is 0x2db000, which becomes 0x300000 --- start of .data becomes 0xc0b00000 instead of 0xc09a8000 25 .data 00062728 c09a8000 c09a8000 009b0000 2**6 CONTENTS, ALLOC, LOAD, DATA 26 .bss 00754870 c0a0a740 c0a0a740 00a12728 2**6 ALLOC --- which means the kernel image finishes at 0xC12B6FB0 whereas it used to finish at 0xC115EFB0. 27 .comment 00000011 00000000 00000000 00a12728 2**0 CONTENTS, READONLY 28 .ARM.attributes 00000010 00000000 00000000 00a12739 2**0 CONTENTS, READONLY That's almost 1.5MB larger on an image size of 18MB. Percentage wise, that sounds small, but the thing to realise is that growth is independent of the image size, so a smaller image sees a larger %age wise growth in its size. People have already complained bitterly when I've said that stealing memory and taking out out of memblock should always be 1MB aligned, so /no one/ has the right to say "it's only 1.5MB, it doesn't matter" because quite frankly they should've been saying that and supporting me with the memblock issue. So, I really don't want to hear that argument! However, if you look at where these boundaries are placed, they're not quite in the right place. For example, the .init.data section is writable, and so should be grouped with the .data section. So should .data..percpu. Now, a few other things stand out from the above: (a) .text.head - imx, sunxi and tegra need to fix that. There is no specific meaning to it. (b) .init.text is executable, and can't be in a NX region when it's set as non-executable. (c) we can't free the .init sections (sections 15 through up to and including 23) anymore with this feature enabled because it's setup as read-only memory.
On Fri, Feb 21, 2014 at 5:20 AM, Russell King - ARM Linux <linux@arm.linux.org.uk> wrote: > On Fri, Feb 21, 2014 at 12:37:04PM +0000, Dave Martin wrote: >> It would be good if someone who's more familiar with the parms and >> vmlinux.lds stuff could take a look at it, though I don't see any >> obvious problem yet. > > The biggest issue with it is that we end up with: > > - the .text section rounded up to 1MB > - the .rodata section rounded up to 1MB > > That means we can end up wasting up to 1MB of memory for each (in the > worst case where we encroach into the next 1MB aligned region by a few > bytes) and this memory can't be re-used. > > The alternative is to adjust the maps such that we end up mapping the > .text / .rodata overlap 1MB using 4K pages, taking the additional TLB > hit by doing so. > > The .text is aligned to 1MB, so the majority of the first 0x8000 to > 0x100000 is unused. The end of the .text section is aligned to 1MB, > and the start of the .data section is also aligned to 1MB. > > So, the minimum kernel size is: 0x100000 + MB_ALIGN(sizeof(.text)) + > MB_ALIGN(sizeof(.rodata)) + MB_ALIGN(sizeof(init sections)) + sizeof(.data) > - 0x8000 > > So, looking at this kernel I've recently built: > > Idx Name Size VMA LMA File off Algn > 0 .head.text 00000204 c0008000 c0008000 00008000 2**2 > CONTENTS, ALLOC, LOAD, READONLY, CODE > > --- .text this gets set to 0xc0100000, we lose 0xc0008240 to 0xc0100000 > > 1 .text 006c4530 c0008240 c0008240 00008240 2**6 > CONTENTS, ALLOC, LOAD, READONLY, CODE > 2 .text.head 0000004c c06cc770 c06cc770 006cc770 2**2 > CONTENTS, ALLOC, LOAD, READONLY, CODE > > --- sizeof(.text) + sizeof(.text.head) becomes 0x700000 > --- .rodata starts at 0xc0800000 instead of 0xc06cd000 > > 3 .rodata 0022f568 c06cd000 c06cd000 006cd000 2**6 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 4 __bug_table 0000873c c08fc568 c08fc568 008fc568 2**0 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 5 .pci_fixup 00000030 c0904ca4 c0904ca4 00904ca4 2**2 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 6 __ksymtab 00008158 c0904cd4 c0904cd4 00904cd4 2**2 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 7 __ksymtab_gpl 00006858 c090ce2c c090ce2c 0090ce2c 2**2 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 8 __kcrctab 000040ac c0913684 c0913684 00913684 2**2 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 9 __kcrctab_gpl 0000342c c0917730 c0917730 00917730 2**2 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 10 __ksymtab_strings 00022a08 c091ab5c c091ab5c 0091ab5c 2**0 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 11 __param 00000c70 c093d564 c093d564 0093d564 2**2 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 12 __modver 00000e2c c093e1d4 c093e1d4 0093e1d4 2**2 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 13 __ex_table 00000f18 c093f000 c093f000 0093f000 2**3 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 14 .notes 00000024 c093ff18 c093ff18 0093ff18 2**2 > CONTENTS, ALLOC, LOAD, READONLY, CODE > 15 .vectors 00000020 00000000 c0940000 00940000 2**2 > CONTENTS, ALLOC, LOAD, READONLY, CODE > 16 .stubs 00000240 00001000 c0940020 00941000 2**5 > CONTENTS, ALLOC, LOAD, READONLY, CODE > 17 .init.text 00051760 c0940260 c0940260 00948260 2**5 > CONTENTS, ALLOC, LOAD, READONLY, CODE > 18 .exit.text 00002130 c09919c0 c09919c0 009999c0 2**2 > CONTENTS, ALLOC, LOAD, READONLY, CODE > 19 .init.arch.info 00000108 c0993af0 c0993af0 0099baf0 2**2 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 20 .init.tagtable 00000048 c0993bf8 c0993bf8 0099bbf8 2**2 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 21 .init.smpalt 000032f8 c0993c40 c0993c40 0099bc40 2**2 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 22 .init.pv_table 00000314 c0996f38 c0996f38 0099ef38 2**0 > CONTENTS, ALLOC, LOAD, READONLY, DATA > 23 .init.data 0000c19c c0997250 c0997250 0099f250 2**3 > CONTENTS, ALLOC, LOAD, DATA > 24 .data..percpu 000035c0 c09a4000 c09a4000 009ac000 2**6 > CONTENTS, ALLOC, LOAD, DATA > > --- sizeof previous sections is 0x2db000, which becomes 0x300000 > --- start of .data becomes 0xc0b00000 instead of 0xc09a8000 > > 25 .data 00062728 c09a8000 c09a8000 009b0000 2**6 > CONTENTS, ALLOC, LOAD, DATA > 26 .bss 00754870 c0a0a740 c0a0a740 00a12728 2**6 > ALLOC > > --- which means the kernel image finishes at 0xC12B6FB0 whereas it used > to finish at 0xC115EFB0. > > 27 .comment 00000011 00000000 00000000 00a12728 2**0 > CONTENTS, READONLY > 28 .ARM.attributes 00000010 00000000 00000000 00a12739 2**0 > CONTENTS, READONLY > > That's almost 1.5MB larger on an image size of 18MB. Percentage wise, > that sounds small, but the thing to realise is that growth is independent > of the image size, so a smaller image sees a larger %age wise growth in > its size. > > People have already complained bitterly when I've said that stealing > memory and taking out out of memblock should always be 1MB aligned, so > /no one/ has the right to say "it's only 1.5MB, it doesn't matter" > because quite frankly they should've been saying that and supporting me > with the memblock issue. So, I really don't want to hear that argument! Absolutely, the memory waste is a problem. This is why I made sure to leave the rodata as a separate config item. It seems like if people want this feature, the cost is either going to be this kind of alignment loss, or TLB hit switching to 4K pages. It seems like the latter is significantly more "costly". Regardless, that's why it's all behind config options. > However, if you look at where these boundaries are placed, they're not > quite in the right place. For example, the .init.data section is writable, > and so should be grouped with the .data section. So should .data..percpu. > > Now, a few other things stand out from the above: > > (a) .text.head - imx, sunxi and tegra need to fix that. There is no > specific meaning to it. > > (b) .init.text is executable, and can't be in a NX region when it's > set as non-executable. > > (c) we can't free the .init sections (sections 15 through up to and > including 23) anymore with this feature enabled because it's setup > as read-only memory. Perhaps I've misunderstood something, but I don't think b) and c) or the comments about .init.data and .data..percpu are problems for this series because of when the mapping happens. Until init finishes, these sections are fully RXW. Only after init is done is the init section made NX. -Kees
On Fri, Feb 21, 2014 at 2:09 PM, Kees Cook <keescook@chromium.org> wrote: > On Fri, Feb 21, 2014 at 5:20 AM, Russell King - ARM Linux > <linux@arm.linux.org.uk> wrote: >> On Fri, Feb 21, 2014 at 12:37:04PM +0000, Dave Martin wrote: >>> It would be good if someone who's more familiar with the parms and >>> vmlinux.lds stuff could take a look at it, though I don't see any >>> obvious problem yet. >> >> The biggest issue with it is that we end up with: >> >> - the .text section rounded up to 1MB >> - the .rodata section rounded up to 1MB >> >> That means we can end up wasting up to 1MB of memory for each (in the >> worst case where we encroach into the next 1MB aligned region by a few >> bytes) and this memory can't be re-used. >> >> The alternative is to adjust the maps such that we end up mapping the >> .text / .rodata overlap 1MB using 4K pages, taking the additional TLB >> hit by doing so. >> >> The .text is aligned to 1MB, so the majority of the first 0x8000 to >> 0x100000 is unused. The end of the .text section is aligned to 1MB, >> and the start of the .data section is also aligned to 1MB. >> >> So, the minimum kernel size is: 0x100000 + MB_ALIGN(sizeof(.text)) + >> MB_ALIGN(sizeof(.rodata)) + MB_ALIGN(sizeof(init sections)) + sizeof(.data) >> - 0x8000 >> >> So, looking at this kernel I've recently built: >> >> Idx Name Size VMA LMA File off Algn >> 0 .head.text 00000204 c0008000 c0008000 00008000 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, CODE >> >> --- .text this gets set to 0xc0100000, we lose 0xc0008240 to 0xc0100000 >> >> 1 .text 006c4530 c0008240 c0008240 00008240 2**6 >> CONTENTS, ALLOC, LOAD, READONLY, CODE >> 2 .text.head 0000004c c06cc770 c06cc770 006cc770 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, CODE >> >> --- sizeof(.text) + sizeof(.text.head) becomes 0x700000 >> --- .rodata starts at 0xc0800000 instead of 0xc06cd000 >> >> 3 .rodata 0022f568 c06cd000 c06cd000 006cd000 2**6 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 4 __bug_table 0000873c c08fc568 c08fc568 008fc568 2**0 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 5 .pci_fixup 00000030 c0904ca4 c0904ca4 00904ca4 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 6 __ksymtab 00008158 c0904cd4 c0904cd4 00904cd4 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 7 __ksymtab_gpl 00006858 c090ce2c c090ce2c 0090ce2c 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 8 __kcrctab 000040ac c0913684 c0913684 00913684 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 9 __kcrctab_gpl 0000342c c0917730 c0917730 00917730 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 10 __ksymtab_strings 00022a08 c091ab5c c091ab5c 0091ab5c 2**0 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 11 __param 00000c70 c093d564 c093d564 0093d564 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 12 __modver 00000e2c c093e1d4 c093e1d4 0093e1d4 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 13 __ex_table 00000f18 c093f000 c093f000 0093f000 2**3 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 14 .notes 00000024 c093ff18 c093ff18 0093ff18 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, CODE >> 15 .vectors 00000020 00000000 c0940000 00940000 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, CODE >> 16 .stubs 00000240 00001000 c0940020 00941000 2**5 >> CONTENTS, ALLOC, LOAD, READONLY, CODE >> 17 .init.text 00051760 c0940260 c0940260 00948260 2**5 >> CONTENTS, ALLOC, LOAD, READONLY, CODE >> 18 .exit.text 00002130 c09919c0 c09919c0 009999c0 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, CODE >> 19 .init.arch.info 00000108 c0993af0 c0993af0 0099baf0 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 20 .init.tagtable 00000048 c0993bf8 c0993bf8 0099bbf8 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 21 .init.smpalt 000032f8 c0993c40 c0993c40 0099bc40 2**2 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 22 .init.pv_table 00000314 c0996f38 c0996f38 0099ef38 2**0 >> CONTENTS, ALLOC, LOAD, READONLY, DATA >> 23 .init.data 0000c19c c0997250 c0997250 0099f250 2**3 >> CONTENTS, ALLOC, LOAD, DATA >> 24 .data..percpu 000035c0 c09a4000 c09a4000 009ac000 2**6 >> CONTENTS, ALLOC, LOAD, DATA >> >> --- sizeof previous sections is 0x2db000, which becomes 0x300000 >> --- start of .data becomes 0xc0b00000 instead of 0xc09a8000 >> >> 25 .data 00062728 c09a8000 c09a8000 009b0000 2**6 >> CONTENTS, ALLOC, LOAD, DATA >> 26 .bss 00754870 c0a0a740 c0a0a740 00a12728 2**6 >> ALLOC >> >> --- which means the kernel image finishes at 0xC12B6FB0 whereas it used >> to finish at 0xC115EFB0. >> >> 27 .comment 00000011 00000000 00000000 00a12728 2**0 >> CONTENTS, READONLY >> 28 .ARM.attributes 00000010 00000000 00000000 00a12739 2**0 >> CONTENTS, READONLY >> >> That's almost 1.5MB larger on an image size of 18MB. Percentage wise, >> that sounds small, but the thing to realise is that growth is independent >> of the image size, so a smaller image sees a larger %age wise growth in >> its size. >> >> People have already complained bitterly when I've said that stealing >> memory and taking out out of memblock should always be 1MB aligned, so >> /no one/ has the right to say "it's only 1.5MB, it doesn't matter" >> because quite frankly they should've been saying that and supporting me >> with the memblock issue. So, I really don't want to hear that argument! > > Absolutely, the memory waste is a problem. This is why I made sure to > leave the rodata as a separate config item. It seems like if people > want this feature, the cost is either going to be this kind of > alignment loss, or TLB hit switching to 4K pages. It seems like the > latter is significantly more "costly". Regardless, that's why it's all > behind config options. > >> However, if you look at where these boundaries are placed, they're not >> quite in the right place. For example, the .init.data section is writable, >> and so should be grouped with the .data section. So should .data..percpu. >> >> Now, a few other things stand out from the above: >> >> (a) .text.head - imx, sunxi and tegra need to fix that. There is no >> specific meaning to it. >> >> (b) .init.text is executable, and can't be in a NX region when it's >> set as non-executable. >> >> (c) we can't free the .init sections (sections 15 through up to and >> including 23) anymore with this feature enabled because it's setup >> as read-only memory. > > Perhaps I've misunderstood something, but I don't think b) and c) or > the comments about .init.data and .data..percpu are problems for this > series because of when the mapping happens. Until init finishes, these > sections are fully RXW. Only after init is done is the init section > made NX. I haven't heard anything back, and I think I've answered the questions raised. Since these changes are behind CONFIG items, other folks have successfully tested the series, and at least Google, Linaro, and the codeaurora folks are interested in these features, can I send these to the patch tracker? Thanks, -Kees
On 3/13/2014 12:07 PM, Kees Cook wrote: > On Fri, Feb 21, 2014 at 2:09 PM, Kees Cook <keescook@chromium.org> wrote: >> On Fri, Feb 21, 2014 at 5:20 AM, Russell King - ARM Linux >> <linux@arm.linux.org.uk> wrote: >>> On Fri, Feb 21, 2014 at 12:37:04PM +0000, Dave Martin wrote: >>>> It would be good if someone who's more familiar with the parms and >>>> vmlinux.lds stuff could take a look at it, though I don't see any >>>> obvious problem yet. >>> >>> The biggest issue with it is that we end up with: >>> >>> - the .text section rounded up to 1MB >>> - the .rodata section rounded up to 1MB >>> >>> That means we can end up wasting up to 1MB of memory for each (in the >>> worst case where we encroach into the next 1MB aligned region by a few >>> bytes) and this memory can't be re-used. >>> >>> The alternative is to adjust the maps such that we end up mapping the >>> .text / .rodata overlap 1MB using 4K pages, taking the additional TLB >>> hit by doing so. >>> >>> The .text is aligned to 1MB, so the majority of the first 0x8000 to >>> 0x100000 is unused. The end of the .text section is aligned to 1MB, >>> and the start of the .data section is also aligned to 1MB. >>> >>> So, the minimum kernel size is: 0x100000 + MB_ALIGN(sizeof(.text)) + >>> MB_ALIGN(sizeof(.rodata)) + MB_ALIGN(sizeof(init sections)) + sizeof(.data) >>> - 0x8000 >>> >>> So, looking at this kernel I've recently built: >>> >>> Idx Name Size VMA LMA File off Algn >>> 0 .head.text 00000204 c0008000 c0008000 00008000 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>> >>> --- .text this gets set to 0xc0100000, we lose 0xc0008240 to 0xc0100000 >>> >>> 1 .text 006c4530 c0008240 c0008240 00008240 2**6 >>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>> 2 .text.head 0000004c c06cc770 c06cc770 006cc770 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>> >>> --- sizeof(.text) + sizeof(.text.head) becomes 0x700000 >>> --- .rodata starts at 0xc0800000 instead of 0xc06cd000 >>> >>> 3 .rodata 0022f568 c06cd000 c06cd000 006cd000 2**6 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 4 __bug_table 0000873c c08fc568 c08fc568 008fc568 2**0 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 5 .pci_fixup 00000030 c0904ca4 c0904ca4 00904ca4 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 6 __ksymtab 00008158 c0904cd4 c0904cd4 00904cd4 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 7 __ksymtab_gpl 00006858 c090ce2c c090ce2c 0090ce2c 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 8 __kcrctab 000040ac c0913684 c0913684 00913684 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 9 __kcrctab_gpl 0000342c c0917730 c0917730 00917730 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 10 __ksymtab_strings 00022a08 c091ab5c c091ab5c 0091ab5c 2**0 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 11 __param 00000c70 c093d564 c093d564 0093d564 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 12 __modver 00000e2c c093e1d4 c093e1d4 0093e1d4 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 13 __ex_table 00000f18 c093f000 c093f000 0093f000 2**3 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 14 .notes 00000024 c093ff18 c093ff18 0093ff18 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>> 15 .vectors 00000020 00000000 c0940000 00940000 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>> 16 .stubs 00000240 00001000 c0940020 00941000 2**5 >>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>> 17 .init.text 00051760 c0940260 c0940260 00948260 2**5 >>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>> 18 .exit.text 00002130 c09919c0 c09919c0 009999c0 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>> 19 .init.arch.info 00000108 c0993af0 c0993af0 0099baf0 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 20 .init.tagtable 00000048 c0993bf8 c0993bf8 0099bbf8 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 21 .init.smpalt 000032f8 c0993c40 c0993c40 0099bc40 2**2 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 22 .init.pv_table 00000314 c0996f38 c0996f38 0099ef38 2**0 >>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>> 23 .init.data 0000c19c c0997250 c0997250 0099f250 2**3 >>> CONTENTS, ALLOC, LOAD, DATA >>> 24 .data..percpu 000035c0 c09a4000 c09a4000 009ac000 2**6 >>> CONTENTS, ALLOC, LOAD, DATA >>> >>> --- sizeof previous sections is 0x2db000, which becomes 0x300000 >>> --- start of .data becomes 0xc0b00000 instead of 0xc09a8000 >>> >>> 25 .data 00062728 c09a8000 c09a8000 009b0000 2**6 >>> CONTENTS, ALLOC, LOAD, DATA >>> 26 .bss 00754870 c0a0a740 c0a0a740 00a12728 2**6 >>> ALLOC >>> >>> --- which means the kernel image finishes at 0xC12B6FB0 whereas it used >>> to finish at 0xC115EFB0. >>> >>> 27 .comment 00000011 00000000 00000000 00a12728 2**0 >>> CONTENTS, READONLY >>> 28 .ARM.attributes 00000010 00000000 00000000 00a12739 2**0 >>> CONTENTS, READONLY >>> >>> That's almost 1.5MB larger on an image size of 18MB. Percentage wise, >>> that sounds small, but the thing to realise is that growth is independent >>> of the image size, so a smaller image sees a larger %age wise growth in >>> its size. >>> >>> People have already complained bitterly when I've said that stealing >>> memory and taking out out of memblock should always be 1MB aligned, so >>> /no one/ has the right to say "it's only 1.5MB, it doesn't matter" >>> because quite frankly they should've been saying that and supporting me >>> with the memblock issue. So, I really don't want to hear that argument! >> >> Absolutely, the memory waste is a problem. This is why I made sure to >> leave the rodata as a separate config item. It seems like if people >> want this feature, the cost is either going to be this kind of >> alignment loss, or TLB hit switching to 4K pages. It seems like the >> latter is significantly more "costly". Regardless, that's why it's all >> behind config options. >> >>> However, if you look at where these boundaries are placed, they're not >>> quite in the right place. For example, the .init.data section is writable, >>> and so should be grouped with the .data section. So should .data..percpu. >>> >>> Now, a few other things stand out from the above: >>> >>> (a) .text.head - imx, sunxi and tegra need to fix that. There is no >>> specific meaning to it. >>> >>> (b) .init.text is executable, and can't be in a NX region when it's >>> set as non-executable. >>> >>> (c) we can't free the .init sections (sections 15 through up to and >>> including 23) anymore with this feature enabled because it's setup >>> as read-only memory. >> >> Perhaps I've misunderstood something, but I don't think b) and c) or >> the comments about .init.data and .data..percpu are problems for this >> series because of when the mapping happens. Until init finishes, these >> sections are fully RXW. Only after init is done is the init section >> made NX. > > I haven't heard anything back, and I think I've answered the questions raised. > > Since these changes are behind CONFIG items, other folks have > successfully tested the series, and at least Google, Linaro, and the > codeaurora folks are interested in these features, can I send these to > the patch tracker? I'm getting a section mismatch warning when compiling WARNING: vmlinux.o(.text+0x1257c): Section mismatch in reference from the function free_initmem() to the (unknown reference) .init.data:(unknown) The function free_initmem() references the (unknown reference) __initdata (unknown). This is often because free_initmem lacks a __initdata annotation or the annotation of (unknown) is wrong. free_initmem -> fix_kernmem_perms aren't marked as __init but there is a reference to section_perms which is marked as __initdata. Kernel output looks as expected on a v7 non LPAE system / # cat /sys/kernel/debug/kernel_page_tables ---[ Modules ]--- 0xbfe01000-0xbfe0d000 48K RW NX SHD MEM/CACHED/WBWA ---[ Kernel Mapping ]--- 0xc0000000-0xc0100000 1M RW NX SHD 0xc0100000-0xc0600000 5M ro x SHD 0xc0600000-0xc0900000 3M ro NX SHD 0xc0900000-0xef800000 751M RW NX SHD ---[ vmalloc() Area ]--- 0xf0000000-0xf0001000 4K RW NX SHD DEV/SHARED 0xf0002000-0xf0003000 4K RW NX SHD DEV/SHARED 0xf0006000-0xf0007000 4K RW NX SHD DEV/SHARED 0xf0018000-0xf0058000 256K RW NX SHD MEM/BUFFERABLE/WC 0xf005a000-0xf005b000 4K RW NX SHD DEV/SHARED 0xf005c000-0xf005f000 12K RW NX SHD MEM/CACHED/WBWA 0xf0060000-0xf0064000 16K RW NX SHD DEV/SHARED 0xf0068000-0xf006e000 24K RW NX SHD DEV/SHARED ---[ vmalloc() End ]--- ---[ Fixmap Area ]--- 0xfffae000-0xfffb0000 8K RW NX SHD MEM/CACHED/WBWA 0xfffbe000-0xfffc0000 8K RW NX SHD MEM/CACHED/WBWA 0xfffce000-0xfffd0000 8K RW NX SHD MEM/CACHED/WBWA 0xfffde000-0xfffe0000 8K RW NX SHD MEM/CACHED/WBWA ---[ Vectors ]--- 0xffff0000-0xffff1000 4K USR ro x SHD MEM/CACHED/WBWA 0xffff1000-0xffff2000 4K ro x SHD MEM/CACHED/WBWA ---[ Vectors End ]--- Thanks, Laura
On 2/17/2014 4:34 AM, Dave Martin wrote: > On Fri, Feb 14, 2014 at 11:11:07AM -0800, Kees Cook wrote: >> On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin@arm.com> wrote: >>> On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: >>>> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though >>>> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which >>>> sets rodata read-only (but executable), where as this option additionally >>>> splits rodata from the kernel text (resulting in potentially more memory >>>> lost to padding) and sets it non-executable as well. The end result is >>>> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be >>>> marked purely read-only. >>> >>> This triggers an Oops in kexec, because we have a block of code in .text >>> which is a template for generating baremetal code to relocate the new >>> kernel, and some literal words are written into it before copying. >> >> You're writing into the text area? I would imagine that >> CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the >> right place to be building code -- shouldn't the module area be used >> for that? >> >>> Possibly this should be in .rodata, not .text. >> >> Well, rodata should be neither writable nor executable. > > We're not writing into code exactly. > > This code is never executed in-place in vmlinux. It gets copied, and > only copies are ever executed. > > Some pointers and offsets get poked into the code to configure it. > > I think it would be better simply to put the code in .rodata, and > poke paramaters into the copy, not the original -- but that's a bit > more awkward to code up, since the values can't be poked simply by > writing global variables. > >> >>> There may be a few other instances of this kind of thing. >> >> This config will certainly find them! :) But, that's why it's behind a config. > > I haven't tested exhaustively, but it think this is sufficient for a > Tested-by. The patch does seem to be doing what it is intended to > do, and doesn't seem to be triggering false positives all over the > place. > >> >>> Are you aware of similar situations on other arches? >> >> I think there were some problems a long time ago on x86 for rodata too. > > It would be good to get this kexec case fixed -- I'll try to hack up > a separate patch. > FWIW, we've hit issues not just with kexec but kprobes as well. The same problems exist with this series: / # echo p:nl 0xc01d5c00 >> /sys/kernel/debug/tracing/kprobe_events / # echo 1 > /sys/kernel/debug/tracing/events/kprobes/nl/enable [ 1639.739629] Unable to handle kernel paging request at virtual address c01d5c00 [ 1639.739655] pgd = edbc4000 [ 1639.745730] [c01d5c00] *pgd=0001141e(bad) [ 1639.752413] Internal error: Oops: 80d [#1] PREEMPT SMP ARM [ 1639.752503] Modules linked in: [ 1639.760920] CPU: 0 PID: 58 Comm: sh Not tainted 3.14.0-rc7-next-20140318-00004-ga0191b7-dirty #170 [ 1639.761015] task: edb90d80 ti: ed018000 task.ti: ed018000 [ 1639.769870] PC is at patch_text+0x4/0x10 [ 1639.775333] LR is at arm_kprobe+0x28/0x38 [ 1639.779327] pc : [<c058acec>] lr : [<c058bcc4>] psr: 20000013 [ 1639.779327] sp : ed019f10 ip : a0000013 fp : 01e7fb34 [ 1639.783241] r10: 00000000 r9 : 01e80ab8 r8 : 00000002 [ 1639.794517] r7 : ed019f80 r6 : ed900bc4 r5 : edb19fa0 r4 : edb19f08 [ 1639.799727] r3 : c01d5c00 r2 : ed019f08 r1 : e7f001f8 r0 : c01d5c00 [ 1639.806326] Flags: nzCv IRQs on FIQs on Mode SVC_32 ISA ARM Segment user [ 1639.812837] Control: 10c5787d Table: 2dbc406a DAC: 00000015 [ 1639.820040] Process sh (pid: 58, stack limit = 0xed018240) [ 1639.825768] Stack: (0xed019f10 to 0xed01a000) [ 1639.831150] 9f00: 00000000 c058bd5c ed900ba0 c018b724 [ 1639.835584] 9f20: ed6fab00 00000000 edb2d240 00000002 ed019f80 c018bdc4 ffffffff 00000001 [ 1639.843743] 9f40: edb2d240 01e80ab8 ed019f80 00000002 00000002 c01e1b1c 01e7fb34 c011c3c0 [ 1639.851902] 9f60: 00000003 00000000 00000000 edb2d240 edb2d240 00000002 01e80ab8 c01e2108 [ 1639.860063] 9f80: 00000000 00000000 00200200 00157ecc 00000001 01e80ab8 00000004 c0106e64 [ 1639.868222] 9fa0: ed018000 c0106ce0 00157ecc 00000001 00000001 01e80ab8 00000002 00000000 [ 1639.876381] 9fc0: 00157ecc 00000001 01e80ab8 00000004 00000020 01e7fb48 01e7fb14 01e7fb34 [ 1639.884542] 9fe0: 00000000 bef4562c 0001ee5d 0000a8cc 60000010 00000001 00000000 00000000 [ 1639.892709] [<c058acec>] (patch_text) from [<c058bcc4>] (arm_kprobe+0x28/0x38) [ 1639.900862] [<c058bcc4>] (arm_kprobe) from [<c058bd5c>] (enable_kprobe+0x88/0x94) [ 1639.907983] [<c058bd5c>] (enable_kprobe) from [<c018b724>] (__ftrace_event_enable_disable+0x13c/0x200) [ 1639.915537] [<c018b724>] (__ftrace_event_enable_disable) from [<c018bdc4>] (event_enable_write+0x78/0xd4) [ 1639.924741] [<c018bdc4>] (event_enable_write) from [<c01e1b1c>] (vfs_write+0xac/0x188) [ 1639.934372] [<c01e1b1c>] (vfs_write) from [<c01e2108>] (SyS_write+0x40/0x94) [ 1639.942187] [<c01e2108>] (SyS_write) from [<c0106ce0>] (ret_fast_syscall+0x0/0x30) [ 1639.949386] Code: e4831004 e1a01003 eaee298d e1a03000 (e4831004) [ 1639.956766] ---[ end trace b548269e2c7a3190 ]--- We had some functions that allowed the text to be temporarily made writable but something uniform for kexec would be useful as well (our kexec solution has been 'turn it off') > Cheers > ---Dave > Thanks, Laura
On Sun, Mar 23, 2014 at 12:32 PM, Laura Abbott <lauraa@codeaurora.org> wrote: > On 3/13/2014 12:07 PM, Kees Cook wrote: >> On Fri, Feb 21, 2014 at 2:09 PM, Kees Cook <keescook@chromium.org> wrote: >>> On Fri, Feb 21, 2014 at 5:20 AM, Russell King - ARM Linux >>> <linux@arm.linux.org.uk> wrote: >>>> On Fri, Feb 21, 2014 at 12:37:04PM +0000, Dave Martin wrote: >>>>> It would be good if someone who's more familiar with the parms and >>>>> vmlinux.lds stuff could take a look at it, though I don't see any >>>>> obvious problem yet. >>>> >>>> The biggest issue with it is that we end up with: >>>> >>>> - the .text section rounded up to 1MB >>>> - the .rodata section rounded up to 1MB >>>> >>>> That means we can end up wasting up to 1MB of memory for each (in the >>>> worst case where we encroach into the next 1MB aligned region by a few >>>> bytes) and this memory can't be re-used. >>>> >>>> The alternative is to adjust the maps such that we end up mapping the >>>> .text / .rodata overlap 1MB using 4K pages, taking the additional TLB >>>> hit by doing so. >>>> >>>> The .text is aligned to 1MB, so the majority of the first 0x8000 to >>>> 0x100000 is unused. The end of the .text section is aligned to 1MB, >>>> and the start of the .data section is also aligned to 1MB. >>>> >>>> So, the minimum kernel size is: 0x100000 + MB_ALIGN(sizeof(.text)) + >>>> MB_ALIGN(sizeof(.rodata)) + MB_ALIGN(sizeof(init sections)) + sizeof(.data) >>>> - 0x8000 >>>> >>>> So, looking at this kernel I've recently built: >>>> >>>> Idx Name Size VMA LMA File off Algn >>>> 0 .head.text 00000204 c0008000 c0008000 00008000 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>>> >>>> --- .text this gets set to 0xc0100000, we lose 0xc0008240 to 0xc0100000 >>>> >>>> 1 .text 006c4530 c0008240 c0008240 00008240 2**6 >>>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>>> 2 .text.head 0000004c c06cc770 c06cc770 006cc770 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>>> >>>> --- sizeof(.text) + sizeof(.text.head) becomes 0x700000 >>>> --- .rodata starts at 0xc0800000 instead of 0xc06cd000 >>>> >>>> 3 .rodata 0022f568 c06cd000 c06cd000 006cd000 2**6 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 4 __bug_table 0000873c c08fc568 c08fc568 008fc568 2**0 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 5 .pci_fixup 00000030 c0904ca4 c0904ca4 00904ca4 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 6 __ksymtab 00008158 c0904cd4 c0904cd4 00904cd4 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 7 __ksymtab_gpl 00006858 c090ce2c c090ce2c 0090ce2c 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 8 __kcrctab 000040ac c0913684 c0913684 00913684 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 9 __kcrctab_gpl 0000342c c0917730 c0917730 00917730 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 10 __ksymtab_strings 00022a08 c091ab5c c091ab5c 0091ab5c 2**0 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 11 __param 00000c70 c093d564 c093d564 0093d564 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 12 __modver 00000e2c c093e1d4 c093e1d4 0093e1d4 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 13 __ex_table 00000f18 c093f000 c093f000 0093f000 2**3 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 14 .notes 00000024 c093ff18 c093ff18 0093ff18 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>>> 15 .vectors 00000020 00000000 c0940000 00940000 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>>> 16 .stubs 00000240 00001000 c0940020 00941000 2**5 >>>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>>> 17 .init.text 00051760 c0940260 c0940260 00948260 2**5 >>>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>>> 18 .exit.text 00002130 c09919c0 c09919c0 009999c0 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, CODE >>>> 19 .init.arch.info 00000108 c0993af0 c0993af0 0099baf0 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 20 .init.tagtable 00000048 c0993bf8 c0993bf8 0099bbf8 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 21 .init.smpalt 000032f8 c0993c40 c0993c40 0099bc40 2**2 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 22 .init.pv_table 00000314 c0996f38 c0996f38 0099ef38 2**0 >>>> CONTENTS, ALLOC, LOAD, READONLY, DATA >>>> 23 .init.data 0000c19c c0997250 c0997250 0099f250 2**3 >>>> CONTENTS, ALLOC, LOAD, DATA >>>> 24 .data..percpu 000035c0 c09a4000 c09a4000 009ac000 2**6 >>>> CONTENTS, ALLOC, LOAD, DATA >>>> >>>> --- sizeof previous sections is 0x2db000, which becomes 0x300000 >>>> --- start of .data becomes 0xc0b00000 instead of 0xc09a8000 >>>> >>>> 25 .data 00062728 c09a8000 c09a8000 009b0000 2**6 >>>> CONTENTS, ALLOC, LOAD, DATA >>>> 26 .bss 00754870 c0a0a740 c0a0a740 00a12728 2**6 >>>> ALLOC >>>> >>>> --- which means the kernel image finishes at 0xC12B6FB0 whereas it used >>>> to finish at 0xC115EFB0. >>>> >>>> 27 .comment 00000011 00000000 00000000 00a12728 2**0 >>>> CONTENTS, READONLY >>>> 28 .ARM.attributes 00000010 00000000 00000000 00a12739 2**0 >>>> CONTENTS, READONLY >>>> >>>> That's almost 1.5MB larger on an image size of 18MB. Percentage wise, >>>> that sounds small, but the thing to realise is that growth is independent >>>> of the image size, so a smaller image sees a larger %age wise growth in >>>> its size. >>>> >>>> People have already complained bitterly when I've said that stealing >>>> memory and taking out out of memblock should always be 1MB aligned, so >>>> /no one/ has the right to say "it's only 1.5MB, it doesn't matter" >>>> because quite frankly they should've been saying that and supporting me >>>> with the memblock issue. So, I really don't want to hear that argument! >>> >>> Absolutely, the memory waste is a problem. This is why I made sure to >>> leave the rodata as a separate config item. It seems like if people >>> want this feature, the cost is either going to be this kind of >>> alignment loss, or TLB hit switching to 4K pages. It seems like the >>> latter is significantly more "costly". Regardless, that's why it's all >>> behind config options. >>> >>>> However, if you look at where these boundaries are placed, they're not >>>> quite in the right place. For example, the .init.data section is writable, >>>> and so should be grouped with the .data section. So should .data..percpu. >>>> >>>> Now, a few other things stand out from the above: >>>> >>>> (a) .text.head - imx, sunxi and tegra need to fix that. There is no >>>> specific meaning to it. >>>> >>>> (b) .init.text is executable, and can't be in a NX region when it's >>>> set as non-executable. >>>> >>>> (c) we can't free the .init sections (sections 15 through up to and >>>> including 23) anymore with this feature enabled because it's setup >>>> as read-only memory. >>> >>> Perhaps I've misunderstood something, but I don't think b) and c) or >>> the comments about .init.data and .data..percpu are problems for this >>> series because of when the mapping happens. Until init finishes, these >>> sections are fully RXW. Only after init is done is the init section >>> made NX. >> >> I haven't heard anything back, and I think I've answered the questions raised. >> >> Since these changes are behind CONFIG items, other folks have >> successfully tested the series, and at least Google, Linaro, and the >> codeaurora folks are interested in these features, can I send these to >> the patch tracker? > > I'm getting a section mismatch warning when compiling Ah! Good catch, thanks. That config got turned off for me somewhere along the lines. I've turned it back on and I see the same thing. > WARNING: vmlinux.o(.text+0x1257c): Section mismatch in reference from the > function free_initmem() to the (unknown reference) .init.data:(unknown) > The function free_initmem() references > the (unknown reference) __initdata (unknown). > This is often because free_initmem lacks a __initdata > annotation or the annotation of (unknown) is wrong. > > free_initmem -> fix_kernmem_perms aren't marked as __init but there is a > reference to section_perms which is marked as __initdata. Yeah, I don't see a way to have fix_kernmem_perms marked as __init without continuing to trigger the warnings. Instead, I've now dropped __initdata from section_perms instead. That cleaned up the warning for me; I will send an updated version. > Kernel output looks as expected on a v7 non LPAE system Great, thanks for the confirmation! :) -Kees > > / # cat /sys/kernel/debug/kernel_page_tables > ---[ Modules ]--- > 0xbfe01000-0xbfe0d000 48K RW NX SHD MEM/CACHED/WBWA > ---[ Kernel Mapping ]--- > 0xc0000000-0xc0100000 1M RW NX SHD > 0xc0100000-0xc0600000 5M ro x SHD > 0xc0600000-0xc0900000 3M ro NX SHD > 0xc0900000-0xef800000 751M RW NX SHD > ---[ vmalloc() Area ]--- > 0xf0000000-0xf0001000 4K RW NX SHD DEV/SHARED > 0xf0002000-0xf0003000 4K RW NX SHD DEV/SHARED > 0xf0006000-0xf0007000 4K RW NX SHD DEV/SHARED > 0xf0018000-0xf0058000 256K RW NX SHD MEM/BUFFERABLE/WC > 0xf005a000-0xf005b000 4K RW NX SHD DEV/SHARED > 0xf005c000-0xf005f000 12K RW NX SHD MEM/CACHED/WBWA > 0xf0060000-0xf0064000 16K RW NX SHD DEV/SHARED > 0xf0068000-0xf006e000 24K RW NX SHD DEV/SHARED > ---[ vmalloc() End ]--- > ---[ Fixmap Area ]--- > 0xfffae000-0xfffb0000 8K RW NX SHD MEM/CACHED/WBWA > 0xfffbe000-0xfffc0000 8K RW NX SHD MEM/CACHED/WBWA > 0xfffce000-0xfffd0000 8K RW NX SHD MEM/CACHED/WBWA > 0xfffde000-0xfffe0000 8K RW NX SHD MEM/CACHED/WBWA > ---[ Vectors ]--- > 0xffff0000-0xffff1000 4K USR ro x SHD MEM/CACHED/WBWA > 0xffff1000-0xffff2000 4K ro x SHD MEM/CACHED/WBWA > ---[ Vectors End ]--- > > Thanks, > Laura > > -- > Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, > hosted by The Linux Foundation
On Sun, Mar 23, 2014 at 12:47 PM, Laura Abbott <lauraa@codeaurora.org> wrote: > On 2/17/2014 4:34 AM, Dave Martin wrote: >> On Fri, Feb 14, 2014 at 11:11:07AM -0800, Kees Cook wrote: >>> On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin@arm.com> wrote: >>>> On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: >>>>> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though >>>>> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which >>>>> sets rodata read-only (but executable), where as this option additionally >>>>> splits rodata from the kernel text (resulting in potentially more memory >>>>> lost to padding) and sets it non-executable as well. The end result is >>>>> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be >>>>> marked purely read-only. >>>> >>>> This triggers an Oops in kexec, because we have a block of code in .text >>>> which is a template for generating baremetal code to relocate the new >>>> kernel, and some literal words are written into it before copying. >>> >>> You're writing into the text area? I would imagine that >>> CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the >>> right place to be building code -- shouldn't the module area be used >>> for that? >>> >>>> Possibly this should be in .rodata, not .text. >>> >>> Well, rodata should be neither writable nor executable. >> >> We're not writing into code exactly. >> >> This code is never executed in-place in vmlinux. It gets copied, and >> only copies are ever executed. >> >> Some pointers and offsets get poked into the code to configure it. >> >> I think it would be better simply to put the code in .rodata, and >> poke paramaters into the copy, not the original -- but that's a bit >> more awkward to code up, since the values can't be poked simply by >> writing global variables. >> >>> >>>> There may be a few other instances of this kind of thing. >>> >>> This config will certainly find them! :) But, that's why it's behind a config. >> >> I haven't tested exhaustively, but it think this is sufficient for a >> Tested-by. The patch does seem to be doing what it is intended to >> do, and doesn't seem to be triggering false positives all over the >> place. >> >>> >>>> Are you aware of similar situations on other arches? >>> >>> I think there were some problems a long time ago on x86 for rodata too. >> >> It would be good to get this kexec case fixed -- I'll try to hack up >> a separate patch. >> > > FWIW, we've hit issues not just with kexec but kprobes as well. The same > problems exist with this series: For this stage, how about I make this "depends on KEXEC=n && KPROBES=n"? Then as these areas get fixed, we can drop those requirements. -Kees > > / # echo p:nl 0xc01d5c00 >> /sys/kernel/debug/tracing/kprobe_events > / # echo 1 > /sys/kernel/debug/tracing/events/kprobes/nl/enable > [ 1639.739629] Unable to handle kernel paging request at virtual address c01d5c00 > [ 1639.739655] pgd = edbc4000 > [ 1639.745730] [c01d5c00] *pgd=0001141e(bad) > [ 1639.752413] Internal error: Oops: 80d [#1] PREEMPT SMP ARM > [ 1639.752503] Modules linked in: > [ 1639.760920] CPU: 0 PID: 58 Comm: sh Not tainted 3.14.0-rc7-next-20140318-00004-ga0191b7-dirty #170 > [ 1639.761015] task: edb90d80 ti: ed018000 task.ti: ed018000 > [ 1639.769870] PC is at patch_text+0x4/0x10 > [ 1639.775333] LR is at arm_kprobe+0x28/0x38 > [ 1639.779327] pc : [<c058acec>] lr : [<c058bcc4>] psr: 20000013 > [ 1639.779327] sp : ed019f10 ip : a0000013 fp : 01e7fb34 > [ 1639.783241] r10: 00000000 r9 : 01e80ab8 r8 : 00000002 > [ 1639.794517] r7 : ed019f80 r6 : ed900bc4 r5 : edb19fa0 r4 : edb19f08 > [ 1639.799727] r3 : c01d5c00 r2 : ed019f08 r1 : e7f001f8 r0 : c01d5c00 > [ 1639.806326] Flags: nzCv IRQs on FIQs on Mode SVC_32 ISA ARM Segment user > [ 1639.812837] Control: 10c5787d Table: 2dbc406a DAC: 00000015 > [ 1639.820040] Process sh (pid: 58, stack limit = 0xed018240) > [ 1639.825768] Stack: (0xed019f10 to 0xed01a000) > [ 1639.831150] 9f00: 00000000 c058bd5c ed900ba0 c018b724 > [ 1639.835584] 9f20: ed6fab00 00000000 edb2d240 00000002 ed019f80 c018bdc4 ffffffff 00000001 > [ 1639.843743] 9f40: edb2d240 01e80ab8 ed019f80 00000002 00000002 c01e1b1c 01e7fb34 c011c3c0 > [ 1639.851902] 9f60: 00000003 00000000 00000000 edb2d240 edb2d240 00000002 01e80ab8 c01e2108 > [ 1639.860063] 9f80: 00000000 00000000 00200200 00157ecc 00000001 01e80ab8 00000004 c0106e64 > [ 1639.868222] 9fa0: ed018000 c0106ce0 00157ecc 00000001 00000001 01e80ab8 00000002 00000000 > [ 1639.876381] 9fc0: 00157ecc 00000001 01e80ab8 00000004 00000020 01e7fb48 01e7fb14 01e7fb34 > [ 1639.884542] 9fe0: 00000000 bef4562c 0001ee5d 0000a8cc 60000010 00000001 00000000 00000000 > [ 1639.892709] [<c058acec>] (patch_text) from [<c058bcc4>] (arm_kprobe+0x28/0x38) > [ 1639.900862] [<c058bcc4>] (arm_kprobe) from [<c058bd5c>] (enable_kprobe+0x88/0x94) > [ 1639.907983] [<c058bd5c>] (enable_kprobe) from [<c018b724>] (__ftrace_event_enable_disable+0x13c/0x200) > [ 1639.915537] [<c018b724>] (__ftrace_event_enable_disable) from [<c018bdc4>] (event_enable_write+0x78/0xd4) > [ 1639.924741] [<c018bdc4>] (event_enable_write) from [<c01e1b1c>] (vfs_write+0xac/0x188) > [ 1639.934372] [<c01e1b1c>] (vfs_write) from [<c01e2108>] (SyS_write+0x40/0x94) > [ 1639.942187] [<c01e2108>] (SyS_write) from [<c0106ce0>] (ret_fast_syscall+0x0/0x30) > [ 1639.949386] Code: e4831004 e1a01003 eaee298d e1a03000 (e4831004) > [ 1639.956766] ---[ end trace b548269e2c7a3190 ]--- > > > We had some functions that allowed the text to be temporarily made writable but something > uniform for kexec would be useful as well (our kexec solution has been 'turn it off') > > >> Cheers >> ---Dave >> > > Thanks, > Laura > > -- > Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, > hosted by The Linux Foundation
On Sun, 23 Mar 2014, Kees Cook wrote: > On Sun, Mar 23, 2014 at 12:47 PM, Laura Abbott <lauraa@codeaurora.org> wrote: > > On 2/17/2014 4:34 AM, Dave Martin wrote: > >> On Fri, Feb 14, 2014 at 11:11:07AM -0800, Kees Cook wrote: > >>> On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin@arm.com> wrote: > >>>> On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: > >>>>> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though > >>>>> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which > >>>>> sets rodata read-only (but executable), where as this option additionally > >>>>> splits rodata from the kernel text (resulting in potentially more memory > >>>>> lost to padding) and sets it non-executable as well. The end result is > >>>>> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be > >>>>> marked purely read-only. > >>>> > >>>> This triggers an Oops in kexec, because we have a block of code in .text > >>>> which is a template for generating baremetal code to relocate the new > >>>> kernel, and some literal words are written into it before copying. > >>> > >>> You're writing into the text area? I would imagine that > >>> CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the > >>> right place to be building code -- shouldn't the module area be used > >>> for that? > >>> > >>>> Possibly this should be in .rodata, not .text. > >>> > >>> Well, rodata should be neither writable nor executable. > >> > >> We're not writing into code exactly. > >> > >> This code is never executed in-place in vmlinux. It gets copied, and > >> only copies are ever executed. > >> > >> Some pointers and offsets get poked into the code to configure it. > >> > >> I think it would be better simply to put the code in .rodata, and > >> poke paramaters into the copy, not the original -- but that's a bit > >> more awkward to code up, since the values can't be poked simply by > >> writing global variables. > >> > >>> > >>>> There may be a few other instances of this kind of thing. > >>> > >>> This config will certainly find them! :) But, that's why it's behind a config. > >> > >> I haven't tested exhaustively, but it think this is sufficient for a > >> Tested-by. The patch does seem to be doing what it is intended to > >> do, and doesn't seem to be triggering false positives all over the > >> place. > >> > >>> > >>>> Are you aware of similar situations on other arches? > >>> > >>> I think there were some problems a long time ago on x86 for rodata too. > >> > >> It would be good to get this kexec case fixed -- I'll try to hack up > >> a separate patch. > >> > > > > FWIW, we've hit issues not just with kexec but kprobes as well. The same > > problems exist with this series: > > For this stage, how about I make this "depends on KEXEC=n && > KPROBES=n"? Then as these areas get fixed, we can drop those > requirements. Do they really need "fixing"? The goal here is to increase security by preventing kernel code to be modified. And now it would require hole punching in order to support kprobes. If security is important enough for this option to be attractive to you, then wouldn't you want to keep kprobes firmly turned off as well? Nicolas
On Sun, Mar 23, 2014 at 4:37 PM, Nicolas Pitre <nicolas.pitre@linaro.org> wrote: > On Sun, 23 Mar 2014, Kees Cook wrote: > >> On Sun, Mar 23, 2014 at 12:47 PM, Laura Abbott <lauraa@codeaurora.org> wrote: >> > On 2/17/2014 4:34 AM, Dave Martin wrote: >> >> On Fri, Feb 14, 2014 at 11:11:07AM -0800, Kees Cook wrote: >> >>> On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin@arm.com> wrote: >> >>>> On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: >> >>>>> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though >> >>>>> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which >> >>>>> sets rodata read-only (but executable), where as this option additionally >> >>>>> splits rodata from the kernel text (resulting in potentially more memory >> >>>>> lost to padding) and sets it non-executable as well. The end result is >> >>>>> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be >> >>>>> marked purely read-only. >> >>>> >> >>>> This triggers an Oops in kexec, because we have a block of code in .text >> >>>> which is a template for generating baremetal code to relocate the new >> >>>> kernel, and some literal words are written into it before copying. >> >>> >> >>> You're writing into the text area? I would imagine that >> >>> CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the >> >>> right place to be building code -- shouldn't the module area be used >> >>> for that? >> >>> >> >>>> Possibly this should be in .rodata, not .text. >> >>> >> >>> Well, rodata should be neither writable nor executable. >> >> >> >> We're not writing into code exactly. >> >> >> >> This code is never executed in-place in vmlinux. It gets copied, and >> >> only copies are ever executed. >> >> >> >> Some pointers and offsets get poked into the code to configure it. >> >> >> >> I think it would be better simply to put the code in .rodata, and >> >> poke paramaters into the copy, not the original -- but that's a bit >> >> more awkward to code up, since the values can't be poked simply by >> >> writing global variables. >> >> >> >>> >> >>>> There may be a few other instances of this kind of thing. >> >>> >> >>> This config will certainly find them! :) But, that's why it's behind a config. >> >> >> >> I haven't tested exhaustively, but it think this is sufficient for a >> >> Tested-by. The patch does seem to be doing what it is intended to >> >> do, and doesn't seem to be triggering false positives all over the >> >> place. >> >> >> >>> >> >>>> Are you aware of similar situations on other arches? >> >>> >> >>> I think there were some problems a long time ago on x86 for rodata too. >> >> >> >> It would be good to get this kexec case fixed -- I'll try to hack up >> >> a separate patch. >> >> >> > >> > FWIW, we've hit issues not just with kexec but kprobes as well. The same >> > problems exist with this series: >> >> For this stage, how about I make this "depends on KEXEC=n && >> KPROBES=n"? Then as these areas get fixed, we can drop those >> requirements. > > Do they really need "fixing"? > > The goal here is to increase security by preventing kernel code to be > modified. And now it would require hole punching in order to support > kprobes. > > If security is important enough for this option to be attractive to you, > then wouldn't you want to keep kprobes firmly turned off as well? I couldn't agree more. :) I don't build with these options, but if someone wants both of these, we'll have to deal with that. Until then, we should keep it disabled with the negative "depends on". -Kees
On Sun, 2014-03-23 at 16:21 -0600, Kees Cook wrote: > On Sun, Mar 23, 2014 at 12:47 PM, Laura Abbott <lauraa@codeaurora.org> wrote: > > On 2/17/2014 4:34 AM, Dave Martin wrote: > >> On Fri, Feb 14, 2014 at 11:11:07AM -0800, Kees Cook wrote: > >>> On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin@arm.com> wrote: > >>>> On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: > >>>>> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though > >>>>> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which > >>>>> sets rodata read-only (but executable), where as this option additionally > >>>>> splits rodata from the kernel text (resulting in potentially more memory > >>>>> lost to padding) and sets it non-executable as well. The end result is > >>>>> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be > >>>>> marked purely read-only. > >>>> > >>>> This triggers an Oops in kexec, because we have a block of code in .text > >>>> which is a template for generating baremetal code to relocate the new > >>>> kernel, and some literal words are written into it before copying. > >>> > >>> You're writing into the text area? I would imagine that > >>> CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the > >>> right place to be building code -- shouldn't the module area be used > >>> for that? > >>> > >>>> Possibly this should be in .rodata, not .text. > >>> > >>> Well, rodata should be neither writable nor executable. > >> > >> We're not writing into code exactly. > >> > >> This code is never executed in-place in vmlinux. It gets copied, and > >> only copies are ever executed. > >> > >> Some pointers and offsets get poked into the code to configure it. > >> > >> I think it would be better simply to put the code in .rodata, and > >> poke paramaters into the copy, not the original -- but that's a bit > >> more awkward to code up, since the values can't be poked simply by > >> writing global variables. > >> > >>> > >>>> There may be a few other instances of this kind of thing. > >>> > >>> This config will certainly find them! :) But, that's why it's behind a config. > >> > >> I haven't tested exhaustively, but it think this is sufficient for a > >> Tested-by. The patch does seem to be doing what it is intended to > >> do, and doesn't seem to be triggering false positives all over the > >> place. > >> > >>> > >>>> Are you aware of similar situations on other arches? > >>> > >>> I think there were some problems a long time ago on x86 for rodata too. > >> > >> It would be good to get this kexec case fixed -- I'll try to hack up > >> a separate patch. > >> > > > > FWIW, we've hit issues not just with kexec but kprobes as well. The same > > problems exist with this series: > > For this stage, how about I make this "depends on KEXEC=n && > KPROBES=n"? There's also ftrace (CONFIG_DYNAMIC_FTRACE I believe) which modifies kernel code with a call to probe_kernel_write(), which GDB uses as well. And grepping for the patch_text() function also shows __arch_jump_label_transform() modifies kernel code. Not sure how and when that gets used.
On Sun, Mar 23, 2014 at 06:47:36PM +0000, Laura Abbott wrote: > On 2/17/2014 4:34 AM, Dave Martin wrote: > > On Fri, Feb 14, 2014 at 11:11:07AM -0800, Kees Cook wrote: > >> On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin@arm.com> wrote: > >>> On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote: > >>>> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though > >>>> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which > >>>> sets rodata read-only (but executable), where as this option additionally > >>>> splits rodata from the kernel text (resulting in potentially more memory > >>>> lost to padding) and sets it non-executable as well. The end result is > >>>> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be > >>>> marked purely read-only. > >>> > >>> This triggers an Oops in kexec, because we have a block of code in .text > >>> which is a template for generating baremetal code to relocate the new > >>> kernel, and some literal words are written into it before copying. > >> > >> You're writing into the text area? I would imagine that > >> CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the > >> right place to be building code -- shouldn't the module area be used > >> for that? > >> [...] > FWIW, we've hit issues not just with kexec but kprobes as well. The same > problems exist with this series: [...] > We had some functions that allowed the text to be temporarily made writable but something > uniform for kexec would be useful as well (our kexec solution has been 'turn it off') kexec doesn't rely on poking the kernel text: the fact that it does this is just a side-effect of the way it is currently implemented. I would like to fix it -- it's currently on my todo list, but I consider it non-urgent. depending on !KEXEC seems reasonable for now. People building a hardened system may choose do disable kexec for other reasons, but that's a separate issue entirely. kprobes is a different matter: getting it to work with strict permissions is likely to be complex and costly. kprobes is already unavoidably complex and costly, so that is probably better viewed as a kprobes problem to be solved later, rather than a strict permissions problem. Cheers ---Dave
2014-03-24 11:47 GMT+01:00 Jon Medhurst (Tixy) <tixy@linaro.org>: > On Sun, 2014-03-23 at 16:21 -0600, Kees Cook wrote: >> For this stage, how about I make this "depends on KEXEC=n && >> KPROBES=n"? > > There's also ftrace (CONFIG_DYNAMIC_FTRACE I believe) which modifies > kernel code with a call to probe_kernel_write(), which GDB uses as well. x86 handles this by making all kernel text R/W around the place where ftrace does the modifications. This is called under stop_machine(). See 16239630974516a887 ("ftrace, x86: make kernel text writable only for conversions"). > And grepping for the patch_text() function also shows > __arch_jump_label_transform() modifies kernel code. Not sure how and > when that gets used. It gets used when "Optimize very unlikely/likely branches" (CONFIG_JUMP_LABEL) is enabled. These "very unlikely/likely branches" are used, among other things, for controlling tracepoints and some scheduler/networking features. x86 handles jump labels (and kprobes) by mapping the page being modified read-write around the modification. See text_poke(). See also this ARM patch from Kyle Martin which I don't think has been merged: http://marc.info/?l=linux-arm-kernel&m=139291862727863&w=2
On Mon, Mar 24, 2014 at 3:47 AM, Jon Medhurst (Tixy) <tixy@linaro.org> wrote: > On Sun, 2014-03-23 at 16:21 -0600, Kees Cook wrote: >> For this stage, how about I make this "depends on KEXEC=n && >> KPROBES=n"? > > There's also ftrace (CONFIG_DYNAMIC_FTRACE I believe) which modifies > kernel code with a call to probe_kernel_write(), which GDB uses as well. > > And grepping for the patch_text() function also shows > __arch_jump_label_transform() modifies kernel code. Not sure how and > when that gets used. Right, so, I'm trying to fix ftrace now, and I've hit a wall. It is as if changes to the kernel text PMD aren't being noticed after the kernel is running. Does anyone know why this might be happening? Code and details here: https://lkml.org/lkml/2014/4/1/674 -Kees
On 4/1/2014 3:34 PM, Kees Cook wrote: > On Mon, Mar 24, 2014 at 3:47 AM, Jon Medhurst (Tixy) <tixy@linaro.org> wrote: >> On Sun, 2014-03-23 at 16:21 -0600, Kees Cook wrote: >>> For this stage, how about I make this "depends on KEXEC=n && >>> KPROBES=n"? >> >> There's also ftrace (CONFIG_DYNAMIC_FTRACE I believe) which modifies >> kernel code with a call to probe_kernel_write(), which GDB uses as well. >> >> And grepping for the patch_text() function also shows >> __arch_jump_label_transform() modifies kernel code. Not sure how and >> when that gets used. > > Right, so, I'm trying to fix ftrace now, and I've hit a wall. It is as > if changes to the kernel text PMD aren't being noticed after the > kernel is running. Does anyone know why this might be happening? > > Code and details here: > https://lkml.org/lkml/2014/4/1/674 > > -Kees > We had a flush_tlb_kernel_page after the pmd_flush in our out of tree code which makes the text writeable in __patch_text. Laura
On Tue, Apr 1, 2014 at 3:54 PM, Laura Abbott <lauraa@codeaurora.org> wrote: > On 4/1/2014 3:34 PM, Kees Cook wrote: >> On Mon, Mar 24, 2014 at 3:47 AM, Jon Medhurst (Tixy) <tixy@linaro.org> wrote: >>> On Sun, 2014-03-23 at 16:21 -0600, Kees Cook wrote: >>>> For this stage, how about I make this "depends on KEXEC=n && >>>> KPROBES=n"? >>> >>> There's also ftrace (CONFIG_DYNAMIC_FTRACE I believe) which modifies >>> kernel code with a call to probe_kernel_write(), which GDB uses as well. >>> >>> And grepping for the patch_text() function also shows >>> __arch_jump_label_transform() modifies kernel code. Not sure how and >>> when that gets used. >> >> Right, so, I'm trying to fix ftrace now, and I've hit a wall. It is as >> if changes to the kernel text PMD aren't being noticed after the >> kernel is running. Does anyone know why this might be happening? >> >> Code and details here: >> https://lkml.org/lkml/2014/4/1/674 >> >> -Kees >> > > We had a flush_tlb_kernel_page after the pmd_flush in our out of tree code > which makes the text writeable in __patch_text. I tried flush_tlb_kernel_range(), which I'd expect to do the same thing. I can try with _page() too. -Kees
diff --git a/arch/arm/include/asm/cacheflush.h b/arch/arm/include/asm/cacheflush.h index e9a49fe0284e..2b058fc7a188 100644 --- a/arch/arm/include/asm/cacheflush.h +++ b/arch/arm/include/asm/cacheflush.h @@ -486,4 +486,9 @@ int set_memory_rw(unsigned long addr, int numpages); int set_memory_x(unsigned long addr, int numpages); int set_memory_nx(unsigned long addr, int numpages); +#ifdef CONFIG_DEBUG_RODATA +/* This has already happened during free_initmem. */ +static inline void mark_rodata_ro(void) { } +#endif + #endif diff --git a/arch/arm/kernel/vmlinux.lds.S b/arch/arm/kernel/vmlinux.lds.S index 08fa667ef2f1..ec79e7268e09 100644 --- a/arch/arm/kernel/vmlinux.lds.S +++ b/arch/arm/kernel/vmlinux.lds.S @@ -120,6 +120,9 @@ SECTIONS ARM_CPU_KEEP(PROC_INFO) } +#ifdef CONFIG_DEBUG_RODATA + . = ALIGN(1<<SECTION_SHIFT); +#endif RO_DATA(PAGE_SIZE) . = ALIGN(4); diff --git a/arch/arm/mm/Kconfig b/arch/arm/mm/Kconfig index 999eb505faee..7c8bbe7e2769 100644 --- a/arch/arm/mm/Kconfig +++ b/arch/arm/mm/Kconfig @@ -968,3 +968,15 @@ config ARM_KERNMEM_PERMS region is padded to section-size (1MiB) boundaries (because their permissions are different and splitting the 1M pages into 4K ones causes TLB performance problems), wasting memory. + +config DEBUG_RODATA + bool "Split rodata from text and set it read-only/non-executable" + depends on ARM_KERNMEM_PERMS + default y + help + If this is set, rodata will be split from kernel text and made + non-executable. (This option already depends on the option + CONFIG_STRICT_KERNMEM_PERMS which makes rodata read-only, though + still executable.) This creates another section-size padded + region, so it can waste more memory space while gaining a pure + read-only rodata region. diff --git a/arch/arm/mm/init.c b/arch/arm/mm/init.c index f0b1df53f436..5b1b049501b9 100644 --- a/arch/arm/mm/init.c +++ b/arch/arm/mm/init.c @@ -656,6 +656,14 @@ struct section_perm __initdata section_perms[] = { .prot = PMD_SECT_APX | PMD_SECT_AP_WRITE, #endif }, +#ifdef CONFIG_DEBUG_RODATA + /* Make rodata RO (set NX). */ + { + .start = (unsigned long)__start_rodata, + .end = (unsigned long)__init_begin, + .prot = PMD_SECT_XN, + } +#endif }; static inline void section_update(unsigned long addr, pmdval_t prot)
Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though the behavior is different: it depends on STRICT_KERNMEM_PERMS, which sets rodata read-only (but executable), where as this option additionally splits rodata from the kernel text (resulting in potentially more memory lost to padding) and sets it non-executable as well. The end result is that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be marked purely read-only. Signed-off-by: Kees Cook <keescook@chromium.org> --- arch/arm/include/asm/cacheflush.h | 5 +++++ arch/arm/kernel/vmlinux.lds.S | 3 +++ arch/arm/mm/Kconfig | 12 ++++++++++++ arch/arm/mm/init.c | 8 ++++++++ 4 files changed, 28 insertions(+)