| Message ID | cover.1635383031.git.pawan.kumar.gupta@linux.intel.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Unprivileged BPF default | expand |
On 28.10.2021 14:49, Mark Rutland wrote: >On Wed, Oct 27, 2021 at 06:33:22PM -0700, Pawan Gupta wrote: >> Borrow CONFIG_CPU_SPECTRE from ARM to be available for all >> architectures. This will help in configuration of features that depend >> on CPU being affected by spectre class of vulnerabilities. >> >> Signed-off-by: Pawan Gupta <pawan.kumar.gupta@linux.intel.com> > >Given that spectre isn't one specific issue, biut rather a blanket term >for a bunch of things that can have variable overlap, I don't think this >makes much sense unless we're going to add finer-grained options for all >the variants, and IMO it'd make more sene for the architectures to >directly select the things that'd otherwise be dependent on this. Isn't ARM already using CPU_SPECTRE for selecting things: config HARDEN_BRANCH_PREDICTOR bool "Harden the branch predictor against aliasing attacks" if EXPERT depends on CPU_SPECTRE This was the whole motivation for doing the same for x86. Adding a condition for all architectures is also okay, but its going to a little messier: config BPF_UNPRIV_DEFAULT_OFF default y if X86 || ARM || ... This approach would make sense if architectures wants to explicitly select the defaults irrespective of architecture being affected by spectre. If that's the case I will change the BPF_UNPRIV_DEFAULT_OFF default to depend on architecture. I hope BPF maintainer, Daniel is okay with it? Pawan Added BPF maintainers and bpf@vger to cc.
On Thu, Oct 28, 2021 at 12:36:58PM -0700, Pawan Gupta wrote: > On 28.10.2021 14:49, Mark Rutland wrote: > > On Wed, Oct 27, 2021 at 06:33:22PM -0700, Pawan Gupta wrote: > > > Borrow CONFIG_CPU_SPECTRE from ARM to be available for all > > > architectures. This will help in configuration of features that depend > > > on CPU being affected by spectre class of vulnerabilities. > > > > > > Signed-off-by: Pawan Gupta <pawan.kumar.gupta@linux.intel.com> > > > > Given that spectre isn't one specific issue, biut rather a blanket term > > for a bunch of things that can have variable overlap, I don't think this > > makes much sense unless we're going to add finer-grained options for all > > the variants, and IMO it'd make more sene for the architectures to > > directly select the things that'd otherwise be dependent on this. > > Isn't ARM already using CPU_SPECTRE for selecting things: > > config HARDEN_BRANCH_PREDICTOR > bool "Harden the branch predictor against aliasing attacks" if EXPERT > depends on CPU_SPECTRE It's true that arch/arm does, but that's not true for other architectures, e.g. powerpc or arm64, and and as above I don't think it makes sense to make this generic in its current form because "spectre" is a somewhat vague generic term. > This was the whole motivation for doing the same for x86. > > Adding a condition for all architectures is also okay, but its going to > a little messier: > > config BPF_UNPRIV_DEFAULT_OFF > default y if X86 || ARM || ... > > This approach would make sense if architectures wants to explicitly > select the defaults irrespective of architecture being affected by > spectre. If we're going to change the default for some architectures, I think it'd make much more sense to just do that for all, without any arch-specific conditionality, i.e. config BPF_UNPRIV_DEFAULT_OFF default y ... so that the behaviour is consistent across all architectures, and we don't have to play a whack-a-mole game as/when we realise architectures are affected by some variant of an issue relating to speculation. Thanks, Mark.
On Thu, Oct 28, 2021 at 12:36:58PM -0700, Pawan Gupta wrote: > Isn't ARM already using CPU_SPECTRE for selecting things: > > config HARDEN_BRANCH_PREDICTOR > bool "Harden the branch predictor against aliasing attacks" if EXPERT > depends on CPU_SPECTRE > > This was the whole motivation for doing the same for x86. > > Adding a condition for all architectures is also okay, but its going to > a little messier: > > config BPF_UNPRIV_DEFAULT_OFF > default y if X86 || ARM || ... It doesn't have to be (but sadly we end up repeating "DEFAULT"): config BPF_UNPRIV_DEFAULT_OFF_DEFAULT bool config BPF_UNPRIV_DEFAULT_OFF bool "Disable unprivileged BPF by default" default BPF_UNPRIV_DEFAULT_OFF_DEFAULT Then architectures can select BPF_UNPRIV_DEFAULT_OFF_DEFAULT if they wish this to be defaulted to "yes". However, please note that this has limited use given that the BPF_UNPRIV_DEFAULT_OFF option has been around for a while now. Any existing configuration that mentions this symbol will override any default specified in the Kconfig files if the option is user-visible. So, IMHO, defaults need to be set correctly from the point in time that the option is introduced.
On 10/29/21 11:22 AM, Mark Rutland wrote: > On Thu, Oct 28, 2021 at 12:36:58PM -0700, Pawan Gupta wrote: >> On 28.10.2021 14:49, Mark Rutland wrote: >>> On Wed, Oct 27, 2021 at 06:33:22PM -0700, Pawan Gupta wrote: >>>> Borrow CONFIG_CPU_SPECTRE from ARM to be available for all >>>> architectures. This will help in configuration of features that depend >>>> on CPU being affected by spectre class of vulnerabilities. >>>> >>>> Signed-off-by: Pawan Gupta <pawan.kumar.gupta@linux.intel.com> >>> >>> Given that spectre isn't one specific issue, biut rather a blanket term >>> for a bunch of things that can have variable overlap, I don't think this >>> makes much sense unless we're going to add finer-grained options for all >>> the variants, and IMO it'd make more sene for the architectures to >>> directly select the things that'd otherwise be dependent on this. >> >> Isn't ARM already using CPU_SPECTRE for selecting things: >> >> config HARDEN_BRANCH_PREDICTOR >> bool "Harden the branch predictor against aliasing attacks" if EXPERT >> depends on CPU_SPECTRE > > It's true that arch/arm does, but that's not true for other > architectures, e.g. powerpc or arm64, and and as above I don't think it > makes sense to make this generic in its current form because "spectre" > is a somewhat vague generic term. > >> This was the whole motivation for doing the same for x86. >> >> Adding a condition for all architectures is also okay, but its going to >> a little messier: >> >> config BPF_UNPRIV_DEFAULT_OFF >> default y if X86 || ARM || ... >> >> This approach would make sense if architectures wants to explicitly >> select the defaults irrespective of architecture being affected by >> spectre. > > If we're going to change the default for some architectures, I think > it'd make much more sense to just do that for all, without any > arch-specific conditionality, i.e. > > config BPF_UNPRIV_DEFAULT_OFF > default y Lets just go with 'default y'. The main rationale for this change was motivated by spectre, so would have been good to indicate this also with an explicit dependency for broken HW, not just help description. Pretty much agreeing with Greg here [0]. Eventually, we might need some arch generic way to determine arch- common spectre type bugs, so that for unaffected HW we don't need to apply some of them from verifier, but that's still tbd. Thanks, Daniel [0] https://lore.kernel.org/bpf/YXrTev6WMXry9pFI@kroah.com/
On 29.10.2021 18:05, Daniel Borkmann wrote: >On 10/29/21 11:22 AM, Mark Rutland wrote: >>On Thu, Oct 28, 2021 at 12:36:58PM -0700, Pawan Gupta wrote: >>>On 28.10.2021 14:49, Mark Rutland wrote: >>>>On Wed, Oct 27, 2021 at 06:33:22PM -0700, Pawan Gupta wrote: >>>>>Borrow CONFIG_CPU_SPECTRE from ARM to be available for all >>>>>architectures. This will help in configuration of features that depend >>>>>on CPU being affected by spectre class of vulnerabilities. >>>>> >>>>>Signed-off-by: Pawan Gupta <pawan.kumar.gupta@linux.intel.com> >>>> >>>>Given that spectre isn't one specific issue, biut rather a blanket term >>>>for a bunch of things that can have variable overlap, I don't think this >>>>makes much sense unless we're going to add finer-grained options for all >>>>the variants, and IMO it'd make more sene for the architectures to >>>>directly select the things that'd otherwise be dependent on this. >>> >>>Isn't ARM already using CPU_SPECTRE for selecting things: >>> >>> config HARDEN_BRANCH_PREDICTOR >>> bool "Harden the branch predictor against aliasing attacks" if EXPERT >>> depends on CPU_SPECTRE >> >>It's true that arch/arm does, but that's not true for other >>architectures, e.g. powerpc or arm64, and and as above I don't think it >>makes sense to make this generic in its current form because "spectre" >>is a somewhat vague generic term. >> >>>This was the whole motivation for doing the same for x86. >>> >>>Adding a condition for all architectures is also okay, but its going to >>>a little messier: >>> >>> config BPF_UNPRIV_DEFAULT_OFF >>> default y if X86 || ARM || ... >>> >>>This approach would make sense if architectures wants to explicitly >>>select the defaults irrespective of architecture being affected by >>>spectre. >> >>If we're going to change the default for some architectures, I think >>it'd make much more sense to just do that for all, without any >>arch-specific conditionality, i.e. >> >> config BPF_UNPRIV_DEFAULT_OFF >> default y > >Lets just go with 'default y'. The main rationale for this change was motivated >by spectre, so would have been good to indicate this also with an explicit >dependency for broken HW, not just help description. Pretty much agreeing with >Greg here [0]. Eventually, we might need some arch generic way to determine arch- >common spectre type bugs, so that for unaffected HW we don't need to apply some >of them from verifier, but that's still tbd. I will send a patch soon with 'default y'. Thanks, Pawan
On 29.10.2021 11:08, Russell King (Oracle) wrote: >On Thu, Oct 28, 2021 at 12:36:58PM -0700, Pawan Gupta wrote: >> Isn't ARM already using CPU_SPECTRE for selecting things: >> >> config HARDEN_BRANCH_PREDICTOR >> bool "Harden the branch predictor against aliasing attacks" if EXPERT >> depends on CPU_SPECTRE >> >> This was the whole motivation for doing the same for x86. >> >> Adding a condition for all architectures is also okay, but its going to >> a little messier: >> >> config BPF_UNPRIV_DEFAULT_OFF >> default y if X86 || ARM || ... > >It doesn't have to be (but sadly we end up repeating "DEFAULT"): > >config BPF_UNPRIV_DEFAULT_OFF_DEFAULT > bool > >config BPF_UNPRIV_DEFAULT_OFF > bool "Disable unprivileged BPF by default" > default BPF_UNPRIV_DEFAULT_OFF_DEFAULT > >Then architectures can select BPF_UNPRIV_DEFAULT_OFF_DEFAULT if they >wish this to be defaulted to "yes". Looks like we are settling on unconditional 'default y' for now [1]. I have sent a v3 with 'default y' [2]. >However, please note that this has limited use given that the >BPF_UNPRIV_DEFAULT_OFF option has been around for a while now. Any >existing configuration that mentions this symbol will override any >default specified in the Kconfig files if the option is user-visible. Yes, existing configurations will have to toggle this manually. However, many distros already have BPF_UNPRIV_DEFAULT_OFF=y in their configuration. >So, IMHO, defaults need to be set correctly from the point in time >that the option is introduced. Agree. [1] https://lore.kernel.org/lkml/6130e55f-4d84-5ada-4e86-5b678e3eaf5e@iogearbox.net/ [2] https://lore.kernel.org/lkml/0ace9ce3f97656d5f62d11093ad7ee81190c3c25.1635535215.git.pawan.kumar.gupta@linux.intel.com/