Message ID | 20180130161433.6910-1-julien.grall@arm.com |
---|---|
Headers | show |
Series | xen/arm: Inject an exception to the guest rather than crashing it | expand |
On 30/01/18 16:14, Julien Grall wrote: > Hi all, > > This small series replaces all call to domain_crash_synchronous by injecting > an exception to the guest. > > This will result to a nicer trace from the guest (no need to manually walk > the stack) and give a chance to the guest to give a bit more information on > what it was doing. > > Cheers, > > Julien Grall (3): > xen/arm: io: Distinguish unhandled IO from aborted one > xen/arm: Don't crash domain on bad MMIO emulation > xen/arm: Don't crash the domain on invalid HVC immediate Thanks. I don't feel qualified to review these, but some notes. Patch 1. s/avodi/avoid/ in the commit message Patches 2 and 3. You probably want to convert the printks to gdprintk()s, otherwise guests can choke up the ratelimited log. Doing so will also mean that the vcpu will be identified consistently, which it isn't currently. ~Andrew
On 30/01/18 16:38, Andrew Cooper wrote: > On 30/01/18 16:14, Julien Grall wrote: >> Hi all, >> >> This small series replaces all call to domain_crash_synchronous by injecting >> an exception to the guest. >> >> This will result to a nicer trace from the guest (no need to manually walk >> the stack) and give a chance to the guest to give a bit more information on >> what it was doing. >> >> Cheers, >> >> Julien Grall (3): >> xen/arm: io: Distinguish unhandled IO from aborted one >> xen/arm: Don't crash domain on bad MMIO emulation >> xen/arm: Don't crash the domain on invalid HVC immediate > > Thanks. > > I don't feel qualified to review these, but some notes. > > Patch 1. s/avodi/avoid/ in the commit message > > Patches 2 and 3. You probably want to convert the printks to > gdprintk()s, otherwise guests can choke up the ratelimited log. Doing > so will also mean that the vcpu will be identified consistently, which > it isn't currently. We didn't use g*printk because it would be more confusing to print the current vCPU in some cases (e.g when accessing the re-distributor of another vCPU) or does not matter (e.g for ITS). The problem with the debug version is those information are actually quite useful in non-debug build. We found quite a few issues thanks to them. I think it would make more sense for Xen to provide per-guest ratelimited than hiding those messages in non-debug build. > > ~Andrew >
On 30/01/18 17:00, Julien Grall wrote: > > > On 30/01/18 16:38, Andrew Cooper wrote: >> On 30/01/18 16:14, Julien Grall wrote: >>> Hi all, >>> >>> This small series replaces all call to domain_crash_synchronous by >>> injecting >>> an exception to the guest. >>> >>> This will result to a nicer trace from the guest (no need to >>> manually walk >>> the stack) and give a chance to the guest to give a bit more >>> information on >>> what it was doing. >>> >>> Cheers, >>> >>> Julien Grall (3): >>> xen/arm: io: Distinguish unhandled IO from aborted one >>> xen/arm: Don't crash domain on bad MMIO emulation >>> xen/arm: Don't crash the domain on invalid HVC immediate >> >> Thanks. >> >> I don't feel qualified to review these, but some notes. >> >> Patch 1. s/avodi/avoid/ in the commit message >> >> Patches 2 and 3. You probably want to convert the printks to >> gdprintk()s, otherwise guests can choke up the ratelimited log. Doing >> so will also mean that the vcpu will be identified consistently, which >> it isn't currently. > We didn't use g*printk because it would be more confusing to print the > current vCPU in some cases (e.g when accessing the re-distributor of > another vCPU) or does not matter (e.g for ITS). In the former case, you'd want to print both current, and the target vcpu. The latter still matters what current is if something goes wrong. We have plenty of similar cases in x86, but at the point you are printing an diagnostic message, ignoring current is almost always the wrong think to do. > > The problem with the debug version is those information are actually > quite useful in non-debug build. We found quite a few issues thanks to > them. > > I think it would make more sense for Xen to provide per-guest > ratelimited than hiding those messages in non-debug build. Per guest is quite a lot more complicated than global, and would still require a global limit to prevent a concerted attack from multiple guests to avoid DoSing the system. Debug vs unilateral is your prerogative as a maintainer, but as you've said yourself, the are used for debugging purposes, which proves my point. ~Andrew
Hi, On 30/01/18 18:29, Andrew Cooper wrote: > On 30/01/18 17:00, Julien Grall wrote: >> >> >> On 30/01/18 16:38, Andrew Cooper wrote: >>> On 30/01/18 16:14, Julien Grall wrote: >>>> Hi all, >>>> >>>> This small series replaces all call to domain_crash_synchronous by >>>> injecting >>>> an exception to the guest. >>>> >>>> This will result to a nicer trace from the guest (no need to >>>> manually walk >>>> the stack) and give a chance to the guest to give a bit more >>>> information on >>>> what it was doing. >>>> >>>> Cheers, >>>> >>>> Julien Grall (3): >>>> xen/arm: io: Distinguish unhandled IO from aborted one >>>> xen/arm: Don't crash domain on bad MMIO emulation >>>> xen/arm: Don't crash the domain on invalid HVC immediate >>> >>> Thanks. >>> >>> I don't feel qualified to review these, but some notes. >>> >>> Patch 1. s/avodi/avoid/ in the commit message >>> >>> Patches 2 and 3. You probably want to convert the printks to >>> gdprintk()s, otherwise guests can choke up the ratelimited log. Doing >>> so will also mean that the vcpu will be identified consistently, which >>> it isn't currently. >> We didn't use g*printk because it would be more confusing to print the >> current vCPU in some cases (e.g when accessing the re-distributor of >> another vCPU) or does not matter (e.g for ITS). > > In the former case, you'd want to print both current, and the target > vcpu. The latter still matters what current is if something goes wrong. > > We have plenty of similar cases in x86, but at the point you are > printing an diagnostic message, ignoring current is almost always the > wrong think to do. I will look at it on another series. > >> >> The problem with the debug version is those information are actually >> quite useful in non-debug build. We found quite a few issues thanks to >> them. >> >> I think it would make more sense for Xen to provide per-guest >> ratelimited than hiding those messages in non-debug build. > > Per guest is quite a lot more complicated than global, and would still > require a global limit to prevent a concerted attack from multiple > guests to avoid DoSing the system. > > Debug vs unilateral is your prerogative as a maintainer, but as you've > said yourself, the are used for debugging purposes, which proves my point. So on x86, you always request the user to reproduce it with debug build enable? Stefano, what's your opinion here? Cheers,
On 30/01/18 18:46, Julien Grall wrote: > Hi, > > On 30/01/18 18:29, Andrew Cooper wrote: >> On 30/01/18 17:00, Julien Grall wrote: >>> >>> >>> On 30/01/18 16:38, Andrew Cooper wrote: >>>> On 30/01/18 16:14, Julien Grall wrote: >>>>> Hi all, >>>>> >>>>> This small series replaces all call to domain_crash_synchronous by >>>>> injecting >>>>> an exception to the guest. >>>>> >>>>> This will result to a nicer trace from the guest (no need to >>>>> manually walk >>>>> the stack) and give a chance to the guest to give a bit more >>>>> information on >>>>> what it was doing. >>>>> >>>>> Cheers, >>>>> >>>>> Julien Grall (3): >>>>> xen/arm: io: Distinguish unhandled IO from aborted one >>>>> xen/arm: Don't crash domain on bad MMIO emulation >>>>> xen/arm: Don't crash the domain on invalid HVC immediate >>>> >>>> Thanks. >>>> >>>> I don't feel qualified to review these, but some notes. >>>> >>>> Patch 1. s/avodi/avoid/ in the commit message >>>> >>>> Patches 2 and 3. You probably want to convert the printks to >>>> gdprintk()s, otherwise guests can choke up the ratelimited log. Doing >>>> so will also mean that the vcpu will be identified consistently, which >>>> it isn't currently. >>> We didn't use g*printk because it would be more confusing to print the >>> current vCPU in some cases (e.g when accessing the re-distributor of >>> another vCPU) or does not matter (e.g for ITS). >> >> In the former case, you'd want to print both current, and the target >> vcpu. The latter still matters what current is if something goes wrong. >> >> We have plenty of similar cases in x86, but at the point you are >> printing an diagnostic message, ignoring current is almost always the >> wrong think to do. > > I will look at it on another series. Fair enough. > >> >>> >>> The problem with the debug version is those information are actually >>> quite useful in non-debug build. We found quite a few issues thanks to >>> them. >>> >>> I think it would make more sense for Xen to provide per-guest >>> ratelimited than hiding those messages in non-debug build. >> >> Per guest is quite a lot more complicated than global, and would still >> require a global limit to prevent a concerted attack from multiple >> guests to avoid DoSing the system. >> >> Debug vs unilateral is your prerogative as a maintainer, but as you've >> said yourself, the are used for debugging purposes, which proves my >> point. > > So on x86, you always request the user to reproduce it with debug > build enable? That is very specific to what goes wrong, but no, we generally don't have release-build messages for "the guest screwed up and we hit it expected way for doing so". We do (well - are trying to, which is how I found this domain_crash_sync issue in the first place) get consistent at printing useful release-build messages for abnormal termination of the guest. In terms of making it easier to debug, XenServer always ships with a release and debug hypervisor built from the same source so customers can trivially switch into the debug version and collect logs if we think that is a useful thing to do, but this is only used in a fraction of cases in the first place. Anything more complicated is definitely going to require an experienced human to investigate, at which point they can do whatever they want. ~Andrew
On Tue, 30 Jan 2018, Julien Grall wrote: > Hi, > > On 30/01/18 18:29, Andrew Cooper wrote: > > On 30/01/18 17:00, Julien Grall wrote: > > > > > > > > > On 30/01/18 16:38, Andrew Cooper wrote: > > > > On 30/01/18 16:14, Julien Grall wrote: > > > > > Hi all, > > > > > > > > > > This small series replaces all call to domain_crash_synchronous by > > > > > injecting > > > > > an exception to the guest. > > > > > > > > > > This will result to a nicer trace from the guest (no need to > > > > > manually walk > > > > > the stack) and give a chance to the guest to give a bit more > > > > > information on > > > > > what it was doing. > > > > > > > > > > Cheers, > > > > > > > > > > Julien Grall (3): > > > > > xen/arm: io: Distinguish unhandled IO from aborted one > > > > > xen/arm: Don't crash domain on bad MMIO emulation > > > > > xen/arm: Don't crash the domain on invalid HVC immediate > > > > > > > > Thanks. > > > > > > > > I don't feel qualified to review these, but some notes. > > > > > > > > Patch 1. s/avodi/avoid/ in the commit message > > > > > > > > Patches 2 and 3. You probably want to convert the printks to > > > > gdprintk()s, otherwise guests can choke up the ratelimited log. Doing > > > > so will also mean that the vcpu will be identified consistently, which > > > > it isn't currently. > > > We didn't use g*printk because it would be more confusing to print the > > > current vCPU in some cases (e.g when accessing the re-distributor of > > > another vCPU) or does not matter (e.g for ITS). > > > > In the former case, you'd want to print both current, and the target > > vcpu. The latter still matters what current is if something goes wrong. > > > > We have plenty of similar cases in x86, but at the point you are > > printing an diagnostic message, ignoring current is almost always the > > wrong think to do. > > I will look at it on another series. > > > > > > > > > The problem with the debug version is those information are actually > > > quite useful in non-debug build. We found quite a few issues thanks to > > > them. > > > > > > I think it would make more sense for Xen to provide per-guest > > > ratelimited than hiding those messages in non-debug build. > > > > Per guest is quite a lot more complicated than global, and would still > > require a global limit to prevent a concerted attack from multiple > > guests to avoid DoSing the system. > > > > Debug vs unilateral is your prerogative as a maintainer, but as you've > > said yourself, the are used for debugging purposes, which proves my point. > > So on x86, you always request the user to reproduce it with debug build > enable? > > Stefano, what's your opinion here? I think it would be great to have per-guest rate limiting. On ARM, it would be impossible to ask users to repro with debug enabled, given that many deployments are embedded (set-top boxes, etc). I think it is OK to keep the XENLOG_DEBUG, they should be filtered out by loglvl. But we need to be careful about the others. We might want to convert the XENLOG_WARNING in traps.c to XENLOG_DEBUG: they are warning about guests misbehaving, but from the hypervisor point of view, it's not a problem, guests can shoot themselves if they want to; it's OK.
On 30/01/18 19:21, Stefano Stabellini wrote: > On Tue, 30 Jan 2018, Julien Grall wrote: >> Hi, >> >> On 30/01/18 18:29, Andrew Cooper wrote: >>> On 30/01/18 17:00, Julien Grall wrote: >>>> >>>> On 30/01/18 16:38, Andrew Cooper wrote: >>>>> On 30/01/18 16:14, Julien Grall wrote: >>>>>> Hi all, >>>>>> >>>>>> This small series replaces all call to domain_crash_synchronous by >>>>>> injecting >>>>>> an exception to the guest. >>>>>> >>>>>> This will result to a nicer trace from the guest (no need to >>>>>> manually walk >>>>>> the stack) and give a chance to the guest to give a bit more >>>>>> information on >>>>>> what it was doing. >>>>>> >>>>>> Cheers, >>>>>> >>>>>> Julien Grall (3): >>>>>> xen/arm: io: Distinguish unhandled IO from aborted one >>>>>> xen/arm: Don't crash domain on bad MMIO emulation >>>>>> xen/arm: Don't crash the domain on invalid HVC immediate >>>>> Thanks. >>>>> >>>>> I don't feel qualified to review these, but some notes. >>>>> >>>>> Patch 1. s/avodi/avoid/ in the commit message >>>>> >>>>> Patches 2 and 3. You probably want to convert the printks to >>>>> gdprintk()s, otherwise guests can choke up the ratelimited log. Doing >>>>> so will also mean that the vcpu will be identified consistently, which >>>>> it isn't currently. >>>> We didn't use g*printk because it would be more confusing to print the >>>> current vCPU in some cases (e.g when accessing the re-distributor of >>>> another vCPU) or does not matter (e.g for ITS). >>> In the former case, you'd want to print both current, and the target >>> vcpu. The latter still matters what current is if something goes wrong. >>> >>> We have plenty of similar cases in x86, but at the point you are >>> printing an diagnostic message, ignoring current is almost always the >>> wrong think to do. >> I will look at it on another series. >> >>>> The problem with the debug version is those information are actually >>>> quite useful in non-debug build. We found quite a few issues thanks to >>>> them. >>>> >>>> I think it would make more sense for Xen to provide per-guest >>>> ratelimited than hiding those messages in non-debug build. >>> Per guest is quite a lot more complicated than global, and would still >>> require a global limit to prevent a concerted attack from multiple >>> guests to avoid DoSing the system. >>> >>> Debug vs unilateral is your prerogative as a maintainer, but as you've >>> said yourself, the are used for debugging purposes, which proves my point. >> So on x86, you always request the user to reproduce it with debug build >> enable? >> >> Stefano, what's your opinion here? > I think it would be great to have per-guest rate limiting. > > On ARM, it would be impossible to ask users to repro with debug enabled, > given that many deployments are embedded (set-top boxes, etc). > > I think it is OK to keep the XENLOG_DEBUG, they should be filtered out > by loglvl. But we need to be careful about the others. We might want to > convert the XENLOG_WARNING in traps.c to XENLOG_DEBUG: they are warning > about guests misbehaving, but from the hypervisor point of view, it's not > a problem, guests can shoot themselves if they want to; it's OK. This is a valid argument, but to do it consistently, you'll want to provide an arch-specific version of gdprintk() so that the common code printk()s don't evaporate into /dev/null. ~Andrew
Hi Stefano, On 30/01/18 19:21, Stefano Stabellini wrote: > On Tue, 30 Jan 2018, Julien Grall wrote: >> Hi, >> >> On 30/01/18 18:29, Andrew Cooper wrote: >>> On 30/01/18 17:00, Julien Grall wrote: >>>> >>>> >>>> On 30/01/18 16:38, Andrew Cooper wrote: >>>>> On 30/01/18 16:14, Julien Grall wrote: >>>>>> Hi all, >>>>>> >>>>>> This small series replaces all call to domain_crash_synchronous by >>>>>> injecting >>>>>> an exception to the guest. >>>>>> >>>>>> This will result to a nicer trace from the guest (no need to >>>>>> manually walk >>>>>> the stack) and give a chance to the guest to give a bit more >>>>>> information on >>>>>> what it was doing. >>>>>> >>>>>> Cheers, >>>>>> >>>>>> Julien Grall (3): >>>>>> xen/arm: io: Distinguish unhandled IO from aborted one >>>>>> xen/arm: Don't crash domain on bad MMIO emulation >>>>>> xen/arm: Don't crash the domain on invalid HVC immediate >>>>> >>>>> Thanks. >>>>> >>>>> I don't feel qualified to review these, but some notes. >>>>> >>>>> Patch 1. s/avodi/avoid/ in the commit message >>>>> >>>>> Patches 2 and 3. You probably want to convert the printks to >>>>> gdprintk()s, otherwise guests can choke up the ratelimited log. Doing >>>>> so will also mean that the vcpu will be identified consistently, which >>>>> it isn't currently. >>>> We didn't use g*printk because it would be more confusing to print the >>>> current vCPU in some cases (e.g when accessing the re-distributor of >>>> another vCPU) or does not matter (e.g for ITS). >>> >>> In the former case, you'd want to print both current, and the target >>> vcpu. The latter still matters what current is if something goes wrong. >>> >>> We have plenty of similar cases in x86, but at the point you are >>> printing an diagnostic message, ignoring current is almost always the >>> wrong think to do. >> >> I will look at it on another series. >> >>> >>>> >>>> The problem with the debug version is those information are actually >>>> quite useful in non-debug build. We found quite a few issues thanks to >>>> them. >>>> >>>> I think it would make more sense for Xen to provide per-guest >>>> ratelimited than hiding those messages in non-debug build. >>> >>> Per guest is quite a lot more complicated than global, and would still >>> require a global limit to prevent a concerted attack from multiple >>> guests to avoid DoSing the system. >>> >>> Debug vs unilateral is your prerogative as a maintainer, but as you've >>> said yourself, the are used for debugging purposes, which proves my point. >> >> So on x86, you always request the user to reproduce it with debug build >> enable? >> >> Stefano, what's your opinion here? > > I think it would be great to have per-guest rate limiting. > > On ARM, it would be impossible to ask users to repro with debug enabled, > given that many deployments are embedded (set-top boxes, etc). > > I think it is OK to keep the XENLOG_DEBUG, they should be filtered out > by loglvl. But we need to be careful about the others. We might want to > convert the XENLOG_WARNING in traps.c to XENLOG_DEBUG: they are warning > about guests misbehaving, but from the hypervisor point of view, it's not > a problem, guests can shoot themselves if they want to; it's OK. Not really. Some of the XENLOG_WARNING may happen because Xen does not handle a trap (see the one on do_trap_guest_sync). That has nothing to do with a guest misbehaving and could happen if Xen boots on newer hardware. Cheers,