Message ID | 20191023000547.7831-1-f.fainelli@gmail.com |
---|---|
Headers | show |
Series | irqchip/gic: Allow the use of SGI interrupts | expand |
Hi Florian, Needless to say, I mostly have questions... On 2019-10-23 01:05, Florian Fainelli wrote: > SGI interrupts are a convenient way for trusted firmware to target a > specific set of CPUs. Update the ARM GIC code to allow the > translation > and mapping of SGI interrupts. > > Since the kernel already uses SGIs for various inter-processor > interrupt > activities, we specifically make sure that we do not let users of the > IRQ API to even try to map those. > > Internal IPIs remain dispatched through handle_IPI() while public > SGIs > get promoted to a normal interrupt flow management. > > Signed-off-by: Florian Fainelli <f.fainelli@gmail.com> > --- > drivers/irqchip/irq-gic.c | 41 > +++++++++++++++++++++++++++------------ > 1 file changed, 29 insertions(+), 12 deletions(-) > > diff --git a/drivers/irqchip/irq-gic.c b/drivers/irqchip/irq-gic.c > index 30ab623343d3..dcfdbaacdd64 100644 > --- a/drivers/irqchip/irq-gic.c > +++ b/drivers/irqchip/irq-gic.c > @@ -385,7 +385,10 @@ static void __exception_irq_entry > gic_handle_irq(struct pt_regs *regs) > * Pairs with the write barrier in gic_raise_softirq > */ > smp_rmb(); > - handle_IPI(irqnr, regs); > + if (irqnr < NR_IPI) > + handle_IPI(irqnr, regs); > + else > + handle_domain_irq(gic->domain, irqnr, regs); Double EOI, UNPREDICTABLE territory, your state machine is now dead. > #endif > continue; > } > @@ -1005,20 +1008,34 @@ static int gic_irq_domain_translate(struct > irq_domain *d, > if (fwspec->param_count < 3) > return -EINVAL; > > - /* Get the interrupt number and add 16 to skip over SGIs */ > - *hwirq = fwspec->param[1] + 16; > - > - /* > - * For SPIs, we need to add 16 more to get the GIC irq > - * ID number > - */ > - if (!fwspec->param[0]) > + *hwirq = fwspec->param[1]; > + switch (fwspec->param[0]) { > + case 0: > + /* > + * For SPIs, we need to add 16 more to get the GIC irq > + * ID number > + */ > + *hwirq += 16; > + /* fall through */ > + case 1: > + /* Add 16 to skip over SGIs */ > *hwirq += 16; > + *type = fwspec->param[2] & IRQ_TYPE_SENSE_MASK; > > - *type = fwspec->param[2] & IRQ_TYPE_SENSE_MASK; > + /* Make it clear that broken DTs are... broken */ > + WARN_ON(*type == IRQ_TYPE_NONE); > + break; > + case 2: > + /* Refuse to map internal IPIs */ > + if (*hwirq < NR_IPI) So depending on how the kernel uses SGIs, you can or cannot use these SGIs. That looks like a good way to corner ourselves into not being to change much. Also, do you expect this to work for both Group-0 and Group-1 interrupts (since you imply that this works as a communication medium with the secure side)? Given that the kernel running in NS has no way to enable/disable Group-0 interrupts, this looks terminally flawed. Or is that Group-1 only? How do we describe which SGIs are guaranteed to be available to Linux? > + return -EPERM; > + > + *type = IRQ_TYPE_NONE; Or not. SGI are edge triggered, by definition. > + break; > + default: > + break; > + } > > - /* Make it clear that broken DTs are... broken */ > - WARN_ON(*type == IRQ_TYPE_NONE); Really? M.
Hello marc, On 10/23/19 6:22 AM, Marc Zyngier wrote: > Hi Florian, > > Needless to say, I mostly have questions... > > On 2019-10-23 01:05, Florian Fainelli wrote: >> SGI interrupts are a convenient way for trusted firmware to target a >> specific set of CPUs. Update the ARM GIC code to allow the translation >> and mapping of SGI interrupts. >> >> Since the kernel already uses SGIs for various inter-processor interrupt >> activities, we specifically make sure that we do not let users of the >> IRQ API to even try to map those. >> >> Internal IPIs remain dispatched through handle_IPI() while public SGIs >> get promoted to a normal interrupt flow management. >> >> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com> >> --- >> drivers/irqchip/irq-gic.c | 41 +++++++++++++++++++++++++++------------ >> 1 file changed, 29 insertions(+), 12 deletions(-) >> >> diff --git a/drivers/irqchip/irq-gic.c b/drivers/irqchip/irq-gic.c >> index 30ab623343d3..dcfdbaacdd64 100644 >> --- a/drivers/irqchip/irq-gic.c >> +++ b/drivers/irqchip/irq-gic.c >> @@ -385,7 +385,10 @@ static void __exception_irq_entry >> gic_handle_irq(struct pt_regs *regs) >> * Pairs with the write barrier in gic_raise_softirq >> */ >> smp_rmb(); >> - handle_IPI(irqnr, regs); >> + if (irqnr < NR_IPI) >> + handle_IPI(irqnr, regs); >> + else >> + handle_domain_irq(gic->domain, irqnr, regs); > > Double EOI, UNPREDICTABLE territory, your state machine is now dead. Oh yes, the interrupt flow now also goes through ->irq_eoi (that's the whole point), meh. > >> #endif >> continue; >> } >> @@ -1005,20 +1008,34 @@ static int gic_irq_domain_translate(struct >> irq_domain *d, >> if (fwspec->param_count < 3) >> return -EINVAL; >> >> - /* Get the interrupt number and add 16 to skip over SGIs */ >> - *hwirq = fwspec->param[1] + 16; >> - >> - /* >> - * For SPIs, we need to add 16 more to get the GIC irq >> - * ID number >> - */ >> - if (!fwspec->param[0]) >> + *hwirq = fwspec->param[1]; >> + switch (fwspec->param[0]) { >> + case 0: >> + /* >> + * For SPIs, we need to add 16 more to get the GIC irq >> + * ID number >> + */ >> + *hwirq += 16; >> + /* fall through */ >> + case 1: >> + /* Add 16 to skip over SGIs */ >> *hwirq += 16; >> + *type = fwspec->param[2] & IRQ_TYPE_SENSE_MASK; >> >> - *type = fwspec->param[2] & IRQ_TYPE_SENSE_MASK; >> + /* Make it clear that broken DTs are... broken */ >> + WARN_ON(*type == IRQ_TYPE_NONE); >> + break; >> + case 2: >> + /* Refuse to map internal IPIs */ >> + if (*hwirq < NR_IPI) > > So depending on how the kernel uses SGIs, you can or cannot use these SGIs. > That looks like a good way to corner ourselves into not being to change > much. arch/arm/kernel/smp.c has a forward looking statement about SGI numbering: /* * SGI8-15 can be reserved by secure firmware, and thus may * not be usable by the kernel. Please keep the above limited * to at most 8 entries. */ is this something that can be used as an universal and unbreakable rule for the ARM64 kernel as well in order to ensure SGIs 8-15 can be usable through the IRQ API or is this simply not a guarantee at all? > > Also, do you expect this to work for both Group-0 and Group-1 interrupts > (since you imply that this works as a communication medium with the secure > side)? Given that the kernel running in NS has no way to enable/disable > Group-0 interrupts, this looks terminally flawed. Or is that Group-1 only? That would be Group-1 interrupts only, are you suggesting there is an additional check being done that such SGIs are actually part of Group-1? > > How do we describe which SGIs are guaranteed to be available to Linux? In our case, the Device Tree mailbox node gets populated its interrupts property with the SGI number(s), and that same number is also passed as a configuration parameter to the trusted firmware. Or are you echoing back to your earlier comment about the fact that if the kernel changes its own definition of NR_IPI then we suddenly start breaking IRQ API uses of SGIs in a certain range? > >> + return -EPERM; >> + >> + *type = IRQ_TYPE_NONE; > > Or not. SGI are edge triggered, by definition. > >> + break; >> + default: >> + break; >> + } >> >> - /* Make it clear that broken DTs are... broken */ >> - WARN_ON(*type == IRQ_TYPE_NONE); > > Really? Given the comment in gic_set_type() about SGIs, the WARN_ON() was moved above to continue checking for GIC_SPI and GIC_PPI, but we should extract the type from the Devic eTree and only permit an edge mask.
On 2019-10-23 18:02, Florian Fainelli wrote: > Hello marc, > > On 10/23/19 6:22 AM, Marc Zyngier wrote: >> Hi Florian, >> >> Needless to say, I mostly have questions... >> >> On 2019-10-23 01:05, Florian Fainelli wrote: >>> SGI interrupts are a convenient way for trusted firmware to target >>> a >>> specific set of CPUs. Update the ARM GIC code to allow the >>> translation >>> and mapping of SGI interrupts. >>> >>> Since the kernel already uses SGIs for various inter-processor >>> interrupt >>> activities, we specifically make sure that we do not let users of >>> the >>> IRQ API to even try to map those. >>> >>> Internal IPIs remain dispatched through handle_IPI() while public >>> SGIs >>> get promoted to a normal interrupt flow management. >>> >>> Signed-off-by: Florian Fainelli <f.fainelli@gmail.com> >>> --- >>> drivers/irqchip/irq-gic.c | 41 >>> +++++++++++++++++++++++++++------------ >>> 1 file changed, 29 insertions(+), 12 deletions(-) >>> >>> diff --git a/drivers/irqchip/irq-gic.c b/drivers/irqchip/irq-gic.c >>> index 30ab623343d3..dcfdbaacdd64 100644 >>> --- a/drivers/irqchip/irq-gic.c >>> +++ b/drivers/irqchip/irq-gic.c >>> @@ -385,7 +385,10 @@ static void __exception_irq_entry >>> gic_handle_irq(struct pt_regs *regs) >>> * Pairs with the write barrier in gic_raise_softirq >>> */ >>> smp_rmb(); >>> - handle_IPI(irqnr, regs); >>> + if (irqnr < NR_IPI) >>> + handle_IPI(irqnr, regs); >>> + else >>> + handle_domain_irq(gic->domain, irqnr, regs); >> >> Double EOI, UNPREDICTABLE territory, your state machine is now dead. > > Oh yes, the interrupt flow now also goes through ->irq_eoi (that's > the > whole point), meh. Indeed. But to be honest, we should probably consider moving all the SGI handling to normal interrupts. There's hardly any reason why we should keep SGIs out of the normal interrupt model, other than maybe performance (and that's pretty dubious). > >> >>> #endif >>> continue; >>> } >>> @@ -1005,20 +1008,34 @@ static int gic_irq_domain_translate(struct >>> irq_domain *d, >>> if (fwspec->param_count < 3) >>> return -EINVAL; >>> >>> - /* Get the interrupt number and add 16 to skip over SGIs >>> */ >>> - *hwirq = fwspec->param[1] + 16; >>> - >>> - /* >>> - * For SPIs, we need to add 16 more to get the GIC irq >>> - * ID number >>> - */ >>> - if (!fwspec->param[0]) >>> + *hwirq = fwspec->param[1]; >>> + switch (fwspec->param[0]) { >>> + case 0: >>> + /* >>> + * For SPIs, we need to add 16 more to get the GIC irq >>> + * ID number >>> + */ >>> + *hwirq += 16; >>> + /* fall through */ >>> + case 1: >>> + /* Add 16 to skip over SGIs */ >>> *hwirq += 16; >>> + *type = fwspec->param[2] & IRQ_TYPE_SENSE_MASK; >>> >>> - *type = fwspec->param[2] & IRQ_TYPE_SENSE_MASK; >>> + /* Make it clear that broken DTs are... broken */ >>> + WARN_ON(*type == IRQ_TYPE_NONE); >>> + break; >>> + case 2: >>> + /* Refuse to map internal IPIs */ >>> + if (*hwirq < NR_IPI) >> >> So depending on how the kernel uses SGIs, you can or cannot use >> these SGIs. >> That looks like a good way to corner ourselves into not being to >> change >> much. > > arch/arm/kernel/smp.c has a forward looking statement about SGI > numbering: > > /* > * SGI8-15 can be reserved by secure firmware, and thus may > * not be usable by the kernel. Please keep the above limited > * to at most 8 entries. > */ > > is this something that can be used as an universal and unbreakable > rule > for the ARM64 kernel as well in order to ensure SGIs 8-15 can be > usable > through the IRQ API or is this simply not a guarantee at all? There is no guarantee whatsoever. There's an ARM recommendation about the above split, but that's it. Hardly something that can be enforced. Now, your firmware is the one that gives you the DT, so if it is inconsistent in configuring the interrupt and presenting it to the kernel, tough luck. >> Also, do you expect this to work for both Group-0 and Group-1 >> interrupts >> (since you imply that this works as a communication medium with the >> secure >> side)? Given that the kernel running in NS has no way to >> enable/disable >> Group-0 interrupts, this looks terminally flawed. Or is that Group-1 >> only? > > That would be Group-1 interrupts only, are you suggesting there is an > additional check being done that such SGIs are actually part of > Group-1? You can try and change the configuration of that interrupt (priority, for example), and see if that sticks. If it doesn't, you're in trouble (and nothing you can do about it). >> >> How do we describe which SGIs are guaranteed to be available to >> Linux? > > In our case, the Device Tree mailbox node gets populated its > interrupts > property with the SGI number(s), and that same number is also passed > as > a configuration parameter to the trusted firmware. Or are you echoing > back to your earlier comment about the fact that if the kernel > changes > its own definition of NR_IPI then we suddenly start breaking IRQ API > uses of SGIs in a certain range? That's indeed my worry. There is also the fact that the kernel itself will never expose such reservation in DT, so we'd have to tread carefully here. We probably need to specify that *only* SGI8-15 are allowed to be described as such. Another thing: why don't you use a PPI? If you use a GICv2, you're also using ancient cores, and they have PPIs to spare. Or do you rely on being able to inject interrupts from one core to another? Thanks, M.