genirq: __setup_irq(): fix type of literal 1 in shift
PATCH 1/1 genirq: __setup_irq(): fix type of literal 1 in shift
If we ever get a value >= 31 from ffz(), we'd be invoking UB; in the > 31 case, probably assigning the same thread_mask to to multiple irqactions (at least on x86_64, where the shift count is implicitly truncated to 5 bits).
In practice, I think the bug is mostly harmless, since when we first get ffz == 31, the RHS is - ignoring UB - (int)(0x80000000), and sign-extension means that the 32nd irqaction just ends up eating the top 33 bits of thread_mask, so all subsequent __setup_irq calls would just end up hitting the -EBUSY branch. (AFAICT, no code seems to rely on ->thread_mask being a single bit; it's all whole-sale |= and &=).
However, a sufficiently aggressive optimizer may use the UB of 1<<31 to decide that doesn't happen, and hence elide the sign-extension code, so that subsequent calls can indeed get ffz > 31. In any case, this is the right thing to do.
--- kernel/irq/manage.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/irq/manage.c +1 -1
@@ -1305,7 +1305,7 @@ __setup_irq(unsigned int irq, struct irq_desc *desc, struct irqaction *new)* thread_mask assigned. See the loop above which or's * all existing action->thread_mask bits. */ - new->thread_mask =1 <<ffz (thread_mask); + new->thread_mask =1UL <<ffz (thread_mask); }else if (new->handler == irq_default_primary_handler && !(desc->irq_data.chip->flags &IRQCHIP_ONESHOT_SAFE )) {
Signature · 1 line
2.11.0
tip-bot for Rasmus Villemoes
merge bot[tip:irq/core] genirq: Fix type of shifting literal 1 in __setup_irq()
genirq: Fix type of shifting literal 1 in __setup_irq()
Commit-ID: ffc661c99f621152d5fdcf53f9df0d48c326318b Gitweb: https://git.kernel.org/tip/ffc661c99f621152d5fdcf53f9df0d48c326318b Author: Rasmus Villemoes <[email protected]> AuthorDate: Mon, 30 Oct 2017 22:35:47 +0100 Committer: Thomas Gleixner <[email protected]> CommitDate: Sun, 12 Nov 2017 23:25:40 +0100
genirq: Fix type of shifting literal 1 in __setup_irq()
If ffz() ever returns a value >= 31 then the following shift is undefined behaviour because the literal 1 which gets shifted is treated as signed integer.
In practice, the bug is probably harmless, since the first undefined shift count is 31 which results - ignoring UB - in (int)(0x80000000). This gets sign extended so bit 32-63 will be set as well and all subsequent __setup_irq() calls would just end up hitting the -EBUSY branch.
However, a sufficiently aggressive optimizer may use the UB of 1<<31 to decide that doesn't happen, and hence elide the sign-extension code, so that subsequent calls can indeed get ffz > 31.
In any case, the right thing to do is to make the literal 1UL.
[ tglx: For this to happen a single interrupt would have to be shared by 32
devices. Hardware like that does not exist and would have way more problems than that. ]
--- kernel/irq/manage.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/irq/manage.c +1 -1
@@ -1289,7 +1289,7 @@ __setup_irq(unsigned int irq, struct irq_desc *desc, struct irqaction *new)* thread_mask assigned. See the loop above which or's * all existing action->thread_mask bits. */ - new->thread_mask =1 <<ffz (thread_mask); + new->thread_mask =1UL <<ffz (thread_mask); }else if (new->handler == irq_default_primary_handler && !(desc->irq_data.chip->flags &IRQCHIP_ONESHOT_SAFE )) {