The IMSIC state variable eistate[] is modified by CSR instructions
within a range dedicated to the local CPU and by MMIO writes from any CPU.
Access to eistate from MMIO accessors is protected by the BQL, but
read-modify-write (RMW) sequences from CSRRW do not acquire the BQL,
making the RMW sequence vulnerable to a race condition with MMIO access
from a remote CPU.
This race can manifest as missing IPI or MSI in multi-CPU systems, eg:
[ 43.008092] watchdog: BUG: soft lockup - CPU#2 stuck for 27s! [kworker/u19:1:52]
[ 43.011723] CPU: 2 UID: 0 PID: 52 Comm: kworker/u19:1 Not tainted 6.11.0-rc6
[ 43.013070] Workqueue: events_unbound deferred_probe_work_func
[ 43.018776] [<ffffffff800b4a86>] smp_call_function_many_cond+0x190/0x5c2
[ 43.019205] [<ffffffff800b4f28>] on_each_cpu_cond_mask+0x20/0x32
[ 43.019447] [<ffffffff8001069a>] __flush_tlb_range+0xf2/0x190
[ 43.019683] [<ffffffff80010914>] flush_tlb_kernel_range+0x20/0x28
The interrupt line raise/lower sequence was changed to prevent a race
between the evaluation of the eistate and the execution of the qemu_irq
raise/lower, ensuring that the interrupt line is not incorrectly
deactivated based on a stale topei check result. To avoid holding BQL
all modifications of eistate are converted to atomic operations.
Signed-off-by: Tomasz Jeznach <tjeznach@rivosinc.com>
Reviewed-by: Alistair Francis <alistair.francis@wdc.com>
Message-ID: <a7604e4d61068ca4d384ae2a1377e1521d4d0235.1725651699.git.tjeznach@rivosinc.com>
Signed-off-by: Alistair Francis <alistair.francis@wdc.com>
1165e30d95