Amit Machhiwal <[email protected]> writes: > A previous change ("KVM: PPC: Book3S HV: Add preempt_disable() around > virtual-mode HPTE bit-lock users") at [1] added preempt_disable() around > guest vCPU HPT hcalls and page fault handlers executed in virtual mode. > This ensures vCPU threads cannot be preempted while holding > HPTE_V_HVLOCK or the reverse-map lock. > > However, several host-side virtual-mode paths in book3s_64_mmu_hv.c also > acquire HPTE_V_HVLOCK and lock_rmap() in process context with preemption > enabled. Specifically, kvm_unmap_rmapp() and kvm_age_rmapp() take both > locks during MMU notifier callbacks and memslot flushes, while > resize_hpt_rehash_hpte() acquires HPTE_V_HVLOCK during HPT resizing. If > a host thread is preempted while holding either bit-lock, any guest vCPU > thread scheduled on the same CPU spinning for the same lock with > preemption disabled will never yield, resulting in a permanent CPU hang. > > Fix this by disabling preemption across these host-side critical > sections. > > For kvm_unmap_rmapp() and kvm_age_rmapp(), place preempt_disable() > before lock_rmap() so that both the rmap chain lock and the subsequent > HPTE_V_HVLOCK bit-lock are held under a single non-preemptible window. > To preserve the ABBA lock ordering between the two locks, re-enable > preemption after dropping the rmap lock on any retry or early-exit path > before spinning on cpu_relax(), allowing the lock holder to be > scheduled. > > For kvm_htab_write(), the HPT migration restore path, replace the direct > calls to kvmppc_do_h_remove() with kvmppc_virtmode_do_h_remove(), a thin > wrapper that brackets kvmppc_do_h_remove() with preempt_disable() and > preempt_enable(), mirroring the existing kvmppc_virtmode_do_h_enter() > helper. > > For resize_hpt_rehash_hpte(), place preempt_disable() before the > try_lock_hpte() spin loop and preempt_enable() after unlock_hpte() at > the function exit. > > [1] https://git.kernel.org/powerpc/c/abbbe5ff83e7c9ed3a4d61b73011f1df7bf87907 > > Fixes: 6165d5dd99db ("KVM: PPC: Book3S HV: add virtual mode handlers for HPT > hcalls and page faults") > Cc: [email protected] # v5.14+ > Signed-off-by: Amit Machhiwal <[email protected]> > --- > Changes in this version: > - Separated host-side fixes in book3s_64_mmu_hv.c into a standalone patch > (guest-side paths were merged from v1 resend). > - Previous discussion: > https://lore.kernel.org/all/[email protected]/ >
Thanks. Looks good to me. Please feel free to add: Reviewed-by: Ritesh Harjani (IBM) <[email protected]>
