powerpc/64s: Fix native_hpte_remove() to be irq-safe
authorMichael Ellerman <mpe@ellerman.id.au>
Wed, 17 May 2023 12:30:33 +0000 (22:30 +1000)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 23 Jul 2023 11:49:33 +0000 (13:49 +0200)
commit 8bbe9fee5848371d4af101be445303cac8d880c5 upstream.

Lockdep warns that the use of the hpte_lock in native_hpte_remove() is
not safe against an IRQ coming in:

  ================================
  WARNING: inconsistent lock state
  6.4.0-rc2-g0c54f4d30ecc #1 Not tainted
  --------------------------------
  inconsistent {IN-SOFTIRQ-W} -> {SOFTIRQ-ON-W} usage.
  qemu-system-ppc/93865 [HC0[0]:SC0[0]:HE1:SE1] takes:
  c0000000021f5180 (hpte_lock){+.?.}-{0:0}, at: native_lock_hpte+0x8/0xd0
  {IN-SOFTIRQ-W} state was registered at:
    lock_acquire+0x134/0x3f0
    native_lock_hpte+0x44/0xd0
    native_hpte_insert+0xd4/0x2a0
    __hash_page_64K+0x218/0x4f0
    hash_page_mm+0x464/0x840
    do_hash_fault+0x11c/0x260
    data_access_common_virt+0x210/0x220
    __ip_select_ident+0x140/0x150
    ...
    net_rx_action+0x3bc/0x440
    __do_softirq+0x180/0x534
    ...
    sys_sendmmsg+0x34/0x50
    system_call_exception+0x128/0x320
    system_call_common+0x160/0x2e4
  ...
   Possible unsafe locking scenario:

         CPU0
         ----
    lock(hpte_lock);
    <Interrupt>
      lock(hpte_lock);

   *** DEADLOCK ***
  ...
  Call Trace:
    dump_stack_lvl+0x98/0xe0 (unreliable)
    print_usage_bug.part.0+0x250/0x278
    mark_lock+0xc9c/0xd30
    __lock_acquire+0x440/0x1ca0
    lock_acquire+0x134/0x3f0
    native_lock_hpte+0x44/0xd0
    native_hpte_remove+0xb0/0x190
    kvmppc_mmu_map_page+0x650/0x698 [kvm_pr]
    kvmppc_handle_pagefault+0x534/0x6e8 [kvm_pr]
    kvmppc_handle_exit_pr+0x6d8/0xe90 [kvm_pr]
    after_sprg3_load+0x80/0x90 [kvm_pr]
    kvmppc_vcpu_run_pr+0x108/0x270 [kvm_pr]
    kvmppc_vcpu_run+0x34/0x48 [kvm]
    kvm_arch_vcpu_ioctl_run+0x340/0x470 [kvm]
    kvm_vcpu_ioctl+0x338/0x8b8 [kvm]
    sys_ioctl+0x7c4/0x13e0
    system_call_exception+0x128/0x320
    system_call_common+0x160/0x2e4

I suspect kvm_pr is the only caller that doesn't already have IRQs
disabled, which is why this hasn't been reported previously.

Fix it by disabling IRQs in native_hpte_remove().

Fixes: 35159b5717fa ("powerpc/64s: make HPTE lock and native_tlbie_lock irq-safe")
Cc: stable@vger.kernel.org # v6.1+
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://msgid.link/20230517123033.18430-1-mpe@ellerman.id.au
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
arch/powerpc/mm/book3s64/hash_native.c

index 9342e79870dfd408bb8220ed1865c8f9e1959f9e..430d1d935a7cb5c6355aeac0c0037b00cec84a17 100644 (file)
@@ -328,10 +328,12 @@ static long native_hpte_insert(unsigned long hpte_group, unsigned long vpn,
 
 static long native_hpte_remove(unsigned long hpte_group)
 {
+       unsigned long hpte_v, flags;
        struct hash_pte *hptep;
        int i;
        int slot_offset;
-       unsigned long hpte_v;
+
+       local_irq_save(flags);
 
        DBG_LOW("    remove(group=%lx)\n", hpte_group);
 
@@ -356,13 +358,16 @@ static long native_hpte_remove(unsigned long hpte_group)
                slot_offset &= 0x7;
        }
 
-       if (i == HPTES_PER_GROUP)
-               return -1;
+       if (i == HPTES_PER_GROUP) {
+               i = -1;
+               goto out;
+       }
 
        /* Invalidate the hpte. NOTE: this also unlocks it */
        release_hpte_lock();
        hptep->v = 0;
-
+out:
+       local_irq_restore(flags);
        return i;
 }