Re: [Qemu-devel] [PATCH 07/11] exec: add a reference to the region returned by address_space_translate

[Paolo Bonzini <pbonzini@redhat.com>]

Il 28/06/2013 15:41, Jan Kiszka ha scritto:
> On 2013-06-25 13:13, Paolo Bonzini wrote:
>> Once address_space_translate will only be protected by RCU, the returned
>> MemoryRegion can disappear as soon as the RCU read-side critical
>> section ends.  Avoid this by adding a reference to the region, and
>> dropping it in the caller of address_space_translate.
>>
>> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
>> ---
>>  exec.c                | 16 ++++++++++++++--
>>  include/exec/memory.h |  3 ++-
>>  2 files changed, 16 insertions(+), 3 deletions(-)
>>
>> diff --git a/exec.c b/exec.c
>> index 5454e4d..b7f032d 100644
>> --- a/exec.c
>> +++ b/exec.c
>> @@ -286,6 +286,7 @@ MemoryRegion *address_space_translate(AddressSpace *as, hwaddr addr,
>>  
>>      *plen = len;
>>      *xlat = addr;
>> +    memory_region_ref(mr);
>>      return mr;
>>  }
>>  
>> @@ -1960,6 +1961,7 @@ bool address_space_rw(AddressSpace *as, hwaddr addr, uint8_t *buf,
>>                  memcpy(buf, ptr, l);
>>              }
>>          }
>> +        memory_region_unref(mr);
>>          len -= l;
>>          buf += l;
>>          addr += l;
>> @@ -2125,8 +2127,10 @@ void *address_space_map(AddressSpace *as,
>>              raddr = memory_region_get_ram_addr(mr) + xlat;
>>          } else {
>>              if (memory_region_get_ram_addr(mr) + xlat != raddr + todo) {
>> +                memory_region_unref(mr);
>>                  break;
>>              }
>> +            memory_region_unref(mr);
>>          }
>>  
>>          len -= l;
> 
> This stage of the implementation looks inconsistent for
> memory_region_get_ram_addr. Isn't the idea so far that the mr reference
> is dropped again before leaving the function?

memory_region_get_ram_addr is not taking a reference.
address_space_map's address_space_translate is matched by the unref in
address_space_unmap.  You're getting the same memory region multiple
times, and you have to unref it every time except the first.

But there is a missing unref at the end of address_space_map (bug
introduced when separating this patch from patch 11).  I'll resend the
series.

Paolo

>> @@ -2226,6 +2230,7 @@ static inline uint32_t ldl_phys_internal(hwaddr addr,
>>              break;
>>          }
>>      }
>> +    memory_region_unref(mr);
>>      return val;
>>  }
>>  
>> @@ -2285,6 +2290,7 @@ static inline uint64_t ldq_phys_internal(hwaddr addr,
>>              break;
>>          }
>>      }
>> +    memory_region_unref(mr);
>>      return val;
>>  }
>>  
>> @@ -2352,6 +2358,7 @@ static inline uint32_t lduw_phys_internal(hwaddr addr,
>>              break;
>>          }
>>      }
>> +    memory_region_unref(mr);
>>      return val;
>>  }
>>  
>> @@ -2399,6 +2406,7 @@ void stl_phys_notdirty(hwaddr addr, uint32_t val)
>>              }
>>          }
>>      }
>> +    memory_region_unref(mr);
>>  }
>>  
>>  /* warning: addr must be aligned */
>> @@ -2440,6 +2448,7 @@ static inline void stl_phys_internal(hwaddr addr, uint32_t val,
>>          }
>>          invalidate_and_set_dirty(addr1, 4);
>>      }
>> +    memory_region_unref(mr);
>>  }
>>  
>>  void stl_phys(hwaddr addr, uint32_t val)
>> @@ -2503,6 +2512,7 @@ static inline void stw_phys_internal(hwaddr addr, uint32_t val,
>>          }
>>          invalidate_and_set_dirty(addr1, 2);
>>      }
>> +    memory_region_unref(mr);
>>  }
>>  
>>  void stw_phys(hwaddr addr, uint32_t val)
>> @@ -2592,11 +2602,13 @@ bool cpu_physical_memory_is_io(hwaddr phys_addr)
>>  {
>>      MemoryRegion*mr;
>>      hwaddr l = 1;
>> +    bool res;
>>  
>>      mr = address_space_translate(&address_space_memory,
>>                                   phys_addr, &phys_addr, &l, false);
>>  
>> -    return !(memory_region_is_ram(mr) ||
>> -             memory_region_is_romd(mr));
>> +    res = !(memory_region_is_ram(mr) || memory_region_is_romd(mr));
>> +    memory_region_unref(mr);
>> +    return res;
>>  }
>>  #endif
> 
> Unless I'm missing something ATM, address_space_access_valid,
> cpu_physical_memory_write_rom and tb_invalidate_phys_addr also require
> unref calls.
> 
>> diff --git a/include/exec/memory.h b/include/exec/memory.h
>> index c842d48..2a37b2c 100644
>> --- a/include/exec/memory.h
>> +++ b/include/exec/memory.h
>> @@ -1003,7 +1003,8 @@ bool address_space_write(AddressSpace *as, hwaddr addr,
>>  bool address_space_read(AddressSpace *as, hwaddr addr, uint8_t *buf, int len);
>>  
>>  /* address_space_translate: translate an address range into an address space
>> - * into a MemoryRegion and an address range into that section
>> + * into a MemoryRegion and an address range into that section.  Add a reference
>> + * to that region.
>>   *
>>   * @as: #AddressSpace to be accessed
>>   * @addr: address within that address space
>>
> 
> Jan
>