Re: [PATCH v1 10/13] migrate/ram: Handle RAM block resizes during postcopy

[David Hildenbrand]

On 20.02.20 21:54, Peter Xu wrote:
> On Wed, Feb 19, 2020 at 05:17:22PM +0100, David Hildenbrand wrote:
>> Resizing while migrating is dangerous and does not work as expected.
>> The whole migration code works on the usable_length of ram blocks and does
>> not expect this to change at random points in time.
>>
>> In the case of postcopy, relying on used_length is racy as soon as the
>> guest is running. Also, when used_length changes we might leave the
>> uffd handler registered for some memory regions, reject valid pages
>> when migrating and fail when sending the recv bitmap to the source.
>>
>> Resizing can be trigger *after* (but not during) a reset in
>> ACPI code by the guest
>> - hw/arm/virt-acpi-build.c:acpi_ram_update()
>> - hw/i386/acpi-build.c:acpi_ram_update()
>>
>> Let's remember the original used_length in a separate variable and
>> use it in relevant postcopy code. Make sure to update it when we resize
>> during precopy, when synchronizing the RAM block sizes with the source.
>>
>> Cc: "Dr. David Alan Gilbert" <address@hidden>
>> Cc: Juan Quintela <address@hidden>
>> Cc: Eduardo Habkost <address@hidden>
>> Cc: Paolo Bonzini <address@hidden>
>> Cc: Igor Mammedov <address@hidden>
>> Cc: "Michael S. Tsirkin" <address@hidden>
>> Cc: Richard Henderson <address@hidden>
>> Cc: Shannon Zhao <address@hidden>
>> Cc: Alex Bennée <address@hidden>
>> Cc: Peter Xu <address@hidden>
>> Signed-off-by: David Hildenbrand <address@hidden>
>> ---
>>  include/exec/ramblock.h  |  9 +++++++++
>>  migration/postcopy-ram.c | 15 ++++++++++++---
>>  migration/ram.c          | 11 +++++++++--
>>  3 files changed, 30 insertions(+), 5 deletions(-)
>>
>> diff --git a/include/exec/ramblock.h b/include/exec/ramblock.h
>> index 07d50864d8..0e9e9b346b 100644
>> --- a/include/exec/ramblock.h
>> +++ b/include/exec/ramblock.h
>> @@ -59,6 +59,15 @@ struct RAMBlock {
>>       */
>>      unsigned long *clear_bmap;
>>      uint8_t clear_bmap_shift;
>> +
>> +    /*
>> +     * RAM block used_length before the guest started running while postcopy
>> +     * was active. Once the guest is running, used_length can change. Used 
>> to
>> +     * register/unregister uffd handlers and as the size of the recv bitmap.
>> +     * Receiving any page beyond this length will bail out, as it could not 
>> have
>> +     * been valid on the source.
>> +     */
>> +    ram_addr_t postcopy_length;
>>  };
>>  #endif
>>  #endif
>> diff --git a/migration/postcopy-ram.c b/migration/postcopy-ram.c
>> index a36402722b..c68caf4e42 100644
>> --- a/migration/postcopy-ram.c
>> +++ b/migration/postcopy-ram.c
>> @@ -17,6 +17,7 @@
>>   */
>>  
>>  #include "qemu/osdep.h"
>> +#include "qemu/rcu.h"
>>  #include "exec/target_page.h"
>>  #include "migration.h"
>>  #include "qemu-file.h"
>> @@ -31,6 +32,7 @@
>>  #include "qemu/error-report.h"
>>  #include "trace.h"
>>  #include "hw/boards.h"
>> +#include "exec/ramblock.h"
>>  
>>  /* Arbitrary limit on size of each discard command,
>>   * keeps them around ~200 bytes
>> @@ -456,6 +458,13 @@ static int init_range(RAMBlock *rb, void *opaque)
>>      ram_addr_t length = qemu_ram_get_used_length(rb);
>>      trace_postcopy_init_range(block_name, host_addr, offset, length);
>>  
>> +    /*
>> +     * Save the used_length before running the guest. In case we have to
>> +     * resize RAM blocks when syncing RAM block sizes from the source during
>> +     * precopy, we'll update it manually via the ram block notifier.
>> +     */
>> +    rb->postcopy_length = length;
>> +
>>      /*
>>       * We need the whole of RAM to be truly empty for postcopy, so things
>>       * like ROMs and any data tables built during init must be zero'd
>> @@ -478,7 +487,7 @@ static int cleanup_range(RAMBlock *rb, void *opaque)
>>      const char *block_name = qemu_ram_get_idstr(rb);
>>      void *host_addr = qemu_ram_get_host_addr(rb);
>>      ram_addr_t offset = qemu_ram_get_offset(rb);
>> -    ram_addr_t length = qemu_ram_get_used_length(rb);
>> +    ram_addr_t length = rb->postcopy_length;
>>      MigrationIncomingState *mis = opaque;
>>      struct uffdio_range range_struct;
>>      trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
>> @@ -600,7 +609,7 @@ static int nhp_range(RAMBlock *rb, void *opaque)
>>      const char *block_name = qemu_ram_get_idstr(rb);
>>      void *host_addr = qemu_ram_get_host_addr(rb);
>>      ram_addr_t offset = qemu_ram_get_offset(rb);
>> -    ram_addr_t length = qemu_ram_get_used_length(rb);
>> +    ram_addr_t length = rb->postcopy_length;
>>      trace_postcopy_nhp_range(block_name, host_addr, offset, length);
>>  
>>      /*
>> @@ -644,7 +653,7 @@ static int ram_block_enable_notify(RAMBlock *rb, void 
>> *opaque)
>>      struct uffdio_register reg_struct;
>>  
>>      reg_struct.range.start = (uintptr_t)qemu_ram_get_host_addr(rb);
>> -    reg_struct.range.len = qemu_ram_get_used_length(rb);
>> +    reg_struct.range.len = rb->postcopy_length;
>>      reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
>>  
>>      /* Now tell our userfault_fd that it's responsible for this area */
>> diff --git a/migration/ram.c b/migration/ram.c
>> index ab1f5534cf..6d1dcb362c 100644
>> --- a/migration/ram.c
>> +++ b/migration/ram.c
>> @@ -244,7 +244,7 @@ int64_t ramblock_recv_bitmap_send(QEMUFile *file,
>>          return -1;
>>      }
>>  
>> -    nbits = block->used_length >> TARGET_PAGE_BITS;
>> +    nbits = block->postcopy_length >> TARGET_PAGE_BITS;
>>  
>>      /*
>>       * Make sure the tmp bitmap buffer is big enough, e.g., on 32bit
>> @@ -3159,7 +3159,13 @@ static int ram_load_postcopy(QEMUFile *f)
>>                  break;
>>              }
>>  
>> -            if (!offset_in_ramblock(block, addr)) {
>> +            /*
>> +             * Relying on used_length is racy and can result in false 
>> positives.
>> +             * We might place pages beyond used_length in case RAM was 
>> shrunk
>> +             * while in postcopy, which is fine - trying to place via
>> +             * UFFDIO_COPY/UFFDIO_ZEROPAGE will never segfault.
>> +             */
>> +            if (!block->host || addr >= block->postcopy_length) {
>>                  error_report("Illegal RAM offset " RAM_ADDR_FMT, addr);
>>                  ret = -EINVAL;
>>                  break;
>> @@ -3744,6 +3750,7 @@ static void ram_mig_ram_block_resized(RAMBlockNotifier 
>> *n, void *host,
>>                               rb->idstr);
>>              }
>>          }
>> +        rb->postcopy_length = new_size;
> 
> With this change, postcopy_length will be the same as used_length?
> 
> I thought you wanted to cache that value when starting the postcopy
> phase so postcopy_length should be constant after set once.  Did I
> misunderstood?

So, my understanding on the migration target:

1. Source VM started and initialized.
- Theoretically we could have resizes here already in the future (e.g.,
  virtio-mem). Right now, not.
2. Precopy data is loaded
- RAM block sizes will be synchronized to the source.
3. Guest starts running
- Any RAM block resize will differ to the source.


For postcopy, we have to "freeze" the RAM block size after 2, but before
3. So rb->postcopy_length will always correspond to used_length on the
migration source.

Interestingly, userfaultfd handlers etc. are registered just before 3,
before any "resizes of interest" were handled.

So, postcopy_length will match used_length until we start our guest and
do resizes (shrink/grow) that are out of sync with our source.

Makes sense?

-- 
Thanks,

David / dhildenb