OSDN Git Service

multifd: Copy pages before compressing them with zlib
authorIlya Leoshkevich <iii@linux.ibm.com>
Tue, 5 Jul 2022 20:35:59 +0000 (22:35 +0200)
committerDr. David Alan Gilbert <dgilbert@redhat.com>
Wed, 20 Jul 2022 11:15:08 +0000 (12:15 +0100)
commit007e179ef0e97eafda4c9ff2a9d665a1947c7c6d
treef581c54112616acef952812961af1c5a702b89d1
parent8aff6f501d3561127e3425bbf0ebd936b3b92d0c
multifd: Copy pages before compressing them with zlib

zlib_send_prepare() compresses pages of a running VM. zlib does not
make any thread-safety guarantees with respect to changing deflate()
input concurrently with deflate() [1].

One can observe problems due to this with the IBM zEnterprise Data
Compression accelerator capable zlib [2]. When the hardware
acceleration is enabled, migration/multifd/tcp/plain/zlib test fails
intermittently [3] due to sliding window corruption. The accelerator's
architecture explicitly discourages concurrent accesses [4]:

    Page 26-57, "Other Conditions":

    As observed by this CPU, other CPUs, and channel
    programs, references to the parameter block, first,
    second, and third operands may be multiple-access
    references, accesses to these storage locations are
    not necessarily block-concurrent, and the sequence
    of these accesses or references is undefined.

Mark Adler pointed out that vanilla zlib performs double fetches under
certain circumstances as well [5], therefore we need to copy data
before passing it to deflate().

[1] https://zlib.net/manual.html
[2] https://github.com/madler/zlib/pull/410
[3] https://lists.nongnu.org/archive/html/qemu-devel/2022-03/msg03988.html
[4] http://publibfp.dhe.ibm.com/epubs/pdf/a227832c.pdf
[5] https://lists.gnu.org/archive/html/qemu-devel/2022-07/msg00889.html

Signed-off-by: Ilya Leoshkevich <iii@linux.ibm.com>
Message-Id: <20220705203559.2960949-1-iii@linux.ibm.com>
Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
migration/multifd-zlib.c