migrate.c: migrate_fd_put_buffer(): Do not busyloop: stop writing if EWOULDBLOCK
PATCH 1/2 migrate.c: migrate_fd_put_buffer(): Do not busyloop: stop writing if EWOULDBLOCK
Theoretically the migration code is non-blocking, designed for live migration.
Practically migrate_fd_put_buffer busy-loops trying to write. That's since on many machines EWOULDBLOCK==EAGAIN (e.g. look in <linux>/include/asm-generic/errno.h).
Busy looping is much worse then blocking (in case a write actually blocks).
Fixed by removing "(s->get_error(s)) == EWOULDBLOCK)" from the while condition.
--- migration.c | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-)
migration.c +1 -1
@@ -168,7 +168,7 @@ ssize_t migrate_fd_put_buffer(void *opaque, const void *data, size_t size)do { ret = s->write (s, data, size); - }while (ret ==-1 && ((s->get_error (s)) ==EINTR || (s->get_error (s)) ==EWOULDBLOCK )); + }while (ret ==-1 && s->get_error (s) ==EINTR );if (ret ==-1 ) ret = -(s->get_error (s));
Signature · 1 line
1.6.0.6
PATCH 2/2 : debugging migration busy-looping
Try it with/without fix for migration busy-loop, and migrate a VM. (without the fix, patch fails just manually fix it).
To see the problem clearly try to suspend the destination while a migration process is active (e.g. with CTRL-z and fg).
--- migration.c | 3 +++ 1 files changed, 3 insertions(+), 0 deletions(-)
migration.c +3 -0
@@ -165,11 +165,14 @@ ssize_t migrate_fd_put_buffer(void *opaque, const void *data, size_t size) {FdMigrationState *s = opaque;ssize_t ret; +int n =0 ;do { ret = s->write (s, data, size); +if (++n %1000 ==0 )printf ("LOOP size=%lu n=%d\n" , size, n); }while (ret ==-1 && s->get_error (s) ==EINTR ); +if (n >4 )printf ("OUT size=%lu n=%d\n" , size, n);if (ret ==-1 ) ret = -(s->get_error (s));
Signature · 1 line
1.6.0.6