[CRIU] [PATCH v5 22/33] eventpoll: Make post_open stage may fail

Kirill Tkhai ktkhai at virtuozzo.com
Wed Jan 18 02:16:16 PST 2017



On 18.01.2017 13:11, Pavel Emelyanov wrote:
> On 12/26/2016 05:28 PM, Kirill Tkhai wrote:
>> Currently, it's just an additional check. But later this will be used.
>>
>> v5: Use "1" for return
> 
> Why 1?

1 means "the fle is not ready for restore". Where should I add a comment
about that?

Starting from this set there's still a mix between explicit numbers
> (in code) and ORV_something (in comments).

Ah, sure, there is ORV in comment to patch "files: Kill struct file_desc_ops::post_open",
which should sounds as

"    files: Kill struct file_desc_ops::post_open
    
    This make open fds asynchronous. Now we restore them in a poll-style.
    Every ->open callback operates like a poll. It returns 1, i.e.
    "call me again", if some events have not received yet, OR 0,
    if restoring of a fle has finished.
    
    We iterate over list of file descriptors and look at its return value.
    If there were not progress in restoring of a fle, we sleep on task_st
    futex. As all events are sent together with changing task_st, we will
    be woken up, when there is a some new work for us.	"

Should I resend it or series?
 
>> v2: Use generic FLE_OPEN and FLE_RESTORED to determ if a fle is ready
>>
>> Signed-off-by: Kirill Tkhai <ktkhai at virtuozzo.com>
>> ---
>>  criu/eventpoll.c |   27 +++++++++++++++++++++++++++
>>  1 file changed, 27 insertions(+)
>>
>> diff --git a/criu/eventpoll.c b/criu/eventpoll.c
>> index 88076db8a..42a656f31 100644
>> --- a/criu/eventpoll.c
>> +++ b/criu/eventpoll.c
>> @@ -21,6 +21,7 @@
>>  #include "image.h"
>>  #include "util.h"
>>  #include "log.h"
>> +#include "pstree.h"
>>  
>>  #include "protobuf.h"
>>  #include "images/eventpoll.pb-c.h"
>> @@ -140,6 +141,24 @@ static int eventpoll_open(struct file_desc *d)
>>  	close(tmp);
>>  	return -1;
>>  }
>> +
>> +static int epoll_not_ready_tfd(EventpollTfdEntry *tdefe)
>> +{
>> +	struct fdinfo_list_entry *fle;
>> +
>> +	list_for_each_entry(fle, &rsti(current)->used, used_list) {
>> +		if (tdefe->tfd != fle->fe->fd)
>> +			continue;
>> +
>> +		if (fle->desc->ops->type == FD_TYPES__EVENTPOLL)
>> +			return (fle->stage != FLE_OPEN);
>> +		else
>> +			return (fle->stage != FLE_RESTORED);
>> +	}
>> +	BUG();
>> +	return 0;
>> +}
>> +
>>  static int eventpoll_retore_tfd(int fd, int id, EventpollTfdEntry *tdefe)
>>  {
>>  	struct epoll_event event;
>> @@ -165,11 +184,19 @@ static int eventpoll_post_open(struct file_desc *d, int fd)
>>  	info = container_of(d, struct eventpoll_file_info, d);
>>  
>>  	for (i = 0; i < info->efe->n_tfd; i++) {
>> +		if (epoll_not_ready_tfd(info->efe->tfd[i]))
>> +			return 1;
>> +	}
>> +	for (i = 0; i < info->efe->n_tfd; i++) {
>>  		if (eventpoll_retore_tfd(fd, info->efe->id, info->efe->tfd[i]))
>>  			return -1;
>>  	}
>>  
>>  	list_for_each_entry(td_info, &eventpoll_tfds, list) {
>> +		if (epoll_not_ready_tfd(td_info->tdefe))
>> +			return 1;
>> +	}
>> +	list_for_each_entry(td_info, &eventpoll_tfds, list) {
>>  		if (td_info->tdefe->id != info->efe->id)
>>  			continue;
>>  
>>
>> .
>>
> 


More information about the CRIU mailing list