| Re: [mosquitto-dev] Question about storing messages in the heap of broker |
Hi Sergey,
I've just tried to reproduce this using Mosquitto 1.6.9 running under massif, and a simple client that sends qos 2 messages to topic1, topic2, topic3, ...
In the first case, the heap usage increases to a low level just after startup, then remains essentially constant.
In the second case, I changed the client behaviour to publish with the retained bit set. The behaviour there is increasing heap usage.
A retained message requires a number of things to be stored: the topic on which it was published, the tree items, the message itself, and a few other bits. This is entirely consistent with what you are seeing, and may be the cause. There are other situations where the same outcome occurs - like if a client makes a subscription, then the topic tree items will remain in memory. Or if a persistent client is disconnected, then queued messages for it will also exhibit this behaviour, even if the messages aren't retained.
If your situation is different, please let us know where it differs and we can see what is happening.
Regards,
Roger
On Wed, 15 Apr 2020, 16:40 Sergey Grekhov, <grekhss@xxxxxxxxx> wrote:
_______________________________________________Hello, everyone!
I am investigating a possible memory leak issue in mosquitto and would like to ask for some help: I need to understand how it manages the published messages.
The use case I am testing is relatively simple: mosquitto broker without bridge, one client (both publisher and subscriber). Client subscribes for the topic, sends in it 100 identical messages with QOS = 2 and falls into busyloop (with responding to MQTT-pings).
In order to track memory allocations I had to modify the code in lib/memory_mosq.c. Particularly, I have added a list where I add/remove allocations. Based on this list I am calculating the total used memory. The items in this list also store the number of the line where mosquitto__malloc/calloc/realloc/strdup was invocated and a kind of a lifetime counter which provides understanding of how long this chunk is in the heap. (And also I found that some allocations are created using malloc/calloc/strdup bypassing mosquitto__*alloc functions. So, I also changed them.)
Using this approach I have found following strange behavior: after client finishes its publishing, broker still keeps some allocated objects in heap. Particularly:
- allocations from sub__add_hier_entry():
- mosquitto__calloc(1, sizeof(struct mosquitto__subhier))
- mosquitto__malloc(len+1)
- mosquitto__malloc(sizeof(UT_hash_table)) hidden in macro HASH_ADD_KEYPTR
- allocations from db__messages_easy_queue()/db__message_store():
- mosquitto__strdup(topic)
- mosquitto__calloc(1, sizeof(struct mosquitto_msg_store))
The tracking list shows that each of these non-freed allocations appears at least 50 times which is quite a lot for 100 publications. I also tried to change QOS and found that total memory is decreasing when one changes QOS from 2 to 1 or 0. The decrease is about 1K per each QOS downgrade. This looks like a memory leak.
At the same time, I believe that mosquitto is a project developed by professionals which knows how to manage memory. So, probably, I am missing something - maybe there is a memory reclaiming/re-using procedure which frees/re-uses these resources after some time?
Thank you in advance for help!
Kind regards,
Sergey.
mosquitto-dev mailing list
mosquitto-dev@xxxxxxxxxxx
To unsubscribe from this list, visit https://www.eclipse.org/mailman/listinfo/mosquitto-dev
_______________________________________________ mosquitto-dev mailing list mosquitto-dev@xxxxxxxxxxx To unsubscribe from this list, visit https://www.eclipse.org/mailman/listinfo/mosquitto-dev