>On Fri, Apr 18, 2014 at 09:22:26AM +0000, Atsushi Kumagai wrote: >> >On 04/17/14 at 12:52pm, Baoquan He wrote: >> >> On 04/17/14 at 04:01am, Atsushi Kumagai wrote: >> >> > Hello Baoquan, >> >> > >> >> > >Hi Atsushi, >> >> > > >> >> > >I have got the test machine where bug reported and did a test. The >> >> > >changed code can make elf dump successful. >> >> > >> >> > Great, thanks for your help! >> >> > However, I still have questions. >> >> > >> >> > First, what is the difference between yours and mine? >> >> > >> >> > http://lists.infradead.org/pipermail/kexec/2014-April/011535.html >> >> >> >> Yeah, you are right, it's the same on changing the code bug. I mush >> >> haven't read your patch carefully. >> > must<-- >> >> >> >> > >> >> > My patch includes renaming some values, but the purpose looks >> >> > the same as yours. >> >> > Further, you described as below, >> >> > >> >> > >On 04/14/14 at 04:02pm, Baoquan He wrote: >> >> > but I still don't think this bug causes OOM. >> >> > Even if needed_size is calculated as so much size wrongly, bufsize_cyclic >> >> > will not exceed 40% of free memory by the check below: >> >> > >> >> > info->bufsize_cyclic = (free_size <= needed_size) ? free_size : needed_size; >> >> > >> >> > So it looks that bitmap1(40%) and bitmap2(40%) will fit in 80% of free >> >> > memory in any case. >> >> > >> >> > I may misunderstand something since your patch has an effect on this >> >> > issue in practice, could you correct me? >> >> >> >> It definitely will cause OOM. On my test machine, it has 100G memory. So >> >> per old code, its needed_size is 3200K*2 == 6.4M, if currently free >> >> memory is only 15M left, the free_size will be 15M*0.4 which is 6M. So >> >> info->bufsize_cyclic is assigned to be 6M. and only 3M is left for other >> >> use, e.g page cache, dynamic allocation. OOM will happen. >> >> >> > >> >BTW, in our case, there's about 30M free memory when we started saving >> >dump. It should be caused by my coarse estimation above. >> >> Thanks for your description, I understand that situation and >> the nature of the problem. >> >> That is, the assumption that 20% of free memory is enough for >> makedumpfile can be broken if free memory is too small. >> If your machine has 200GB memory, OOM will happen even after fix >> the too allocation bug. > >Why? In cyclic mode, shouldn't makedumpfile's memory usage be fixed >and should not be dependent on amount of RAM present in the system? Strictly speaking, it's not fixed but just restricted by the safe limit(80% of free memory) like below: - bitmap size: used for 1st and 2nd bitmaps - remains: can be used for the other works of makedumpfile (e.g. I/O buffer) pattern | bitmap size | remains ----------------------------------------------+---------------+------------- A. 100G memory with the too allocation bug | 12.8 MB | 17.2 MB B. 100G memory with fixed makedumpfile | 6.4 MB | 23.6 MB C. 200G memory with fixed makedumpfile | 12.8 MB | 17.2 MB D. 300G memory with fixed makedumpfile | 19.2 MB | 10.8 MB E. 400G memory with fixed makedumpfile | 24.0 MB | 6.0 MB F. 500G memory with fixed makedumpfile | 24.0 MB | 6.0 MB ... Baoquan got OOM in A pattern and didn't get it in B, so C must also fail due to OOM. This is just what I wanted to say. >Also, so even 30MB is not sufficient to run makedumpfile. That looks >like a lot of free memory to me. According to the table above and Baoquan's report, 23.6MB was enough but 17.2 MB was short for the other works, it sounds too much requirement also to me. >> I don't think this is a problem, it's natural that a lack of memory >> causes OOM. However, there is a thing we can do for improvement. >> >> What I think is: >> >> 1. Use a constant value as safe limit to calculate bufsize_cyclic >> instead of 80% of free memory. This value must be enough for >> makedumpfile's work except bitmap. >> >> 2. If free memory is smaller than the value, makedumpfile gives up >> to work early. > >What do we gain by makedumpfile giving up. System will reboot. System >will reboot anyway after OOM. Oh, you got a point. >> >> This change may reduce the possibility of lack of memory, but the >> required memory size will be changing every version, so maintaining >> it sounds tough to me. > >I think we need to dive deeper to figure out why 30MB of free memory >is not sufficient. To me something looks wrong here. I agree. I'm trying to get the peak size of memory footprint except cyclic_buffer with memory cgroup (memory.max_usage_in_bytes), but I don't make it yet. >Secondly, I think using absolute values is not a good idea. It will be >very hard to keep track and udpate that value. > >At max we can warn saying makedumpfile needs X MB of memory and only >Y MB of free memory is available. Yes, I don't want to choose that way, and to maintain the constant value just to inform possibility of OOM sounds worthless. Thanks Atsushi Kumagai
