Hi Alasdaia, 2014-07-20 3:13 GMT+09:00 Alasdair G Kergon <agk@xxxxxxxxxx>: > On Sat, Jul 19, 2014 at 08:40:52PM +0900, Satoru Takeuchi wrote: >> In addition, I'd like to know your opinion about whether setting >> the upper limit of nr_rambuf_pool argument is neccesary or not. > > 1) If it's a parameter passed from userspace under the control of the > sysadmin, then you should just trust the sysadmin to set a sensible > value, and only return an error if the value they ask for is unobtainable. Thank you for your comment. I don't say anything if this function returns an ENOMEM without any effect to whole system. In this case, it fails *after* evicting massive amount of memory. Since linux is used by very large amount of system, I consider someone will eventually make a mistake. So I consider setting upper limit is better way from the perspective of fool proof. > > Example: dm-ioctl.c copy_params() > > 2) If certain large values would be useless, leaving no memory for > anything else, then you could apply a sensible limit based on the amount > of memory available. I agree with you. It's difficult to guess the appropriate upper limit from the perspective of dm-writeboost at this point since it's depend on system and requires a number of performance measurements. So, the second best way is limiting the rambuf size based on the available memory as you say. It can avoid the above-mentioned problem anyway and very easy to implement. Thanks, Satoru > > Example: dm_bufio.c > > * Memory management policy: > * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory > * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower). > * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers. > > Alasdair > -- dm-devel mailing list dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
