On Thu, Oct 24, 2019 at 11:05 PM Olivier Matz <olivier.m...@6wind.com> wrote: > > Hi, > > On Wed, Oct 23, 2019 at 08:32:08PM +0530, Jerin Jacob wrote: > > On Wed, Oct 23, 2019 at 8:17 PM Olivier Matz <olivier.m...@6wind.com> wrote: > > > > > > Hi, > > > > > > On Wed, Oct 23, 2019 at 03:42:39PM +0530, Jerin Jacob wrote: > > > > On Tue, Oct 22, 2019 at 7:01 PM Vamsi Krishna Attunuru > > > > <vattun...@marvell.com> wrote: > > > > > > > > > > Hi Ferruh, > > > > > > > > > > Can you please explain the problems in using kni dedicated mbuf alloc > > > > > routines while enabling kni iova=va mode. Please see the below > > > > > discussion with Andrew. He wanted to know the problems in having > > > > > newer APIs. > > > > > > > > > > > > While waiting for the Ferruh reply, I would like to summarise the > > > > current status > > > > > > > > # In order to make KNI work with IOVA as VA, We need to make sure > > > > mempool pool _object_ should not span across two huge pages > > > > > > > > # This problem can be fixed by, either of: > > > > > > > > a) Introduce a flag in mempool to define this constraint, so that, > > > > when only needed, this constraint enforced and this is in line > > > > with existing semantics of addressing such problems in mempool > > > > > > > > b) Instead of creating a flag, Make this behavior by default in > > > > mempool for IOVA as VA case > > > > > > > > Upside: > > > > b1) There is no need for specific mempool_create for KNI. > > > > > > > > Downside: > > > > b2) Not align with existing mempool API semantics > > > > b3) There will be a trivial amount of memory waste as we can not > > > > allocate from the edge. Considering the normal huge > > > > page memory size is 1G or 512MB this not a real issue. > > > > > > > > c) Make IOVA as PA when KNI kernel module is loaded > > > > > > > > Upside: > > > > c1) Doing option (a) would call for new KNI specific mempool create > > > > API i.e existing KNI applications need a one-line change in > > > > application to make it work with release 19.11 or later. > > > > > > > > Downslide: > > > > c2) Driver which needs RTE_PCI_DRV_NEED_IOVA_AS_VA can not work with KNI > > > > c3) Need root privilege to run KNI as IOVA as PA need root privilege > > > > > > > > For the next year, we expect applications to work 19.11 without any > > > > code change. My personal opinion to make go with option (a) > > > > and update the release notes to document the change any it simple > > > > one-line change. > > > > > > > > The selection of (a) vs (b) is between KNI and Mempool maintainers. > > > > Could we please reach a consensus? Or can we discuss this TB meeting? > > > > > > > > We are going back and forth on this feature on for the last 3 > > > > releases. Now that, we solved all the technical problems, please help > > > > us > > > > to decide (a) vs (b) to make forward progress. > > > > > > Thank you for the summary. > > > What is not clear to me is if (a) or (b) may break an existing > > > application, and if yes, in which case. > > > > Thanks for the reply. > > > > To be clear we are talking about out of tree KNI tree application. > > Which they don't want to > > change rte_pktmbuf_pool_create() to rte_kni_pktmbuf_pool_create() and > > build for v19.11 > > > > So in case (b) there is no issue as It will be using > > rte_pktmbuf_pool_create (). > > But in case of (a) it will create an issue if out of tree KNI > > application is using rte_pktmbuf_pool_create() which is not using the > > NEW flag. > > Following yesterday's discussion at techboard, I looked at the mempool > code and at my previous RFC patch. It took some time to remind me what > was my worries.
Thanks for the review Olivier. Just to make sure the correct one is reviewed. 1) v7 had similar issue mentioned http://patches.dpdk.org/patch/56585/ 2) v11 addressed the review comments and you have given the Acked-by for mempool change http://patches.dpdk.org/patch/61559/ My thought process in the TB meeting was, since rte_mempool_populate_from_pg_sz_chunks() reviwed replace rte_pktmbuf_pool_create's rte_mempool_populate_default() with rte_mempool_populate_from_pg_sz_chunks() in IOVA == VA case to avoid a new KNI mempool_create API. > > Currently, in rte_mempool_populate_default(), when the mempool is > populated, we first try to allocate one iova-contiguous block of (n * > elt_size). On success, we use this memory to fully populate the mempool > without taking care of crossing page boundaries. > > If we change the behavior to prevent objects from crossing pages, the > assumption that allocating (n * elt_size) is always enough becomes > wrong. By luck, there is no real impact, because if the mempool is not > fully populated after this first iteration, it will allocate a new > chunk. > > To be rigorous, we need to better calculate the amount of memory to > allocate, according to page size. > > Looking at the code, I found another problem in the same area: let's say > we populate a mempool that requires 1.1GB (and we use 1G huge pages): > > 1/ mempool code will first tries to allocate an iova-contiguous zone > of 1.1G -> fail > 2/ it then tries to allocate a page-aligned non iova-contiguous > zone of 1.1G, which is 2G. On success, a lot of memory is wasted. > 3/ on error, we try to allocate the biggest zone, it can still return > a zone between 1.1G and 2G, which can also waste memory. > > I will rework my mempool patchset to properly address these issues, > hopefully tomorrow. OK. > > Also, I thought about another idea to solve your issue, not sure it is > better but it would not imply to change the mempool behavior. If I > understood the problem, when a mbuf is accross 2 pages, the copy of the > data can fail in kni because the mbuf is not virtually contiguous in the For KNI use case, we would need _physically_ contiguous to make sure that using, get_user_pages_remote() we get physically contiguous memory map, so that both KNI kernel thread and KNI kernel context and DPDK userspace can use the same memory in different contexts. > kernel. So why not in this case splitting the memcpy() into several, > each of them being on a single page (and calling phys2virt() for each > page)? The same would have to be done when accessing the fields of the > mbuf structure if it crosses a page boundary. Would that work? This If the above is the requirement, Does this logic need to be in slow path or fast path? > could be a B plan. OK. > > Olivier
