Bruce Richardson <bruce.richard...@intel.com> writes: > On Sat, Sep 23, 2023 at 10:21:04AM +0200, Thomas Monjalon wrote: >> 22/09/2023 15:23, Bruce Richardson: >> > On Fri, Sep 22, 2023 at 02:57:32PM +0200, Thomas Monjalon wrote: >> > > 20/09/2023 12:09, Bruce Richardson: >> > > > On Wed, Sep 20, 2023 at 12:00:08PM +0200, David Marchand wrote: >> > > > > On Thu, Sep 14, 2023 at 12:42 PM Bruce Richardson >> > > > > <bruce.richard...@intel.com> wrote: >> > > > > > >> > > > > > When examining the IOL testing failures for patch series [1], I >> > > > > > observed >> > > > > > that the failures reported were in the eal_flags_file_prefix unit >> > > > > > test. >> > > > > > I was able to reproduce this on my system by passing an additional >> > > > > > "--on-pci" flag to the test run, since the log to the test has >> > > > > > errors >> > > > > > about device availability. Adding the "no-pci" flag to the >> > > > > > individual >> > > > > >> > > > > Something is not clear to me. >> > > > > >> > > > > While I understand that passing "no-pci" helps avoiding the issue (as >> > > > > described below), I have some trouble understanding this passage >> > > > > (above) with "--on-pci". >> > > > >> > > > That's a typo for no-pci. When I ran the test on my system with the >> > > > main >> > > > process using no-pci, I was able to reproduce the issue seen in the IOL >> > > > lab. Otherwise I couldn't reproduce it. >> > > > >> > > > > How did you reproduce the issue? >> > > > > >> > > > > >> > > > > > test commands used by the unit tests fixed the issue thereafter, >> > > > > > allowing the test to pass in all cases for me. Therefore, I am >> > > > > > submitting this patch in the hopes of making the test more robust, >> > > > > > since >> > > > > > the observed failures seem unrelated to the original patchset [1] I >> > > > > > submitted. >> > > > > > >> > > > > > [1] http://patches.dpdk.org/project/dpdk/list/?series=29406 >> > > > > > >> > > > > > Bruce Richardson (1): >> > > > > > app/test: skip PCI bus scan when testing prefix flags >> > > > > > >> > > > > > app/test/test_eal_flags.c | 20 ++++++++++---------- >> > > > > > 1 file changed, 10 insertions(+), 10 deletions(-) >> > > > > >> > > > > Iiuc, the problem is that the file_prefix unit test can fail if any >> > > > > DPDK subsystem forgets to release some memory and some hugepages are >> > > > > left behind at the cleanup step. >> > > > > Passing --no-pci as you suggest hides issues coming from PCI drivers. >> > > > > >> > > > > This is something I tried to fix too, with >> > > > > https://patchwork.dpdk.org/project/dpdk/list/?series=29288 though my >> > > > > fix only handles a part of the issue (here, the ethdev drivers). >> > > > > >> > > > > Another way to make the file prefix more robust would be to remove >> > > > > the >> > > > > check on released memory, or move it to another test. >> > > > > >> > > > I actually think the test is a good one to have. Also, taking in your >> > > > patch >> > > > to help with the issue is a good idea also. >> > > > >> > > > I'd still suggest that this patch be considered anyway, as there is no >> > > > need >> > > > to do PCI bus scanning as part of this test. Therefore I'd view it as a >> > > > harmless addition that may help things. >> > > >> > > I'm hesitating. >> > > This test is checking if some memory is left, and I think it is sane. >> > > If we add --no-pci, we reduce the coverage of this check. >> > > >> > > Now that the root cause is fixed by David in ethdev >> > > (https://patches.dpdk.org/project/dpdk/patch/20230821085806.3062613-4-david.march...@redhat.com/) >> > > we could continue checking memory freeing with PCI drivers. >> > > So I tend to reject this patch. >> > > >> > > Other opinions? >> > > >> > No objection to this patch being rejected if not necessary. >> > >> > However, I'd question if the normal case is actually checking for freeing >> > memory in PCI drivers. I suspect that in EAL cleanup we delete all files we >> > use, irrespective of whether the mappings are still in use. Then when the >> > process exits the hugepages will be completely freed back - even if some >> > components leaked memory. I believe this case is checking for correct EAL >> > cleanup of hugepage files, not for any memory leaks, and in that regard >> > omitting some components should make no difference. >> >> You're right, that's why I'm hesitating. >> Fortunately it helped to discover a memory leak. >> Do we want to add a new specific test for memory leaks, >> or is it OK to have it in this one? >> > > Not really sure. I'd tend towards saying that special memory leak checkers > like valgrind are better to use than trying to detect them in unit tests > directly. However, not an expert in this area.
I do tend to agree that we should rely on more generic memory infra like valgrind. However, the way we use mempools doesn't always lend itself to leak checkers like valgrind which usually expect to own all the individual blocks. Maybe newer versions can work with our mempools though? > /Bruce
