From: Joerg Vehlow Date: Mon, 16 Nov 2020 14:09:14 +0100 Subject: [LTP] [PATCH 0/1] overcommit_memory: Remove unstable subtest Message-ID: <20201116130915.18264-1-lkml@jv-coder.de> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: ltp@lists.linux.it Hi, this is something like an RFC. (I think I mixed my thoughts between this and the patch description. Maybe read the patch description first). I found the overcommit_memory test, that tries to allocate all alocatable memory for overcommit policy never, failed a lot and a lot more often, if the system has more memory. When looking at the kernel source I found the reason: The percpu counter that counts the used memory uses a counter for every cpu, if the allocation or deallocations are very small. The more memory the system has, the bigger "small" is defined. See mm_compute_batch. I started seeing this issue a lot after upgrading to 20200930 comming from 20190115. Some changes in the framework may have led to this. I don't think this is a kernel bug, but a result from switching between overcommit modes. In overcommit mode never, the batch size is a lot smaller than in the other modes (ram_pages/cpus/256 instead of ra,_pages/cpus/4). This leads to allocations done before switching the mode to be accounted in the per cpu counters, and deallocated after in the global counter, making the global counter negative. If the overcommit mode was the same all the time, it should all have been accounted in the same counters and the global counter wouldn't be negative. J?rg