Re: recent head having significantly less "avail memory"
Date: Mon, 13 Sep 2021 22:25:55 UTC
On 14/09/21 00:12, Konstantin Belousov wrote: > On Mon, Sep 13, 2021 at 10:07:46PM +0200, Guido Falsi wrote: >> On 13/09/21 19:08, Konstantin Belousov wrote: >>> On Mon, Sep 13, 2021 at 02:59:25PM +0200, Guido Falsi via freebsd-current wrote: >>>> Hi, >>>> >>>> I updated head recently and today I noticed a difference which looks wrong. >>>> >>>> At boodt the new head shows signifcantly less avail memory than before, >>>> around 3 GiB less. >>>> >>>> I moved from commit 71fbc6faed6 [1] where I got: >>>> >>>> Aug 28 22:03:03 marvin kernel: real memory = 17179869184 (16384 MB) >>>> Aug 28 22:03:03 marvin kernel: avail memory = 16590352384 (15821 MB) >>>> >>>> to commit 7955efd574b [2] where I get: >>>> >>>> Sep 13 10:44:40 marvin kernel: real memory = 17179869184 (16384 MB) >>>> Sep 13 10:44:40 marvin kernel: avail memory = 13298876416 (12682 MB) >>>> >>>> I'm seeing this on multiple machines. >>>> >>>> Unluckily bisecting and trying an older loader.efi in sseparate tests did >>>> not give me any more insight. >>>> >>>> The recent changes to efi loader, starting with commit 6032b6ba9596 [3] look >>>> like a possible trigger to this, but I have been unable to confirm it. >>>> >>>> Any suggesstions on how to proceed to debug thiss? ANy idea what a fix could >>>> be? >>> >>> Is this UEFI or bios boot? >>> Provide verbose dmesg for old and new boots on the same machine. >>> For UEFI boot, show output of 'sysctl machdep.efi_map', again for old >>> and new boots. >>> >> >> I shared the data you request here: >> >> https://www.madpilot.net/cloud/s/ENW5zF7jfmrmFeG > > Thanks. > > If you do on the loader prompt for the new (AKA bad) kernel > copy_staging enable > and then boot, does the report of avail memory becomes good? > Yes, it works as expected, that is reports the amount of memory I expect: Sep 14 00:24:50 marvin kernel: real memory = 17179869184 (16384 MB) Sep 14 00:24:50 marvin kernel: avail memory = 16590356480 (15821 MB) -- Guido Falsi <mad@madpilot.net>