Hi, > From: Al Stone [mailto:ahs3(a)redhat.com] > Sent: Tuesday, June 16, 2015 5:37 AM > > On 06/14/2015 06:39 PM, Zheng, Lv wrote: > > Hi, > > > >> From: Al Stone [mailto:ahs3(a)redhat.com] > >> Sent: Saturday, June 13, 2015 6:00 AM > >> > >> On 06/09/2015 12:16 AM, Zheng, Lv wrote: > >>> Hi, > >>> > >>> The patch is wrong. > >>> If we want to dump tables without accessing /dev/mem, we have acpidump -c. > >>> The only problem here is: > >>> There is a bug that even with -c specified, acpidump cannot work without accessing /dev/mem. > >>> So we shouldn't introduce a new portable acpidump OSL but should fix the very problem. > >>> > >>> We've noticed the similar issue in handling this kernel bug: > >>> https://bugzilla.kernel.org/show_bug.cgi?id=87971 > >>> The fix for this issue can be found here: > >>> https://github.com/acpica/acpica/pull/77 > >>> > >>> Thanks and best regards > >>> -Lv > >> > >> The problem I have with this fix for the issue is that this command: > >> > >> # acpidump -n APIC > >> > >> will work on x86 but fail on arm64. Getting the -c option to work properly > >> is a good thing, and should be done regardless. But, this means we have to > >> use "acpidump -c -n APIC" on systems without /dev/mem. For my part: > >> > >> (a) these are not really customized tables, but the standard ACPI tables > >> used to boot the system, so using -c is a little confusing, and > >> > >> (b) we now have to use different commands on different architectures to > >> get the same result. > >> > >> My preference would be that the same command has the same behavior everywhere, > >> which is why I patched it the way I did. > > > > This is not our preference, though... > > I've been in the trouble of being easily confused by the bug reporters. > > And I cannot distinguish if the table has been customized or root cause the bugs if the reporter did what I asked for but hid what he > silently did. > > If we cannot distinguish the BIOS provided tables and the user customized tables from users' output, we'll likely lose the capability of > taking evidence... > > Understood; I was in technical support roles for many years. > > Perhaps I am misunderstanding this use of the word "customized." On ARM, the > only way I can be sure to get the tables supplied by UEFI is to use the UEFI > shell and dump them. Once the OS has booted, the tables in /sys/firmware/acpi > are what the system is currently using, and those are the ones I'm interested > in using for debugging when something breaks. So, unless I have replaced the > DSDT in the initrd, or added SSDTs, /sys/firmware/acpi will be exactly the same > as what UEFI provided. Even then, those are the tables I want to see if there > is a problem since those are the ones being used. Is that what "customized" > means, in this case: the actual run-time tables and not the tables provided by > UEFI? Yes, so the "customized" tables should be dumped by default as you've suggested. > > >> This happens to be very useful on arm64 where /dev/mem is not reliable, but it > >> could also be usable on any system that does not want to expose /dev/mem for > >> some reason. I think the advantage of the patch sent is that the user command > >> stays the same no matter what the situation. > > > > My suggestion would be to support this via kernel. > > Let the kernel exposes 2 kind of tables (customized or not customized) into the /sys/firmware/acpi/tables. > > Which can meet both of the requirements. > > Okay, this sort of answers my questions. "Customized" is not the word I would > have chosen for this -- I think I would have made it > /sys/firmware/acpi/tables/current (i.e., "customized") and > /sys/firmware/acpi/tables/boot (i.e., > not "customized"). When I was thinking of tables as being customized, I was > thinking of new versions of all kinds of tables, not just the DSDT and SSDTs. > On ARM, going in and out of ACPI mode is just not allowed so there's really no > opportunity to customize the tables much, other than adding SSDTs. > > Creating new sysfs entries in a kernel driver would be simple enough to do; > the issue would be getting it accepted. On x86, there's no guarantee that > the BIOS tables and run-time tables are the same, so customized and not > customized may be very different, and these kernel changes would make sense. > On ARMv8, the boot and run-time tables will be essentially identical most of > the time so it's not clear there's a lot of value in adding more stuff to > sysfs. If I create this kernel code, though, it seems to me it would make > sense to change acpidump so it does not use /dev/mem for any architecture, but > always uses sysfs on all architectures -- do you agree? IMO, we could do this better. Users won't want to see the original BIOS tables. So the BIOS tables needn't appear in the userspace unless required, thus we could export them via debugfs. Or even better: We could provide some ioctl or sysctl from kernel to allow acpidump to obtain what it want, for example: 1. Find RSDP 2. Validate and dump a table from any address (we need to support Load/LoadTable opcodes) > > I'll take a look at doing this on the kernel side... Thanks. Best regards -Lv > > > Thanks and best regards > > -Lv > > > >> > >>>> From: Devel [mailto:devel-bounces(a)acpica.org] On Behalf Of Al Stone > >>>> Sent: Thursday, June 04, 2015 8:43 AM > >>>> > >>>> On arm64 systems in particular, the use of /dev/mem is not recommended. The > >>>> contents may or may not be valid depending on the memory map being used, since > >>>> they are not standardized. > >>>> > >>>> The attached patch will cause arm64 (aka ARMv8 or AArch64) Linux systems to > >>>> use a new file called source/os_specific/service_layers/oslinuxtbl_nodevmem.c > >>>> (a subset of the oslinuxtbl.c code) that allows acpidump to read all ACPI > >>>> tables from /sys/firmware/acpi instead of from /dev/mem. This will help ensure > >>>> that the tables retrieved are the ones actually being used and that their > >>>> content is correct. > >>>> > >>>> This patch applies on top of the 20150515 version, and I have included it in > >>>> the 20150515-2 versions of the Fedora and Debian packages. I've tried the > >>>> resulting acpidump on x86, x86_64 and arm64 systems and it seems to work well. > >>>> > >>>> Signed-off-by: Al Stone > >>>> > >>>> -- > >>>> ciao, > >>>> al > >>>> ----------------------------------- > >>>> Al Stone > >>>> Software Engineer > >>>> Red Hat, Inc. > >>>> ahs3(a)redhat.com > >>>> ----------------------------------- > >> > >> > >> -- > >> ciao, > >> al > >> ----------------------------------- > >> Al Stone > >> Software Engineer > >> Red Hat, Inc. > >> ahs3(a)redhat.com > >> ----------------------------------- > > > -- > ciao, > al > ----------------------------------- > Al Stone > Software Engineer > Red Hat, Inc. > ahs3(a)redhat.com > -----------------------------------