> Currently, kernel enumerates the possible CPUs by parsing both ACPI MADT > Local APIC entries and x2APIC entries. So CPUs with "valid" APIC IDs, > even if they have duplicated APIC IDs in Local APIC and x2APIC, are > always enumerated. > > Below is what ACPI MADT Local APIC and x2APIC describes on an > Ivebridge-EP system, > > [02Ch 0044 1] Subtable Type : 00 [Processor Local APIC] > [02Fh 0047 1] Local Apic ID : 00 > ... > [164h 0356 1] Subtable Type : 00 [Processor Local APIC] > [167h 0359 1] Local Apic ID : 39 > [16Ch 0364 1] Subtable Type : 00 [Processor Local APIC] > [16Fh 0367 1] Local Apic ID : FF > ... > [3ECh 1004 1] Subtable Type : 09 [Processor Local x2APIC] > [3F0h 1008 4] Processor x2Apic ID : 00000000 > ... > [B5Ch 2908 1] Subtable Type : 09 [Processor Local x2APIC] > [B60h 2912 4] Processor x2Apic ID : 00000077 > > As a result, kernel shows "smpboot: Allowing 168 CPUs, 120 hotplug CPUs". > And this wastes significant amount of memory for the per-cpu data. > Plus this also breaks https://lore.kernel.org/all/87edm36qqb.ffs@tglx/, > because __max_logical_packages is over-estimated by the APIC IDs in > the x2APIC entries. > > According to https://uefi.org/specs/ACPI/6.5/05_ACPI_Software_Programming_Model.html#processor-local-x2apic-structure, > "[Compatibility note] On some legacy OSes, Logical processors with APIC > ID values less than 255 (whether in XAPIC or X2APIC mode) must use the > Processor Local APIC structure to convey their APIC information to OSPM, > and those processors must be declared in the DSDT using the Processor() > keyword. Logical processors with APIC ID values 255 and greater must use > the Processor Local x2APIC structure and be declared using the Device() > keyword.". > > Enumerate CPUs from x2APIC enties with APIC ID values 255 or greater, > when valid CPU from Local APIC is already detected. > > Signed-off-by: Zhang Rui <rui.zhang@xxxxxxxxx> > --- > I didn't find any clear statement in the ACPI spec about if a mixture of > Local APIC and x2APIC entries is allowed or not. So it would be great if > this can be clarified. Has this been clarified? The reason that I ask is that Google Cloud has a 360 vCPU Zen4 VM occupying two virtual sockets, and the corresponding MADT table has a mixture of Local APIC and X2APIC entries. All of the LPUs in virtual socket 0 have extended APIC IDs below 255, and they have Local APIC entries. All of the LPUs in virtual socket 1 have extended APIC IDs above 255, and they have X2APIC entries. Prior to this change, Linux assigned CPU numbers to all even-numbered LPUs on virtual socket 0, followed by all even-numbered LPUs on virtual socket 1, followed by all odd-numbered LPUs on virtual socket 0, followed by all odd-numbered LPUs on virtual socket 1. node #0, CPUs: #1 #2 ... #87 #88 #89 node #1, CPUs: #90 #91 #92 ... #177 #178 #179 node #0, CPUs: #180 #181 #182 ... #267 #268 #269 node #1, CPUs: #270 #271 #272 ... #357 #358 #359 After this change, however, Linux assigns CPU numbers to all LPUs on virtual socket 0 before assigning any CPU numbers to LPUs on virtual socket 1. node #0, CPUs: #1 #2 ... #87 #88 #89 node #1, CPUs: #180 #181 #182 ... #267 #268 #269 node #0, CPUs: #90 #91 #92 ... #177 #178 #179 node #1, CPUs: #270 #271 #272 ... #357 #358 #359 I suspect that this is because all Local APIC MADT entries are now processed before all X2APIC MADT entries, whereas they may have been interleaved before. TBH, I'm not sure that there is actually anything wrong with the new numbering scheme. The topology is reported correctly (e.g. in /sys/devices/system/cpu/cpu0/topology/thread_siblings_list). Yet, the new enumeration does seem to contradict user expectations. Thanks, --jim
