On Wed, Mar 03, 2021 at 05:09:16PM +0100, Igor Mammedov wrote: > On Wed, 3 Mar 2021 16:03:36 +0100 > Laszlo Ersek <ler...@redhat.com> wrote: > > > On 03/02/21 19:43, David Hildenbrand wrote: > > > > > We are dealing with different blobs here (tables_blob vs. cmd_blob). > > > > OK, thanks -- this was the important bit I was missing. Over time I've > > lost track of the actual set of fw_cfg blobs that QEMU exposes, for the > > purposes of the ACPI linker/loader. > > > > I've looked up the acpi_add_rom_blob() calls in "hw/i386/acpi-build.c" > > and "hw/arm/virt-acpi-build.c": > > > > hw name max_size > > notes > > ------- ------------------------------------------- > > ------------------------------------ ------ > > > > virt ACPI_BUILD_TABLE_FILE ("etc/acpi/tables") > > ACPI_BUILD_TABLE_MAX_SIZE (0x200000) n/a > > virt ACPI_BUILD_LOADER_FILE ("etc/table-loader") 0 > > n/a > > virt ACPI_BUILD_RSDP_FILE ("etc/acpi/rsdp") 0 > > simply modeled on i386 (below) > > > > i386 ACPI_BUILD_TABLE_FILE ("etc/acpi/tables") > > ACPI_BUILD_TABLE_MAX_SIZE (0x200000) n/a > > i386 ACPI_BUILD_LOADER_FILE ("etc/table-loader") 0 > > n/a > > i386 ACPI_BUILD_RSDP_FILE ("etc/acpi/rsdp") 0 > > d70414a5788c, 358774d780ee8 > > > > microvm ACPI_BUILD_TABLE_FILE ("etc/acpi/tables") > > ACPI_BUILD_TABLE_MAX_SIZE (0x200000) n/a > > microvm "etc/table-loader" 0 > > no macro for name??? > > microvm ACPI_BUILD_RSDP_FILE ("etc/acpi/rsdp") 0 > > simply modeled on i386 (above) > > > > (I notice there are some other (optional) fw_cfg blobs too, related TPM, > > vmgenid, nvdimm etc, using fw_cfg_add_file() rather than > > acpi_add_rom_blob() -- so those are immutable (never regenerated). I > > definitely needed this reminder...) > > most of them are just guest RAM reservations (guest/hose exchange buffer) > and "etc/tpm/config" seems to immutable for specific configuration > > > > So, my observations: > > > > (1) microvm open-codes "etc/table-loader", rather than using the macro > > ACPI_BUILD_LOADER_FILE. > > > > The proposed patch corrects it, which I welcome per se. However, it > > should arguably be a separate patch. I found it distracting, in spite of > > the commit message highlighting it. I don't insist though, I'm > > admittedly rusty on this code. > > > > > > (2) The proposed patch sets "max_size" to ACPI_BUILD_LOADER_MAX_SIZE for > > each ACPI_BUILD_LOADER_FILE. Makes sense, upon constructing / reviewing > > the above table. > > > > (I'm no longer sure if tweaking the alignment were the preferable path > > forward.) > > > > Either way, I'd request including the above table in the commit message. > > (Maybe drop the "notes" column.) > > > > > > (3) The above 9 invocations are *all* of the acpi_add_rom_blob() > > invocations. I find the interface brittle. It's not helpful to have so > > many macros for the names and the max sizes. We should have a table with > > three entries and -- minimally -- two columns, specifying name and > > max_size -- possibly some more call arguments, if such can be extracted. > > We should also have an enum type for selecting a row in this table, and > > then acpi_add_rom_blob() should be called with an enum constant. > > > > Of course, talk is cheap. :) > > > > > > (4) When do we plan to introduce a nonzero "max_size" for > > ACPI_BUILD_RSDP_FILE ("etc/acpi/rsdp")? > > > > Is the current zero value a time bomb? > > it's not likely to go over 4k, but if we enforce max_size!=0 we may set it 4k, > which it's aligned to anyways.
Right. BTW there is an alternative I did not think of earlier. Lots of tables are actually fixed. We currently let guest calculate the checksum for all tables but that is not a must. We could prefill the checksum for most of them and cut the size by almost half. This fixes the issues in a way that seems cleaner to me as it migrates both ways for all configs and saves some resources. I'm not against making it resizeable too though. > > > Put differently: acpi_add_rom_blob() should be *impossible* to call with > > "max_size=0", arguably. *Whenever* we call acpi_add_rom_blob(), we do > > that because the blob is resizable (mutable) -- but that also means we > > should have a safety margin, does it not? So calling acpi_add_rom_blob() > > with "max_size=0" looks self-contradictory. > > main use-case for using acpi_add_rom_blob() is for mutable blobs, > so that all these blobs were transferred during migration to the destination, > to ensure that guest sees consistent data set (from source instead of mix of > source/dst blobs). > > Resize came later on, when we got sick of ad-hock (align)/size bumping of > "etc/acpi/tables" in configurations where size was on verge of crossing > border to the next aligned size and related knobs to keep that mess > migratable. > > > > > FWIW, this could be covered by the table proposed in point (3). > > > > > > In total, I don't disagree with the patch (beyond the fact that the new > > macro's value doesn't match the commit message), functionally speaking. > > However, wrt. readability, I think the patch further complicates the > > code. I'd suggest five patches: > > > > #1 -- use "etc/table-loader" via the proper macro name in "microvm", > > > > #2 -- rework acpi_add_rom_blob() for using a table of constants + an > > enum type, > > > > #3 -- bump the "max_size" field for ACPI_BUILD_LOADER_FILE, for the > > current symptom, > > > > #4 -- set a nonzero "max_size" for the remaining ACPI_BUILD_RSDP_FILE, > > for "future-proofing", > > > > #5 -- in the new acpi_add_rom_blob() implementation, taking the enum, > > assert(max_size != 0). > > > > (I haven't thought through what this would mean for migration, forward > > or backward; I'm just brain-storming.) > > > > Thanks > > Laszlo