SchOS/kernel/arch/x86_64/acpi.c

// Find the processors, the way firmware from this century describes them.
//
// mpinit() walks the Intel MP table, which is from 1994. Modern firmware fills
// it in minimally or omits it: measured 2026-08-08, QEMU/SeaBIOS with -smp 4
// publishes a 208-byte MP table containing exactly *one* processor entry, then
// buses and interrupt assignments. Everything since has expected the operating
// system to read the ACPI MADT instead.
//
// The consequence was not that SMP was slow, it was that ncpu was 1 and SMP was
// silently never exercised. Machines whose firmware does publish a full MP
// table -- VMware, older hardware -- took the other path, started their APs,
// and triple faulted. So this file is a prerequisite for debugging that: until
// the CPUs are visible under QEMU there is no way to reproduce the crash except
// on the hardware it happens on.
//
// Written from the ACPI specification. No table layout here is copied from
// another kernel; they are all in chapter 5 of the published standard.

#include "types.h"
#include "defs.h"
#include "schboot.h"   // SCHBOOT_RSDP_ADDR: the loader/kernel hand-off
#include "param.h"
#include "memlayout.h"
#include "mmu.h"
#include "proc.h"
#include "x86.h"

extern struct cpu cpus[NCPU];
extern int ncpu;
extern uchar ioapicid;
extern volatile uint *lapic;

// The Root System Description Pointer. Found by signature, not by being told
// where it is -- it sits either in the Extended BIOS Data Area or in the BIOS
// read-only region, on a 16-byte boundary.
struct acpi_rsdp {
  char   sig[8];              // "RSD PTR "
  uint8  checksum;            // over the first 20 bytes
  char   oemid[6];
  uint8  revision;            // 0 = ACPI 1.0 (RSDT), 2+ = 2.0 (XSDT)
  uint32 rsdt_addr;
  // ACPI 2.0 and later only, and only valid when revision >= 2.
  uint32 length;
  uint64 xsdt_addr;
  uint8  ext_checksum;        // over `length` bytes
  uint8  reserved[3];
} __attribute__((packed));

// Every ACPI table starts with this.
struct acpi_sdt {
  char   sig[4];
  uint32 length;
  uint8  revision;
  uint8  checksum;
  char   oemid[6];
  char   oem_table_id[8];
  uint32 oem_revision;
  uint32 creator_id;
  uint32 creator_revision;
} __attribute__((packed));

// The Multiple APIC Description Table: signature "APIC".
struct acpi_madt {
  struct acpi_sdt h;
  uint32 lapic_addr;
  uint32 flags;               // bit 0: legacy 8259 PICs are present
} __attribute__((packed));

#define MADT_PCAT_COMPAT   1

// MADT entry types. Only these three are read; the rest are interrupt
// routing overrides that ioapic.c does not consult yet.
#define MADT_LAPIC          0
#define MADT_IOAPIC         1
#define MADT_LAPIC_OVERRIDE 5

struct madt_entry {
  uint8 type;
  uint8 length;
} __attribute__((packed));

struct madt_lapic {
  struct madt_entry e;
  uint8  acpi_id;
  uint8  apic_id;
  uint32 flags;               // bit 0: enabled, bit 1: online-capable
} __attribute__((packed));

#define LAPIC_ENABLED        1
#define LAPIC_ONLINE_CAPABLE 2

struct madt_ioapic {
  struct madt_entry e;
  uint8  id;
  uint8  reserved;
  uint32 addr;
  uint32 gsi_base;
} __attribute__((packed));

struct madt_lapic_override {
  struct madt_entry e;
  uint16 reserved;
  uint64 addr;
} __attribute__((packed));

static int
checksum_ok(const void *p, uint32 len)
{
  const uint8 *b = p;
  uint8 sum = 0;
  for (uint32 i = 0; i < len; i++)
    sum += b[i];
  return sum == 0;
}

// Scan a physical range for the RSDP signature. It is 16-byte aligned by
// specification, which makes this cheap enough to do twice.
static struct acpi_rsdp *
scan_rsdp(uint64 base, uint64 len)
{
  for (uint64 off = 0; off + sizeof(struct acpi_rsdp) <= len; off += 16) {
    struct acpi_rsdp *r = (struct acpi_rsdp *)P2V(base + off);
    if (memcmp(r->sig, "RSD PTR ", 8) != 0)
      continue;
    // Signature alone is not enough: the string can appear in unrelated data,
    // and following a bad pointer here means faulting before there is anything
    // to report the fault with.
    if (!checksum_ok(r, 20))
      continue;
    return r;
  }
  return 0;
}

static struct acpi_rsdp *
find_rsdp(void)
{
  // A UEFI loader hands the address over directly, because the scan below is a
  // BIOS convention that UEFI firmware need not honour and OVMF does not: the
  // kernel found no RSDP, so no MADT, so no IOAPIC, and nothing routed IRQ 4 --
  // it reached a shell prompt on polled serial output and then ignored every
  // key.
  //
  // Validated, never trusted. The BIOS path does not write this slot, so there
  // it holds whatever happened to be in memory, and following a wild pointer
  // faults before there is anything to report the fault with. The bound is what
  // the direct map covers; the signature and checksum below do the rest.
  //
  // Deliberately *not* requiring 16-byte alignment. That rule applies to
  // searching the BIOS area, not to a pointer handed over by firmware -- OVMF's
  // sits at 0x3f77e014, and demanding alignment threw away the very address
  // this code was added to use.
  uint64 handed = *(uint64 *)P2V(SCHBOOT_RSDP_ADDR);
  if (handed && handed < 0x100000000ULL) {
    struct acpi_rsdp *r = (struct acpi_rsdp *)P2V(handed);
    if (memcmp(r->sig, "RSD PTR ", 8) == 0 && checksum_ok(r, 20))
      return r;
  }

  // The EBDA, whose segment address the BIOS leaves at 0x40E.
  uint16 ebda_seg = *(uint16 *)P2V(0x40E);
  if (ebda_seg) {
    struct acpi_rsdp *r = scan_rsdp((uint64)ebda_seg << 4, 1024);
    if (r)
      return r;
  }
  return scan_rsdp(0xE0000, 0x20000);
}

// Walk the RSDT or XSDT looking for a table by signature.
//
// Two nearly identical loops rather than one clever one: the only difference is
// the pointer width -- 32-bit entries in the RSDT, 64-bit in the XSDT -- and
// reading a 64-bit pointer out of a 32-bit array is exactly the sort of bug
// that produces a plausible address pointing at nothing.
static struct acpi_sdt *
find_table(struct acpi_rsdp *rsdp, const char *sig)
{
  struct acpi_sdt *root;
  int wide;

  if (rsdp->revision >= 2 && rsdp->xsdt_addr) {
    root = (struct acpi_sdt *)P2V(rsdp->xsdt_addr);
    wide = 1;
  } else {
    root = (struct acpi_sdt *)P2V((uint64)rsdp->rsdt_addr);
    wide = 0;
  }

  if (!checksum_ok(root, root->length))
    return 0;

  uint32 n = (root->length - sizeof(struct acpi_sdt)) / (wide ? 8 : 4);
  for (uint32 i = 0; i < n; i++) {
    uint64 pa;
    if (wide) {
      // Byte-wise, because the array is packed and 8-byte entries in it are
      // not guaranteed to be 8-byte aligned.
      uint8 *p = (uint8 *)(root + 1) + i * 8;
      pa = 0;
      for (int b = 7; b >= 0; b--)
        pa = (pa << 8) | p[b];
    } else {
      pa = ((uint32 *)(root + 1))[i];
    }
    struct acpi_sdt *t = (struct acpi_sdt *)P2V(pa);
    if (memcmp(t->sig, sig, 4) == 0)
      return t;
  }
  return 0;
}

// Returns 1 if the processors were enumerated, 0 to fall back to mpinit().
int
acpi_init(void)
{
  struct acpi_rsdp *rsdp = find_rsdp();
  if (!rsdp)
    return 0;

  struct acpi_madt *madt = (struct acpi_madt *)find_table(rsdp, "APIC");
  if (!madt || !checksum_ok(madt, madt->h.length))
    return 0;

  lapic = (uint *)P2V((uint64)madt->lapic_addr);

  uint8 *p = (uint8 *)madt + sizeof(struct acpi_madt);
  uint8 *e = (uint8 *)madt + madt->h.length;
  int found_ioapic = 0;

  while (p + sizeof(struct madt_entry) <= e) {
    struct madt_entry *ent = (struct madt_entry *)p;
    // A zero length would spin here forever. Firmware tables are not always
    // well formed, and an infinite loop this early is indistinguishable from a
    // hang with no output at all.
    if (ent->length < sizeof(struct madt_entry))
      break;

    switch (ent->type) {
    case MADT_LAPIC: {
      struct madt_lapic *l = (struct madt_lapic *)p;
      // Skip processors that are neither enabled nor online-capable: those are
      // sockets the firmware is describing but not offering.
      if (!(l->flags & (LAPIC_ENABLED | LAPIC_ONLINE_CAPABLE)))
        break;
      if (ncpu < NCPU) {
        cpus[ncpu].apicid = l->apic_id;
        ncpu++;
      }
      break;
    }
    case MADT_IOAPIC: {
      struct madt_ioapic *io = (struct madt_ioapic *)p;
      // The first one only. Multi-I/O-APIC routing needs the interrupt source
      // overrides too, which nothing here reads yet, so claiming to support it
      // would be worse than not.
      if (!found_ioapic) {
        ioapicid = io->id;
        found_ioapic = 1;
      }
      break;
    }
    case MADT_LAPIC_OVERRIDE: {
      struct madt_lapic_override *o = (struct madt_lapic_override *)p;
      lapic = (uint *)P2V(o->addr);
      break;
    }
    }
    p += ent->length;
  }

  if (ncpu == 0)
    return 0;

  // The 8259s are still there and still wired to the CPU. On a system with an
  // IMCR, this switches interrupt delivery from the PICs to the APIC; on one
  // without, the ports are unclaimed and the writes go nowhere. mpinit() does
  // the same thing guarded by the MP table's imcrp flag.
  if (madt->flags & MADT_PCAT_COMPAT) {
    outb(0x22, 0x70);
    outb(0x23, inb(0x23) | 1);
  }

  cprintf("acpi: %d cpu%s, lapic at %p, ioapic id %d\n",
          ncpu, ncpu == 1 ? "" : "s", lapic, ioapicid);
  return 1;
}