SchOS/ldls/kpi/pci.c
Nexoniarz 071499a378 ldls: a second driver, and everything it found
e1000 -- Linux's, unmodified -- probes, opens, gets an address by DHCP and
carries ping traffic in a SchOS process. The kernel was driving that NIC and
gave it up, which makes it the first real handover as well as the second
driver.

None of what it needed was a missing feature. Every one was a call that
succeeded, returned zero and did nothing:

  _find_next_bit          "bit 0 is set"          -- for_each_cpu never ended
  netif_carrier_off       carrier already up      -- TCTL.EN never set
  __x86_indirect_thunk_r8 the jump discarded      -- 255 rx buffers, none made
  dma_map_page_attrs      bus address 0           -- NIC aimed at physical 0

so the rule from memcpy is now general: a stub returning zero is unsafe
wherever zero is a meaningful answer.

The loader grew ko_fixup_percpu(). The stack guard turned out to be one
per-CPU variable among several; `current` cannot be defused because the driver
dereferences it. Taking the %gs off instead leaves the same opcode reading the
same width from storage the host supplies -- a per-CPU variable in a
single-threaded host is just a variable. Eight bytes become seven and a nop,
and the displacement is not always last, so there is now just enough of an x86
decoder in ko.c to know how long the instruction is.

ldls/kpi/thunk.c is fifteen lines of assembly and the highest-value file here:
retpolines are control flow, and a stub cannot stand in for a jump.

The kernel side:

  - ioperm.c, the TSS I/O permission bitmap. A driver's outl is an instruction,
    not a call, and e1000's reset path picks it deliberately. IOPL=3 would hand
    over the PIC and the disk with it, so it is per-port. Writing it found that
    the I/O map base had been written at offset 64 rather than 102, leaving the
    CPU checking ports 0-831 against the TSS's own RSP0.
  - claim masks the device's IRQ line: 450,000 a second, otherwise.
  - net_detach(), the missing half of net_attach_late(). An interface left in
    the stack keeps its route and sends every frame to a driver that refuses
    it, which looks exactly like a userspace driver that does not work.
  - e1000_release() gives up the interface, not just the hardware.

8139too still passes 4/4; ldls-check still passes 27; four CPUs, no panic.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VQ7Qx8E1q1M1bbCq8QLHS2
2026-09-14 15:20:07 +02:00

208 lines
6.5 KiB
C

// SPDX-License-Identifier: Apache-2.0
//
// The PCI half of the KPI. See ldls/kpi/kpi.h for the provenance rule.
//
// Every one of these lands on `/dev/pci`, which is the interface
// docs/plan/hybrid.md built and `vgpud` proved. That it fits a foreign driver
// without alteration is the point the hybrid plan was making when it insisted
// the interface be about *PCI devices* rather than about a driver model:
//
// "an interface whose first consumers are native drivers is a general
// mechanism that something foreign can use later, rather than a hole shaped
// like something foreign."
//
// This file is that claim being collected.
#include "types.h"
#include "user.h"
#include "fcntl.h"
#include "posix.h"
#include "pciuser.h"
#include "kpi.h"
// The PCI resource flag bits, which are ABI between the kernel and every
// driver written against it. Not layout, so BTF does not carry them.
#define LDLS_IORESOURCE_IO 0x00000100
#define LDLS_IORESOURCE_MEM 0x00000200
// The device is claimed by the host before the driver runs, because claiming
// is what enables memory space and bus mastering -- and a BAR read before that
// can legitimately be zero. So by the time a driver calls this, it has already
// happened, and the honest answer is success.
int
kpi_pci_enable_device(void *pdev)
{
(void)pdev;
return 0;
}
// Regions are not reserved against anyone: this process holds the whole
// device, exclusively, because /dev/pci gives out one device per process.
// There is nobody to arbitrate with.
int
kpi_pci_request_regions(void *pdev, const char *name)
{
(void)pdev; (void)name;
return 0;
}
void
kpi_pci_release_regions(void *pdev)
{
(void)pdev;
}
// Also already done, by PCIIOC_CLAIM, which sets PCI_CMD_MEMORY and
// PCI_CMD_BUS_MASTER together -- see do_claim() in kernel/dev/pciuser.c, and
// the comment there about why a driver that has to ask separately will forget.
void
kpi_pci_set_master(void *pdev)
{
(void)pdev;
}
int
kpi_pci_read_config_word(void *pdev, int where, unsigned short *v)
{
struct pciuser_cfg c;
(void)pdev;
c.index = (unsigned int)kpi.index;
c.offset = (unsigned int)where & ~3u;
c.value = 0;
if (ioctl(kpi.pcifd, PCIIOC_CFG, &c) < 0)
return -1;
if (v)
*v = (unsigned short)(c.value >> ((where & 2) * 8));
return 0;
}
// Refused, deliberately and loudly.
//
// /dev/pci has no config *write* at all, and that is the line that matters:
// a write is how a BAR gets moved out from under the kernel. A driver that
// needs one has found a real limit of this interface rather than a bug, and
// silently succeeding would turn that into a mystery somewhere else.
int
kpi_pci_write_config_word(void *pdev, int where, unsigned short v)
{
(void)pdev; (void)v;
kpi_log("kpi: the driver tried to WRITE config space at 0x");
kpi_loghex((unsigned long long)where, 2);
kpi_log(" -- /dev/pci does not allow that\n");
return -1;
}
// Map a BAR.
//
// The offset passed to mmap is the BAR's *physical address*, which is what
// makes the kernel able to check it: an offset that is not inside a BAR of a
// device this process holds is refused, and that check is the whole difference
// between /dev/pci and /dev/mem with extra steps.
void *
kpi_pci_iomap(void *pdev, int bar, unsigned long max)
{
unsigned long long len;
void *p;
(void)pdev;
if (bar < 0 || bar >= 6 || kpi.bar_base[bar] == 0) {
kpi_log("kpi: pci_iomap of BAR "); kpi_lognum(bar);
kpi_log(", which this device does not have\n");
return 0;
}
// An I/O-space BAR has no address to map. rtl8139 offers both and its driver
// prefers MMIO when it can get it; saying no here is what makes it take that
// path rather than issuing port instructions a process cannot execute.
if (kpi.bar_flags[bar] & PCIUSER_BAR_IO) {
kpi_log("kpi: BAR "); kpi_lognum(bar);
kpi_log(" is I/O space, which cannot be mapped\n");
return 0;
}
len = kpi.bar_size[bar];
if (max && max < len)
len = max;
p = mmap(0, len, PROT_READ | PROT_WRITE, MAP_SHARED, kpi.pcifd,
kpi.bar_base[bar]);
// Any small negative is an error, not an address.
//
// SchOS returns -errno from mmap rather than MAP_FAILED, so checking only
// for -1 lets every other failure through as a pointer. That is not a
// theoretical worry: this returned -EACCES, the driver took -13 as its
// register base, and the first write landed at 0x4e -- thirteen below where
// it meant to be, which is how the real fault below was eventually found.
if (p == 0 || ((long)p < 0 && (long)p > -4096)) {
kpi_log("kpi: the kernel refused to map BAR "); kpi_lognum(bar);
kpi_log(", error "); kpi_lognum(-(long)p); kpi_log("\n");
return 0;
}
kpi.io_base = (unsigned long)p;
kpi.io_len = len;
kpi_log(" kpi: mapped BAR "); kpi_lognum(bar);
kpi_log(" -- phys 0x"); kpi_loghex(kpi.bar_base[bar], 16);
kpi_log(", "); kpi_lognum((long long)len); kpi_log(" bytes\n");
return p;
}
// pci_ioremap_bar(pdev, bar) -- the same thing as pci_iomap with no cap. A
// second spelling, which is the sort of thing a second driver turns up.
void *
kpi_pci_ioremap_bar(void *pdev, int bar)
{
return kpi_pci_iomap(pdev, bar, 0);
}
// pci_select_bars(pdev, flags) -- which BARs match a resource kind.
//
// A driver asks this before requesting regions, so the answer decides which
// BARs it thinks exist. Reporting one that is not there gets it refused later;
// omitting one it needs gets it to give up here.
int
kpi_pci_select_bars(void *pdev, unsigned long flags)
{
int bars = 0, b;
(void)pdev;
for (b = 0; b < 6; b++) {
if (!kpi.bar_base[b])
continue;
if ((kpi.bar_flags[b] & PCIUSER_BAR_IO) && (flags & LDLS_IORESOURCE_IO))
bars |= 1 << b;
if ((kpi.bar_flags[b] & PCIUSER_BAR_MMIO) && (flags & LDLS_IORESOURCE_MEM))
bars |= 1 << b;
}
return bars;
}
int
kpi_pci_request_selected_regions(void *pdev, int bars, const char *name)
{
(void)pdev; (void)bars; (void)name;
return 0; // nobody to arbitrate with; see above
}
void
kpi_pci_release_selected_regions(void *pdev, int bars)
{
(void)pdev; (void)bars;
}
// Saving and restoring config space is what a driver does around a reset it
// expects to clear the BARs. LDLS never resets the device at that level and
// has no config write to restore with, so there is nothing to save -- and
// saying so by succeeding is right, because a driver that cannot save state
// refuses to probe at all.
int kpi_pci_save_state(void *pdev) { (void)pdev; return 0; }
void kpi_pci_restore_state(void *pdev) { (void)pdev; }
void
kpi_pci_iounmap(void *pdev, void *addr)
{
(void)pdev; (void)addr;
}