Files
lix/src/libstore/platform/darwin.cc
T
Artemis TosiniandArtemis Tosini c03de0df62 gc: Find roots using libproc on Darwin
Previously, the garbage collector found runtime roots on Darwin by
shelling out to `lsof -n -w -F n` then parsing the result.
However, this requires an lsof binary and can be extremely slow.

The official Apple lsof returns in a reasonable amount of time,
about 250ms in my tests, but the lsof packaged in nixpkgs is quite slow,
taking about 40 seconds to run the command.

Using libproc directly is about the same speed as Apple lsof,
and allows us to reënable several tests that were disabled on Darwin.

Change-Id: Ifa0adda7984e13c15535693baba835aae79a3577
2024-04-25 23:24:21 -04:00

224 lines
9.0 KiB
C++

#include "gc-store.hh"
#include "signals.hh"
#include "platform/darwin.hh"
#include "regex.hh"
#include <sys/proc_info.h>
#include <sys/sysctl.h>
#include <libproc.h>
#include <regex>
namespace nix {
void DarwinLocalStore::findPlatformRoots(UncheckedRoots & unchecked)
{
auto storePathRegex = regex::storePathRegex(storeDir);
std::vector<int> pids;
int pidBufSize = 1;
while (pidBufSize > pids.size() * sizeof(int)) {
// Reserve some extra size so we don't fail too much
pids.resize((pidBufSize + pidBufSize / 8) / sizeof(int));
pidBufSize = proc_listpids(PROC_ALL_PIDS, 0, pids.data(), pids.size() * sizeof(int));
if (pidBufSize <= 0) {
throw SysError("Listing PIDs");
}
}
pids.resize(pidBufSize / sizeof(int));
for (auto pid : pids) {
// It doesn't make sense to ask about the kernel
if (pid == 0) {
continue;
}
try {
// Process cwd/root directory
struct proc_vnodepathinfo vnodeInfo;
if (proc_pidinfo(pid, PROC_PIDVNODEPATHINFO, 0, &vnodeInfo, sizeof(vnodeInfo)) <= 0) {
throw SysError("Getting pid %1% working directory", pid);
}
unchecked[std::string(vnodeInfo.pvi_cdir.vip_path)].emplace(fmt("{libproc/%d/cwd}", pid)
);
unchecked[std::string(vnodeInfo.pvi_rdir.vip_path)].emplace(
fmt("{libproc/%d/rootdir}", pid)
);
// File descriptors
std::vector<struct proc_fdinfo> fds;
int fdBufSize = 1;
while (fdBufSize > fds.size() * sizeof(struct proc_fdinfo)) {
// Reserve some extra size so we don't fail too much
fds.resize((fdBufSize + fdBufSize / 8) / sizeof(struct proc_fdinfo));
fdBufSize = proc_pidinfo(
pid, PROC_PIDLISTFDS, 0, fds.data(), fds.size() * sizeof(struct proc_fdinfo)
);
if (fdBufSize <= 0) {
throw SysError("Listing pid %1% file descriptors", pid);
}
}
fds.resize(fdBufSize / sizeof(struct proc_fdinfo));
for (auto fd : fds) {
// By definition, only a vnode is on the filesystem
if (fd.proc_fdtype != PROX_FDTYPE_VNODE) {
continue;
}
struct vnode_fdinfowithpath fdInfo;
if (proc_pidfdinfo(
pid, fd.proc_fd, PROC_PIDFDVNODEPATHINFO, &fdInfo, sizeof(fdInfo)
)
<= 0)
{
// They probably just closed this fd, no need to cancel looking at ranges and
// arguments
if (errno == EBADF) {
continue;
}
throw SysError("Getting pid %1% fd %2% path", pid, fd.proc_fd);
}
unchecked[std::string(fdInfo.pvip.vip_path)].emplace(
fmt("{libproc/%d/fd/%d}", pid, fd.proc_fd)
);
}
// Regions (e.g. mmapped files, executables, shared libraries)
uint64_t nextAddr = 0;
while (true) {
// Seriously, what are you doing XNU?
// There's 3 flavors of PROC_PIDREGIONPATHINFO:
// * PROC_PIDREGIONPATHINFO includes all regions
// * PROC_PIDREGIONPATHINFO2 includes regions backed by a vnode
// * PROC_PIDREGIONPATHINFO3 includes regions backed by a vnode on a specified
// filesystem Only PROC_PIDREGIONPATHINFO is documented. Unfortunately, using it
// would make finding gcroots take about 100x as long and tests would fail from
// timeout. According to the Frida source code, PROC_PIDREGIONPATHINFO2 has been
// available since XNU 2782.1.97 in OS X 10.10
//
// 22 means PROC_PIDREGIONPATHINFO2
struct proc_regionwithpathinfo regionInfo;
if (proc_pidinfo(pid, 22, nextAddr, &regionInfo, sizeof(regionInfo)) <= 0) {
// PROC_PIDREGIONPATHINFO signals we're done with an error,
// so we're expected to hit this once per process
if (errno == ESRCH || errno == EINVAL) {
break;
}
throw SysError("Getting pid %1% region path", pid);
}
unchecked[std::string(regionInfo.prp_vip.vip_path)].emplace(
fmt("{libproc/%d/region}", pid)
);
nextAddr = regionInfo.prp_prinfo.pri_address + regionInfo.prp_prinfo.pri_size;
}
// Arguments and environment variables
// We can't read environment variables of binaries with entitlements unless
// nix has the `com.apple.private.read-environment-variables` entitlement or SIP is off
// We can read arguments for all applications though.
// Yes, it's a sysctl, the proc_info and sysctl APIs are mostly similar,
// but both have exclusive capabilities
int sysctlName[3] = {CTL_KERN, KERN_PROCARGS2, pid};
size_t argsSize = 0;
if (sysctl(sysctlName, 3, nullptr, &argsSize, nullptr, 0) < 0) {
throw SysError("Reading pid %1% arguments", pid);
}
std::vector<char> args(argsSize);
if (sysctl(sysctlName, 3, args.data(), &argsSize, nullptr, 0) < 0) {
throw SysError("Reading pid %1% arguments", pid);
}
if (argsSize < args.size()) {
args.resize(argsSize);
}
// We have these perfectly nice arguments, but have to ignore them because
// otherwise we'd see arguments to nix-store commands and
// `nix-store --delete /nix/store/whatever` would always fail
// First 4 bytes are an int of argc.
if (args.size() < sizeof(int)) {
continue;
}
auto argc = reinterpret_cast<int *>(args.data())[0];
auto argsIter = args.begin();
std::advance(argsIter, sizeof(int));
// Executable then argc args, each separated by some number of null bytes
for (int i = 0; argsIter != args.end() && i < argc + 1; i++) {
argsIter = std::find(argsIter, args.end(), '\0');
argsIter = std::find_if(argsIter, args.end(), [](char ch) { return ch != '\0'; });
}
if (argsIter != args.end()) {
auto env_end = std::sregex_iterator{};
for (auto i = std::sregex_iterator{argsIter, args.end(), storePathRegex};
i != env_end;
++i)
{
unchecked[i->str()].emplace(fmt("{libproc/%d/environ}", pid));
}
};
// Per-thread working directories
struct proc_taskallinfo taskAllInfo;
if (proc_pidinfo(pid, PROC_PIDTASKALLINFO, 0, &taskAllInfo, sizeof(taskAllInfo)) <= 0) {
throw SysError("Reading pid %1% tasks", pid);
}
// If the process doesn't have the per-thread cwd flag then we already have the
// process-wide cwd from PROC_PIDVNODEPATHINFO
if (taskAllInfo.pbsd.pbi_flags & PROC_FLAG_THCWD) {
std::vector<uint64_t> tids(taskAllInfo.ptinfo.pti_threadnum);
int tidBufSize = proc_pidinfo(
pid, PROC_PIDLISTTHREADS, 0, tids.data(), tids.size() * sizeof(uint64_t)
);
if (tidBufSize <= 0) {
throw SysError("Listing pid %1% threads", pid);
}
for (auto tid : tids) {
struct proc_threadwithpathinfo threadPathInfo;
if (proc_pidinfo(
pid,
PROC_PIDTHREADPATHINFO,
tid,
&threadPathInfo,
sizeof(threadPathInfo)
)
<= 0)
{
throw SysError("Reading pid %1% thread %2% cwd", pid, tid);
}
unchecked[std::string(threadPathInfo.pvip.vip_path)].emplace(
fmt("{libproc/%d/thread/%d/cwd}", pid, tid)
);
}
}
} catch (SysError & e) {
// ENOENT/ESRCH: Process no longer exists (proc_info)
// EINVAL: Process no longer exists (sysctl)
// EACCESS/EPERM: We don't have permission to read this field (proc_info)
// EIO: Kernel failed to read from target process memory during KERN_PROCARGS2 (sysctl)
if (errno == ENOENT || errno == ESRCH || errno == EINVAL || errno == EACCES
|| errno == EPERM || errno == EIO)
{
continue;
}
throw;
}
}
}
}