ceph-mirror/src/rgw/rgw_process.cc
Jesse F. Williamson fa53c743f7 rgw: group Lua request hook improvements
Run postauth hooks through the bytecode-capable Lua read path and keep the Lua request metadata adapters aligned with transparent HTTP argument lookup.

The regression test verifies removed scripts stop affecting requests; the benchmark path showed a 500-statement hook dropping from 1961-1998ms to 372-393ms over 20K load/execute iterations, roughly an 80-81% reduction.

Assisted-by: Codex:GPT-5

Signed-off-by: Jesse F. Williamson <jfw@ibm.com>
2026-07-29 09:15:19 -07:00

589 lines
18 KiB
C++

// -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:nil -*-
// vim: ts=8 sw=2 sts=2 expandtab ft=cpp
#include "common/errno.h"
#include "common/Throttle.h"
#include "common/WorkQueue.h"
#include "include/scope_guard.h"
#include <utility>
#include "rgw_auth_registry.h"
#include "rgw_dmclock_scheduler.h"
#include "rgw_rest.h"
#include "rgw_frontend.h"
#include "rgw_request.h"
#include "rgw_process.h"
#include "rgw_loadgen.h"
#include "rgw_client_io.h"
#include "rgw_opa.h"
#include "rgw_perf_counters.h"
#include "rgw_lua.h"
#include "rgw_lua_request.h"
#include "rgw_tracer.h"
#include "rgw_ratelimit.h"
#include "rgw_bucket_logging.h"
#include "services/svc_zone_utils.h"
#define dout_subsys ceph_subsys_rgw
using namespace std;
using rgw::dmclock::Scheduler;
void RGWProcess::RGWWQ::_dump_queue()
{
if (!g_conf()->subsys.should_gather<ceph_subsys_rgw, 20>()) {
return;
}
deque<RGWRequest *>::iterator iter;
if (process->m_req_queue.empty()) {
dout(20) << "RGWWQ: empty" << dendl;
return;
}
dout(20) << "RGWWQ:" << dendl;
for (iter = process->m_req_queue.begin();
iter != process->m_req_queue.end(); ++iter) {
dout(20) << "req: " << hex << *iter << dec << dendl;
}
} /* RGWProcess::RGWWQ::_dump_queue */
auto schedule_request(Scheduler *scheduler, req_state *s, RGWOp *op)
{
using rgw::dmclock::SchedulerCompleter;
if (!scheduler)
return std::make_pair(0,SchedulerCompleter{});
const auto client = op->dmclock_client();
const auto cost = op->dmclock_cost();
if (s->cct->_conf->subsys.should_gather(ceph_subsys_rgw, 10)) {
ldpp_dout(op,10) << "scheduling with "
<< s->cct->_conf.get_val<std::string>("rgw_scheduler_type")
<< " client=" << static_cast<int>(client)
<< " cost=" << cost << dendl;
}
return scheduler->schedule_request(client, {},
req_state::Clock::to_double(s->time),
cost,
s->yield);
}
bool RGWProcess::RGWWQ::_enqueue(RGWRequest* req) {
process->m_req_queue.push_back(req);
perfcounter->inc(l_rgw_qlen);
dout(20) << "enqueued request req=" << hex << req << dec << dendl;
_dump_queue();
return true;
}
RGWRequest* RGWProcess::RGWWQ::_dequeue() {
if (process->m_req_queue.empty())
return NULL;
RGWRequest *req = process->m_req_queue.front();
process->m_req_queue.pop_front();
dout(20) << "dequeued request req=" << hex << req << dec << dendl;
_dump_queue();
perfcounter->inc(l_rgw_qlen, -1);
return req;
}
void RGWProcess::RGWWQ::_process(RGWRequest *req, ThreadPool::TPHandle &) {
perfcounter->inc(l_rgw_qactive);
process->handle_request(this, req);
process->req_throttle.put(1);
perfcounter->inc(l_rgw_qactive, -1);
}
bool rate_limit(rgw::sal::Driver* driver, req_state* s) {
// we dont want to limit health check or system or admin requests
if ((s->op_type == RGW_OP_GET_HEALTH_CHECK) || s->system_request)
return false;
std::string userfind;
RGWRateLimitInfo global_user;
RGWRateLimitInfo global_bucket;
RGWRateLimitInfo global_anon;
RGWRateLimitInfo* bucket_ratelimit;
RGWRateLimitInfo* user_ratelimit;
driver->get_ratelimit(global_bucket, global_user, global_anon);
bucket_ratelimit = &global_bucket;
user_ratelimit = &global_user;
s->user->get_id().to_str(userfind);
userfind = "u" + userfind;
s->ratelimit_user_name = userfind;
std::string bucketfind = !rgw::sal::Bucket::empty(s->bucket.get()) ? "b" + s->bucket->get_marker() : "";
s->ratelimit_bucket_marker = bucketfind;
const char *method = s->info.method;
bool is_sts_user = (s->auth.identity && s->auth.identity->get_identity_type() == TYPE_ROLE);
if (is_sts_user) {
ldpp_dout(s, 21) << "STS user detected: uid=" << std::quoted(s->user->get_id().to_str()) << dendl;
auto op_ret = s->user->read_attrs(s, s->yield);
if (op_ret < 0) {
ldpp_dout(s, 0) << "checking rate_limit: uid=" << std::quoted(s->user->get_id().to_str()) << " failed to read user attrs" << dendl;
}
}
auto iter = s->user->get_attrs().find(RGW_ATTR_RATELIMIT);
if(iter != s->user->get_attrs().end()) {
try {
RGWRateLimitInfo user_ratelimit_temp;
bufferlist& bl = iter->second;
auto biter = bl.cbegin();
decode(user_ratelimit_temp, biter);
// override global rate limiting only if local rate limiting is enabled
if (user_ratelimit_temp.enabled)
*user_ratelimit = user_ratelimit_temp;
} catch (buffer::error& err) {
ldpp_dout(s, 0) << "ERROR: failed to decode rate limit" << dendl;
return -EIO;
}
} else {
ldpp_dout(s, 21) << "checking rate_limit: uid=" << std::quoted(s->user->get_id().to_str()) << " does not have a rate limit attribute" << dendl;
}
if (s->user->get_id().id == RGW_USER_ANON_ID && global_anon.enabled) {
*user_ratelimit = global_anon;
}
int64_t limit_bucket = 0;
int64_t limit_user = s->ratelimit_data->should_rate_limit(method, s->ratelimit_user_name, s->time, user_ratelimit, s->info.request_params);
if(!rgw::sal::Bucket::empty(s->bucket.get()))
{
iter = s->bucket->get_attrs().find(RGW_ATTR_RATELIMIT);
if(iter != s->bucket->get_attrs().end()) {
try {
RGWRateLimitInfo bucket_ratelimit_temp;
bufferlist& bl = iter->second;
auto biter = bl.cbegin();
decode(bucket_ratelimit_temp, biter);
// override global rate limiting only if local rate limiting is enabled
if (bucket_ratelimit_temp.enabled)
*bucket_ratelimit = bucket_ratelimit_temp;
} catch (buffer::error& err) {
ldpp_dout(s, 0) << "ERROR: failed to decode rate limit" << dendl;
return -EIO;
}
}
if (!limit_user) {
limit_bucket = s->ratelimit_data->should_rate_limit(method, s->ratelimit_bucket_marker, s->time, bucket_ratelimit, s->info.request_params);
}
}
if(limit_bucket && !limit_user) {
s->ratelimit_data->giveback_tokens(method, s->ratelimit_user_name, s->info.request_params, user_ratelimit);
}
s->user_ratelimit = *user_ratelimit;
s->bucket_ratelimit = *bucket_ratelimit;
int64_t delay = limit_user ? limit_user : limit_bucket;
if (delay > 0) {
s->ratelimit_retry_after = delay;
}
return (limit_user || limit_bucket);
}
static int execute_post_auth_lua_script(RGWOp * const op, req_state * const s)
{
if (op->get_type() == RGW_OP_GET_HEALTH_CHECK) {
return 0;
}
const auto [script, rc] = rgw::lua::read_script_or_bytecode(
s, s->penv.lua.manager.get(), s->bucket_tenant, s->yield,
rgw::lua::context::postAuth);
if (rc == -ENOENT) {
return 0;
}
if (rc < 0) {
ldpp_dout(op, 5) <<
"WARNING: failed to execute post authorization script. "
"error: " << rc << dendl;
return 0;
}
int script_return_code = 0;
const auto execute_rc = rgw::lua::request::execute(
s->penv.rest, s->penv.olog.get(), s, op, script, script_return_code);
if (execute_rc < 0) {
ldpp_dout(op, 5) <<
"WARNING: failed to execute post authorization script. "
"error: " << execute_rc << dendl;
}
if (script_return_code == -EPERM) {
return script_return_code;
}
return 0;
}
int rgw_process_authenticated(RGWHandler_REST * const handler,
RGWOp *& op,
RGWRequest * const req,
req_state * const s,
optional_yield y,
rgw::sal::Driver* driver,
const bool skip_retarget)
{
ldpp_dout(op, 2) << "init permissions" << dendl;
int ret = handler->init_permissions(op, y);
if (ret < 0) {
return ret;
}
/**
* Only some accesses support website mode, and website mode does NOT apply
* if you are using the REST endpoint either (ergo, no authenticated access)
*/
if (! skip_retarget) {
ldpp_dout(op, 2) << "recalculating target" << dendl;
ret = handler->retarget(op, &op, y);
if (ret < 0) {
return ret;
}
req->op = op;
} else {
ldpp_dout(op, 2) << "retargeting skipped because of SubOp mode" << dendl;
}
/* If necessary extract object ACL and put them into req_state. */
ldpp_dout(op, 2) << "reading permissions" << dendl;
ret = handler->read_permissions(op, y);
if (ret < 0) {
return ret;
}
ldpp_dout(op, 2) << "init op" << dendl;
ret = op->init_processing(y);
if (op->get_type() == RGW_OP_OPTIONS_CORS && ret == -EINVAL) {
ldpp_dout(op, 0) << "NOTICE: RGW_OP_OPTIONS_CORS shouldn't return -EINVAL in case we have a global CORS!" << dendl;
ret = 0;
}
if (ret < 0) {
return ret;
}
ldpp_dout(op, 2) << "verifying op mask" << dendl;
ret = op->verify_op_mask();
if (ret < 0) {
return ret;
}
/* Check if OPA is used to authorize requests */
if (s->cct->_conf->rgw_use_opa_authz) {
ret = rgw_opa_authorize(op, s);
if (ret < 0) {
return ret;
}
}
ldpp_dout(op, 2) << "verifying op permissions" << dendl;
{
auto span = tracing::rgw::tracer.add_span("verify_permission", s->trace);
std::swap(span, s->trace);
ret = op->verify_permission(y);
std::swap(span, s->trace);
}
if (ret == -EACCES || ret == -EPERM || ret == -ERR_AUTHORIZATION) {
// system requests may impersonate another user/role for permission checks
// so only rely on is_admin() to override permissions
if (s->auth.identity->is_admin()) {
dout(2) << "overriding permissions due to admin operation" << dendl;
} else {
return ret;
}
} else if (ret < 0) {
// other errors are not overridden as they might be invalid input
return ret;
}
ldpp_dout(op, 2) << "verifying op params" << dendl;
ret = op->verify_params();
if (ret < 0) {
return ret;
}
ldpp_dout(op, 2) << "pre-executing" << dendl;
op->pre_exec();
ldpp_dout(op, 2) << "check rate limiting" << dendl;
if (rate_limit(driver, s)) {
return -ERR_RATE_LIMITED;
}
ret = execute_post_auth_lua_script(op, s);
if (ret < 0) {
return ret;
}
ldpp_dout(op, 2) << "executing" << dendl;
{
auto span = tracing::rgw::tracer.add_span("execute", s->trace);
std::swap(span, s->trace);
op->execute(y);
std::swap(span, s->trace);
}
ldpp_dout(op, 2) << "completing" << dendl;
op->complete();
return 0;
}
int process_request(const RGWProcessEnv& penv,
RGWRequest* const req,
const std::string& frontend_prefix,
RGWRestfulIO* const client_io,
optional_yield yield,
rgw::dmclock::Scheduler *scheduler,
string* user,
ceph::coarse_real_clock::duration* latency,
int* http_ret)
{
int ret = client_io->init(g_ceph_context);
rgw::sal::Driver* driver = penv.driver;
auto trans_id = driver->zone_unique_trans_id(req->id);
dout(1) << "====== starting new request req=" << hex << req << dec
<< " request_id=" << trans_id << " =====" << dendl;
perfcounter->inc(l_rgw_req);
RGWEnv& rgw_env = client_io->get_env();
req_state rstate(g_ceph_context, penv, &rgw_env, req->id);
req_state *s = &rstate;
s->ratelimit_data = penv.ratelimiting->get_active();
std::unique_ptr<rgw::sal::User> u = driver->get_user(rgw_user());
s->set_user(u);
if (ret < 0) {
s->cio = client_io;
abort_early(s, nullptr, ret, nullptr, yield);
return ret;
}
s->req_id = driver->zone_unique_id(req->id);
s->trans_id = trans_id;
s->host_id = driver->get_host_id();
s->yield = yield;
RGWOp* op = nullptr;
int init_error = 0;
bool should_log = false;
RGWREST* rest = penv.rest;
RGWRESTMgr *mgr;
bool is_health_request = false;
RGWHandler_REST *handler = rest->get_handler(driver, s,
*penv.auth_registry,
frontend_prefix,
client_io, &mgr, &init_error);
// RAII handlers to keep these around during process_request():
auto put_handler_guard = make_scope_guard([rest, &handler] {
rest->put_handler(handler);
});
auto put_op_guard = make_scope_guard([&handler, &op] {
if (handler) {
handler->put_op(op);
}
});
rgw::dmclock::SchedulerCompleter c;
if (init_error != 0) {
abort_early(s, nullptr, init_error, nullptr, yield);
goto done;
}
ldpp_dout(s, 10) << "handler=" << typeid(*handler).name() << dendl;
should_log = mgr->get_logging();
ldpp_dout(s, 2) << "getting op " << s->op << dendl;
op = handler->get_op();
if (!op) {
abort_early(s, NULL, -ERR_METHOD_NOT_ALLOWED, handler, yield);
goto done;
}
std::tie(ret,c) = schedule_request(scheduler, s, op);
if (ret < 0) {
if (ret == -EAGAIN) {
ret = -ERR_RATE_LIMITED;
}
ldpp_dout(op,0) << "Scheduling request failed with " << ret << dendl;
abort_early(s, op, ret, handler, yield);
goto done;
}
req->op = op;
ldpp_dout(op, 10) << "op=" << typeid(*op).name() << " " << dendl;
s->op_type = op->get_type();
try {
ldpp_dout(op, 2) << "verifying requester" << dendl;
ret = op->verify_requester(*penv.auth_registry, yield);
if (ret < 0) {
dout(10) << "failed to authorize request" << dendl;
abort_early(s, op, ret, handler, yield);
goto done;
}
/* FIXME: remove this after switching all handlers to the new authentication
* infrastructure. */
if (nullptr == s->auth.identity) {
auto result = rgw::auth::transform_old_authinfo(
op, yield, driver, s->user.get());
if (!result) {
abort_early(s, op, result.error(), handler, yield);
goto done;
}
s->auth.identity = std::move(result).value();
}
ldpp_dout(op, 2) << "normalizing buckets and tenants" << dendl;
ret = handler->postauth_init(yield);
if (ret < 0) {
dout(10) << "failed to run post-auth init" << dendl;
abort_early(s, op, ret, handler, yield);
goto done;
}
if (s->user->get_info().suspended) {
dout(10) << "user is suspended, uid=" << s->user->get_id() << dendl;
abort_early(s, op, -ERR_USER_SUSPENDED, handler, yield);
goto done;
}
is_health_request = (op->get_type() == RGW_OP_GET_HEALTH_CHECK);
{
s->trace_enabled = tracing::rgw::tracer.is_enabled();
if (!is_health_request) {
auto [lua_script, rc] = rgw::lua::read_script_or_bytecode(s, penv.lua.manager.get(),
s->bucket_tenant, s->yield,
rgw::lua::context::preRequest);
if (rc == -ENOENT) {
// no script, nothing to do
} else if (rc < 0) {
ldpp_dout(op, 5) <<
"WARNING: failed to execute pre request script. "
"error: " << rc << dendl;
} else {
int script_return_code = 0;
rc = rgw::lua::request::execute(rest, penv.olog.get(), s, op, lua_script, script_return_code);
if (rc < 0) {
ldpp_dout(op, 5) <<
"WARNING: failed to execute pre request script. "
"error: " << rc << dendl;
}
if (script_return_code == -EPERM) {
abort_early(s, op, script_return_code, handler, yield);
goto done;
}
}
}
}
s->trace = tracing::rgw::tracer.start_trace(op->name(), s->trace_enabled);
s->trace->SetAttribute(tracing::rgw::TRANS_ID, s->trans_id);
ret = rgw_process_authenticated(handler, op, req, s, yield, driver);
if (ret < 0) {
abort_early(s, op, ret, handler, yield);
goto done;
}
} catch (const ceph::crypto::DigestException& e) {
dout(0) << "authentication failed" << e.what() << dendl;
abort_early(s, op, -ERR_INVALID_SECRET_KEY, handler, yield);
}
done:
if (op) {
if (s->trace) {
s->trace->SetAttribute(tracing::rgw::OP_RESULT, op->get_ret());
s->trace->SetAttribute(tracing::rgw::HOST_ID, driver->get_host_id());
if (!rgw::sal::User::empty(s->user)) {
s->trace->SetAttribute(tracing::rgw::USER_ID, s->user->get_id().id);
}
if (!rgw::sal::Bucket::empty(s->bucket)) {
s->trace->SetAttribute(tracing::rgw::BUCKET_NAME, s->bucket->get_name());
}
if (!rgw::sal::Object::empty(s->object)) {
s->trace->SetAttribute(tracing::rgw::OBJECT_NAME, s->object->get_name());
}
}
if (!is_health_request) {
auto [lua_script, rc] = rgw::lua::read_script_or_bytecode(s, penv.lua.manager.get(),
s->bucket_tenant, s->yield,
rgw::lua::context::postRequest);
if (rc == -ENOENT) {
// no script, nothing to do
} else if (rc < 0) {
ldpp_dout(op, 5) <<
"WARNING: failed to read post request script. "
"error: " << rc << dendl;
} else {
rc = rgw::lua::request::execute(rest, penv.olog.get(), s, op, lua_script);
if (rc < 0) {
ldpp_dout(op, 5) <<
"WARNING: failed to execute post request script. "
"error: " << rc << dendl;
}
}
}
}
try {
client_io->complete_request();
} catch (rgw::io::Exception& e) {
dout(0) << "ERROR: client_io->complete_request() returned "
<< e.what() << dendl;
perfcounter->inc(l_rgw_qlen, -1);
perfcounter->inc(l_rgw_qactive, -1);
}
if (should_log) {
rgw_log_op(rest, s, op, penv.olog.get());
}
if (op && op->always_do_bucket_logging()) {
std::ignore = rgw::bucketlogging::log_record(driver,
rgw::bucketlogging::LoggingType::Standard,
s->object.get(),
s,
op->canonical_name(),
"",
(s->src_object ? s->src_object->get_size() : (s->object ? s->object->get_size() : 0)),
op,
yield,
true,
false);
}
if (http_ret != nullptr) {
*http_ret = s->err.http_ret;
}
int op_ret = 0;
if (user && !rgw::sal::User::empty(s->user.get())) {
*user = s->user->get_id().to_str();
}
if (op) {
op_ret = op->get_ret();
ldpp_dout(op, 2) << "op status=" << op_ret << dendl;
ldpp_dout(op, 2) << "http status=" << s->err.http_ret << dendl;
} else {
ldpp_dout(s, 2) << "http status=" << s->err.http_ret << dendl;
}
const auto lat = s->time_elapsed();
if (latency) {
*latency = lat;
}
dout(1) << "====== req done req=" << hex << req << dec
<< " op=" << (op ? op->name() : "unknown")
<< " bucket=" << s->bucket_name
<< " status=" << op_ret
<< " http_status=" << s->err.http_ret
<< " latency=" << lat
<< " request_id=" << s->trans_id
<< " ======"
<< dendl;
return (ret < 0 ? ret : s->err.ret);
} /* process_request */