[Tarantool-patches] [PATCH v2 luajit 5/5] Use generic trace error for OOM during trace stitching.

Sergey Kaplun skaplun at tarantool.org
Sun Jun 16 13:00:13 MSK 2024


From: Mike Pall <mike>

Thanks to Sergey Kaplun.

(cherry picked from commit b8b49bf3954b23e32e34187a6ada00021c26e172)

The previous commit doesn't handle the case when the error code is
`LUA_ERRMEM`. This patch adds a workaround by using the generic error
message.

Sergey Kaplun:
* added the description and the test for the problem

Part of tarantool/tarantool#9924
---
 src/lj_ffrecord.c                             |  2 +
 .../lj-1166-error-stitch-oom-ir-buff.test.lua | 40 ++++++++++++++++++-
 ...j-1166-error-stitch-oom-snap-buff.test.lua | 35 +++++++++++++++-
 3 files changed, 73 insertions(+), 4 deletions(-)

diff --git a/src/lj_ffrecord.c b/src/lj_ffrecord.c
index ff14e9e4..d5fc081e 100644
--- a/src/lj_ffrecord.c
+++ b/src/lj_ffrecord.c
@@ -150,6 +150,8 @@ static void recff_stitch(jit_State *J)
   if (errcode) {
     if (errcode == LUA_ERRRUN)
       copyTV(L, L->top-1, L->top + (1 + LJ_FR2));
+    else
+      setintV(L->top-1, (int32_t)LJ_TRERR_RECERR);
     lj_err_throw(L, errcode);  /* Propagate errors. */
   }
 }
diff --git a/test/tarantool-tests/lj-1166-error-stitch-oom-ir-buff.test.lua b/test/tarantool-tests/lj-1166-error-stitch-oom-ir-buff.test.lua
index 85afea11..354f5975 100644
--- a/test/tarantool-tests/lj-1166-error-stitch-oom-ir-buff.test.lua
+++ b/test/tarantool-tests/lj-1166-error-stitch-oom-ir-buff.test.lua
@@ -10,10 +10,18 @@ local test = tap.test('lj-1166-error-stitch-oom-ir-buff'):skipcond({
   ['Disabled on *BSD due to #4819'] = jit.os == 'BSD',
 })
 
-test:plan(1)
-
+local jparse = require('utils').jit.parse
 local allocinject = require('allocinject')
 
+local IS_DUALNUM = tostring(tonumber('-0')) ~= tostring(-0)
+
+-- XXX: Avoid other traces compilation due to hotcount collisions
+-- for predictable results.
+jit.off()
+jit.flush()
+
+test:plan(2)
+
 -- Generate the following Lua chunk:
 -- local s1
 -- ...
@@ -43,6 +51,10 @@ end
 -- XXX: amount of slots is empirical.
 local tracef = assert(loadstring(create_chunk(175)))
 
+-- We only need the abort reason in the test.
+jparse.start('t')
+
+jit.on()
 jit.opt.start('hotloop=1', '-loop', '-fold')
 
 allocinject.enable()
@@ -51,6 +63,30 @@ tracef()
 
 allocinject.disable()
 
+local _, aborted_traces = jparse.finish()
+
+jit.off()
+
 test:ok(true, 'stack is balanced')
 
+-- Tarantool may compile traces on the startup. These traces
+-- already exceed the maximum IR amount before the trace in this
+-- test is compiled. Hence, there is no need to reallocate the IR
+-- buffer, so the check for the IR size is not triggered.
+test:skipcond({
+  ['Impossible to predict the number of IRs for Tarantool'] = _TARANTOOL,
+  -- The amount of IR for traces is different for non x86/x64
+  -- arches and DUALNUM mode.
+  ['Disabled for non-x86_64 arches'] = jit.arch ~= 'x64' and jit.arch ~= 'x86',
+  ['Disabled for DUALNUM mode'] = IS_DUALNUM,
+})
+
+assert(aborted_traces and aborted_traces[1], 'aborted trace is persisted')
+
+-- We tried to compile only one trace.
+local reason = aborted_traces[1][1].abort_reason
+
+test:like(reason, 'error thrown or hook called during recording',
+          'abort reason is correct')
+
 test:done(true)
diff --git a/test/tarantool-tests/lj-1166-error-stitch-oom-snap-buff.test.lua b/test/tarantool-tests/lj-1166-error-stitch-oom-snap-buff.test.lua
index ea50ebc4..c8e13760 100644
--- a/test/tarantool-tests/lj-1166-error-stitch-oom-snap-buff.test.lua
+++ b/test/tarantool-tests/lj-1166-error-stitch-oom-snap-buff.test.lua
@@ -10,10 +10,16 @@ local test = tap.test('lj-1166-error-stitch-oom-snap-buff'):skipcond({
   ['Disabled on *BSD due to #4819'] = jit.os == 'BSD',
 })
 
-test:plan(1)
-
+local jparse = require('utils').jit.parse
 local allocinject = require('allocinject')
 
+-- XXX: Avoid other traces compilation due to hotcount collisions
+-- for predictable results.
+jit.off()
+jit.flush()
+
+test:plan(2)
+
 -- Generate the following Lua chunk:
 -- for i = 1, 2 do
 --   if i < 1 then end
@@ -34,6 +40,7 @@ local function create_chunk(n_conds)
   return chunk
 end
 
+jit.on()
 -- XXX: Need to compile the cycle in the `create_chunk()` to
 -- preallocate the snapshot buffer.
 jit.opt.start('hotloop=1', '-loop', '-fold')
@@ -45,6 +52,9 @@ local tracef = assert(loadstring(create_chunk(6)))
 jit.off()
 jit.flush()
 
+-- We only need the abort reason in the test.
+jparse.start('t')
+
 -- XXX: Update hotcounts to avoid hash collisions.
 jit.opt.start('hotloop=1')
 jit.on()
@@ -55,6 +65,27 @@ tracef()
 
 allocinject.disable()
 
+local _, aborted_traces = jparse.finish()
+
+jit.off()
+
 test:ok(true, 'stack is balanced')
 
+-- Tarantool may compile traces on the startup. These traces
+-- already exceed the maximum snapshot amount before the trace in
+-- this test is compiled. Hence, there is no need to reallocate
+-- the snapshot buffer, so the check for the snap size is not
+-- triggered.
+test:skipcond({
+  ['Impossible to predict the number of snapshots for Tarantool'] = _TARANTOOL,
+})
+
+assert(aborted_traces and aborted_traces[1], 'aborted trace is persisted')
+
+-- We tried to compile only one trace.
+local reason = aborted_traces[1][1].abort_reason
+
+test:like(reason, 'error thrown or hook called during recording',
+          'abort reason is correct')
+
 test:done(true)
-- 
2.45.1



More information about the Tarantool-patches mailing list