file: Change nowait_works mode detection

In the end of the day the nowait mode to use depends on three sources of
information: the filesystem itself, the kernel version and the reactor
option (CLI switch). By now the selection was between yes and no modes,
this patch wires the read_only mode detection and activation.

Signed-off-by: Pavel Emelyanov <[email protected]>

Author: xemul

include/seastar/core/reactor_config.hh

@@ -42,7 +42,7 @@ struct reactor_config {
     bool strict_o_direct = true;
     bool bypass_fsync = false;
     bool no_poll_aio = false;
-    bool aio_nowait_works = false;
+    std::optional<bool> aio_nowait_works = false;
     bool abort_on_too_long_task_queue = false;
 };
 /// \endcond

src/core/file.cc

@@ -1107,10 +1107,22 @@ make_file_impl(int fd, file_open_options options, int flags, struct stat st) noe
                 fsi.fsync_is_exclusive = true;
             }
 
-            if (fs_nowait_works && engine()._cfg.aio_nowait_works) {
-                fsi.nowait_works = nowait_mode::yes;
-            } else {
+            if (!fs_nowait_works) {
                 fsi.nowait_works = nowait_mode::no;
+            } else if (engine()._cfg.aio_nowait_works.has_value()) {
+                if (*engine()._cfg.aio_nowait_works) {
+                    fsi.nowait_works = nowait_mode::yes;
+                } else {
+                    fsi.nowait_works = nowait_mode::no;
+                }
+            } else {
+                if (internal::kernel_uname().whitelisted({"6.0"})) {
+                    fsi.nowait_works = nowait_mode::read_only; // seastar issue #2974
+                } else if (internal::kernel_uname().whitelisted({"4.13"})) {
+                    fsi.nowait_works = nowait_mode::yes;
+                } else {
+                    fsi.nowait_works = nowait_mode::no;
+                }
             }
             s_fstype.insert(std::make_pair(st.st_dev, std::move(fsi)));
             return make_file_impl(fd, std::move(options), flags, std::move(st));

src/core/reactor.cc

@@ -1493,7 +1493,7 @@ reactor::test::get_stall_detector_report_function() {
 }
 
 bool reactor::test::linux_aio_nowait() {
-    return engine()._cfg.aio_nowait_works;
+    return engine()._cfg.aio_nowait_works.value_or(true);
 }
 
 reactor::test::long_task_queue_state
@@ -3905,7 +3905,7 @@ reactor_options::reactor_options(program_options::option_group* parent_group)
     , blocked_reactor_reports_per_minute(*this, "blocked-reactor-reports-per-minute", 5, "Maximum number of backtraces reported by stall detector per minute")
     , blocked_reactor_report_format_oneline(*this, "blocked-reactor-report-format-oneline", true, "Print a simplified backtrace on a single line")
     , relaxed_dma(*this, "relaxed-dma", "allow using buffered I/O if DMA is not available (reduces performance)")
-    , linux_aio_nowait(*this, "linux-aio-nowait", internal::kernel_uname().whitelisted({"4.13"}), // base version where this works
+    , linux_aio_nowait(*this, "linux-aio-nowait", {},
                 "use the Linux NOWAIT AIO feature, which reduces reactor stalls due to aio (autodetected)")
     , unsafe_bypass_fsync(*this, "unsafe-bypass-fsync", false, "Bypass fsync(), may result in data loss. Use for testing on consumer drives")
     , kernel_page_cache(*this, "kernel-page-cache", false,
@@ -4450,7 +4450,7 @@ void smp::configure(const smp_options& smp_opts, const reactor_options& reactor_
         .strict_o_direct = !reactor_opts.relaxed_dma,
         .bypass_fsync = reactor_opts.unsafe_bypass_fsync.get_value(),
         .no_poll_aio = !reactor_opts.poll_aio.get_value() || (reactor_opts.poll_aio.defaulted() && reactor_opts.overprovisioned),
-        .aio_nowait_works = reactor_opts.linux_aio_nowait.get_value(), // Mixed in with filesystem-provided values later
+        .aio_nowait_works = reactor_opts.linux_aio_nowait.defaulted() ? std::optional<bool>(std::nullopt) : std::optional<bool>(reactor_opts.linux_aio_nowait.get_value()), // Mixed in with filesystem-provided values later
         .abort_on_too_long_task_queue = reactor_opts.abort_on_too_long_task_queue.get_value(),
     };