[v3,2/2] Write the DWARF index in the background
Commit Message
The new DWARF cooked indexer interacts poorly with the DWARF index
cache. In particular, the cache will require gdb to wait for the
cooked index to be finalized. As this happens in the foreground, it
means that users with this setting enabled will see a slowdown.
This patch changes gdb to write the cache entry a worker thread. (As
usual, in the absence of threads, this work is simply done immediately
in the main thread.)
Some care is taken to ensure that this can't crash, and that gdb will
not exit before the task is complete.
To avoid use-after-free problems, the DWARF per-BFD object explicitly
waits for the index cache task to complete.
To avoid gdb exiting early, an exit observer is used to wait for all
such pending tasks.
In normal use, neither of these waits will be very visible. For users
using "-batch" to pre-generate the index, though, it would be.
However I don't think there is much to be done about this, as it was
the status quo ante.
---
gdb/dwarf2/cooked-index.c | 73 ++++++++++++++++++++++++++++++++++++++-
gdb/dwarf2/cooked-index.h | 31 ++++++++++-------
gdb/dwarf2/mapped-index.h | 9 +++++
gdb/dwarf2/read.c | 11 +++---
4 files changed, 107 insertions(+), 17 deletions(-)
@@ -20,11 +20,20 @@
#include "defs.h"
#include "dwarf2/cooked-index.h"
#include "dwarf2/read.h"
+#include "dwarf2/index-cache.h"
#include "cp-support.h"
#include "c-lang.h"
#include "ada-lang.h"
#include "split-name.h"
+#include "observable.h"
+#include "run-on-main-thread.h"
#include <algorithm>
+#include <unordered_set>
+
+/* We don't want gdb to exit while it is in the process of writing to
+ the index cache. So, all live cooked index vectors are stored
+ here, and then these are all waited for before exit proceeds. */
+static std::unordered_set<cooked_index_vector *> active_vectors;
/* Hash function for cooked_index_entry. */
@@ -282,11 +291,46 @@ cooked_index::find (gdb::string_view name, bool completing)
return range (lower, upper);
}
-cooked_index_vector::cooked_index_vector (vec_type &&vec)
+cooked_index_vector::cooked_index_vector (vec_type &&vec,
+ dwarf2_per_bfd *per_bfd)
: m_vector (std::move (vec))
{
for (auto &idx : m_vector)
idx->finalize ();
+
+ /* This must be set after all the finalization tasks have been
+ started, because it may call 'wait'. */
+ m_write_future
+ = gdb::thread_pool::g_thread_pool->post_task ([this, per_bfd] ()
+ {
+ maybe_write_index (per_bfd);
+ });
+
+ /* ACTIVE_VECTORS is not locked, and this assert ensures that this
+ will be caught if ever moved to the background. */
+ gdb_assert (is_main_thread ());
+ active_vectors.insert (this);
+}
+
+cooked_index_vector::~cooked_index_vector ()
+{
+ /* The 'finalize' method may be run in a different thread. If
+ this object is destroyed before this completes, then the method
+ will end up writing to freed memory. Waiting for this to
+ complete avoids this problem; and the cost seems ignorable
+ because creating and immediately destroying the debug info is a
+ relatively rare thing to do. */
+ wait ();
+
+ /* Likewise for the index-creating future, though this one must also
+ waited for by the per-BFD object to ensure the required data
+ remains live. */
+ wait_completely ();
+
+ /* Remove our entry from the global list. See the assert in the
+ constructor to understand this. */
+ gdb_assert (is_main_thread ());
+ active_vectors.erase (this);
}
/* See cooked-index.h. */
@@ -345,3 +389,30 @@ cooked_index_vector::get_main () const
return result;
}
+
+void
+cooked_index_vector::maybe_write_index (dwarf2_per_bfd *per_bfd)
+{
+ /* Wait for finalization. */
+ wait ();
+
+ /* (maybe) store an index in the cache. */
+ global_index_cache.store (per_bfd);
+}
+
+/* Wait for all the index cache entries to be written before gdb
+ exits. */
+static void
+wait_for_index_cache (int)
+{
+ gdb_assert (is_main_thread ());
+ for (cooked_index_vector *item : active_vectors)
+ item->wait_completely ();
+}
+
+void _initialize_cooked_index ();
+void
+_initialize_cooked_index ()
+{
+ gdb::observers::gdb_exiting.attach (wait_for_index_cache, "cooked-index");
+}
@@ -36,6 +36,7 @@
#include "gdbsupport/range-chain.h"
struct dwarf2_per_cu_data;
+struct dwarf2_per_bfd;
/* Flags that describe an entry in the index. */
enum cooked_index_flag_enum : unsigned char
@@ -307,7 +308,8 @@ class cooked_index_vector : public dwarf_scanner_base
object. */
typedef std::vector<std::unique_ptr<cooked_index>> vec_type;
- explicit cooked_index_vector (vec_type &&vec);
+ cooked_index_vector (vec_type &&vec, dwarf2_per_bfd *per_bfd);
+ ~cooked_index_vector () override;
DISABLE_COPY_AND_ASSIGN (cooked_index_vector);
/* Wait until the finalization of the entire cooked_index_vector is
@@ -318,17 +320,6 @@ class cooked_index_vector : public dwarf_scanner_base
item->wait ();
}
- ~cooked_index_vector ()
- {
- /* The 'finalize' methods may be run in a different thread. If
- this object is destroyed before these complete, then one will
- end up writing to freed memory. Waiting for finalization to
- complete avoids this problem; and the cost seems ignorable
- because creating and immediately destroying the debug info is a
- relatively rare thing to do. */
- wait ();
- }
-
/* A range over a vector of subranges. */
typedef range_chain<cooked_index::range> range;
@@ -367,11 +358,27 @@ class cooked_index_vector : public dwarf_scanner_base
quick_symbol_functions_up make_quick_functions () const override;
+ /* Wait for the index to be completely finished. For ordinary uses,
+ the index code ensures this itself -- e.g., 'all_entries' will
+ wait on the 'finalize' future. However, on destruction, if an
+ index is being written, it's also necessary to wait for that to
+ complete. */
+ void wait_completely () override
+ {
+ m_write_future.wait ();
+ }
+
private:
+ /* Maybe write the index to the index cache. */
+ void maybe_write_index (dwarf2_per_bfd *per_bfd);
+
/* The vector of cooked_index objects. This is stored because the
entries are stored on the obstacks in those objects. */
vec_type m_vector;
+
+ /* A future that tracks when the 'index_write' method is done. */
+ std::future<void> m_write_future;
};
#endif /* GDB_DWARF2_COOKED_INDEX_H */
@@ -73,6 +73,15 @@ struct dwarf_scanner_base
will return 'this' as a cooked index. For other forms, it will
throw an exception with an appropriate error message. */
virtual cooked_index_vector *index_for_writing () = 0;
+
+ /* Wait for reading of the debuginfo to be completely finished.
+ This normally has a trivial implementation, but if a subclass
+ does any background reading, it's needed to ensure that the
+ reading is completed before destroying the containing per-BFD
+ object. */
+ virtual void wait_completely ()
+ {
+ }
};
/* Base class containing bits shared by both .gdb_index and
@@ -1467,6 +1467,11 @@ dwarf2_per_bfd::dwarf2_per_bfd (bfd *obfd, const dwarf2_debug_sections *names,
dwarf2_per_bfd::~dwarf2_per_bfd ()
{
+ /* Data from the per-BFD may be needed when finalizing the cooked
+ index table, so wait here while this happens. */
+ if (index_table != nullptr)
+ index_table->wait_completely ();
+
for (auto &per_cu : all_units)
{
per_cu->imported_symtabs_free ();
@@ -5432,9 +5437,6 @@ dwarf2_build_psymtabs (struct objfile *objfile)
try
{
dwarf2_build_psymtabs_hard (per_objfile);
-
- /* (maybe) store an index in the cache. */
- global_index_cache.store (per_objfile->per_bfd);
}
catch (const gdb_exception_error &except)
{
@@ -7182,7 +7184,8 @@ dwarf2_build_psymtabs_hard (dwarf2_per_objfile *per_objfile)
indexes.end ());
indexes.shrink_to_fit ();
- cooked_index_vector *vec = new cooked_index_vector (std::move (indexes));
+ cooked_index_vector *vec = new cooked_index_vector (std::move (indexes),
+ per_bfd);
per_bfd->index_table.reset (vec);
const cooked_index_entry *main_entry = vec->get_main ();