From: Venky Shankar <vshankar@redhat.com>
Date: Thu, 8 Nov 2018 12:58:37 +0000 (+0530)
Subject: mds: disallow dumping huge caches to formatter
X-Git-Tag: v14.1.0~685^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=6be2ce98c629d2cfe9d8443659e59600148b7675;p=ceph.git

mds: disallow dumping huge caches to formatter

Dumping huge caches (~ 1G) to formatter causes mds to
hang or get terminated. Until the underlying issues is
fixed, disallow dumping cache if cache usage exceeds this
threshold. Also, Patrick feels a similar issue might be
hiding there for cache dumps to file when cache sizes
are really huge. This case has not been been reproduced
yet, hence, the limit for cache dumps to file is currently
kept not capped.

Fixes: http://tracker.ceph.com/issues/36703
Signed-off-by: Venky Shankar <vshankar@redhat.com>
---

diff --git a/src/common/options.cc b/src/common/options.cc
index c10f459884da..4b84646f1915 100644
--- a/src/common/options.cc
+++ b/src/common/options.cc
@@ -7353,6 +7353,16 @@ std::vector<Option> get_mds_options() {
     Option("mds_max_retries_on_remount_failure", Option::TYPE_UINT, Option::LEVEL_ADVANCED)
      .set_default(5)
      .set_description("number of consecutive failed remount attempts for invalidating kernel dcache after which client would abort."),
+
+    Option("mds_dump_cache_threshold_formatter", Option::TYPE_SIZE, Option::LEVEL_DEV)
+     .set_default(1_G)
+     .set_description("threshold for cache usage to disallow \"dump cache\" operation to formatter")
+     .set_long_description("Disallow MDS from dumping caches to formatter via \"dump cache\" command if cache usage exceeds this threshold."),
+
+    Option("mds_dump_cache_threshold_file", Option::TYPE_SIZE, Option::LEVEL_DEV)
+     .set_default(0)
+     .set_description("threshold for cache usage to disallow \"dump cache\" operation to file")
+     .set_long_description("Disallow MDS from dumping caches to file via \"dump cache\" command if cache usage exceeds this threshold."),
   });
 }
 
diff --git a/src/mds/MDCache.cc b/src/mds/MDCache.cc
index 41c652cfea56..cd4c74867070 100644
--- a/src/mds/MDCache.cc
+++ b/src/mds/MDCache.cc
@@ -12099,6 +12099,32 @@ int MDCache::dump_cache(Formatter *f)
 int MDCache::dump_cache(std::string_view fn, Formatter *f)
 {
   int r = 0;
+
+  // dumping large caches may cause mds to hang or worse get killed.
+  // so, disallow the dump if the cache size exceeds the configured
+  // threshold, which is 1G for formatter and unlimited for file (note
+  // that this can be jacked up by the admin... and is nothing but foot
+  // shooting, but the option itself is for devs and hence dangerous to
+  // tune). TODO: remove this when fixed.
+  uint64_t threshold = f ?
+    g_conf().get_val<Option::size_t>("mds_dump_cache_threshold_formatter") :
+    g_conf().get_val<Option::size_t>("mds_dump_cache_threshold_file");
+
+  if (threshold && cache_size() > threshold) {
+    if (f) {
+      std::stringstream ss;
+      ss << "cache usage exceeds dump threshold";
+      f->open_object_section("result");
+      f->dump_string("error", ss.str());
+      f->close_section();
+    } else {
+      derr << "cache usage exceeds dump threshold" << dendl;
+      r = -EINVAL;
+    }
+    return r;
+  }
+
+  r = 0;
   int fd = -1;
 
   if (f) {
diff --git a/src/mds/MDSDaemon.cc b/src/mds/MDSDaemon.cc
index 079227cb63ec..578d0b90502f 100644
--- a/src/mds/MDSDaemon.cc
+++ b/src/mds/MDSDaemon.cc
@@ -358,6 +358,8 @@ const char** MDSDaemon::get_tracked_conf_keys() const
     "mds_cache_reservation",
     "mds_health_cache_threshold",
     "mds_cache_mid",
+    "mds_dump_cache_threshold_formatter",
+    "mds_dump_cache_threshold_file",
     // MDBalancer
     "mds_bal_fragment_dirs",
     "mds_bal_fragment_interval",