mon/MDSMonitor: warn when creating fs with default EC data pool

Rationale can be found in [1]. Point is that EC pools incur a
significant performance penalty when dealing with small files and xattr
updates. This is because _every_ inode has a corresponding data pool
object with backtrace information stored in its xattr.

[1] doc/cephfs/createfs.rst

Fixes: https://tracker.ceph.com/issues/42450
Signed-off-by: Patrick Donnelly <pdonnell@redhat.com>

diff --git a/src/mon/FSCommands.cc b/src/mon/FSCommands.cc
index 0cb0f9a994f8e6f4ffac6a1fedc9186bf34592e9..d45afb84072434145bd070a7ecf180bf51a24cb8 100644 (file)
--- a/src/mon/FSCommands.cc
+++ b/src/mon/FSCommands.cc
@@ -238,12 +238,12 @@ class FsNewHandler : public FileSystemCommandHandler
     pg_pool_t const *metadata_pool = mon->osdmon()->osdmap.get_pg_pool(metadata);
     ceph_assert(metadata_pool != NULL);  // Checked it existed above
 
-    int r = _check_pool(mon->osdmon()->osdmap, data, false, force, &ss);
+    int r = _check_pool(mon->osdmon()->osdmap, data, POOL_DATA_DEFAULT, force, &ss);
     if (r < 0) {
       return r;
     }
 
-    r = _check_pool(mon->osdmon()->osdmap, metadata, true, force, &ss);
+    r = _check_pool(mon->osdmon()->osdmap, metadata, POOL_METADATA, force, &ss);
     if (r < 0) {
       return r;
     }
@@ -679,7 +679,7 @@ class AddDataPoolHandler : public FileSystemCommandHandler
       }
     }
 
-    int r = _check_pool(mon->osdmon()->osdmap, poolid, false, false, &ss);
+    int r = _check_pool(mon->osdmon()->osdmap, poolid, POOL_DATA_EXTRA, false, &ss);
     if (r != 0) {
       return r;
     }
@@ -986,7 +986,7 @@ FileSystemCommandHandler::load(Paxos *paxos)
 int FileSystemCommandHandler::_check_pool(
     OSDMap &osd_map,
     const int64_t pool_id,
-    bool metadata,
+    int type,
     bool force,
     std::stringstream *ss) const
 {
@@ -1000,32 +1000,41 @@ int FileSystemCommandHandler::_check_pool(
 
   const string& pool_name = osd_map.get_pool_name(pool_id);
 
-  if (pool->is_erasure() && metadata) {
+  if (pool->is_erasure()) {
+    if (type == POOL_METADATA) {
       *ss << "pool '" << pool_name << "' (id '" << pool_id << "')"
          << " is an erasure-coded pool.  Use of erasure-coded pools"
          << " for CephFS metadata is not permitted";
-    return -EINVAL;
-  } else if (pool->is_erasure() && !pool->allows_ecoverwrites()) {
-    // non-overwriteable EC pools are only acceptable with a cache tier overlay
-    if (!pool->has_tiers() || !pool->has_read_tier() || !pool->has_write_tier()) {
+      return -EINVAL;
+    } else if (type == POOL_DATA_DEFAULT && !force) {
       *ss << "pool '" << pool_name << "' (id '" << pool_id << "')"
-         << " is an erasure-coded pool, with no overwrite support";
+             " is an erasure-coded pool."
+             " Use of an EC pool for the default data pool is discouraged;"
+             " see the online CephFS documentation for more information."
+             " Use --force to override.";
       return -EINVAL;
-    }
+    } else if (!pool->allows_ecoverwrites()) {
+      // non-overwriteable EC pools are only acceptable with a cache tier overlay
+      if (!pool->has_tiers() || !pool->has_read_tier() || !pool->has_write_tier()) {
+        *ss << "pool '" << pool_name << "' (id '" << pool_id << "')"
+            << " is an erasure-coded pool, with no overwrite support";
+        return -EINVAL;
+      }
 
-    // That cache tier overlay must be writeback, not readonly (it's the
-    // write operations like modify+truncate we care about support for)
-    const pg_pool_t *write_tier = osd_map.get_pg_pool(
-        pool->write_tier);
-    ceph_assert(write_tier != NULL);  // OSDMonitor shouldn't allow DNE tier
-    if (write_tier->cache_mode == pg_pool_t::CACHEMODE_FORWARD
-        || write_tier->cache_mode == pg_pool_t::CACHEMODE_READONLY) {
-      *ss << "EC pool '" << pool_name << "' has a write tier ("
-          << osd_map.get_pool_name(pool->write_tier)
-          << ") that is configured "
-             "to forward writes.  Use a cache mode such as 'writeback' for "
-             "CephFS";
-      return -EINVAL;
+      // That cache tier overlay must be writeback, not readonly (it's the
+      // write operations like modify+truncate we care about support for)
+      const pg_pool_t *write_tier = osd_map.get_pg_pool(
+          pool->write_tier);
+      ceph_assert(write_tier != NULL);  // OSDMonitor shouldn't allow DNE tier
+      if (write_tier->cache_mode == pg_pool_t::CACHEMODE_FORWARD
+          || write_tier->cache_mode == pg_pool_t::CACHEMODE_READONLY) {
+        *ss << "EC pool '" << pool_name << "' has a write tier ("
+            << osd_map.get_pool_name(pool->write_tier)
+            << ") that is configured "
+               "to forward writes.  Use a cache mode such as 'writeback' for "
+               "CephFS";
+        return -EINVAL;
+      }
     }
   }
 

diff --git a/src/mon/FSCommands.h b/src/mon/FSCommands.h
index 69662d2b29e5c881e49f4e0ece954200806c6bb1..66e0286a465fdfa132bcdde28371abc113ff0453 100644 (file)
--- a/src/mon/FSCommands.h
+++ b/src/mon/FSCommands.h
@@ -30,6 +30,11 @@ class FileSystemCommandHandler : protected CommandHandler
 protected:
   std::string prefix;
 
+  enum {
+    POOL_METADATA,
+    POOL_DATA_DEFAULT,
+    POOL_DATA_EXTRA,
+  };
   /**
    * Return 0 if the pool is suitable for use with CephFS, or
    * in case of errors return a negative error code, and populate
@@ -40,7 +45,7 @@ protected:
   int _check_pool(
       OSDMap &osd_map,
       const int64_t pool_id,
-      bool metadata,
+      int type,
       bool force,
       std::stringstream *ss) const;