journal: consider file deletion errors a reason for rotation

[elogind.git] / src / journal / journal-file.c

diff --git a/src/journal/journal-file.c b/src/journal/journal-file.c
index 1736ff57af795b08b69cf4a943f1c4278061fd6c..c9030c56addb5e7dac5f6d19c23b9b70d9fae234 100644 (file)
--- a/src/journal/journal-file.c
+++ b/src/journal/journal-file.c
@@ -27,6 +27,7 @@
 #include <fcntl.h>
 #include <stddef.h>
 
+#include "btrfs-util.h"
 #include "journal-def.h"
 #include "journal-file.h"
 #include "journal-authenticate.h"
@@ -67,6 +68,9 @@
 /* How much to increase the journal file size at once each time we allocate something new. */
 #define FILE_SIZE_INCREASE (8ULL*1024ULL*1024ULL)              /* 8MB */
 
+/* Reread fstat() of the file for detecting deletions at least this often */
+#define LAST_STAT_REFRESH_USEC (5*USEC_PER_SEC)
+
 /* The mmap context to use for the header we pick as one above the last defined typed */
 #define CONTEXT_HEADER _OBJECT_TYPE_MAX
 
@@ -137,6 +141,9 @@ void journal_file_close(JournalFile *f) {
         if (f->mmap && f->fd >= 0)
                 mmap_cache_close_fd(f->mmap, f->fd);
 
+        if (f->fd >= 0 && f->defrag_on_close)
+                btrfs_defrag_fd(f->fd);
+
         safe_close(f->fd);
         free(f->path);
 
@@ -319,6 +326,22 @@ static int journal_file_verify_header(JournalFile *f) {
         return 0;
 }
 
+static int journal_file_fstat(JournalFile *f) {
+        assert(f);
+        assert(f->fd >= 0);
+
+        if (fstat(f->fd, &f->last_stat) < 0)
+                return -errno;
+
+        f->last_stat_usec = now(CLOCK_MONOTONIC);
+
+        /* Refuse appending to files that are already deleted */
+        if (f->last_stat.st_nlink <= 0)
+                return -EIDRM;
+
+        return 0;
+}
+
 static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size) {
         uint64_t old_size, new_size;
         int r;
@@ -340,8 +363,21 @@ static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size)
         if (new_size < le64toh(f->header->header_size))
                 new_size = le64toh(f->header->header_size);
 
-        if (new_size <= old_size)
-                return 0;
+        if (new_size <= old_size) {
+
+                /* We already pre-allocated enough space, but before
+                 * we write to it, let's check with fstat() if the
+                 * file got deleted, in order make sure we don't throw
+                 * away the data immediately. Don't check fstat() for
+                 * all writes though, but only once ever 10s. */
+
+                if (f->last_stat_usec + LAST_STAT_REFRESH_USEC > now(CLOCK_MONOTONIC))
+                        return 0;
+
+                return journal_file_fstat(f);
+        }
+
+        /* Allocate more space. */
 
         if (f->metrics.max_size > 0 && new_size > f->metrics.max_size)
                 return -E2BIG;
@@ -376,12 +412,9 @@ static int journal_file_allocate(JournalFile *f, uint64_t offset, uint64_t size)
         if (r != 0)
                 return -r;
 
-        if (fstat(f->fd, &f->last_stat) < 0)
-                return -errno;
-
         f->header->arena_size = htole64(new_size - le64toh(f->header->header_size));
 
-        return 0;
+        return journal_file_fstat(f);
 }
 
 static unsigned type_to_context(ObjectType type) {
@@ -392,6 +425,8 @@ static unsigned type_to_context(ObjectType type) {
 }
 
 static int journal_file_move_to(JournalFile *f, ObjectType type, bool keep_always, uint64_t offset, uint64_t size, void **ret) {
+        int r;
+
         assert(f);
         assert(ret);
 
@@ -403,8 +438,11 @@ static int journal_file_move_to(JournalFile *f, ObjectType type, bool keep_alway
                 /* Hmm, out of range? Let's refresh the fstat() data
                  * first, before we trust that check. */
 
-                if (fstat(f->fd, &f->last_stat) < 0 ||
-                    offset + size > (uint64_t) f->last_stat.st_size)
+                r = journal_file_fstat(f);
+                if (r < 0)
+                        return r;
+
+                if (offset + size > (uint64_t) f->last_stat.st_size)
                         return -EADDRNOTAVAIL;
         }
 
@@ -2548,10 +2586,9 @@ int journal_file_open(
                 goto fail;
         }
 
-        if (fstat(f->fd, &f->last_stat) < 0) {
-                r = -errno;
+        r = journal_file_fstat(f);
+        if (r < 0)
                 goto fail;
-        }
 
         if (f->last_stat.st_size == 0 && f->writable) {
                 /* Let's attach the creation time to the journal file,
@@ -2580,10 +2617,9 @@ int journal_file_open(
                 if (r < 0)
                         goto fail;
 
-                if (fstat(f->fd, &f->last_stat) < 0) {
-                        r = -errno;
+                r = journal_file_fstat(f);
+                if (r < 0)
                         goto fail;
-                }
 
                 newly_created = true;
         }
@@ -2700,12 +2736,20 @@ int journal_file_rotate(JournalFile **f, bool compress, bool seal) {
         if (r < 0)
                 return -ENOMEM;
 
+        /* Try to rename the file to the archived version. If the file
+         * already was deleted, we'll get ENOENT, let's ignore that
+         * case. */
         r = rename(old_file->path, p);
-        if (r < 0)
+        if (r < 0 && errno != ENOENT)
                 return -errno;
 
         old_file->header->state = STATE_ARCHIVED;
 
+        /* Currently, btrfs is not very good with out write patterns
+         * and fragments heavily. Let's defrag our journal files when
+         * we archive them */
+        old_file->defrag_on_close = true;
+
         r = journal_file_open(old_file->path, old_file->flags, old_file->mode, compress, seal, NULL, old_file->mmap, old_file, &new_file);
         journal_file_close(old_file);
 
@@ -2736,7 +2780,8 @@ int journal_file_open_reliably(
             r != -EPROTONOSUPPORT && /* incompatible feature */
             r != -EBUSY && /* unclean shutdown */
             r != -ESHUTDOWN && /* already archived */
-            r != -EIO /* IO error, including SIGBUS on mmap */)
+            r != -EIO && /* IO error, including SIGBUS on mmap */
+            r != -EIDRM /* File has been deleted */)
                 return r;
 
         if ((flags & O_ACCMODE) == O_RDONLY)
@@ -2761,6 +2806,10 @@ int journal_file_open_reliably(
         if (r < 0)
                 return -errno;
 
+        /* btrfs doesn't cope well with our write pattern and
+         * fragments heavily. Let's defrag all files we rotate */
+        (void) btrfs_defrag(p);
+
         log_warning("File %s corrupted or uncleanly shut down, renaming and replacing.", fname);
 
         return journal_file_open(fname, flags, mode, compress, seal,