| Message ID | ff17414d261065d9eff01335040f5aca3a048059.1722862822.git.ps@pks.im (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | reftable: improvements and fixes for compaction | expand |
Patrick Steinhardt <ps@pks.im> writes: > The locking employed by compaction uses the following schema: > > 1. Lock "tables.list" and verify that it matches the version we have > loaded in core. > > 2. Lock each of the tables in the user-supplied range of tables that > we are supposed to compact. These locks prohibit any concurrent > process to compact those tables while we are doing that. > > 3. Unlock "tables.list". This enables concurrent processes to add new > tables to the stack, but also allows them to compact tables outside > of the range of tables that we have locked. > > 4. Perform the compaction. > > 5. Lock "tables.list" again. > > 6. Move the compacted table into place. > > 7. Write the new order of tables, including the compacted table, into > the lockfile. > > 8. Commit the lockfile into place. > This summary helps a lot, thanks! [snip] > @@ -1123,6 +1125,100 @@ static int stack_compact_range(struct reftable_stack *st, > } > } > > + /* > + * As we have unlocked the stack while compacting our slice of tables > + * it may have happened that a concurrently running process has updated > + * the stack while we were compacting. In that case, we need to check > + * whether the tables that we have just compacted still exist in the > + * stack in the exact same order as we have compacted them. > + * But as per the current implementation, the tables we compacted would always exist in tables.list, since we've obtained a lock on them. Looking at the code below, wouldn't it be more ideal to talk about how there are two scenarios we need to handle? 1. Stack is upto date, there we simply overwrite the stack with our modified version. 2. Stack is not upto date, in this scenario, we need to amend the stack without loosing out information. An extra check here is that we also see that the tables we compact are still existing. (I don't really get, why they wouldn't be though). > + * If they do exist, then it is fine to continue and replace those > + * tables with our compacted version. If they don't, then we need to > + * abort. > + */ > + err = stack_uptodate(st); > + if (err < 0) > + goto done; > + if (err > 0) { So this is the scenario where the stack is no longer upto date. > + ssize_t new_offset = -1; > + int fd; > + > + fd = open(st->list_file, O_RDONLY); > + if (fd < 0) { > + err = REFTABLE_IO_ERROR; > + goto done; > + } > + > + err = fd_read_lines(fd, &names); > + close(fd); > + if (err < 0) > + goto done; > + > + /* > + * Search for the offset of the first table that we have > + * compacted in the updated "tables.list" file. > + */ > + for (size_t i = 0; names[i]; i++) { > + if (strcmp(names[i], st->readers[first]->name)) > + continue; > + > + /* > + * We have found the first entry. Verify that all the > + * subsequent tables we have compacted still exist in > + * the modified stack in the exact same order as we > + * have compacted them. > + */ > + for (size_t j = 1; j < last - first + 1; j++) { > + const char *old = first + j < st->merged->stack_len ? > + st->readers[first + j]->name : NULL; > + const char *new = names[i + j]; > + > + /* > + * If some entries are missing or in case the tables > + * have changed then we need to bail out. Again, this > + * shouldn't ever happen because we have locked the > + * tables we are compacting. > + */ Okay, this is exactly what I was saying above. It still does makes sense to keep this check to ensure future versions don't break it. > + if (!old || !new || strcmp(old, new)) { > + err = REFTABLE_OUTDATED_ERROR; > + goto done; > + } > + } > + > + new_offset = i; > + break; > + } > + > + /* > + * In case we didn't find our compacted tables in the stack we > + * need to bail out. In theory, this should have never happened > + * because we locked the tables we are compacting. > + */ > + if (new_offset < 0) { > + err = REFTABLE_OUTDATED_ERROR; > + goto done; > + } > + > + /* > + * We have found the new range that we want to replace, so > + * let's update the range of tables that we want to replace. > + */ > + last_to_replace = last + (new_offset - first); > + first_to_replace = new_offset; Nit: might be easier to read as first_to_replace = new_offset; last_to_replace = first_to_replace + (last - first); [snip]
On Thu, Aug 08, 2024 at 07:14:15AM -0500, Karthik Nayak wrote: > Patrick Steinhardt <ps@pks.im> writes: > > + /* > > + * We have found the first entry. Verify that all the > > + * subsequent tables we have compacted still exist in > > + * the modified stack in the exact same order as we > > + * have compacted them. > > + */ > > + for (size_t j = 1; j < last - first + 1; j++) { > > + const char *old = first + j < st->merged->stack_len ? > > + st->readers[first + j]->name : NULL; > > + const char *new = names[i + j]; > > + > > + /* > > + * If some entries are missing or in case the tables > > + * have changed then we need to bail out. Again, this > > + * shouldn't ever happen because we have locked the > > + * tables we are compacting. > > + */ > > Okay, this is exactly what I was saying above. It still does makes sense > to keep this check to ensure future versions don't break it. Yeah, exactly. It's mostly about defense in depth, but should in theory never be needed. > > + if (!old || !new || strcmp(old, new)) { > > + err = REFTABLE_OUTDATED_ERROR; > > + goto done; > > + } > > + } > > + > > + new_offset = i; > > + break; > > + } > > + > > + /* > > + * In case we didn't find our compacted tables in the stack we > > + * need to bail out. In theory, this should have never happened > > + * because we locked the tables we are compacting. > > + */ > > + if (new_offset < 0) { > > + err = REFTABLE_OUTDATED_ERROR; > > + goto done; > > + } > > + > > + /* > > + * We have found the new range that we want to replace, so > > + * let's update the range of tables that we want to replace. > > + */ > > + last_to_replace = last + (new_offset - first); > > + first_to_replace = new_offset; > > Nit: might be easier to read as > > first_to_replace = new_offset; > last_to_replace = first_to_replace + (last - first); True. Initially I didn't have the `first_to_replace` variables, but now that I do it certainly makes it easier to order it naturally. Patrick
diff --git a/reftable/stack.c b/reftable/stack.c index 54982e0f7d..51eb4470c7 100644 --- a/reftable/stack.c +++ b/reftable/stack.c @@ -1020,7 +1020,9 @@ static int stack_compact_range(struct reftable_stack *st, struct lock_file *table_locks = NULL; struct tempfile *new_table = NULL; int is_empty_table = 0, err = 0; + size_t first_to_replace, last_to_replace; size_t i, nlocks = 0; + char **names = NULL; if (first > last || (!expiry && first == last)) { err = 0; @@ -1123,6 +1125,100 @@ static int stack_compact_range(struct reftable_stack *st, } } + /* + * As we have unlocked the stack while compacting our slice of tables + * it may have happened that a concurrently running process has updated + * the stack while we were compacting. In that case, we need to check + * whether the tables that we have just compacted still exist in the + * stack in the exact same order as we have compacted them. + * + * If they do exist, then it is fine to continue and replace those + * tables with our compacted version. If they don't, then we need to + * abort. + */ + err = stack_uptodate(st); + if (err < 0) + goto done; + if (err > 0) { + ssize_t new_offset = -1; + int fd; + + fd = open(st->list_file, O_RDONLY); + if (fd < 0) { + err = REFTABLE_IO_ERROR; + goto done; + } + + err = fd_read_lines(fd, &names); + close(fd); + if (err < 0) + goto done; + + /* + * Search for the offset of the first table that we have + * compacted in the updated "tables.list" file. + */ + for (size_t i = 0; names[i]; i++) { + if (strcmp(names[i], st->readers[first]->name)) + continue; + + /* + * We have found the first entry. Verify that all the + * subsequent tables we have compacted still exist in + * the modified stack in the exact same order as we + * have compacted them. + */ + for (size_t j = 1; j < last - first + 1; j++) { + const char *old = first + j < st->merged->stack_len ? + st->readers[first + j]->name : NULL; + const char *new = names[i + j]; + + /* + * If some entries are missing or in case the tables + * have changed then we need to bail out. Again, this + * shouldn't ever happen because we have locked the + * tables we are compacting. + */ + if (!old || !new || strcmp(old, new)) { + err = REFTABLE_OUTDATED_ERROR; + goto done; + } + } + + new_offset = i; + break; + } + + /* + * In case we didn't find our compacted tables in the stack we + * need to bail out. In theory, this should have never happened + * because we locked the tables we are compacting. + */ + if (new_offset < 0) { + err = REFTABLE_OUTDATED_ERROR; + goto done; + } + + /* + * We have found the new range that we want to replace, so + * let's update the range of tables that we want to replace. + */ + last_to_replace = last + (new_offset - first); + first_to_replace = new_offset; + } else { + /* + * `fd_read_lines()` uses a `NULL` sentinel to indicate that + * the array is at its end. As we use `free_names()` to free + * the array, we need to include this sentinel value here and + * thus have to allocate `stack_len + 1` many entries. + */ + REFTABLE_CALLOC_ARRAY(names, st->merged->stack_len + 1); + for (size_t i = 0; i < st->merged->stack_len; i++) + names[i] = xstrdup(st->readers[i]->name); + last_to_replace = last; + first_to_replace = first; + } + /* * If the resulting compacted table is not empty, then we need to move * it into place now. @@ -1145,12 +1241,12 @@ static int stack_compact_range(struct reftable_stack *st, * have just written. In case the compacted table became empty we * simply skip writing it. */ - for (i = 0; i < first; i++) - strbuf_addf(&tables_list_buf, "%s\n", st->readers[i]->name); + for (i = 0; i < first_to_replace; i++) + strbuf_addf(&tables_list_buf, "%s\n", names[i]); if (!is_empty_table) strbuf_addf(&tables_list_buf, "%s\n", new_table_name.buf); - for (i = last + 1; i < st->merged->stack_len; i++) - strbuf_addf(&tables_list_buf, "%s\n", st->readers[i]->name); + for (i = last_to_replace + 1; names[i]; i++) + strbuf_addf(&tables_list_buf, "%s\n", names[i]); err = write_in_full(get_lock_file_fd(&tables_list_lock), tables_list_buf.buf, tables_list_buf.len); @@ -1203,9 +1299,10 @@ static int stack_compact_range(struct reftable_stack *st, delete_tempfile(&new_table); strbuf_release(&new_table_name); strbuf_release(&new_table_path); - strbuf_release(&tables_list_buf); strbuf_release(&table_name); + free_names(names); + return err; }
The locking employed by compaction uses the following schema: 1. Lock "tables.list" and verify that it matches the version we have loaded in core. 2. Lock each of the tables in the user-supplied range of tables that we are supposed to compact. These locks prohibit any concurrent process to compact those tables while we are doing that. 3. Unlock "tables.list". This enables concurrent processes to add new tables to the stack, but also allows them to compact tables outside of the range of tables that we have locked. 4. Perform the compaction. 5. Lock "tables.list" again. 6. Move the compacted table into place. 7. Write the new order of tables, including the compacted table, into the lockfile. 8. Commit the lockfile into place. Letting concurrent processes modify the "tables.list" file while we are doing the compaction is very much part of the design and thus expected. After all, it may take some time to compact tables in the case where we are compacting a lot of very large tables. But there is a bug in the code. Suppose we have two processes which are compacting two slices of the table. Given that we lock each of the tables before compacting them, we know that the slices must be disjunct from each other. But regardless of that, compaction performed by one process will always impact what the other process needs to write to the "tables.list" file. Right now, we do not check whether the "tables.list" has been changed after we have locked it for the second time in (5). This has the consequence that we will always commit the old, cached in-core tables to disk without paying to respect what the other process has written. This scenario would then lead to data loss and corruption. This can even happen in the simpler case of one compacting process and one writing process. The newly-appended table by the writing process would get discarded by the compacting process because it never sees the new table. Fix this bug by re-checking whether our stack is still up to date after locking for the second time. If it isn't, then we adjust the indices of tables to replace in the updated stack. Signed-off-by: Patrick Steinhardt <ps@pks.im> --- reftable/stack.c | 107 ++++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 102 insertions(+), 5 deletions(-)