GDB is a source level debugger for C, C++ and more languages. It allows inspecting the internal state of a program as it is running as well the post-mortem inspection of crashed programs.
You can attach GDB to a running process, run a process inside GDB or examine a coredump.
The two most common usages of GDB for scylla is running a process inside it (e.g. a unit test):
gdb /path/to/executable
You can specify command-line arguments that gdb will forward to the executable:
gdb /path/to/executable --args arg1 arg2 arg3
Another prevalent usage is to examine coredumps:
gdb --core=/path/to/coredump /path/to/executable
You can also attach it to an already running process:
gdb -p $pid
Where $pid is the PID of the running process you wish to attach GDB
to.
GDB has excellent online documentation that you can find here.
Some of the more important topics:
- Starting GDB
- Setting breakpoints
- Setting catchpoints
- Stepping through the code
- Examining the stack
- Examining data
In general Scylla is quite hard to debug in GDB due to its asynchronous nature. You will soon find that backtraces always lead to the reactor's event loop and stepping through the code will not work as you expect as soon as you leave or enter an asynchronous function. That said GDB is an indispensable tool in debugging coredumps and when used right can be of great help.
Over the years we have collected a set of tools for helping with debugging
scylla. These are collected in scylla-gdb.py and are in
the form of commands,
conveninence functions
and pretty printers.
To load the file issue the following command (inside gdb):
(gdb) source /path/to/scylla-gdb.py
You should be now ready to use all of the tools contained therein. To list all available commands do:
(gdb) help scylla
To read the documentation of an individual command do:
(gdb) help scylla $commandname
Some commands have self explanatory names, some have documentation, and some have neither :( (contributions are welcome).
To get the list of the available convenience functions do:
(gdb) help function
Note that this will list GDB internal functions as well as those added
by scylla-gdb.py.
Again, just like before, to see the documentation of an individual
function do:
(gdb) help function $functionname
When running scylla (or any seastar application for that matter) inside GDB it will get interrupted often due to catching some signals used by seastar internally. This makes debugging almost impossible. To avoid this, instruct GDB to not stop on these signals:
(gdb) handle SIG34 SIG35 SIGUSR1 nostop noprint pass
GDB is known to crash when parsing some of scylla's symbols (especially those related to futures). Usually telling it to not print static members of classes and structs helps:
(gdb) set print static-members no
When using the facilities from scylla-gdb.py it is very useful to know
the full stack of a failure in some of the provided tools, so that you
can fix it or report it. To enable this run:
(gdb) set python print-stack full
Often you find yourself debugging an executable, whose internal source paths don't match those where they can be found on your machine. There is an easy workaround for this:
(gdb) set substitute-path /path/to/src/in/executable /path/to/src/on/your/machine
Note that the pattern that you supply to set substitute-path just has
to be a common prefix of the paths. Example: if the source location
inside the executable to some file is /opt/src/scylla/database.hh and
on your machine it is /home/joe/work/scylla/database.hh, you can make
GDB find the sources on your machine via:
(gdb) set substitute-path /opt/src/scylla /home/joe/work/scylla
This method might not work if the sources do not have a prefix, e.g.
they are relative to the source tree root directory. In this case you can use the
set directories command to set the search path of sources for gdb:
(gdb) set directories /path/to/scylla/source/tree
Multiple directories can be listed, separated with :.
GDB supports writing arbitrary GDB commands in any file and sourcing it.
One can use this to place commands that one would have to issue every
time when debugging in a file, instead of typing them each time GDB is
started.
Conventionally this file is called .gdbinit and GDB in fact will look
for it in you current directory, in your $HOME directory and some other
places. You can always load it by hand if GDB refuses or fails to load it:
(gdb) source /path/to/your/.gdbinit
Scylla provides a gdbinit file helpful for debugging scylla
at the root of the source tree. You can source it from your
local .gdbinit file if you wish.
GDB has a terminal based GUI called
TUI.
This is extremely useful when you wish to see the source code while you
are debugging. The TUI mode can be activated by passing -tui to GDB
on the command line, or any time by executing the tui enable to
activate it and tui disable to deactivate it respectively.
By default the source window has the focus in TUI mode, meaning that command
completion, searching history and line editing doesn't work, e.g. if you use
the up and down keys, you will scroll the source file up and down respectively,
instead of moving in the command history. To focus the command window, issue
focus cmd. To move the focus to the source window again, issue focus src.
Thread local variables are saved in a special area of memory, at a negative
offset from $fs_base. Let's look at an example TLS variable, given the
following C++ code from seastar:
namespace seastar::internal {
inline
scheduling_group*
current_scheduling_group_ptr() noexcept {
// Slow unless constructor is constexpr
static thread_local scheduling_group sg;
return &sg;
}
}
Let's have a look in GDB:
(gdb) p &'seastar::internal::current_scheduling_group_ptr()::sg'
$1 = (<thread local variable, no debug info> *) 0x7fc1f11e7c0c
(gdb) p/x $fs_base
$2 = 0x7fc1f11ff700
(gdb) p/x 0x7fc1f11e7c0c - $fs_base
$3 = 0xfffffffffffe850c
(gdb) p/x -0xfffffffffffe850c
$4 = 0x17af4
The variable sg is located at offset 0x17af4 beneath $fs_base. We
can also calculate the offset (and hence address) of a known TLS
variable in memory as follows:
$fs_offset = $tls_entry - $sizeof_TLS_header
$sizeof_TLS_header can be obtained by listing the program headers of the binary:
$ eu-readelf -l ./a.out
Type Offset VirtAddr PhysAddr FileSiz MemSiz Flg Align
[...]
TLS 0x31ead40 0x00000000033ecd40 0x00000000033ecd40 0x000058 0x017bf0 R 0x40
[...]
We are interested in the size of the TLS header, which is in the
MemSiz column and is 0x017bf0 in this example. The value of the
$tls_entry can be found in the process' symbol table:
$ eu-readelf -s ./a.out
Symbol table [ 5] '.dynsym' contains 1288 entries:
1 local symbol String table: [ 9] '.dynstr'
Num: Value Size Type Bind Vis Ndx Name
[...]
1282: 000000000000010c 4 TLS LOCAL HIDDEN 23 _ZZN7seastar8internal28current_scheduling_group_ptrEvE2sg
[...]
If we substitute these values in we can verify our theory:
(gdb) set $tls_entry = 0x000000000000010c
(gdb) set $sizeof_TLS_header = 0x017bf0
(gdb) p/x $tls_entry - $sizeof_TLS_header
$5 = 0xfffe851c
(gdb) p/x -($tls_entry - $sizeof_TLS_header)
$6 = 0x17ae4
We can also identify a TLS variable based on its address. We know the
value of $sizeof_TLS_header and we can easily calculate $fs_offset.
To identify the variable we need to calculate its $tls_entry based on
which we can find the matching entry in the symbol table. Remaining with
the above example of the address being 0x7fc1f11e7c0c, we can
calculate this as:
$tls_entry = $sizeof_TLS_header + $fs_offset
Do note however that $fs_offset is negative so this is in effect a
substitution:
$tls_entry = 0x017bf0 - 0x17ae4
This yields 0x10c which is exactly the value of the Value column in
the matching symbol table entry. This should work also if you don't have
the address to the start of the object. In this case you have to locate
the entry in the symbol table, whose value range includes the
calculated value. This can be made easier by sorting the symbol table by
the Value column.
In release builds one will find that a significant portion of variables and function parameters are optimized out. This is very annoying but often one can find a way around to inspect the desired variables.
For non-local variables, there is a good chance that a few frames up one can find another reference that wasn't optimized out. Or one can try to find another object, which is not optimized out, and which is also known to hold a reference to the variable.
If the variable is local, one can try to look at the registers (info registers) and try to identify which one holds its value. This is relatively
easy for pointers to objects as heap pointers are easily identifiable (they
start with 0x6 and are composed of 12 digits, e.g.: 0x60f146e3fbc0) and one
can check the size of the object they point to with scylla ptr. If the
pointed-to object is an instance of a class with virtual methods, the object can
be easily identified by looking at its vtable pointer (x/1a $ptr).
Up until release 3.0 we used to build and package Scylla separately for each supported distribution. Starting with 3.1 we moved to relocatable binaries. These are built with a common frozen toolchain and packages are bundled with all dependencies. This means that post 3.1 there is just one build across all supported distros and that the exact environment the binaries were built with is available in the form of a Docker image. This makes debugging cores generated from relocatable binaries much easier. As of now, all releases except 2019.1 ship via relocatable packages, so in this chapter we will focus on how to debug cores generated from relocatable binaries, with a subsection later explaining how to debug cores generated by 2019.1 binaries.
The most convenient way to open a coredump is scripts/open-coredump.sh. Just point it to a coredump and after some time you should get a shell inside the appropriate dbuild container, with a suggested gdb invocation line to open the coredump.
If you prefer to open the coredump manually or the script fails for you, continue below.
Cores produced by relocatable binaries can be simply opened in the dbuild container they were built with. To do that, two things (apart from the core itself of course) are needed:
- The exact frozen toolchain (dbuild container).
- The exact relocatable package the binary was part of.
The frozen toolchain is obtained based on the commit id of the version of the scylla executable the core was produced with. The exact commit hash can be obtained by running:
$ scylla --version
5.2.0~dev-0.20221210.e47794ed9832
The version can be divided into 4 parts:
- The version identifier, in this case: 5.2.0~dev; the ~dev suffix means this is an uncut, development branch (master), for releases this suffix is missing.
- The build identifier, in this case: 0.
- The date, in this case: 20221210.
- The commit hash, in this case: e47794ed9832.
Based on the latter, you can obtain the right frozen toolchain:
$ cd /path/to/scylla
$ git checkout $commit_hash
Once we have the right toolchain, we have to obtain the relocatable package. This is obtained based on the build-id, which can be obtained from the coredump like this:
$ eu-unstrip -n --core $corefile
Or from the executable like this:
$ eu-unstrip -n --exec $executable
You can find the relocatable using the http://backtrace.scylladb.com/index.html search form using either the scylla Build ID or the Release number (e.g. 5.0.0 or 2022.1) to search the packages.
The form can also be used to decode backtraces generated by the corresponding scylla binary.
NOTE: Use the normal relocatable package, usually called
scylla-package.tar.gz, not not the debuginfo one usually called
scylla-debug-package.tar.gz.
Build-id:s for all official releases are listed on http://backtrace.scylladb.com/releases.html.
Move the coredump and the unpackaged relocatable package into some dir
$dir on your system, then:
(host)$ cd /path/to/scylla # with the right commit checked out
(host)$ ./tools/toolchain/dbuild -it -v $dir:/workdir -- bash -l
(dbuild)$ cd /workdir
(dbuild)$ ln -s /path/to/unpackaged-relocatable-package /opt/scylladb # symlink the scylla subdir if you have the unified tarball
(dbuild)$ gdb --core=$corefile /opt/scylladb/libexec/scylla
You might need to add
-ex 'set auto-load safe-path /opt/scylladb/libreloc'
to the command line, see No thread debugging.
GDB complaints that it can't find namespace seastar or some other Scylla
or Seastar symbol that you know exists. This usually happens when GDB is in
the wrong context i.e. a frame is selected which is not in the Scylla executable
but in some other library. A typical situation is opening a coredump and
attempting to access Scylla symbols when the initial frame is in libc.
Move up the stack, or select a frame which is a Scylla or Seastar function to
fix.
Unable to access thread-local variables. Example:
(gdb) p seastar::local_engine
Cannot find thread-local storage for LWP 22604, executable file /usr/lib/debug/usr/bin/scylla.debug:
Cannot find thread-local variables on this target
The first step in finding out why thread debugging doesn't work is enabling additional information about why thread debugging is not working:
(gdb) set debug libthread-db 1
This has to be done right after starting GDB, before the core and the
executable are loaded. You can do this by adding
-iex "set debug libthread-db 1" to your gdb command line.
The usual cause is that GDB failed to find some libraries or that the library versions of those libraries GDB loaded don't match those the core was generated with.
Of special note is the libthread_db.so library, which is crucial for
thread debugging to work. Common causes of failing to find or load this library
are discussed below.
You might see a message like this:
warning: File "/opt/scylladb/libreloc/libthread_db.so.1" auto-loading has been declined by your `auto-load safe-path' set to "$debugdir:$datadir/auto-load"
thread_db_load_search returning 0
To declare the directory this library is found at as safe to load from, do:
set auto-load safe-path /opt/scylladb/libreloc
Use the path that is appropriate for your setup. Alternatively you can use /
as the path to declare your entire file-system as safe to load stuff from.
Note that libthread_db.so is packaged together with libc. So if you have the
build-id appropriate libc package, you can be sure you have the correct
libthread_db.so too.
If you see a message like this:
warning: Expected absolute pathname for libpthread in the inferior, but got .gnu_debugdata for /lib64/libpthread.so.0.
Trying host libthread_db library: /lib64/libthread_db.so.1.
td_ta_new failed: application not linked with libthread
thread_db_load_search returning 0
Installing debug symbols for glibc might solve this. This issue was seen on Fedora 34, for more details see the bug report. Debug symbols can be installed with:
sudo dnf debuginfo-install glibc-2.32-6.fc33.x86_64
Adjust for the exact version you have installed.
If you ensured libthread_db.so is present and is successfully loaded by GDB
but thread debugging still doesn't work, inspect the other libraries loaded by
GDB:
(gdb) info sharedlibrary
The listing will contain the path of the loaded libraries. If a library wasn't
found by GDB that will also be visible in the listing. You can then use the
file utility to obtain the build-id of the libraries:
file /path/to/libsomething.so
This build-id must match the one obtained from the core. The library build-ids from the core can be obtained with:
eu-unstrip -n --core=/path/to/core
In general you can get away some non-core libraries missing or having the wrong
version, but the core libraries like libc.so, libgcc_s.so, librt.so and
ld.so (often called something like ld-linux-x86-64.so.2) etc. must have the
correct version. Best to ensure all libraries are correct to minimize the chance
of something not working. Also, make sure the build-id of the executable matches
that the core was generated with. Again, you can use file to obtain the
build-id of the executable, then compare it with the build-id obtained from the
eu-unstrip listing.
For more information on how to obtain the correct version of libraries and how
to override the path GDB loads them from, see Collecting libraries
and Opening the core on another OS.
Make sure you are using the normal scylla package, not the debuginfo one, see Obtaining the relocatable package.
See Avoid (some) symbol parsing related crashes.
GDB has trouble with frames inlined into the outermost frame in a seastar thread,
or any green threads in general -- where the outermost frame is annotated with
.cfi_undefined rip. See
GDB#26387.
To work around this, pass a limit to bt, such that it excludes the problematic
frame. E.g. if bt prints 10 frames before GDB crashing, use bt 9 to avoid the
crash.
See Tell GDB to not stop on signals used by seastar.
Guides focusing on different aspects of debugging Scylla. These guides assume a release build of Scylla.
Seastar has its own memory allocator optimized for Seastar's thread-per-core architecture. Memory, just like CPU, is sharded among threads, meaning that each shard has its equally-sized exclusive memory area.
The seastar allocator operates on three levels:
- pages (4KB)
- spans of pages (2^N pages, N ∈ [0, 31])
- objects managed by small pools
Small allocations (<= 16KB) are served by small memory pools. There is exactly one pool per supported size, and there is a limited number of sizes available between 1B (8B in practice) and 16KB. Allocations are served by the pool with the closest, but larger than equal size to the requested allocation, but alignments complicate this. Pools allocate spans themselves for the space to allocate objects in.
Large allocations are served by allocating an entire span.
The allocator keeps a description of all the pages and pools in memory in thread-local variables. Using these, it is possible to arrive at the metadata describing any allocation with a few steps:
address -> page -> span -> pool?
This is exploited by the scylla ptr command which, given an address,
prints the following information:
- The allocation (small or large) this address is part of.
- The offset of the address from the beginning of the allocation.
- Is the object dead or live.
Example:
(gdb) scylla ptr 0x6000000f3830
thread 1, small (size <= 512), live (0x6000000f3800 +48)
It is possible to dump the state of the seastar allocator with the
scylla memory command. This prints a report containing the state of
all small pools as well as the availability of spans. It also prints
other Scylla specific information.
Continuation chains are everywhere in Scylla. Every execution flow takes the
form of a continuation chain. This makes debugging very hard because the normal
GDB commands for inspecting and controlling execution flow (backtrace, up,
down, step, next, return, etc.) quickly fail to fulfill their purpose.
One will quickly find that every backtrace just leads to the same event loop in
the seastar reactor.
Continuation chains are formed by tasks. The tasks form an intrusive forward linked list in memory. Each tasks links to the task that depends on it. Example:
future<> bar() {
return sleep(std::chrono::seconds(10)).then([] {
});
}
future<> foo() {
return bar().then([] {
});
}
When foo() is called a continuation chain of 3 tasks will be created:
- T0: sleep()
- T1: bar()::lambda#1
- T2: foo()::lambda#1
T1 depends on T0, and T2 depends on T1. In memory they form a forward linked list like:
T0 -> T1 -> T2
The links are provided by the promise-future pairs in these tasks. Each task contains a future half of one such pair and a promise half of another one. The future is for the value arriving from the previous task and the promise is for the value calculated in the local task, that the next task waits on.
The task object have the following interface:
class task {
scheduling_group _sg;
public:
virtual void run_and_dispose() noexcept = 0;
virtual task* waiting_task() noexcept = 0;
scheduling_group group() const { return _sg; }
shared_backtrace get_backtrace() const;
};
The only thing stored at a known offset is the scheduling group. Each task
object also has an associated action that is executed when the task runs
(run_and_dispose()), as well as pointer to the next task (returned by
waiting_task()). However task being a polymorphic object the layout of the
different kind of tasks is not known, so all we can say about them is that
somewhere they contain a promise object, which has a pointer to the future object
of the task that depends on them. Also note that continuations are just one kind
of task, there are other kinds of tasks as well. Many seastar primites,
like do_with(), repeat(), do_until(), etc. have their own task
types.
Or in other words finding out what are the continuations waited on by this one as
well as the ones waiting on this one.
This involves searching for inbound and outbound references in the task and identifying
the one which is also a task. As this is quite a labour-intensive task, there is
a command in scylla-gdb.py which automates it, called scylla fiber. Example
usage:
(gdb) scylla fiber 0x0000600016217c80
#-1 (task*) 0x000060001a305910 0x0000000004aa5260 vtable for seastar::continuation<...> + 16
#0 (task*) 0x0000600016217c80 0x0000000004aa5288 vtable for seastar::continuation<...> + 16
#1 (task*) 0x000060000ac42940 0x0000000004aa2aa0 vtable for seastar::continuation<...> + 16
#2 (task*) 0x0000600023f59a50 0x0000000004ac1b30 vtable for seastar::continuation<...> + 16
This is somewhat similar to a backtrace, in that it shows tasks that are waited
on by this continuation and tasks that are waiting
for this continuation to finish, similar to how upstream functions are waiting
for the called function to finish before continuing their own execution.
See help scylla fiber and scylla fiber --help for more information on usage.
Seastar threads are a special kind of continuation. Each seastar thread hosts a
stack but it can also be linked into a continuation chain. The stack of seastar
threads is a regular stack and all the normal stack related GDB commands can be
used in it. This can be used to inspect where exactly the seastar thread
stopped when it was suspended to wait on some future. Local variables can be
inspected too. The catch is how to make GDB context switch into the
stack of the seastar stack. Unfortunately there is no method that works
with GDB as of now, the scylla thread command crashes GDB and even if it
didn't, it'd only works in live processes.
To get this working a patched GDB is needed, see
https://github.com/denesb/seastar-gdb for instructions on how to use.
Assert failures are the easiest (easiest but not easy) coredumps to debug because we know the condition that failed, we know where and thus the investigation has a clear scope -- to find out why. This is not always easy though, especially if the root cause happened much earlier, and thus the state the node was in at that time is not observable in the coredump. The root cause might even be on another node altogether. In this case we try to gather as much information as possible and write debug patches that hopefully catch the problem earlier, and try to reproduce with them, hoping to get a new coredump that has more information.
Segmentation faults are usually caused by use-after-free,
use-after-move, dangling pointer/reference or memory corruption.
Unfortunately, coredumps often contain very little immediate information on what
exactly was wrong. It is rare to find something as obvious as a null pointer
trying to be dereferenced. So one has to dig a little to find out what
exactly triggered the SEGFAULT.
The most useful command in this is scylla ptr, as it allows
determining whether the address the current function is working with
belongs to a live object or not.
Once the immediate cause is found, only the "easy" part remains, finding out how it happened. In some cases this can be very difficult. For example in the case of a memory corruption overwriting memory belonging to another object, the overwrite could have happened much earlier, with no traces of what it was in the coredump. In this case the same method has to be used that was mentioned in the case of debugging assert failures: adding additional debug code and trying to reproduce, hoping to catch the perpetrator red-handed.
If the process that is stuck is known, start from there. Try to identify the continuation-chain that is stuck, then follow it to find the future that is blocking it all. There is no way to differentiate a stuck continuation chain from one that is making progress unfortunately, so there are not tried-and-proven methods here either.
OOM crashes are usually the hardest to debug issues. Not only one has to determine the immediate cause which is often already hard enough, as usual one also has to determine what lead to this state, how did it happen.
That said, finding the immediate cause has a pretty standard procedure.
The first step is always issuing a scylla memory command and
determining where the memory is. Lets look at a concrete example:
(gdb) scylla memory
Used memory: 7452069888
Free memory: 20082688
Total memory: 7472152576
LSA:
allocated: 1067712512
used: 1065353216
free: 2359296
Cache:
total: 393216
used: 160704
free: 232512
Memtables:
total: 1067319296
Regular:
real dirty: 1064566784
unspooled: 811568656
System:
real dirty: 393216
unspooled: 393216
Streaming:
real dirty: 0
unspooled: 0
Coordinator:
bg write bytes: 42133 B
hints: 0 B
view hints: 0 B
00 "main"
fg writes: 0
bg writes: 0
fg reads: 0
bg reads: -7
05 "statement"
fg writes: 14
bg writes: 5
fg reads: 94
bg reads: 2352
Replica:
Read Concurrency Semaphores:
user sstable reads: 84/100, remaining mem: 138033377 B, queued: 0
streaming sstable reads: 0/ 10, remaining mem: 149443051 B, queued: 0
system sstable reads: 0/ 10, remaining mem: 149443051 B, queued: 0
Execution Stages:
data query stage:
Total 0
mutation query stage:
Total 0
apply stage:
02 "streaming" 287
Total 287
Tables - Ongoing Operations:
pending writes phaser (top 10):
12 cqlstress_lwt_example.blogposts
2 system.paxos
14 Total (all)
pending reads phaser (top 10):
1863 system.paxos
809 cqlstress_lwt_example.blogposts
2672 Total (all)
pending streams phaser (top 10):
0 Total (all)
Small pools:
objsz spansz usedobj memory unused wst%
1 4096 0 0 0 0.0
1 4096 0 0 0 0.0
1 4096 0 0 0 0.0
1 4096 0 0 0 0.0
2 4096 0 0 0 0.0
2 4096 0 0 0 0.0
3 4096 0 0 0 0.0
3 4096 0 0 0 0.0
4 4096 0 0 0 0.0
5 4096 0 0 0 0.0
6 4096 0 0 0 0.0
7 4096 0 0 0 0.0
8 4096 15285 126976 4696 3.7
10 4096 0 8192 8192 99.9
12 4096 173 8192 6116 74.6
14 4096 0 8192 8192 99.8
16 4096 11151 184320 5904 1.0
20 4096 3570 77824 6424 7.9
24 4096 19131 462848 3704 0.4
28 4096 2572 77824 5808 7.3
32 4096 27021 868352 3680 0.4
40 4096 14680 593920 6720 0.7
48 4096 3318 163840 4576 2.4
56 4096 12077 692224 15912 0.9
64 4096 52719 3375104 1088 0.0
80 4096 16382 1323008 12448 0.6
96 4096 17045 1667072 30752 0.3
112 4096 3402 397312 16288 2.5
128 4096 17767 2281472 7296 0.3
160 4096 17722 2912256 76736 0.3
192 4096 8094 1585152 31104 0.4
224 4096 17087 3891200 63712 0.1
256 4096 77945 21274624 1320704 0.1
320 8192 13232 4366336 132096 0.7
384 8192 5571 2203648 64384 1.0
448 4096 4290 1986560 64640 1.7
512 4096 2830 1503232 54272 2.8
640 12288 960 655360 40960 3.9
768 12288 5751 4489216 72448 0.1
896 8192 326 311296 19200 4.6
1024 4096 4320 5677056 1253376 0.7
1280 20480 251 425984 104704 22.2
1536 12288 3818 6373376 508928 1.7
1792 16384 2711 4980736 122624 0.9
2048 8192 594 1343488 126976 9.5
2560 20480 122 458752 146432 25.7
3072 12288 6596 21823488 1560576 0.9
3584 28672 6 294912 273408 91.1
4096 16384 2039 8372224 20480 0.2
5120 20480 7885 43220992 2849792 0.3
6144 24576 8188 54099968 3792896 0.8
7168 28672 30 622592 407552 53.0
8192 32768 8091 66781184 499712 0.7
10240 40960 15058 165216256 11022336 0.4
12288 49152 7034 92471296 6037504 0.3
14336 57344 6815 111935488 14235648 0.2
16384 65536 14046 230555648 425984 0.2
Small allocations: 872148992 [B]
Page spans:
index size [B] free [B] large [B] [spans]
0 4096 1888256 0 0
1 8192 663552 0 0
2 16384 0 0 0
3 32768 32768 2320334848 70811
4 65536 65536 3031105536 46251
5 131072 131072 1161822208 8864
6 262144 3145728 0 0
7 524288 524288 0 0
8 1048576 1048576 0 0
9 2097152 0 2097152 1
10 4194304 12582912 0 0
11 8388608 0 0 0
12 16777216 0 0 0
13 33554432 0 0 0
14 67108864 0 67108864 1
15 134217728 0 0 0
16 268435456 0 0 0
17 536870912 0 0 0
18 1073741824 0 0 0
19 2147483648 0 0 0
20 4294967296 0 0 0
21 8589934592 0 0 0
22 17179869184 0 0 0
23 34359738368 0 0 0
24 68719476736 0 0 0
25 137438953472 0 0 0
26 274877906944 0 0 0
27 549755813888 0 0 0
28 1099511627776 0 0 0
29 2199023255552 0 0 0
30 4398046511104 0 0 0
31 8796093022208 0 0 0
Large allocations: 6582468608 [B]
We can see a couple of things at glance here: free memory is very low,
cache is fully evicted. These are sure signs of a real OOM. Note that
free memory doesn't have to be 0 in the case of an OOM. It is enough for
a size pool to not be able to allocate more memory spans and thus fail
a critical allocations we cannot recover from. Also cache is fully
evicted doesn't mean it has 0 memory, but when it has just a couple of
KB, it is considered fully evicted. Cache is evicted by the seastar memory
allocator's memory reclamation mechanism, which is hooked up with the
cache and will start trying to free up memory by evicting the cache,
once memory runs low.
The cause of the OOM in this case is too many reads (1863) on
system.paxos. This can be seen in the replica section of the report.
The Coordinator and replica sections contain high level stats of the
state of the coordinator and replica respectively. These stats summarize
the usual suspects. Sometimes just looking at these is enough to
determine what is the cause of the OOM. If not, one has to look at the
last section: the dump of the state of the small pools and the page
spans. What we are looking for is a small pool or a span size that owns
excessive amounts of memory. Once found (there can be more than one) the
next task is to identify what the objects owning that memory are. Note
that in the case of smaller allocations, the memory is usually occupied
directly by some C++ object, why in the case of larger allocations, these are
usually potentially fragmented buffers, owned by some other object.
If the scylla memory output alone is not enough to explain what
exactly is eating up all the memory, there are some further usual
suspects that should be examined.
Look for exploded task- and smp-queues with:
scylla task-queues # lists all task queues on the local shard
scylla smp-queues # lists smp queues
Look for lots of objects with:
scylla task_histogram -a # histogram of all objects with a vtable
A huge number in any of these reports can indicate problems of exploding concurrency, or a shard not being to keep up. This can easily lead to work accumulating, in the form of tasks and associated objects, to the point of OOM.
Another usual suspect is the number of sstables. This can be queried via
scylla sstables. A large number of sstables for a single table (in the
hundreds or more) can cause an otherwise non-problematic amount of reads to use
excessive amount of memory, potentially leading to OOM.
Reversed- and unpaged-reads (or both, combined) can also consume a huge amount of memory, to the point of a few of such reads causing OOM. The way to find these is to inspect readers in memory, trying to locate their partition slice and having a look at their respective options:
partition_slice::option::reversedis set for a reversed querypartition_slice::option::allow_short_readis cleared for an unpaged query
Note that scylla have protections against reverse queries since 4.0, and against unpaged queries since 4.3.
If none of the usual suspects are present then all bets are off and one has to try to identify who the objects of the exploded size-class or span-size belong to. Unfortunately there are no proven methods here: some try to inspect the memory patterns and try to make sense of it, some try to build an object graph out of these objects and make sense of that. For the latter, the following commands might be of help:
scylla small-objects # lists objects from a small pool
scylla generate-object-graph # generates and visualizes an object graph
Good luck, you are off the charted path here.
GDB is well integrated with Python, and besides using the predefined set of
commands from scylla-gdb.py, you can also run an interactive prompt by typing
python-interactive or pi in gdb console. That allows you to benefit from
the rich set of helper classes and wrappers defined in scylla-gdb.py, e.g. in
order to iterate over one of the supported collection types (vectors, small
vectors, maps, tuples, etc.) and print only the interesting bits.
Example:
(gdb) source /path/to/scylla-gdb.py
(gdb) python-interactive
>>> db = gdb.parse_and_eval("*(seastar::sharded<replica::database>*)debug::the_database")
>>> instances = std_vector(db["_instances"])
>>> for i, instance in enumerate(instances):
... print(i, instance["service"])
...
0 {
_b = 0x60000403e000,
_p = 0x60000403e010
}
1 {
_b = 0x60100435e000,
_p = 0x60100435e010
}
You can also easily evaluate single expressions by using the python (or py) command:
(gdb) py print("\n".join([f" I/O queue {key}: {value}" for key, value in std_unordered_map(next(reactors())["_io_queues"])]))
I/O queue 0: std::unique_ptr<seastar::io_queue> = {
get() = 0x60100012ae00
}
In order to define and use variables between gdb and the Python interface,
one can use gdb.convenience_variable() and gdb.set_convenience_variable()
helpers. That can be clearer than relying on gdb.parse_and_eval() for
accessing objects.
Example:
(gdb) p debug::the_database
$1 = (seastar::sharded<replica::database> *) 0x7fffffffbdc0
(gdb) set $db=debug::the_database
(gdb) pi
>>> local_db = sharded(gdb.convenience_variable('db')).local()
>>> gdb.set_convenience_variable('local_db', local_db)
>>>
(gdb) p $local_db
$2 = (replica::database *) 0x60000575e010