On Mon, 30 Oct 2023 12:42:23 -0400 Steven Rostedt rostedt@goodmis.org wrote:
I still get the splat about the trace_array_put when running user_event's ftrace selftest:
[ 26.665931] ------------[ cut here ]------------ [ 26.666663] WARNING: CPU: 12 PID: 291 at kernel/trace/trace.c:516 tracing_release_file_tr+0x46/0x50 [ 26.667470] Modules linked in: [ 26.667808] CPU: 12 PID: 291 Comm: ftrace_test Not tainted 6.6.0-rc7-next-20231026 #3 [ 26.668665] RIP: 0010:tracing_release_file_tr+0x46/0x50 [ 26.669093] Code: d1 03 01 8b 83 c0 1e 00 00 85 c0 74 1d 83 e8 01 48 c7 c7 80 5b ef bc 89 83 c0 1e 00 00 e8 f2 b5 03 01 31 c0 5b e9 75 ee 27 01 <0f> 0b eb df 66 0f 1f 44 00 00 90 90 90 90 90 90 90 90 90 90 90 90 [ 26.670580] RSP: 0018:ffffb6ef858ffee8 EFLAGS: 00010246 [ 26.671128] RAX: 0000000000000000 RBX: ffff9d7ae2364058 RCX: 0000000000000000 [ 26.671793] RDX: 0000000000000000 RSI: ffffffffbcb6b38b RDI: 00000000ffffffff [ 26.672444] RBP: ffff9d7ac3e72200 R08: 0000000000000000 R09: 0000000000000000 [ 26.673072] R10: ffffb6ef858ffee8 R11: ffffffffbb28526f R12: 00000000000f801f [ 26.673705] R13: ffff9d7b661a2020 R14: ffff9d7ac6057728 R15: 0000000000000000 [ 26.674339] FS: 00007fa852fa6740(0000) GS:ffff9d81a6300000(0000) knlGS:0000000000000000 [ 26.674978] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 26.675506] CR2: 00007fa852c2a250 CR3: 0000000105d92001 CR4: 0000000000370eb0 [ 26.676142] Call Trace: [ 26.676357] <TASK> [ 26.676572] ? __warn+0x7f/0x160 [ 26.677092] ? tracing_release_file_tr+0x46/0x50 [ 26.677540] ? report_bug+0x1c3/0x1d0 [ 26.677871] ? handle_bug+0x3c/0x70 [ 26.678196] ? exc_invalid_op+0x14/0x70 [ 26.678520] ? asm_exc_invalid_op+0x16/0x20 [ 26.678845] ? tracing_release_file_tr+0x1f/0x50 [ 26.679268] ? tracing_release_file_tr+0x46/0x50 [ 26.679691] ? tracing_release_file_tr+0x1f/0x50 [ 26.680105] __fput+0xab/0x300 [ 26.680437] __x64_sys_close+0x38/0x80
Hmm, this doesn't tell me much. Let me go play with the user_event self tests.
I added a bunch of printk()s and I'm thinking there's a race in user event (or dynamic event) code.
I put a printk in the open, write and release call to record the filp and the file in the open and this is what I hit:
[ 32.603954] open ffff8d05488bf000 file=ffff8d0484f7a688 [ 32.607026] write ffff8d05488bf000 [ 32.608829] update file = ffff8d0484f7a688 [ 32.610100] update tr = ffffffffb2bebda0 [ 32.622203] write ffff8d05488bf000 [ 32.623231] update file = ffff8d0484f7a688 [ 32.624397] update tr = ffffffffb2bebda0 [ 32.625975] call delayed destroy [ 32.627241] open ffff8d048510fc00 file=ffff8d0484f7a688
Another open with a different filp, but has the same file pointer (which is the meta data that matches the eventfs files, but not part of eventfs).
[ 32.628720] release ffff8d048510fc00
Only one release is called and then we call:
[ 32.630785] Remove event call ffff8d04809d7c58
The above "Remove event call" came from user_event_set_call_visible()
if (visible) { printk("show event call %px\n", &user->call); ret = trace_add_event_call(&user->call); } else { printk("Remove event call %px\n", &user->call); ret = trace_remove_event_call(&user->call); }
Where trace_remove_event_call() calls:
probe_remove_event_call() { __trace_remove_event_call() { event_remove() { remove_event_from_tracers() { remove_event_file_dir() { remove_event_file_dir() { kmem_cache_free(file_cachep, file);
That is, call->file is freed at this point. Now any access to the file pointer is going to be garbage, which a write to enable will cause.
So I see it is freed without seeing a release called.
[ 32.632323] Removing __test_event
The above is the dentry being released (unlinked).
[ 32.633618] Removing enable [ 32.634453] Removing user_events [ 32.634948] write ffff8d05488bf000
Another write is happening to the user event file (it may have been unlinked, but the release was never called, so it is still valid.
[ 32.636440] update file = ffff8d0484f7a688 [ 32.637685] update tr = dc64cc323d943921
The above is the file pointer that we freed, and you can see the file->tr is now garbage.
[ 32.638827] general protection fault, probably for non-canonical address 0xdc64cc323d9457f9: 0000 [#1] PREEMPT SMP PTI [ 32.641712] CPU: 4 PID: 911 Comm: ftrace_test Not tainted 6.6.0-rc4-test-00024-gd402dc722a1b-dirty #158 [ 32.644220] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.16.2-debian-1.16.2-1 04/01/2014 [ 32.646717] RIP: 0010:tracing_update_buffers+0x19/0x50 [ 32.648133] Code: 90 90 90 90 90 90 90 90 90 90 90 90 90 90 90 90 90 66 0f 1f 00 55 31 f6 31 ed 53 48 89 fb 48 c7 c7 e0 bc be b2 e8 d7 b6 cb 00 <80> bb d8 1e 00 00 00 74 15 48 c7 c7 e0 bc be b2 e8 42 9f cb 00 89 [ 32.653125] RSP: 0018:ffffb7b34172be20 EFLAGS: 00010246 [ 32.654573] RAX: 0000000000000000 RBX: dc64cc323d943921 RCX: 0000000000000000 [ 32.656503] RDX: 0000000000000000 RSI: ffffffffb1e09128 RDI: ffffffffb1e09128 [ 32.658445] RBP: 0000000000000000 R08: 000000000000002f R09: 0000000000000000 [ 32.660378] R10: ffffb7b34172be20 R11: 0000000000000001 R12: ffffb7b34172bf08 [ 32.662322] R13: ffff8d0484f7a688 R14: ffffb7b34172bf08 R15: 0000000000000000 [ 32.664249] FS: 00007f1769a10740(0000) GS:ffff8d05f7c00000(0000) knlGS:0000000000000000 [ 32.666438] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033 [ 32.668009] CR2: 000055c0cd72fc70 CR3: 000000010e3c6003 CR4: 0000000000170ee0 [ 32.669954] Call Trace: [ 32.670653] <TASK> [ 32.671263] ? die_addr+0x36/0x90 [ 32.672203] ? exc_general_protection+0x236/0x4a0 [ 32.673521] ? asm_exc_general_protection+0x26/0x30 [ 32.674872] ? __mutex_lock+0x1c8/0xb40 [ 32.675942] ? __mutex_lock+0x1c8/0xb40 [ 32.677009] ? tracing_update_buffers+0x19/0x50 [ 32.678259] event_enable_write+0xb4/0x140 [ 32.679311] vfs_write+0xf2/0x530 [ 32.680172] ? find_held_lock+0x2b/0x80 [ 32.681159] ? _raw_spin_unlock+0x2d/0x50 [ 32.682711] ? rcu_is_watching+0x11/0x50 [ 32.684159] ? _raw_spin_unlock+0x2d/0x50 [ 32.685642] ? trace_preempt_on+0x78/0x80 [ 32.687027] ksys_write+0x75/0x100 [ 32.688257] do_syscall_64+0x3f/0xc0 [ 32.689561] entry_SYSCALL_64_after_hwframe+0x6e/0xd8 [ 32.691048] RIP: 0033:0x7f1769b0ab00
Note, I think I can simplify some of the code (but not fix this bug) by letting the dput free the eventfs_inode as well (I think that is required). But the above looks to be a bug in the implementation of user_events.
-- Steve