Message ID | 20201211171138.63819-2-jonathan.lemon@gmail.com (mailing list archive) |
---|---|
State | Changes Requested |
Delegated to: | BPF |
Headers | show |
Series | bpf: increment and use correct thread iterator | expand |
Context | Check | Description |
---|---|---|
netdev/cover_letter | success | Link |
netdev/fixes_present | success | Link |
netdev/patch_count | success | Link |
netdev/tree_selection | success | Clearly marked for bpf-next |
netdev/subject_prefix | success | Link |
netdev/source_inline | success | Was 0 now: 0 |
netdev/verify_signedoff | success | Link |
netdev/module_param | success | Was 0 now: 0 |
netdev/build_32bit | success | Errors and warnings before: 1 this patch: 1 |
netdev/kdoc | success | Errors and warnings before: 0 this patch: 0 |
netdev/verify_fixes | fail | Link |
netdev/checkpatch | warning | WARNING: From:/Signed-off-by: email address mismatch: 'From: Jonathan Lemon <bsd@fb.com>' != 'Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com>' WARNING: Unknown commit id '67b6b863e6ab', maybe rebased or not pulled? WARNING: Unknown commit id 'a650da2ee52a', maybe rebased or not pulled? |
netdev/build_allmodconfig_warn | success | Errors and warnings before: 1 this patch: 1 |
netdev/header_inline | success | Link |
netdev/stable | success | Stable not CCed |
On Fri, Dec 11, 2020 at 10:56 AM Jonathan Lemon <jonathan.lemon@gmail.com> wrote: > > From: Jonathan Lemon <bsd@fb.com> > > On some systems, some variant of the following splat is > repeatedly seen. The common factor in all traces seems > to be the entry point to task_file_seq_next(). With the > patch, all warnings go away. > > rcu: INFO: rcu_sched self-detected stall on CPU > rcu: \x0926-....: (20992 ticks this GP) idle=d7e/1/0x4000000000000002 softirq=81556231/81556231 fqs=4876 > \x09(t=21033 jiffies g=159148529 q=223125) > NMI backtrace for cpu 26 > CPU: 26 PID: 2015853 Comm: bpftool Kdump: loaded Not tainted 5.6.13-0_fbk4_3876_gd8d1f9bf80bb #1 > Hardware name: Quanta Twin Lakes MP/Twin Lakes Passive MP, BIOS F09_3A12 10/08/2018 > Call Trace: > <IRQ> > dump_stack+0x50/0x70 > nmi_cpu_backtrace.cold.6+0x13/0x50 > ? lapic_can_unplug_cpu.cold.30+0x40/0x40 > nmi_trigger_cpumask_backtrace+0xba/0xca > rcu_dump_cpu_stacks+0x99/0xc7 > rcu_sched_clock_irq.cold.90+0x1b4/0x3aa > ? tick_sched_do_timer+0x60/0x60 > update_process_times+0x24/0x50 > tick_sched_timer+0x37/0x70 > __hrtimer_run_queues+0xfe/0x270 > hrtimer_interrupt+0xf4/0x210 > smp_apic_timer_interrupt+0x5e/0x120 > apic_timer_interrupt+0xf/0x20 > </IRQ> > RIP: 0010:get_pid_task+0x38/0x80 > Code: 89 f6 48 8d 44 f7 08 48 8b 00 48 85 c0 74 2b 48 83 c6 55 48 c1 e6 04 48 29 f0 74 19 48 8d 78 20 ba 01 00 00 00 f0 0f c1 50 20 <85> d2 74 27 78 11 83 c2 01 78 0c 48 83 c4 08 c3 31 c0 48 83 c4 08 > RSP: 0018:ffffc9000d293dc8 EFLAGS: 00000202 ORIG_RAX: ffffffffffffff13 > RAX: ffff888637c05600 RBX: ffffc9000d293e0c RCX: 0000000000000000 > RDX: 0000000000000001 RSI: 0000000000000550 RDI: ffff888637c05620 > RBP: ffffffff8284eb80 R08: ffff88831341d300 R09: ffff88822ffd8248 > R10: ffff88822ffd82d0 R11: 00000000003a93c0 R12: 0000000000000001 > R13: 00000000ffffffff R14: ffff88831341d300 R15: 0000000000000000 > ? find_ge_pid+0x1b/0x20 > task_seq_get_next+0x52/0xc0 > task_file_seq_get_next+0x159/0x220 > task_file_seq_next+0x4f/0xa0 > bpf_seq_read+0x159/0x390 > vfs_read+0x8a/0x140 > ksys_read+0x59/0xd0 > do_syscall_64+0x42/0x110 > entry_SYSCALL_64_after_hwframe+0x44/0xa9 > RIP: 0033:0x7f95ae73e76e > Code: Bad RIP value. > RSP: 002b:00007ffc02c1dbf8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000 > RAX: ffffffffffffffda RBX: 000000000170faa0 RCX: 00007f95ae73e76e > RDX: 0000000000001000 RSI: 00007ffc02c1dc30 RDI: 0000000000000007 > RBP: 00007ffc02c1ec70 R08: 0000000000000005 R09: 0000000000000006 > R10: fffffffffffff20b R11: 0000000000000246 R12: 00000000019112a0 > R13: 0000000000000000 R14: 0000000000000007 R15: 00000000004283c0 > > The attached patch does 3 things: > > 1) If unable to obtain the file structure for the current task, > proceed to the next task number after the one returned from > task_seq_get_next(), instead of the next task number from the > original iterator. > > 2) Use thread_group_leader() instead of the open-coded comparision > of tgid vs pid. > > 3) Only obtain the task reference count at the end of the RCU section > instead of repeatedly obtaining/releasing it when iterathing though > a thread group. > > Fixes: a650da2ee52a ("bpf: Add task and task/file iterator targets") > Fixes: 67b6b863e6ab ("bpf: Avoid iterating duplicated files for task_file iterator") > > Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com> > --- > kernel/bpf/task_iter.c | 8 ++++---- > 1 file changed, 4 insertions(+), 4 deletions(-) > > diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c > index 0458a40edf10..66a52fcf589a 100644 > --- a/kernel/bpf/task_iter.c > +++ b/kernel/bpf/task_iter.c > @@ -33,17 +33,17 @@ static struct task_struct *task_seq_get_next(struct pid_namespace *ns, > pid = find_ge_pid(*tid, ns); > if (pid) { > *tid = pid_nr_ns(pid, ns); > - task = get_pid_task(pid, PIDTYPE_PID); > + task = pid_task(pid, PIDTYPE_PID); > if (!task) { > ++*tid; > goto retry; > - } else if (skip_if_dup_files && task->tgid != task->pid && > + } else if (skip_if_dup_files && !thread_group_leader(task) && > task->files == task->group_leader->files) { > - put_task_struct(task); > task = NULL; > ++*tid; > goto retry; > } > + get_task_struct(task); > } This part looks good. I'd say it deserves a separate patch, but it's minor. > rcu_read_unlock(); > > @@ -164,7 +164,7 @@ task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info) > curr_files = get_files_struct(curr_task); > if (!curr_files) { > put_task_struct(curr_task); > - curr_tid = ++(info->tid); > + curr_tid = curr_tid + 1; Yonghong might know definitively, but it seems like we need to update info->tid here as well: info->tid = curr_tid; If the search eventually yields no task, then info->tid will stay at some potentially much smaller value, and we'll keep re-searching tasks from the same TID on each subsequent read (if user keeps reading the file). So corner case, but good to have covered. > info->fd = 0; > goto again; > } > -- > 2.24.1 >
On Fri, Dec 11, 2020 at 12:23:34PM -0800, Andrii Nakryiko wrote: > > @@ -164,7 +164,7 @@ task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info) > > curr_files = get_files_struct(curr_task); > > if (!curr_files) { > > put_task_struct(curr_task); > > - curr_tid = ++(info->tid); > > + curr_tid = curr_tid + 1; > > Yonghong might know definitively, but it seems like we need to update > info->tid here as well: > > info->tid = curr_tid; > > If the search eventually yields no task, then info->tid will stay at > some potentially much smaller value, and we'll keep re-searching tasks > from the same TID on each subsequent read (if user keeps reading the > file). So corner case, but good to have covered. That applies earlier as well: curr_task = task_seq_get_next(ns, &curr_tid, true); if (!curr_task) { info->task = NULL; info->files = NULL; return NULL; } The logic seems to be "if task == NULL, then return NULL and stop". Is the seq_iterator allowed to continue/restart if seq_next returns NULL? -- Jonathan
On Fri, Dec 11, 2020 at 3:01 PM Jonathan Lemon <jonathan.lemon@gmail.com> wrote: > > On Fri, Dec 11, 2020 at 12:23:34PM -0800, Andrii Nakryiko wrote: > > > @@ -164,7 +164,7 @@ task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info) > > > curr_files = get_files_struct(curr_task); > > > if (!curr_files) { > > > put_task_struct(curr_task); > > > - curr_tid = ++(info->tid); > > > + curr_tid = curr_tid + 1; > > > > Yonghong might know definitively, but it seems like we need to update > > info->tid here as well: > > > > info->tid = curr_tid; > > > > If the search eventually yields no task, then info->tid will stay at > > some potentially much smaller value, and we'll keep re-searching tasks > > from the same TID on each subsequent read (if user keeps reading the > > file). So corner case, but good to have covered. > > That applies earlier as well: > > curr_task = task_seq_get_next(ns, &curr_tid, true); > if (!curr_task) { > info->task = NULL; > info->files = NULL; > return NULL; > } > True, info->tid = curr_tid + 1; seems to be needed here? > The logic seems to be "if task == NULL, then return NULL and stop". > Is the seq_iterator allowed to continue/restart if seq_next returns NULL? I don't think we allow seeking, so no restarts. But nothing will prevent the user to keep calling read() after it returns 0 byte, so yes, continuation is possible. > -- > Jonathan
On 12/11/20 3:01 PM, Jonathan Lemon wrote: > On Fri, Dec 11, 2020 at 12:23:34PM -0800, Andrii Nakryiko wrote: >>> @@ -164,7 +164,7 @@ task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info) >>> curr_files = get_files_struct(curr_task); >>> if (!curr_files) { >>> put_task_struct(curr_task); >>> - curr_tid = ++(info->tid); >>> + curr_tid = curr_tid + 1; >> >> Yonghong might know definitively, but it seems like we need to update >> info->tid here as well: >> >> info->tid = curr_tid; >> >> If the search eventually yields no task, then info->tid will stay at >> some potentially much smaller value, and we'll keep re-searching tasks >> from the same TID on each subsequent read (if user keeps reading the >> file). So corner case, but good to have covered. > > That applies earlier as well: > > curr_task = task_seq_get_next(ns, &curr_tid, true); > if (!curr_task) { > info->task = NULL; > info->files = NULL; > return NULL; > } > > The logic seems to be "if task == NULL, then return NULL and stop". > Is the seq_iterator allowed to continue/restart if seq_next returns NULL? If seq_next() returns NULL, bpf_seq_read() will end and the control will return to user space. There are two cases here: - there are something in the buffer and user will get non-zero-length return data and after this typically user will call read() syscall again. In such cases case, the search will be from last info->tid. - the buffer is empty and user will get a "0" return value for read() system. Typically, user will not call read() syscall any more. But if it does, the search will start from last info->tid. Agree with Andrii, in general, this should not be a big problem. But it is good to get this fixed too. > -- > Jonathan >
On 12/11/20 9:11 AM, Jonathan Lemon wrote: > From: Jonathan Lemon <bsd@fb.com> > > On some systems, some variant of the following splat is > repeatedly seen. The common factor in all traces seems > to be the entry point to task_file_seq_next(). With the > patch, all warnings go away. > > rcu: INFO: rcu_sched self-detected stall on CPU > rcu: \x0926-....: (20992 ticks this GP) idle=d7e/1/0x4000000000000002 softirq=81556231/81556231 fqs=4876 > \x09(t=21033 jiffies g=159148529 q=223125) > NMI backtrace for cpu 26 > CPU: 26 PID: 2015853 Comm: bpftool Kdump: loaded Not tainted 5.6.13-0_fbk4_3876_gd8d1f9bf80bb #1 > Hardware name: Quanta Twin Lakes MP/Twin Lakes Passive MP, BIOS F09_3A12 10/08/2018 > Call Trace: > <IRQ> > dump_stack+0x50/0x70 > nmi_cpu_backtrace.cold.6+0x13/0x50 > ? lapic_can_unplug_cpu.cold.30+0x40/0x40 > nmi_trigger_cpumask_backtrace+0xba/0xca > rcu_dump_cpu_stacks+0x99/0xc7 > rcu_sched_clock_irq.cold.90+0x1b4/0x3aa > ? tick_sched_do_timer+0x60/0x60 > update_process_times+0x24/0x50 > tick_sched_timer+0x37/0x70 > __hrtimer_run_queues+0xfe/0x270 > hrtimer_interrupt+0xf4/0x210 > smp_apic_timer_interrupt+0x5e/0x120 > apic_timer_interrupt+0xf/0x20 > </IRQ> > RIP: 0010:get_pid_task+0x38/0x80 > Code: 89 f6 48 8d 44 f7 08 48 8b 00 48 85 c0 74 2b 48 83 c6 55 48 c1 e6 04 48 29 f0 74 19 48 8d 78 20 ba 01 00 00 00 f0 0f c1 50 20 <85> d2 74 27 78 11 83 c2 01 78 0c 48 83 c4 08 c3 31 c0 48 83 c4 08 > RSP: 0018:ffffc9000d293dc8 EFLAGS: 00000202 ORIG_RAX: ffffffffffffff13 > RAX: ffff888637c05600 RBX: ffffc9000d293e0c RCX: 0000000000000000 > RDX: 0000000000000001 RSI: 0000000000000550 RDI: ffff888637c05620 > RBP: ffffffff8284eb80 R08: ffff88831341d300 R09: ffff88822ffd8248 > R10: ffff88822ffd82d0 R11: 00000000003a93c0 R12: 0000000000000001 > R13: 00000000ffffffff R14: ffff88831341d300 R15: 0000000000000000 > ? find_ge_pid+0x1b/0x20 > task_seq_get_next+0x52/0xc0 > task_file_seq_get_next+0x159/0x220 > task_file_seq_next+0x4f/0xa0 > bpf_seq_read+0x159/0x390 > vfs_read+0x8a/0x140 > ksys_read+0x59/0xd0 > do_syscall_64+0x42/0x110 > entry_SYSCALL_64_after_hwframe+0x44/0xa9 > RIP: 0033:0x7f95ae73e76e > Code: Bad RIP value. > RSP: 002b:00007ffc02c1dbf8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000 > RAX: ffffffffffffffda RBX: 000000000170faa0 RCX: 00007f95ae73e76e > RDX: 0000000000001000 RSI: 00007ffc02c1dc30 RDI: 0000000000000007 > RBP: 00007ffc02c1ec70 R08: 0000000000000005 R09: 0000000000000006 > R10: fffffffffffff20b R11: 0000000000000246 R12: 00000000019112a0 > R13: 0000000000000000 R14: 0000000000000007 R15: 00000000004283c0 > > The attached patch does 3 things: > > 1) If unable to obtain the file structure for the current task, > proceed to the next task number after the one returned from > task_seq_get_next(), instead of the next task number from the > original iterator. Looks like this fix is the real fix for the above warnings. Basically, say we have info->tid = 10 and returned curr_tid = 3000 and tid 3000 has no files. the current logic will go through - set curr_tid = 11 (info->tid++) and returned curr_tid = 3000 - set curr_tid = 12 and returned curr_tid = 3000 ... - set curr_tid = 3000 and returned curr_tid = 3000 - set curr_tid = 3001 and return curr_tid >= 3001 All the above works are redundant work, and it may cause issues for non preemptable kernel. I suggest you factor out this change plus the following change which suggested by Andrii early to a separate patch carried with the below Fixes tag. diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c index 0458a40edf10..56bcaef72e36 100644 --- a/kernel/bpf/task_iter.c +++ b/kernel/bpf/task_iter.c @@ -158,6 +158,7 @@ task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info) if (!curr_task) { info->task = NULL; info->files = NULL; + info->tid = curr_tid + 1; return NULL; } > > 2) Use thread_group_leader() instead of the open-coded comparision > of tgid vs pid. My experiment show there is no difference between thread_group_leader() and comparing tgid/pid, but indeed using existing function is better, so I am okay with this. > > 3) Only obtain the task reference count at the end of the RCU section > instead of repeatedly obtaining/releasing it when iterathing though > a thread group. The above two changes are not really fixing the rcu warnings, but they are nice to have indeed. Could you put it into separate patch (patch 2)? You can add the following improvement change as well -bash-4.4$ git diff diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c index 0458a40edf10..f61e5ddb38ce 100644 --- a/kernel/bpf/task_iter.c +++ b/kernel/bpf/task_iter.c @@ -148,12 +148,12 @@ task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info) * it held a reference to the task/files_struct/file. * Otherwise, it does not hold any reference. */ -again: if (info->task) { curr_task = info->task; curr_files = info->files; curr_fd = info->fd; } else { +again: curr_task = task_seq_get_next(ns, &curr_tid, true); if (!curr_task) { info->task = NULL; to reduce one branch for searching next task in task_file_seq_get_next() function. > > Fixes: a650da2ee52a ("bpf: Add task and task/file iterator targets") > Fixes: 67b6b863e6ab ("bpf: Avoid iterating duplicated files for task_file iterator") > > Signed-off-by: Jonathan Lemon <jonathan.lemon@gmail.com> > --- > kernel/bpf/task_iter.c | 8 ++++---- > 1 file changed, 4 insertions(+), 4 deletions(-) > > diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c > index 0458a40edf10..66a52fcf589a 100644 > --- a/kernel/bpf/task_iter.c > +++ b/kernel/bpf/task_iter.c > @@ -33,17 +33,17 @@ static struct task_struct *task_seq_get_next(struct pid_namespace *ns, > pid = find_ge_pid(*tid, ns); > if (pid) { > *tid = pid_nr_ns(pid, ns); > - task = get_pid_task(pid, PIDTYPE_PID); > + task = pid_task(pid, PIDTYPE_PID); > if (!task) { > ++*tid; > goto retry; > - } else if (skip_if_dup_files && task->tgid != task->pid && > + } else if (skip_if_dup_files && !thread_group_leader(task) && > task->files == task->group_leader->files) { > - put_task_struct(task); > task = NULL; > ++*tid; > goto retry; > } > + get_task_struct(task); > } > rcu_read_unlock(); > > @@ -164,7 +164,7 @@ task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info) > curr_files = get_files_struct(curr_task); > if (!curr_files) { > put_task_struct(curr_task); > - curr_tid = ++(info->tid); > + curr_tid = curr_tid + 1; > info->fd = 0; > goto again; > } >
On Fri, Dec 18, 2020 at 08:53:22AM -0800, Yonghong Song wrote: > > > On 12/11/20 9:11 AM, Jonathan Lemon wrote: > > From: Jonathan Lemon <bsd@fb.com> > > > > On some systems, some variant of the following splat is > > repeatedly seen. The common factor in all traces seems > > to be the entry point to task_file_seq_next(). With the > > patch, all warnings go away. > > > > rcu: INFO: rcu_sched self-detected stall on CPU > > rcu: \x0926-....: (20992 ticks this GP) idle=d7e/1/0x4000000000000002 softirq=81556231/81556231 fqs=4876 > > \x09(t=21033 jiffies g=159148529 q=223125) > > NMI backtrace for cpu 26 > > CPU: 26 PID: 2015853 Comm: bpftool Kdump: loaded Not tainted 5.6.13-0_fbk4_3876_gd8d1f9bf80bb #1 > > Hardware name: Quanta Twin Lakes MP/Twin Lakes Passive MP, BIOS F09_3A12 10/08/2018 > > Call Trace: > > <IRQ> > > dump_stack+0x50/0x70 > > nmi_cpu_backtrace.cold.6+0x13/0x50 > > ? lapic_can_unplug_cpu.cold.30+0x40/0x40 > > nmi_trigger_cpumask_backtrace+0xba/0xca > > rcu_dump_cpu_stacks+0x99/0xc7 > > rcu_sched_clock_irq.cold.90+0x1b4/0x3aa > > ? tick_sched_do_timer+0x60/0x60 > > update_process_times+0x24/0x50 > > tick_sched_timer+0x37/0x70 > > __hrtimer_run_queues+0xfe/0x270 > > hrtimer_interrupt+0xf4/0x210 > > smp_apic_timer_interrupt+0x5e/0x120 > > apic_timer_interrupt+0xf/0x20 > > </IRQ> > > RIP: 0010:get_pid_task+0x38/0x80 > > Code: 89 f6 48 8d 44 f7 08 48 8b 00 48 85 c0 74 2b 48 83 c6 55 48 c1 e6 04 48 29 f0 74 19 48 8d 78 20 ba 01 00 00 00 f0 0f c1 50 20 <85> d2 74 27 78 11 83 c2 01 78 0c 48 83 c4 08 c3 31 c0 48 83 c4 08 > > RSP: 0018:ffffc9000d293dc8 EFLAGS: 00000202 ORIG_RAX: ffffffffffffff13 > > RAX: ffff888637c05600 RBX: ffffc9000d293e0c RCX: 0000000000000000 > > RDX: 0000000000000001 RSI: 0000000000000550 RDI: ffff888637c05620 > > RBP: ffffffff8284eb80 R08: ffff88831341d300 R09: ffff88822ffd8248 > > R10: ffff88822ffd82d0 R11: 00000000003a93c0 R12: 0000000000000001 > > R13: 00000000ffffffff R14: ffff88831341d300 R15: 0000000000000000 > > ? find_ge_pid+0x1b/0x20 > > task_seq_get_next+0x52/0xc0 > > task_file_seq_get_next+0x159/0x220 > > task_file_seq_next+0x4f/0xa0 > > bpf_seq_read+0x159/0x390 > > vfs_read+0x8a/0x140 > > ksys_read+0x59/0xd0 > > do_syscall_64+0x42/0x110 > > entry_SYSCALL_64_after_hwframe+0x44/0xa9 > > RIP: 0033:0x7f95ae73e76e > > Code: Bad RIP value. > > RSP: 002b:00007ffc02c1dbf8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000 > > RAX: ffffffffffffffda RBX: 000000000170faa0 RCX: 00007f95ae73e76e > > RDX: 0000000000001000 RSI: 00007ffc02c1dc30 RDI: 0000000000000007 > > RBP: 00007ffc02c1ec70 R08: 0000000000000005 R09: 0000000000000006 > > R10: fffffffffffff20b R11: 0000000000000246 R12: 00000000019112a0 > > R13: 0000000000000000 R14: 0000000000000007 R15: 00000000004283c0 > > > > The attached patch does 3 things: > > > > 1) If unable to obtain the file structure for the current task, > > proceed to the next task number after the one returned from > > task_seq_get_next(), instead of the next task number from the > > original iterator. > > Looks like this fix is the real fix for the above warnings. > Basically, say we have > info->tid = 10 and returned curr_tid = 3000 and tid 3000 has no files. > the current logic will go through > - set curr_tid = 11 (info->tid++) and returned curr_tid = 3000 > - set curr_tid = 12 and returned curr_tid = 3000 > ... > - set curr_tid = 3000 and returned curr_tid = 3000 > - set curr_tid = 3001 and return curr_tid >= 3001 > > All the above works are redundant work, and it may cause issues > for non preemptable kernel. > > I suggest you factor out this change plus the following change > which suggested by Andrii early to a separate patch carried with > the below Fixes tag. > > diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c > index 0458a40edf10..56bcaef72e36 100644 > --- a/kernel/bpf/task_iter.c > +++ b/kernel/bpf/task_iter.c > @@ -158,6 +158,7 @@ task_file_seq_get_next(struct > bpf_iter_seq_task_file_info *info) > if (!curr_task) { > info->task = NULL; > info->files = NULL; > + info->tid = curr_tid + 1; > return NULL; > } Sure this isn't supposed to be 'curr_tid'? task_seq_get_next() stops when there are no more threads found. This increments the thread id past the search point, and would seem to introduce a potential off-by-one error. That is: curr_tid = 3000. call task_seq_get_next() --> return NULL, curr_tid = 3000. (so there is no tid >= 3000) set curr_tid = 3001. next restart (if there is one) skips a newly created 3000.
On 12/18/20 10:06 AM, Jonathan Lemon wrote: > On Fri, Dec 18, 2020 at 08:53:22AM -0800, Yonghong Song wrote: >> >> >> On 12/11/20 9:11 AM, Jonathan Lemon wrote: >>> From: Jonathan Lemon <bsd@fb.com> >>> >>> On some systems, some variant of the following splat is >>> repeatedly seen. The common factor in all traces seems >>> to be the entry point to task_file_seq_next(). With the >>> patch, all warnings go away. >>> >>> rcu: INFO: rcu_sched self-detected stall on CPU >>> rcu: \x0926-....: (20992 ticks this GP) idle=d7e/1/0x4000000000000002 softirq=81556231/81556231 fqs=4876 >>> \x09(t=21033 jiffies g=159148529 q=223125) >>> NMI backtrace for cpu 26 >>> CPU: 26 PID: 2015853 Comm: bpftool Kdump: loaded Not tainted 5.6.13-0_fbk4_3876_gd8d1f9bf80bb #1 >>> Hardware name: Quanta Twin Lakes MP/Twin Lakes Passive MP, BIOS F09_3A12 10/08/2018 >>> Call Trace: >>> <IRQ> >>> dump_stack+0x50/0x70 >>> nmi_cpu_backtrace.cold.6+0x13/0x50 >>> ? lapic_can_unplug_cpu.cold.30+0x40/0x40 >>> nmi_trigger_cpumask_backtrace+0xba/0xca >>> rcu_dump_cpu_stacks+0x99/0xc7 >>> rcu_sched_clock_irq.cold.90+0x1b4/0x3aa >>> ? tick_sched_do_timer+0x60/0x60 >>> update_process_times+0x24/0x50 >>> tick_sched_timer+0x37/0x70 >>> __hrtimer_run_queues+0xfe/0x270 >>> hrtimer_interrupt+0xf4/0x210 >>> smp_apic_timer_interrupt+0x5e/0x120 >>> apic_timer_interrupt+0xf/0x20 >>> </IRQ> >>> RIP: 0010:get_pid_task+0x38/0x80 >>> Code: 89 f6 48 8d 44 f7 08 48 8b 00 48 85 c0 74 2b 48 83 c6 55 48 c1 e6 04 48 29 f0 74 19 48 8d 78 20 ba 01 00 00 00 f0 0f c1 50 20 <85> d2 74 27 78 11 83 c2 01 78 0c 48 83 c4 08 c3 31 c0 48 83 c4 08 >>> RSP: 0018:ffffc9000d293dc8 EFLAGS: 00000202 ORIG_RAX: ffffffffffffff13 >>> RAX: ffff888637c05600 RBX: ffffc9000d293e0c RCX: 0000000000000000 >>> RDX: 0000000000000001 RSI: 0000000000000550 RDI: ffff888637c05620 >>> RBP: ffffffff8284eb80 R08: ffff88831341d300 R09: ffff88822ffd8248 >>> R10: ffff88822ffd82d0 R11: 00000000003a93c0 R12: 0000000000000001 >>> R13: 00000000ffffffff R14: ffff88831341d300 R15: 0000000000000000 >>> ? find_ge_pid+0x1b/0x20 >>> task_seq_get_next+0x52/0xc0 >>> task_file_seq_get_next+0x159/0x220 >>> task_file_seq_next+0x4f/0xa0 >>> bpf_seq_read+0x159/0x390 >>> vfs_read+0x8a/0x140 >>> ksys_read+0x59/0xd0 >>> do_syscall_64+0x42/0x110 >>> entry_SYSCALL_64_after_hwframe+0x44/0xa9 >>> RIP: 0033:0x7f95ae73e76e >>> Code: Bad RIP value. >>> RSP: 002b:00007ffc02c1dbf8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000 >>> RAX: ffffffffffffffda RBX: 000000000170faa0 RCX: 00007f95ae73e76e >>> RDX: 0000000000001000 RSI: 00007ffc02c1dc30 RDI: 0000000000000007 >>> RBP: 00007ffc02c1ec70 R08: 0000000000000005 R09: 0000000000000006 >>> R10: fffffffffffff20b R11: 0000000000000246 R12: 00000000019112a0 >>> R13: 0000000000000000 R14: 0000000000000007 R15: 00000000004283c0 >>> >>> The attached patch does 3 things: >>> >>> 1) If unable to obtain the file structure for the current task, >>> proceed to the next task number after the one returned from >>> task_seq_get_next(), instead of the next task number from the >>> original iterator. >> >> Looks like this fix is the real fix for the above warnings. >> Basically, say we have >> info->tid = 10 and returned curr_tid = 3000 and tid 3000 has no files. >> the current logic will go through >> - set curr_tid = 11 (info->tid++) and returned curr_tid = 3000 >> - set curr_tid = 12 and returned curr_tid = 3000 >> ... >> - set curr_tid = 3000 and returned curr_tid = 3000 >> - set curr_tid = 3001 and return curr_tid >= 3001 >> >> All the above works are redundant work, and it may cause issues >> for non preemptable kernel. >> >> I suggest you factor out this change plus the following change >> which suggested by Andrii early to a separate patch carried with >> the below Fixes tag. >> >> diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c >> index 0458a40edf10..56bcaef72e36 100644 >> --- a/kernel/bpf/task_iter.c >> +++ b/kernel/bpf/task_iter.c >> @@ -158,6 +158,7 @@ task_file_seq_get_next(struct >> bpf_iter_seq_task_file_info *info) >> if (!curr_task) { >> info->task = NULL; >> info->files = NULL; >> + info->tid = curr_tid + 1; >> return NULL; >> } > > Sure this isn't supposed to be 'curr_tid'? task_seq_get_next() stops > when there are no more threads found. This increments the thread id > past the search point, and would seem to introduce a potential off-by-one > error. This is for the case where read() syscall return length 0, but user space still keep read(). You are right, we may skip one newly created one indeed. Although such a usecase is not common, but info->tid = curr_tid certainly more correct than info->tid = curr_tid + 1. So thanks for suggestion, LGTM. > > That is: > curr_tid = 3000. > call task_seq_get_next() --> return NULL, curr_tid = 3000. > (so there is no tid >= 3000) > set curr_tid = 3001. > > next restart (if there is one) skips a newly created 3000. >
On Fri, Dec 18, 2020 at 10:08 AM Jonathan Lemon <jonathan.lemon@gmail.com> wrote: > > On Fri, Dec 18, 2020 at 08:53:22AM -0800, Yonghong Song wrote: > > > > > > On 12/11/20 9:11 AM, Jonathan Lemon wrote: > > > From: Jonathan Lemon <bsd@fb.com> > > > > > > On some systems, some variant of the following splat is > > > repeatedly seen. The common factor in all traces seems > > > to be the entry point to task_file_seq_next(). With the > > > patch, all warnings go away. > > > > > > rcu: INFO: rcu_sched self-detected stall on CPU > > > rcu: \x0926-....: (20992 ticks this GP) idle=d7e/1/0x4000000000000002 softirq=81556231/81556231 fqs=4876 > > > \x09(t=21033 jiffies g=159148529 q=223125) > > > NMI backtrace for cpu 26 > > > CPU: 26 PID: 2015853 Comm: bpftool Kdump: loaded Not tainted 5.6.13-0_fbk4_3876_gd8d1f9bf80bb #1 > > > Hardware name: Quanta Twin Lakes MP/Twin Lakes Passive MP, BIOS F09_3A12 10/08/2018 > > > Call Trace: > > > <IRQ> > > > dump_stack+0x50/0x70 > > > nmi_cpu_backtrace.cold.6+0x13/0x50 > > > ? lapic_can_unplug_cpu.cold.30+0x40/0x40 > > > nmi_trigger_cpumask_backtrace+0xba/0xca > > > rcu_dump_cpu_stacks+0x99/0xc7 > > > rcu_sched_clock_irq.cold.90+0x1b4/0x3aa > > > ? tick_sched_do_timer+0x60/0x60 > > > update_process_times+0x24/0x50 > > > tick_sched_timer+0x37/0x70 > > > __hrtimer_run_queues+0xfe/0x270 > > > hrtimer_interrupt+0xf4/0x210 > > > smp_apic_timer_interrupt+0x5e/0x120 > > > apic_timer_interrupt+0xf/0x20 > > > </IRQ> > > > RIP: 0010:get_pid_task+0x38/0x80 > > > Code: 89 f6 48 8d 44 f7 08 48 8b 00 48 85 c0 74 2b 48 83 c6 55 48 c1 e6 04 48 29 f0 74 19 48 8d 78 20 ba 01 00 00 00 f0 0f c1 50 20 <85> d2 74 27 78 11 83 c2 01 78 0c 48 83 c4 08 c3 31 c0 48 83 c4 08 > > > RSP: 0018:ffffc9000d293dc8 EFLAGS: 00000202 ORIG_RAX: ffffffffffffff13 > > > RAX: ffff888637c05600 RBX: ffffc9000d293e0c RCX: 0000000000000000 > > > RDX: 0000000000000001 RSI: 0000000000000550 RDI: ffff888637c05620 > > > RBP: ffffffff8284eb80 R08: ffff88831341d300 R09: ffff88822ffd8248 > > > R10: ffff88822ffd82d0 R11: 00000000003a93c0 R12: 0000000000000001 > > > R13: 00000000ffffffff R14: ffff88831341d300 R15: 0000000000000000 > > > ? find_ge_pid+0x1b/0x20 > > > task_seq_get_next+0x52/0xc0 > > > task_file_seq_get_next+0x159/0x220 > > > task_file_seq_next+0x4f/0xa0 > > > bpf_seq_read+0x159/0x390 > > > vfs_read+0x8a/0x140 > > > ksys_read+0x59/0xd0 > > > do_syscall_64+0x42/0x110 > > > entry_SYSCALL_64_after_hwframe+0x44/0xa9 > > > RIP: 0033:0x7f95ae73e76e > > > Code: Bad RIP value. > > > RSP: 002b:00007ffc02c1dbf8 EFLAGS: 00000246 ORIG_RAX: 0000000000000000 > > > RAX: ffffffffffffffda RBX: 000000000170faa0 RCX: 00007f95ae73e76e > > > RDX: 0000000000001000 RSI: 00007ffc02c1dc30 RDI: 0000000000000007 > > > RBP: 00007ffc02c1ec70 R08: 0000000000000005 R09: 0000000000000006 > > > R10: fffffffffffff20b R11: 0000000000000246 R12: 00000000019112a0 > > > R13: 0000000000000000 R14: 0000000000000007 R15: 00000000004283c0 > > > > > > The attached patch does 3 things: > > > > > > 1) If unable to obtain the file structure for the current task, > > > proceed to the next task number after the one returned from > > > task_seq_get_next(), instead of the next task number from the > > > original iterator. > > > > Looks like this fix is the real fix for the above warnings. > > Basically, say we have > > info->tid = 10 and returned curr_tid = 3000 and tid 3000 has no files. > > the current logic will go through > > - set curr_tid = 11 (info->tid++) and returned curr_tid = 3000 > > - set curr_tid = 12 and returned curr_tid = 3000 > > ... > > - set curr_tid = 3000 and returned curr_tid = 3000 > > - set curr_tid = 3001 and return curr_tid >= 3001 > > > > All the above works are redundant work, and it may cause issues > > for non preemptable kernel. > > > > I suggest you factor out this change plus the following change > > which suggested by Andrii early to a separate patch carried with > > the below Fixes tag. > > > > diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c > > index 0458a40edf10..56bcaef72e36 100644 > > --- a/kernel/bpf/task_iter.c > > +++ b/kernel/bpf/task_iter.c > > @@ -158,6 +158,7 @@ task_file_seq_get_next(struct > > bpf_iter_seq_task_file_info *info) > > if (!curr_task) { > > info->task = NULL; > > info->files = NULL; > > + info->tid = curr_tid + 1; > > return NULL; > > } > > Sure this isn't supposed to be 'curr_tid'? task_seq_get_next() stops > when there are no more threads found. This increments the thread id > past the search point, and would seem to introduce a potential off-by-one > error. > > That is: > curr_tid = 3000. > call task_seq_get_next() --> return NULL, curr_tid = 3000. > (so there is no tid >= 3000) > set curr_tid = 3001. > > next restart (if there is one) skips a newly created 3000. Seems fine to me to skip 3000 in such case. 3000 didn't exist at the time of iteration. If there was >=3001 it would have been skipped as well. > > -- > Jonathan
diff --git a/kernel/bpf/task_iter.c b/kernel/bpf/task_iter.c index 0458a40edf10..66a52fcf589a 100644 --- a/kernel/bpf/task_iter.c +++ b/kernel/bpf/task_iter.c @@ -33,17 +33,17 @@ static struct task_struct *task_seq_get_next(struct pid_namespace *ns, pid = find_ge_pid(*tid, ns); if (pid) { *tid = pid_nr_ns(pid, ns); - task = get_pid_task(pid, PIDTYPE_PID); + task = pid_task(pid, PIDTYPE_PID); if (!task) { ++*tid; goto retry; - } else if (skip_if_dup_files && task->tgid != task->pid && + } else if (skip_if_dup_files && !thread_group_leader(task) && task->files == task->group_leader->files) { - put_task_struct(task); task = NULL; ++*tid; goto retry; } + get_task_struct(task); } rcu_read_unlock(); @@ -164,7 +164,7 @@ task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info) curr_files = get_files_struct(curr_task); if (!curr_files) { put_task_struct(curr_task); - curr_tid = ++(info->tid); + curr_tid = curr_tid + 1; info->fd = 0; goto again; }