Hi Waiman,
On 2/28/19 7:47 PM, Waiman Long wrote:
For ipcmni_extend mode, the sequence number space is only 7 bits. So
the chance of id reuse is relatively high compared with the non-extended
mode.
To alleviate this id reuse problem, the id allocation will be done
cyclically to cycle through all the 24-bit id space before wrapping
around when in ipcmni_extend mode. This may cause the use of more memory
in term of the number of xa_nodes allocated as well as potentially more
cachelines used as the xa_nodes may be spread more sparsely in this case.
There is probably a slight memory and performance cost in doing cyclic
id allocation. For applications that really need more than 32k unique IPC
identifiers, this is a small price to pay to avoid the id reuse problem.
Have you measured it?
I have observed -3% for semop() for a 4 level radix tree compared to a
1-level radix tree, and I'm a bit reluctant to accept that.
Especially as the percentage will increase if the syscall overhead goes
down again (-> less spectre impact).
[...]
--- a/ipc/util.c
+++ b/ipc/util.c
@@ -221,7 +221,12 @@ static inline int ipc_idr_alloc(struct ipc_ids *ids, struct kern_ipc_perm *new)
*/
if (next_id < 0) { /* !CHECKPOINT_RESTORE or next_id is unset */
- idx = idr_alloc(&ids->ipcs_idr, new, 0, 0, GFP_NOWAIT);
+ if (ipc_mni_extended)
+ idx = idr_alloc_cyclic(&ids->ipcs_idr, new, 0, ipc_mni,
+ GFP_NOWAIT);
+ else
+ idx = idr_alloc(&ids->ipcs_idr, new, 0, 0, GFP_NOWAIT);
+
if ((idx <= ids->last_idx) && (++ids->seq > IPCID_SEQ_MAX))
ids->seq = 0;
I don't like it that there are two different codepaths.
Attached is a different proposal:
Always use cyclic allocation, with some logic to minimize the additional
radix tree levels.
What do you think?
--
Manfred
>From 491ea87cc3022e50c02caae009f9aeba2b6ddcb4 Mon Sep 17 00:00:00 2001
From: Manfred Spraul <manfred@xxxxxxxxxxxxxxxx>
Date: Sun, 17 Mar 2019 06:29:00 +0100
Subject: [PATCH 2/2] ipc: Do cyclic id allocation for the ipc objects.
For ipcmni_extend mode, the sequence number space is only 7 bits. So
the chance of id reuse is relatively high compared with the non-extended
mode.
To alleviate this id reuse problem, this patch enables cyclic allocation
for the index to the radix tree (idx).
The disadvantage is that this can cause a slight slow-down of the fast
path, as the radix tree could be higher than necessary.
To limit the radix tree height, I have chosen the following limits:
- 1) The cycling is done over in_use*1.5.
- 2) At least, the cycling is done over
"normal" ipcnmi mode: RADIX_TREE_MAP_SIZE elements
"ipcmni_extended": 4096 elements
Result:
- for normal mode:
No change for <= 42 active ipc elements. With more than 42
active ipc elements, a 2nd level would be added to the radix
tree.
Without cyclic allocation, a 2nd level would be added only with
more than 63 active elements.
- for extended mode:
Cycling creates always at least a 2-level radix tree.
With more than 2730 active objects, a 3rd level would be
added, instead of > 4095 active objects until the 3rd level
is added without cyclic allocation.
For a 2-level radix tree compared to a 1-level radix tree, I have
observed < 1% performance impact. I consider this as a good
compromise.
Notes:
1) Normal "x=semget();y=semget();" is unaffected: Then the idx
is e.g. a and a+1, regardless of idr_alloc() or idr_alloc_cyclic()
is used.
2) The -1% happens in a microbenchmark after this situation:
x=semget();
for(i=0;i<4000;i++) {t=semget();semctl(t,0,IPC_RMID);}
y=semget();
Now perform semget calls on x and y that do not sleep.
3) The worst-case reuse cycle time is unfortunately unaffected:
If you have 2^24-1 ipc objects allocated, and get/remove the last
possible element in a loop, then the id is reused after 128
get/remove pairs.
Performance check:
A microbenchmark that performes no-op semop() randomly on two IDs,
with only these two IDs allocated.
The IDs were set using /proc/sys/kernel/sem_next_id.
The test was run 5 times, averages are shown.
1 & 2: Base (6.22 seconds for 10.000.000 semops)
1 & 40: -0.2%
1 & 3348: - 0.8%
1 & 27348: - 1.6%
1 & 15777204: - 3.2%
Or: ~12.6 cpu cycles per additional radix tree level.
The cpu is an Intel I3-5010U. ~1300 cpu cycles/syscall is slower
than what I remember (spectre impact?).
Signed-off-by: Manfred Spraul <manfred@xxxxxxxxxxxxxxxx>
---
ipc/ipc_sysctl.c | 2 ++
ipc/util.c | 10 +++++++++-
ipc/util.h | 3 +++
3 files changed, 14 insertions(+), 1 deletion(-)
diff --git a/ipc/ipc_sysctl.c b/ipc/ipc_sysctl.c
index 73b7782eccf4..bfaae457810c 100644
--- a/ipc/ipc_sysctl.c
+++ b/ipc/ipc_sysctl.c
@@ -122,6 +122,7 @@ static int one = 1;
static int int_max = INT_MAX;
int ipc_mni = IPCMNI;
int ipc_mni_shift = IPCMNI_SHIFT;
+int ipc_min_cycle = RADIX_TREE_MAP_SIZE;
static struct ctl_table ipc_kern_table[] = {
{
@@ -252,6 +253,7 @@ static int __init ipc_mni_extend(char *str)
{
ipc_mni = IPCMNI_EXTEND;
ipc_mni_shift = IPCMNI_EXTEND_SHIFT;
+ ipc_min_cycle = IPCMNI_EXTEND_MIN_CYCLE;
pr_info("IPCMNI extended to %d.\n", ipc_mni);
return 0;
}
diff --git a/ipc/util.c b/ipc/util.c
index 6e0fe3410423..90764b67af51 100644
--- a/ipc/util.c
+++ b/ipc/util.c
@@ -221,9 +221,17 @@ static inline int ipc_idr_alloc(struct ipc_ids *ids, struct kern_ipc_perm *new)
*/
if (next_id < 0) { /* !CHECKPOINT_RESTORE or next_id is unset */
+ int max_idx;
+
+ max_idx = ids->in_use*3/2;
+ if (max_idx > ipc_mni)
+ max_idx = ipc_mni;
+ if (max_idx < ipc_min_cycle)
+ max_idx = ipc_min_cycle;
/* allocate the idx, with a NULL struct kern_ipc_perm */
- idx = idr_alloc(&ids->ipcs_idr, NULL, 0, 0, GFP_NOWAIT);
+ idx = idr_alloc_cyclic(&ids->ipcs_idr, NULL, 0, max_idx,
+ GFP_NOWAIT);
if (idx >= 0) {
/*
diff --git a/ipc/util.h b/ipc/util.h
index 8c834ed39012..ef4e86bb2db8 100644
--- a/ipc/util.h
+++ b/ipc/util.h
@@ -27,12 +27,14 @@
*/
#define IPCMNI_SHIFT 15
#define IPCMNI_EXTEND_SHIFT 24
+#define IPCMNI_EXTEND_MIN_CYCLE (2 << 12)
#define IPCMNI (1 << IPCMNI_SHIFT)
#define IPCMNI_EXTEND (1 << IPCMNI_EXTEND_SHIFT)
#ifdef CONFIG_SYSVIPC_SYSCTL
extern int ipc_mni;
extern int ipc_mni_shift;
+extern int ipc_min_cycle;
#define ipcmni_seq_shift() ipc_mni_shift
#define IPCMNI_IDX_MASK ((1 << ipc_mni_shift) - 1)
@@ -40,6 +42,7 @@ extern int ipc_mni_shift;
#else /* CONFIG_SYSVIPC_SYSCTL */
#define ipc_mni IPCMNI
+#define ipc_min_cycle RADIX_TREE_MAP_SIZE
#define ipcmni_seq_shift() IPCMNI_SHIFT
#define IPCMNI_IDX_MASK ((1 << IPCMNI_SHIFT) - 1)
#endif /* CONFIG_SYSVIPC_SYSCTL */
--
2.17.2
