Skip to contentSkip to conversation

kernfs: fix memleak in kernel_ops_readdir()

PATCH In review Tejun Heo 3 participants 4 messages Aug 5, 2019 – Aug 8, 2019 kernfs

lore
Share & tools
lore

Tejun Heo

author maintainer
#0 lore raw

PATCH 1/1 kernfs: fix memleak in kernel_ops_readdir()

From: Andrea Arcangeli <[email protected]>

If getdents64 is killed or hits on segfault, it'll leave cgroups directories in sysfs pinned leaking memory because the kernfs node won't be freed on rmdir and the parent neither.

Repro:

  # for i in `seq 1000`; do mkdir $i; done
  # rmdir *
  # for i in `seq 1000`; do mkdir $i; done
  # rmdir *
  # for i in `seq 1000`; do while :; do ls $i/ >/dev/null; done & done
  # while :; do killall ls; done

kernfs_node_cache in /proc/slabinfo keeps going up as expected.

Cc [email protected] # goes way back to original sysfs days
---
 fs/kernfs/dir.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)
fs/kernfs/dir.c +5 -2
index 1e7a74b8e064..82b6c699fa34 100644
@@ -1683,11 +1683,14 @@ static int kernfs_fop_readdir(struct file *file, struct dir_context *ctx)
 		kernfs_get(pos);
 
 		mutex_unlock(&kernfs_mutex);
-		if (!dir_emit(ctx, name, len, ino, type))
-			return 0;
+		if (unlikely(!dir_emit(ctx, name, len, ino, type))) {
+			kernfs_put(pos);
+			goto out;
+		}
 		mutex_lock(&kernfs_mutex);
 	}
 	mutex_unlock(&kernfs_mutex);
+out:
 	file->private_data = NULL;
 	ctx->pos = INT_MAX;
 	return 0;
#1 lore raw

Hi,

Quoted 5 lines · Tejun Heo

From: Andrea Arcangeli <[email protected]>

If getdents64 is killed or hits on segfault, it'll leave cgroups directories in sysfs pinned leaking memory because the kernfs node won't be freed on rmdir and the parent neither.

Somehow this causes a regression in Linux next for me where I'm seeing lots of sysfs entries now missing under /sys/bus/platform/devices.

For example, I now only see one .serial entry show up in sysfs. Things work again if I revert commit cc798c83898e ("kernfs: fix memleak inkernel_ops_readdir()"). Any ideas why that would be?

Below is a diff -u of ls /sys/bus/platform/devices for reference showing the missing entries with cc798c83898e.

Regards,

Tony

8< -----------------------

fails renamed from works +0 -183
@@ -1,227 +1,44 @@
-2c000000.ethernet
 40100000.interconnect
-40100000.interconnect:segment@0
-40122000.mcbsp
-4012208c.target-module
-4012408c.target-module
 4012608c.target-module
-40128000.target-module
-4012c000.target-module
-4012e000.target-module
-40130000.target-module
-40130000.wdt
-40132000.target-module
 40138000.target-module
-40138000.timer
-4013a000.target-module
-4013a000.timer
-4013c000.target-module
-4013c000.timer
 4013e000.target-module
-4013e000.timer
-401f1000.target-module
-40304000.ocmcram
-44000000.ocp
-48000000.interconnect
-48000000.interconnect:segment@0
-48000000.interconnect:segment@200000
-48020000.serial
-48020050.target-module
-48032000.target-module
-48032000.timer
-48034000.target-module
-48034000.timer
-48036000.target-module
 48036000.timer
-4803e000.target-module
-4803e000.timer
 48055000.gpio
-48055000.target-module
-48057000.gpio
-48057000.target-module
-48059000.gpio
 48059000.target-module
-4805b000.gpio
-4805b000.target-module
-4805d000.gpio
 4805d000.target-module
-48060000.i2c
-48060000.target-module
-4806a000.serial
-4806a050.target-module
-4806c000.serial
-4806c050.target-module
 4806e000.serial
 4806e050.target-module
-48070000.i2c
 48070000.i2c:twl@48:gpadc
-48070000.i2c:twl@48:pwm
-48070000.i2c:twl@48:pwmled
-48070000.i2c:twl@48:regulator-v1v8
-48070000.i2c:twl@48:regulator-v2v1
 48070000.i2c:twl@48:regulator-vana
-48070000.i2c:twl@48:regulator-vaux1
-48070000.i2c:twl@48:regulator-vaux2
-48070000.i2c:twl@48:regulator-vaux3
 48070000.i2c:twl@48:regulator-vcxio
-48070000.i2c:twl@48:regulator-vdac
-48070000.i2c:twl@48:regulator-vmmc
-48070000.i2c:twl@48:regulator-vpp
-48070000.i2c:twl@48:regulator-vusb
-48070000.i2c:twl@48:regulator-vusim
-48070000.i2c:twl@48:rtc
 48070000.i2c:twl@48:usb-comparator
-48070000.target-module
-48072000.i2c
-48072000.target-module
-48076000.target-module
-48078000.target-module
-48086000.target-module
 48086000.timer
-48088000.target-module
-48088000.timer
-4809608c.target-module
-48098000.spi
-48098000.target-module
-4809a000.spi
 4809a000.target-module
-4809c000.mmc
 4809c000.target-module
-480a5000.des
-480ad000.target-module
-480b2000.1w
-480b2000.target-module
-480b4000.target-module
 480b8000.spi
-480b8000.target-module
-480ba000.spi
-480ba000.target-module
-480d1000.target-module
 480d5000.mmc
-480d5000.target-module
-48240600.local-timer
 48242000.l2-cache-controller
-48350000.i2c
-48350000.target-module
-4a000000.interconnect
-4a000000.interconnect:segment@0
-4a000000.interconnect:segment@100000
-4a000000.interconnect:segment@180000
-4a000000.interconnect:segment@200000
-4a000000.interconnect:segment@280000
 4a000000.interconnect:segment@300000
 4a000000.interconnect:segment@80000
-4a002000.scm
-4a002000.scm_conf
-4a002000.target-module
-4a002260.bandgap
-4a002300.control-phy
-4a00233c.control-phy
-4a004000.cm1
-4a004000.target-module
-4a004300.mpuss_cm
-4a004400.tesla_cm
-4a004500.abe_cm
-4a008000.cm2
-4a008000.target-module
-4a008600.l4_ao_cm
-4a008700.l3_1_cm
-4a008800.l3_2_cm
-4a008900.ducati_cm
-4a008a00.l3_dma_cm
 4a008b00.l3_emif_cm
 4a008c00.d2d_cm
-4a008d00.l4_cfg_cm
 4a008e00.l3_instr_cm
-4a008f00.ivahd_cm
-4a009000.iss_cm
-4a009100.l3_dss_cm
-4a009200.l3_gfx_cm
-4a009300.l3_init_cm
-4a009400.l4_per_cm
-4a056000.dma-controller
 4a056000.target-module
-4a058000.hsi
-4a058000.target-module
 4a062000.target-module
-4a062000.usbhstll
-4a064000.target-module
 4a064000.usbhshost
 4a064800.ohci
-4a064c00.ehci
-4a066000.mmu
-4a0ab000.usb_otg_hs
 4a0ab400.target-module
-4a0ad000.ocp2scp
-4a0ad000.target-module
 4a0ad080.usb2phy
-4a0d9000.smartreflex
-4a0d9038.target-module
 4a0db000.smartreflex
-4a0db038.target-module
-4a0dd000.smartreflex
-4a0dd038.target-module
-4a0f4000.mailbox
-4a0f4000.target-module
-4a0f6000.spinlock
-4a0f6000.target-module
 4a100000.target-module
 4a100040.pinmux
-4a1005a0.omap4_padconf_global
-4a100600.pbias_regulator
-4a10a000.target-module
-4a300000.interconnect
-4a300000.interconnect:segment@0
-4a300000.interconnect:segment@10000
 4a300000.interconnect:segment@20000
-4a306000.prm
-4a306000.target-module
-4a307bd0.regulator-abb-mpu
 4a30a000.scrm
-4a30a000.target-module
-4a30c000.target-module
 4a30c000.target-module:scm@c000
-4a310000.gpio
-4a310000.target-module
 4a314000.target-module
-4a314000.wdt
-4a31c000.keypad
-4a31c000.target-module
-4a31e000.target-module
-4a31e040.pinmux
 4b100000.sham
-4b501000.aes
 4b701000.aes
 4c000000.emif
-4d000000.emif
-4e000000.dmm
-50000000.gpmc
-52000000.target-module
-55082000.mmu
-5601fc00.target-module
-58000000.dss
-Fixed MDIO bus.0
-a0080000.ramoops
-alarmtimer
-connector
-cpufreq-dt
 gpio_keys
-hsusb1_phy
-leds
-omap_dma_system.0
-oprofile-perf.0
-pmu
 reg-dummy
-regulator-vdd33a
-regulator-vddvario
-serial8250
-soc
-soc:dsp
-soc:iva
-soc:mpu
-sound
-twl
-twl6040-codec
-twl6040-gpo
-twl6040-pdmclk
 w2cbw0015_vmmc
author maintainer
#2 lore raw

Hello,

Quoted 12 lines · Tony Lindgren

Hi,

Quoted 5 lines · Tejun Heo

From: Andrea Arcangeli <[email protected]>

If getdents64 is killed or hits on segfault, it'll leave cgroups directories in sysfs pinned leaking memory because the kernfs node won't be freed on rmdir and the parent neither.

Somehow this causes a regression in Linux next for me where I'm seeing lots of sysfs entries now missing under /sys/bus/platform/devices.

Quoted 6 lines

For example, I now only see one .serial entry show up in sysfs. Things work again if I revert commit cc798c83898e ("kernfs: fix memleak inkernel_ops_readdir()"). Any ideas why that would be?

Below is a diff -u of ls /sys/bus/platform/devices for reference showing the missing entries with cc798c83898e.

Ugh, you're right. It can get double-put cuz ctx->pos is put by release too. Greg, sorry about the noise but can you please revert the patch? I'll look into why this looked like memory leak from slabinfo side.

Thanks.

Signature · 1 line
tejun
#3 lore raw
Quoted 19 lines · Tejun Heo

Hello,

Quoted 16 lines · Tony Lindgren

Hi,

Quoted 5 lines · Tejun Heo

From: Andrea Arcangeli <[email protected]>

If getdents64 is killed or hits on segfault, it'll leave cgroups directories in sysfs pinned leaking memory because the kernfs node won't be freed on rmdir and the parent neither.

Somehow this causes a regression in Linux next for me where I'm seeing lots of sysfs entries now missing under /sys/bus/platform/devices.

For example, I now only see one .serial entry show up in sysfs. Things work again if I revert commit cc798c83898e ("kernfs: fix memleak inkernel_ops_readdir()"). Any ideas why that would be?

Quoted 7 lines
Quoted 2 lines

Below is a diff -u of ls /sys/bus/platform/devices for reference showing the missing entries with cc798c83898e.

Ugh, you're right. It can get double-put cuz ctx->pos is put by release too. Greg, sorry about the noise but can you please revert the patch? I'll look into why this looked like memory leak from slabinfo side.

Now reverted, thanks.

greg k-h