2022-09-30 22:45:46

by Rafael Mendonca

[permalink] [raw]
Subject: [PATCH 0/3] uio: uio_dmem_genirq: Fix locking issues in irq config and handling

The goal of this series is to apply the changes from [1] to the
"uio_dmem_genirq" driver. The implementation of "uio_dmem_genirq" was
based on "uio_pdrv_genirq" and it is used in a similar manner to the
"uio_pdrv_genirq" driver with respect to interrupt configuration and
handling. At the time "uio_dmem_genirq" was introduced, both had the
same implementation of the 'uio_info' handlers irqcontrol() and
handler(). Later, [1] was only applied to "uio_pdrv_genirq", even though
both had the same issues addressed by the patch, which ended up making
them a little different.

The original motivation for this series was to fix a bug introduced by
commit [2], which can be solved by copying the implementation of the
interrupt configuration and handling from "uio_pdrv_genirq" to
"uio_dmem_genirq", thus, making their implementation similar again.

Since patch [1] solves the bug introduced in [2] and also addresses
other issues, I broke it into 3 separate changes.

[1] 34cb27528398 ("UIO: Fix concurrency issue")
https://lore.kernel.org/lkml/[email protected]/
[2] b74351287d4b ("uio: fix a sleep-in-atomic-context bug in
uio_dmem_genirq_irqcontrol()")
https://lore.kernel.org/all/[email protected]/

Rafael Mendonca (3):
uio: uio_dmem_genirq: Fix missing unlock in irq configuration
uio: uio_dmem_genirq: Fix deadlock between irq config and handling
uio: uio_dmem_genirq: Use non-atomic bit operations in irq config and handling

drivers/uio/uio_dmem_genirq.c | 22 ++++++++++++++--------
1 file changed, 14 insertions(+), 8 deletions(-)

--
2.34.1


2022-09-30 22:46:04

by Rafael Mendonca

[permalink] [raw]
Subject: [PATCH 1/3] uio: uio_dmem_genirq: Fix missing unlock in irq configuration

Commit b74351287d4b ("uio: fix a sleep-in-atomic-context bug in
uio_dmem_genirq_irqcontrol()") started calling disable_irq() without
holding the spinlock because it can sleep. However, that fix introduced
another bug: if interrupt is already disabled and a new disable request
comes in, then the spinlock is not unlocked:

root@localhost:~# printf '\x00\x00\x00\x00' > /dev/uio0
root@localhost:~# printf '\x00\x00\x00\x00' > /dev/uio0
root@localhost:~# [ 14.851538] BUG: scheduling while atomic: bash/223/0x00000002
[ 14.851991] Modules linked in: uio_dmem_genirq uio myfpga(OE) bochs drm_vram_helper drm_ttm_helper ttm drm_kms_helper drm snd_pcm ppdev joydev psmouse snd_timer snd e1000fb_sys_fops syscopyarea parport sysfillrect soundcore sysimgblt input_leds pcspkr i2c_piix4 serio_raw floppy evbug qemu_fw_cfg mac_hid pata_acpi ip_tables x_tables autofs4 [last unloaded: parport_pc]
[ 14.854206] CPU: 0 PID: 223 Comm: bash Tainted: G OE 6.0.0-rc7 #21
[ 14.854786] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.16.0-0-gd239552ce722-prebuilt.qemu.org 04/01/2014
[ 14.855664] Call Trace:
[ 14.855861] <TASK>
[ 14.856025] dump_stack_lvl+0x4d/0x67
[ 14.856325] dump_stack+0x14/0x1a
[ 14.856583] __schedule_bug.cold+0x4b/0x5c
[ 14.856915] __schedule+0xe81/0x13d0
[ 14.857199] ? idr_find+0x13/0x20
[ 14.857456] ? get_work_pool+0x2d/0x50
[ 14.857756] ? __flush_work+0x233/0x280
[ 14.858068] ? __schedule+0xa95/0x13d0
[ 14.858307] ? idr_find+0x13/0x20
[ 14.858519] ? get_work_pool+0x2d/0x50
[ 14.858798] schedule+0x6c/0x100
[ 14.859009] schedule_hrtimeout_range_clock+0xff/0x110
[ 14.859335] ? tty_write_room+0x1f/0x30
[ 14.859598] ? n_tty_poll+0x1ec/0x220
[ 14.859830] ? tty_ldisc_deref+0x1a/0x20
[ 14.860090] schedule_hrtimeout_range+0x17/0x20
[ 14.860373] do_select+0x596/0x840
[ 14.860627] ? __kernel_text_address+0x16/0x50
[ 14.860954] ? poll_freewait+0xb0/0xb0
[ 14.861235] ? poll_freewait+0xb0/0xb0
[ 14.861517] ? rpm_resume+0x49d/0x780
[ 14.861798] ? common_interrupt+0x59/0xa0
[ 14.862127] ? asm_common_interrupt+0x2b/0x40
[ 14.862511] ? __uart_start.isra.0+0x61/0x70
[ 14.862902] ? __check_object_size+0x61/0x280
[ 14.863255] core_sys_select+0x1c6/0x400
[ 14.863575] ? vfs_write+0x1c9/0x3d0
[ 14.863853] ? vfs_write+0x1c9/0x3d0
[ 14.864121] ? _copy_from_user+0x45/0x70
[ 14.864526] do_pselect.constprop.0+0xb3/0xf0
[ 14.864893] ? do_syscall_64+0x6d/0x90
[ 14.865228] ? do_syscall_64+0x6d/0x90
[ 14.865556] __x64_sys_pselect6+0x76/0xa0
[ 14.865906] do_syscall_64+0x60/0x90
[ 14.866214] ? syscall_exit_to_user_mode+0x2a/0x50
[ 14.866640] ? do_syscall_64+0x6d/0x90
[ 14.866972] ? do_syscall_64+0x6d/0x90
[ 14.867286] ? do_syscall_64+0x6d/0x90
[ 14.867626] entry_SYSCALL_64_after_hwframe+0x63/0xcd
[...] stripped
[ 14.872959] </TASK>

('myfpga' is a simple 'uio_dmem_genirq' driver I wrote to test this)

The implementation of "uio_dmem_genirq" was based on "uio_pdrv_genirq" and
it is used in a similar manner to the "uio_pdrv_genirq" driver with respect
to interrupt configuration and handling. At the time "uio_dmem_genirq" was
introduced, both had the same implementation of the 'uio_info' handlers
irqcontrol() and handler(). Then commit 34cb27528398 ("UIO: Fix concurrency
issue"), which was only applied to "uio_pdrv_genirq", ended up making them
a little different. That commit, among other things, changed disable_irq()
to disable_irq_nosync() in the implementation of irqcontrol(). The
motivation there was to avoid a deadlock between irqcontrol() and
handler(), since it added a spinlock in the irq handler, and disable_irq()
waits for the completion of the irq handler.

By changing disable_irq() to disable_irq_nosync() in irqcontrol(), we also
avoid the sleeping-while-atomic bug that commit b74351287d4b ("uio: fix a
sleep-in-atomic-context bug in uio_dmem_genirq_irqcontrol()") was trying to
fix. Thus, this fixes the missing unlock in irqcontrol() by importing the
implementation of irqcontrol() handler from the "uio_pdrv_genirq" driver.
In the end, it reverts commit b74351287d4b ("uio: fix a
sleep-in-atomic-context bug in uio_dmem_genirq_irqcontrol()") and change
disable_irq() to disable_irq_nosync().

It is worth noting that this still does not address the concurrency issue
fixed by commit 34cb27528398 ("UIO: Fix concurrency issue"). It will be
addressed separately in the next commits.

Split out from commit 34cb27528398 ("UIO: Fix concurrency issue").

Fixes: b74351287d4b ("uio: fix a sleep-in-atomic-context bug in uio_dmem_genirq_irqcontrol()")
Signed-off-by: Rafael Mendonca <[email protected]>
---
drivers/uio/uio_dmem_genirq.c | 8 +++-----
1 file changed, 3 insertions(+), 5 deletions(-)

diff --git a/drivers/uio/uio_dmem_genirq.c b/drivers/uio/uio_dmem_genirq.c
index 1106f3376404..cb283ee36eaa 100644
--- a/drivers/uio/uio_dmem_genirq.c
+++ b/drivers/uio/uio_dmem_genirq.c
@@ -132,13 +132,11 @@ static int uio_dmem_genirq_irqcontrol(struct uio_info *dev_info, s32 irq_on)
if (irq_on) {
if (test_and_clear_bit(0, &priv->flags))
enable_irq(dev_info->irq);
- spin_unlock_irqrestore(&priv->lock, flags);
} else {
- if (!test_and_set_bit(0, &priv->flags)) {
- spin_unlock_irqrestore(&priv->lock, flags);
- disable_irq(dev_info->irq);
- }
+ if (!test_and_set_bit(0, &priv->flags))
+ disable_irq_nosync(dev_info->irq);
}
+ spin_unlock_irqrestore(&priv->lock, flags);

return 0;
}
--
2.34.1

2022-09-30 22:47:18

by Rafael Mendonca

[permalink] [raw]
Subject: [PATCH 3/3] uio: uio_dmem_genirq: Use non-atomic bit operations in irq config and handling

This finishes the port of the irq configuration and handling from
"uio_pdrv_genirq" to "uio_dmem_genirq". It changes the atomic
bit-manipulation routines to their non-atomic counterparts as we are
already guarding the code by spinlock.

Split out from commit 34cb27528398 ("UIO: Fix concurrency issue").

Signed-off-by: Rafael Mendonca <[email protected]>
---
drivers/uio/uio_dmem_genirq.c | 11 ++++++++---
1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/drivers/uio/uio_dmem_genirq.c b/drivers/uio/uio_dmem_genirq.c
index 792c3e9c9ce5..5313307c2754 100644
--- a/drivers/uio/uio_dmem_genirq.c
+++ b/drivers/uio/uio_dmem_genirq.c
@@ -41,6 +41,11 @@ struct uio_dmem_genirq_platdata {
unsigned int refcnt;
};

+/* Bits in uio_dmem_genirq_platdata.flags */
+enum {
+ UIO_IRQ_DISABLED = 0,
+};
+
static int uio_dmem_genirq_open(struct uio_info *info, struct inode *inode)
{
struct uio_dmem_genirq_platdata *priv = info->priv;
@@ -111,7 +116,7 @@ static irqreturn_t uio_dmem_genirq_handler(int irq, struct uio_info *dev_info)
*/

spin_lock(&priv->lock);
- if (!test_and_set_bit(0, &priv->flags))
+ if (!__test_and_set_bit(UIO_IRQ_DISABLED, &priv->flags))
disable_irq_nosync(irq);
spin_unlock(&priv->lock);

@@ -133,10 +138,10 @@ static int uio_dmem_genirq_irqcontrol(struct uio_info *dev_info, s32 irq_on)

spin_lock_irqsave(&priv->lock, flags);
if (irq_on) {
- if (test_and_clear_bit(0, &priv->flags))
+ if (__test_and_clear_bit(UIO_IRQ_DISABLED, &priv->flags))
enable_irq(dev_info->irq);
} else {
- if (!test_and_set_bit(0, &priv->flags))
+ if (!__test_and_set_bit(UIO_IRQ_DISABLED, &priv->flags))
disable_irq_nosync(dev_info->irq);
}
spin_unlock_irqrestore(&priv->lock, flags);
--
2.34.1

2022-09-30 23:04:03

by Rafael Mendonca

[permalink] [raw]
Subject: [PATCH 2/3] uio: uio_dmem_genirq: Fix deadlock between irq config and handling

This fixes a concurrency issue addressed in commit 34cb27528398 ("UIO: Fix
concurrency issue"):

"In a SMP case there was a race condition issue between
Uio_pdrv_genirq_irqcontrol() running on one CPU and irq handler on
another CPU. Fix it by spin_locking shared resources access inside irq
handler."

The implementation of "uio_dmem_genirq" was based on "uio_pdrv_genirq" and
it is used in a similar manner to the "uio_pdrv_genirq" driver with respect
to interrupt configuration and handling. At the time "uio_dmem_genirq" was
merged, both had the same implementation of the 'uio_info' handlers
irqcontrol() and handler(), thus, both had the same concurrency issue
mentioned by the above commit. However, the above patch was only applied to
the "uio_pdrv_genirq" driver.

Split out from commit 34cb27528398 ("UIO: Fix concurrency issue").

Fixes: 0a0c3b5a24bd ("Add new uio device for dynamic memory allocation")
Signed-off-by: Rafael Mendonca <[email protected]>
---
drivers/uio/uio_dmem_genirq.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/drivers/uio/uio_dmem_genirq.c b/drivers/uio/uio_dmem_genirq.c
index cb283ee36eaa..792c3e9c9ce5 100644
--- a/drivers/uio/uio_dmem_genirq.c
+++ b/drivers/uio/uio_dmem_genirq.c
@@ -110,8 +110,10 @@ static irqreturn_t uio_dmem_genirq_handler(int irq, struct uio_info *dev_info)
* remember the state so we can allow user space to enable it later.
*/

+ spin_lock(&priv->lock);
if (!test_and_set_bit(0, &priv->flags))
disable_irq_nosync(irq);
+ spin_unlock(&priv->lock);

return IRQ_HANDLED;
}
@@ -125,7 +127,8 @@ static int uio_dmem_genirq_irqcontrol(struct uio_info *dev_info, s32 irq_on)
* in the interrupt controller, but keep track of the
* state to prevent per-irq depth damage.
*
- * Serialize this operation to support multiple tasks.
+ * Serialize this operation to support multiple tasks and concurrency
+ * with irq handler on SMP systems.
*/

spin_lock_irqsave(&priv->lock, flags);
--
2.34.1