[alsa-devel] [PATCH v4] ALSA: fireworks: accessing to user space outside spinlock
Hi,
This patch is a revised version of a part of my previous patchset:
[alsa-devel] [PATCH v3 0/2] ALSA: fireworks/firewire-tascam: accessing to user space outside spinlock http://mailman.alsa-project.org/pipermail/alsa-devel/2016-August/112366.html
In ALSA fireworks driver, current implementation for hwdep interface have issues for page fault handling. For example, when using single core processor, after executing lock_page() in page fault handler and call task scheduler, then the context never regain CPU again, because any interrupts are disabled by these drivers and the context can't catch waiting event. This patchset fixes the issues.
The issues were reported by Vaishali Thakkar, with a help of coccinelle. Thanks a lot for her care to these minor drivers. http://mailman.alsa-project.org/pipermail/alsa-devel/2016-August/111887.html
Changes in v4: * Give up counting the number of queued response messages to simplify ring-buffer management. Instead, comparing pull/push pointers is used consistently in any parts. * Assign current state to automatic variables in critical section of blocking processing part, then utilize it for later processing.
Changes in v3: * Add a workaround against a case that several tasks are going to read from one response buffer.
Changes in v2: * improve designated initialization for automatic variable of union type
Takashi Sakamoto (1): ALSA: fireworks: accessing to user space outside spinlock
sound/firewire/fireworks/fireworks.h | 1 - sound/firewire/fireworks/fireworks_hwdep.c | 71 +++++++++++++++++------- sound/firewire/fireworks/fireworks_proc.c | 4 +- sound/firewire/fireworks/fireworks_transaction.c | 5 +- 4 files changed, 56 insertions(+), 25 deletions(-)
In hwdep interface of fireworks driver, accessing to user space is in a critical section with disabled local interrupt. Depending on architecture, accessing to user space can cause page fault exception. Then local processor stores machine status and handles the synchronous event. A handler corresponding to the event can call task scheduler to wait for preparing pages. In a case of usage of single core processor, the state to disable local interrupt is worse because it don't handle usual interrupts from hardware.
This commit fixes this bug, performing the accessing outside spinlock. This commit also gives up counting the number of queued response messages to simplify ring-buffer management.
Reported-by: Vaishali Thakkar vaishali.thakkar@oracle.com Cc: stable@vger.kernel.org Fixes: 555e8a8f7f14('ALSA: fireworks: Add command/response functionality into hwdep interface') Signed-off-by: Takashi Sakamoto o-takashi@sakamocchi.jp --- sound/firewire/fireworks/fireworks.h | 1 - sound/firewire/fireworks/fireworks_hwdep.c | 71 +++++++++++++++++------- sound/firewire/fireworks/fireworks_proc.c | 4 +- sound/firewire/fireworks/fireworks_transaction.c | 5 +- 4 files changed, 56 insertions(+), 25 deletions(-)
diff --git a/sound/firewire/fireworks/fireworks.h b/sound/firewire/fireworks/fireworks.h index 03ed352..d73c12b 100644 --- a/sound/firewire/fireworks/fireworks.h +++ b/sound/firewire/fireworks/fireworks.h @@ -108,7 +108,6 @@ struct snd_efw { u8 *resp_buf; u8 *pull_ptr; u8 *push_ptr; - unsigned int resp_queues; };
int snd_efw_transaction_cmd(struct fw_unit *unit, diff --git a/sound/firewire/fireworks/fireworks_hwdep.c b/sound/firewire/fireworks/fireworks_hwdep.c index 33df865..531baab 100644 --- a/sound/firewire/fireworks/fireworks_hwdep.c +++ b/sound/firewire/fireworks/fireworks_hwdep.c @@ -25,6 +25,7 @@ hwdep_read_resp_buf(struct snd_efw *efw, char __user *buf, long remained, { unsigned int length, till_end, type; struct snd_efw_transaction *t; + u8 *pull_ptr; long count = 0;
if (remained < sizeof(type) + sizeof(struct snd_efw_transaction)) @@ -38,8 +39,17 @@ hwdep_read_resp_buf(struct snd_efw *efw, char __user *buf, long remained, buf += sizeof(type);
/* write into buffer as many responses as possible */ - while (efw->resp_queues > 0) { - t = (struct snd_efw_transaction *)(efw->pull_ptr); + spin_lock_irq(&efw->lock); + + /* + * When another task reaches here during this task's access to user + * space, it picks up current position in buffer and can read the same + * series of responses. + */ + pull_ptr = efw->pull_ptr; + + while (efw->push_ptr != pull_ptr) { + t = (struct snd_efw_transaction *)(pull_ptr); length = be32_to_cpu(t->length) * sizeof(__be32);
/* confirm enough space for this response */ @@ -49,26 +59,39 @@ hwdep_read_resp_buf(struct snd_efw *efw, char __user *buf, long remained, /* copy from ring buffer to user buffer */ while (length > 0) { till_end = snd_efw_resp_buf_size - - (unsigned int)(efw->pull_ptr - efw->resp_buf); + (unsigned int)(pull_ptr - efw->resp_buf); till_end = min_t(unsigned int, length, till_end);
- if (copy_to_user(buf, efw->pull_ptr, till_end)) + spin_unlock_irq(&efw->lock); + + if (copy_to_user(buf, pull_ptr, till_end)) return -EFAULT;
- efw->pull_ptr += till_end; - if (efw->pull_ptr >= efw->resp_buf + - snd_efw_resp_buf_size) - efw->pull_ptr -= snd_efw_resp_buf_size; + spin_lock_irq(&efw->lock); + + pull_ptr += till_end; + if (pull_ptr >= efw->resp_buf + snd_efw_resp_buf_size) + pull_ptr -= snd_efw_resp_buf_size;
length -= till_end; buf += till_end; count += till_end; remained -= till_end; } - - efw->resp_queues--; }
+ /* + * All of tasks can read from the buffer nearly simultaneously, but the + * last position for each task is different depending on the length of + * given buffer. Here, for simplicity, a position of buffer is set by + * the latest task. It's better for a listening application to allow one + * thread to read from the buffer. Unless, each task can read different + * sequence of responses depending on variation of buffer length. + */ + efw->pull_ptr = pull_ptr; + + spin_unlock_irq(&efw->lock); + return count; }
@@ -76,14 +99,17 @@ static long hwdep_read_locked(struct snd_efw *efw, char __user *buf, long count, loff_t *offset) { - union snd_firewire_event event; + union snd_firewire_event event = { + .lock_status.type = SNDRV_FIREWIRE_EVENT_LOCK_STATUS, + };
- memset(&event, 0, sizeof(event)); + spin_lock_irq(&efw->lock);
- event.lock_status.type = SNDRV_FIREWIRE_EVENT_LOCK_STATUS; event.lock_status.status = (efw->dev_lock_count > 0); efw->dev_lock_changed = false;
+ spin_unlock_irq(&efw->lock); + count = min_t(long, count, sizeof(event.lock_status));
if (copy_to_user(buf, &event, count)) @@ -98,10 +124,15 @@ hwdep_read(struct snd_hwdep *hwdep, char __user *buf, long count, { struct snd_efw *efw = hwdep->private_data; DEFINE_WAIT(wait); + bool dev_lock_changed; + bool queued;
spin_lock_irq(&efw->lock);
- while ((!efw->dev_lock_changed) && (efw->resp_queues == 0)) { + dev_lock_changed = efw->dev_lock_changed; + queued = efw->push_ptr != efw->pull_ptr; + + while (!dev_lock_changed && !queued) { prepare_to_wait(&efw->hwdep_wait, &wait, TASK_INTERRUPTIBLE); spin_unlock_irq(&efw->lock); schedule(); @@ -109,15 +140,17 @@ hwdep_read(struct snd_hwdep *hwdep, char __user *buf, long count, if (signal_pending(current)) return -ERESTARTSYS; spin_lock_irq(&efw->lock); + dev_lock_changed = efw->dev_lock_changed; + queued = efw->push_ptr != efw->pull_ptr; }
- if (efw->dev_lock_changed) + spin_unlock_irq(&efw->lock); + + if (dev_lock_changed) count = hwdep_read_locked(efw, buf, count, offset); - else if (efw->resp_queues > 0) + else if (queued) count = hwdep_read_resp_buf(efw, buf, count, offset);
- spin_unlock_irq(&efw->lock); - return count; }
@@ -160,7 +193,7 @@ hwdep_poll(struct snd_hwdep *hwdep, struct file *file, poll_table *wait) poll_wait(file, &efw->hwdep_wait, wait);
spin_lock_irq(&efw->lock); - if (efw->dev_lock_changed || (efw->resp_queues > 0)) + if (efw->dev_lock_changed || efw->pull_ptr != efw->push_ptr) events = POLLIN | POLLRDNORM; else events = 0; diff --git a/sound/firewire/fireworks/fireworks_proc.c b/sound/firewire/fireworks/fireworks_proc.c index 0639dcb..beb0a0f 100644 --- a/sound/firewire/fireworks/fireworks_proc.c +++ b/sound/firewire/fireworks/fireworks_proc.c @@ -188,8 +188,8 @@ proc_read_queues_state(struct snd_info_entry *entry, else consumed = (unsigned int)(efw->push_ptr - efw->pull_ptr);
- snd_iprintf(buffer, "%d %d/%d\n", - efw->resp_queues, consumed, snd_efw_resp_buf_size); + snd_iprintf(buffer, "%d/%d\n", + consumed, snd_efw_resp_buf_size); }
static void diff --git a/sound/firewire/fireworks/fireworks_transaction.c b/sound/firewire/fireworks/fireworks_transaction.c index f550808..36a08ba 100644 --- a/sound/firewire/fireworks/fireworks_transaction.c +++ b/sound/firewire/fireworks/fireworks_transaction.c @@ -121,11 +121,11 @@ copy_resp_to_buf(struct snd_efw *efw, void *data, size_t length, int *rcode) size_t capacity, till_end; struct snd_efw_transaction *t;
- spin_lock_irq(&efw->lock); - t = (struct snd_efw_transaction *)data; length = min_t(size_t, be32_to_cpu(t->length) * sizeof(u32), length);
+ spin_lock_irq(&efw->lock); + if (efw->push_ptr < efw->pull_ptr) capacity = (unsigned int)(efw->pull_ptr - efw->push_ptr); else @@ -155,7 +155,6 @@ copy_resp_to_buf(struct snd_efw *efw, void *data, size_t length, int *rcode) }
/* for hwdep */ - efw->resp_queues++; wake_up(&efw->hwdep_wait);
*rcode = RCODE_COMPLETE;
On Wed, 31 Aug 2016 15:58:42 +0200, Takashi Sakamoto wrote:
In hwdep interface of fireworks driver, accessing to user space is in a critical section with disabled local interrupt. Depending on architecture, accessing to user space can cause page fault exception. Then local processor stores machine status and handles the synchronous event. A handler corresponding to the event can call task scheduler to wait for preparing pages. In a case of usage of single core processor, the state to disable local interrupt is worse because it don't handle usual interrupts from hardware.
This commit fixes this bug, performing the accessing outside spinlock. This commit also gives up counting the number of queued response messages to simplify ring-buffer management.
Reported-by: Vaishali Thakkar vaishali.thakkar@oracle.com Cc: stable@vger.kernel.org Fixes: 555e8a8f7f14('ALSA: fireworks: Add command/response functionality into hwdep interface') Signed-off-by: Takashi Sakamoto o-takashi@sakamocchi.jp
Applied, thanks.
Though, now I think whether it could be more easily fixed just by replacing spinlock with a mutex. But this patch is good and simple enough, so I take it.
Takashi
On Aug 31 2016 23:18, Takashi Iwai wrote:
On Wed, 31 Aug 2016 15:58:42 +0200, Takashi Sakamoto wrote:
In hwdep interface of fireworks driver, accessing to user space is in a critical section with disabled local interrupt. Depending on architecture, accessing to user space can cause page fault exception. Then local processor stores machine status and handles the synchronous event. A handler corresponding to the event can call task scheduler to wait for preparing pages. In a case of usage of single core processor, the state to disable local interrupt is worse because it don't handle usual interrupts from hardware.
This commit fixes this bug, performing the accessing outside spinlock. This commit also gives up counting the number of queued response messages to simplify ring-buffer management.
Reported-by: Vaishali Thakkar vaishali.thakkar@oracle.com Cc: stable@vger.kernel.org Fixes: 555e8a8f7f14('ALSA: fireworks: Add command/response functionality into hwdep interface') Signed-off-by: Takashi Sakamoto o-takashi@sakamocchi.jp
Applied, thanks.
Though, now I think whether it could be more easily fixed just by replacing spinlock with a mutex. But this patch is good and simple enough, so I take it.
Copying to ring-buffer is done in software IRQ context, with referring to pull/push pointers, therefore we should use lock primitives without calling task scheduler.
In the other aim to handle several tasks in process context, I've also investigated usage of mutex, but cannot avoid using copy_to_user() in held mutex. In this case, neither usage of spinlock nor mutex has differences.
Thanks
Takashi Sakamoto
participants (2)
-
Takashi Iwai
-
Takashi Sakamoto