On Sun, 2005-09-25 at 09:41 -0500, James Bottomley wrote:
> On Tue, 2005-09-20 at 12:56 -0700, Mark Haverkamp wrote:
> > + && ((hw_fib_pool = kmalloc(sizeof
> > (struct hw_fib *) * num, GFP_ATOMIC|GFP_KERNEL)))
> > + && ((fib_pool = kmalloc(sizeof(struct
> > fib *) * num, GFP_ATOMIC|GFP_KERNEL)))) {
> > + hw_fib_p = hw_fib_pool;
> > + fib_p = fib_pool;
> > + while (hw_fib_p < &hw_fib_pool
> > [num]) {
> > + if (!(*(hw_fib_p++) =
> > kmalloc(sizeof(struct hw_fib), GFP_ATOMIC|GFP_KERNEL))) {
> > + --hw_fib_p;
> > + break;
> > + }
> > + if (!(*(fib_p++) =
> > kmalloc(sizeof(struct fib), GFP_ATOMIC|GFP_KERNEL))) {
>
> all of these allocations should be either GFP_ATOMIC or GFP_KERNEL, but
> not both (it looks like they should be GFP_KERNEL). What GFP_KERNEL|
> GFP_ATOMIC actually gives you is a potentially sleeping allocation that
> will exhaust the emergency pools, which sounds really undesirable.
>
> James
>
James,
Here is the patch again with the fixed kmalloc.
-----
Recevied from Mark Salyzyn from Adaptec.
Aif pre-allocation is used to pull the kmalloc outside of the locks.
Applies to the scsi-misc-2.6 git tree.
Signed-off-by: Mark Haverkamp <markh@xxxxxxxx>
Index: scsi-misc-aac-2/drivers/scsi/aacraid/commsup.c
===================================================================
--- scsi-misc-aac-2.orig/drivers/scsi/aacraid/commsup.c 2005-09-20 13:20:05.000000000 -0700
+++ scsi-misc-aac-2/drivers/scsi/aacraid/commsup.c 2005-09-26 09:14:25.000000000 -0700
@@ -805,7 +805,6 @@
{
struct hw_fib *hw_fib, *hw_newfib;
struct fib *fib, *newfib;
- struct aac_queue_block *queues = dev->queues;
struct aac_fib_context *fibctx;
unsigned long flags;
DECLARE_WAITQUEUE(wait, current);
@@ -825,21 +824,22 @@
* Let the DPC know it has a place to send the AIF's to.
*/
dev->aif_thread = 1;
- add_wait_queue(&queues->queue[HostNormCmdQueue].cmdready, &wait);
+ add_wait_queue(&dev->queues->queue[HostNormCmdQueue].cmdready, &wait);
set_current_state(TASK_INTERRUPTIBLE);
+ dprintk ((KERN_INFO "aac_command_thread start\n"));
while(1)
{
- spin_lock_irqsave(queues->queue[HostNormCmdQueue].lock, flags);
- while(!list_empty(&(queues->queue[HostNormCmdQueue].cmdq))) {
+ spin_lock_irqsave(dev->queues->queue[HostNormCmdQueue].lock, flags);
+ while(!list_empty(&(dev->queues->queue[HostNormCmdQueue].cmdq))) {
struct list_head *entry;
struct aac_aifcmd * aifcmd;
set_current_state(TASK_RUNNING);
-
- entry = queues->queue[HostNormCmdQueue].cmdq.next;
+
+ entry = dev->queues->queue[HostNormCmdQueue].cmdq.next;
list_del(entry);
-
- spin_unlock_irqrestore(queues->queue[HostNormCmdQueue].lock, flags);
+
+ spin_unlock_irqrestore(dev->queues->queue[HostNormCmdQueue].lock, flags);
fib = list_entry(entry, struct fib, fiblink);
/*
* We will process the FIB here or pass it to a
@@ -869,9 +869,54 @@
u32 time_now, time_last;
unsigned long flagv;
+ unsigned num;
+ struct hw_fib ** hw_fib_pool, ** hw_fib_p;
+ struct fib ** fib_pool, ** fib_p;
time_now = jiffies/HZ;
+ /*
+ * Warning: no sleep allowed while
+ * holding spinlock. We take the estimate
+ * and pre-allocate a set of fibs outside the
+ * lock.
+ */
+ num = le32_to_cpu(dev->init->AdapterFibsSize)
+ / sizeof(struct hw_fib); /* some extra */
+ spin_lock_irqsave(&dev->fib_lock, flagv);
+ entry = dev->fib_list.next;
+ while (entry != &dev->fib_list) {
+ entry = entry->next;
+ ++num;
+ }
+ spin_unlock_irqrestore(&dev->fib_lock, flagv);
+ hw_fib_pool = NULL;
+ fib_pool = NULL;
+ if (num
+ && ((hw_fib_pool = kmalloc(sizeof(struct hw_fib *) * num, GFP_KERNEL)))
+ && ((fib_pool = kmalloc(sizeof(struct fib *) * num, GFP_KERNEL)))) {
+ hw_fib_p = hw_fib_pool;
+ fib_p = fib_pool;
+ while (hw_fib_p < &hw_fib_pool[num]) {
+ if (!(*(hw_fib_p++) = kmalloc(sizeof(struct hw_fib), GFP_KERNEL))) {
+ --hw_fib_p;
+ break;
+ }
+ if (!(*(fib_p++) = kmalloc(sizeof(struct fib), GFP_KERNEL))) {
+ kfree(*(--hw_fib_p));
+ break;
+ }
+ }
+ if ((num = hw_fib_p - hw_fib_pool) == 0) {
+ kfree(fib_pool);
+ fib_pool = NULL;
+ kfree(hw_fib_pool);
+ hw_fib_pool = NULL;
+ }
+ } else if (hw_fib_pool) {
+ kfree(hw_fib_pool);
+ hw_fib_pool = NULL;
+ }
spin_lock_irqsave(&dev->fib_lock, flagv);
entry = dev->fib_list.next;
/*
@@ -880,6 +925,8 @@
* fib, and then set the event to wake up the
* thread that is waiting for it.
*/
+ hw_fib_p = hw_fib_pool;
+ fib_p = fib_pool;
while (entry != &dev->fib_list) {
/*
* Extract the fibctx
@@ -912,9 +959,11 @@
* Warning: no sleep allowed while
* holding spinlock
*/
- hw_newfib = kmalloc(sizeof(struct hw_fib), GFP_ATOMIC);
- newfib = kmalloc(sizeof(struct fib), GFP_ATOMIC);
- if (newfib && hw_newfib) {
+ if (hw_fib_p < &hw_fib_pool[num]) {
+ hw_newfib = *hw_fib_p;
+ *(hw_fib_p++) = NULL;
+ newfib = *fib_p;
+ *(fib_p++) = NULL;
/*
* Make the copy of the FIB
*/
@@ -929,15 +978,11 @@
fibctx->count++;
/*
* Set the event to wake up the
- * thread that will waiting.
+ * thread that is waiting.
*/
up(&fibctx->wait_sem);
} else {
printk(KERN_WARNING "aifd: didn't allocate NewFib.\n");
- if(newfib)
- kfree(newfib);
- if(hw_newfib)
- kfree(hw_newfib);
}
entry = entry->next;
}
@@ -947,21 +992,38 @@
*(__le32 *)hw_fib->data = cpu_to_le32(ST_OK);
fib_adapter_complete(fib, sizeof(u32));
spin_unlock_irqrestore(&dev->fib_lock, flagv);
+ /* Free up the remaining resources */
+ hw_fib_p = hw_fib_pool;
+ fib_p = fib_pool;
+ while (hw_fib_p < &hw_fib_pool[num]) {
+ if (*hw_fib_p)
+ kfree(*hw_fib_p);
+ if (*fib_p)
+ kfree(*fib_p);
+ ++fib_p;
+ ++hw_fib_p;
+ }
+ if (hw_fib_pool)
+ kfree(hw_fib_pool);
+ if (fib_pool)
+ kfree(fib_pool);
}
- spin_lock_irqsave(queues->queue[HostNormCmdQueue].lock, flags);
kfree(fib);
+ spin_lock_irqsave(dev->queues->queue[HostNormCmdQueue].lock, flags);
}
/*
* There are no more AIF's
*/
- spin_unlock_irqrestore(queues->queue[HostNormCmdQueue].lock, flags);
+ spin_unlock_irqrestore(dev->queues->queue[HostNormCmdQueue].lock, flags);
schedule();
if(signal_pending(current))
break;
set_current_state(TASK_INTERRUPTIBLE);
}
- remove_wait_queue(&queues->queue[HostNormCmdQueue].cmdready, &wait);
+ if (dev->queues)
+ remove_wait_queue(&dev->queues->queue[HostNormCmdQueue].cmdready, &wait);
dev->aif_thread = 0;
complete_and_exit(&dev->aif_completion, 0);
+ return 0;
}
--
Mark Haverkamp <markh@xxxxxxxx>
-
: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
