evie/android_kernel_oneplus_msm8998
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

coresight: stm: Remove spin lock usage for channel allocation

Browse source 
Current usage of spinlock in channel allocation might result in
live lock issue. In order to avoid this we move to a per cpu
channel allocation scheme. We rely on preemption being disabled
when we allocate channel to avoid any conflict. If we end up
handling an interrupt which logs when preemption is disabled it
is still gauranteed to not interfere with the current logging
as the channel id will be freed up before resuming.

Change-Id: Ia343890dfd9bded84db75e8b13ad1860dec1bbcd
Signed-off-by: Satyajit Desai <sadesai@codeaurora.org>
This commit is contained in:
Satyajit Desai 2016-12-02 17:13:27 -08:00
parent a27b2f1789
commit 9d9a22ef61
1 changed files with 31 additions and 16 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

47
drivers/hwtracing/coresight/coresight-stm.c
View file

@ -141,6 +141,7 @@ struct stm_drvdata {
bool enable;
DECLARE_BITMAP(entities, OST_ENTITY_MAX);
bool data_barrier;
uint32_t ch_alloc_fail_count;
};
static struct stm_drvdata *stmdrvdata;
@ -385,19 +386,26 @@ static const struct coresight_ops stm_cs_ops = {
.source_ops = &stm_source_ops,
};
static uint32_t stm_channel_alloc(uint32_t off)
static uint32_t stm_channel_alloc(void)
{
struct stm_drvdata *drvdata = stmdrvdata;
uint32_t ch;
unsigned long flags;
uint32_t off, ch, num_ch_per_cpu;
int cpu;
spin_lock_irqsave(&drvdata->spinlock, flags);
do {
ch = find_next_zero_bit(drvdata->chs.bitmap,
NR_STM_CHANNEL, off);
} while ((ch < NR_STM_CHANNEL) &&
test_and_set_bit(ch, drvdata->chs.bitmap));
spin_unlock_irqrestore(&drvdata->spinlock, flags);
num_ch_per_cpu = NR_STM_CHANNEL/num_present_cpus();
cpu = get_cpu();
off = num_ch_per_cpu * cpu;
ch = find_next_zero_bit(drvdata->chs.bitmap,
NR_STM_CHANNEL, off);
if (ch > (off + num_ch_per_cpu)) {
put_cpu();
return NR_STM_CHANNEL;
}
set_bit(ch, drvdata->chs.bitmap);
put_cpu();
return ch;
}
@ -405,11 +413,8 @@ static uint32_t stm_channel_alloc(uint32_t off)
static void stm_channel_free(uint32_t ch)
{
struct stm_drvdata *drvdata = stmdrvdata;
unsigned long flags;
spin_lock_irqsave(&drvdata->spinlock, flags);
clear_bit(ch, drvdata->chs.bitmap);
spin_unlock_irqrestore(&drvdata->spinlock, flags);
}
static int stm_send(void *addr, const void *data, uint32_t size)
@ -522,7 +527,14 @@ static inline int __stm_trace(uint32_t options, uint8_t entity_id,
unsigned long ch_addr;
/* allocate channel and get the channel address */
ch = stm_channel_alloc(0);
ch = stm_channel_alloc();
if (ch >= NR_STM_CHANNEL) {
drvdata->ch_alloc_fail_count++;
dev_err_ratelimited(drvdata->dev,
"Channel allocation failed %d",
drvdata->ch_alloc_fail_count);
return 0;
}
ch_addr = (unsigned long)stm_channel_addr(drvdata, ch);
/* send the ost header */
@ -789,15 +801,18 @@ static int stm_probe(struct amba_device *adev, const struct amba_id *id)
if (boot_nr_channel) {
res_size = min((resource_size_t)(boot_nr_channel *
BYTES_PER_CHANNEL), resource_size(&res));
bitmap_size = boot_nr_channel * sizeof(long);
bitmap_size = (boot_nr_channel + sizeof(long) - 1) /
sizeof(long);
} else {
res_size = min((resource_size_t)(NR_STM_CHANNEL *
BYTES_PER_CHANNEL), resource_size(&res));
bitmap_size = NR_STM_CHANNEL * sizeof(long);
bitmap_size = (NR_STM_CHANNEL + sizeof(long) - 1) /
sizeof(long);
}
drvdata->chs.base = devm_ioremap(dev, res.start, res_size);
if (!drvdata->chs.base)
return -ENOMEM;
drvdata->chs.bitmap = devm_kzalloc(dev, bitmap_size, GFP_KERNEL);
if (!drvdata->chs.bitmap)
return -ENOMEM;