[PATCH v9 2/2] iommu/exynos: Add iommu driver for Exynos Platforms
KyongHo Cho
pullip.cho at samsung.com
Wed Feb 29 00:25:45 EST 2012
On Tue, Feb 28, 2012 at 4:53 PM, Kyungmin Park
<kyungmin.park at samsung.com> wrote:
>> +void exynos_sysmmu_set_prefbuf(struct device *owner,
>> + unsigned long base0, unsigned long size0,
>> + unsigned long base1, unsigned long size1)
>> +{
>> + struct sysmmu_drvdata *data = dev_get_drvdata(owner->archdata.iommu);
>> + unsigned long flags;
>> + int i;
>> +
>> + BUG_ON((base0 + (size0 - 1)) <= base0);
>> + BUG_ON((base1 + (size1 - 1)) <= base1);
> Do you want to check size? BUG_ON(size <= 1);?
My mistake. :)
Thank you.
>> +
>> + read_lock_irqsave(&data->lock, flags);
>> + if (!is_sysmmu_active(data))
>> + goto finish;
>> +
>> + for (i = 0; i < data->nsfrs; i++) {
>> + if ((readl(data->sfrbases[i] + S5P_MMU_VERSION) >> 28) == 3) {
>> + sysmmu_block(data->sfrbases[i]);
>> +
>> + if (size1 == 0) {
> Is it possible? if size1 is '0', it can't pass the BUG_ON condition.
Although the above BUG_ON condition is incorrect, it can pass if size1
and base1 are 0
because the type of them is unsigned.
>> +
>> + for (i = 0; i < data->nsfrs; i++) {
>> + __sysmmu_set_ptbase(data->sfrbases[i], pgtable);
>> +
>> + if ((readl(data->sfrbases[i] + S5P_MMU_VERSION) >> 28) == 3) {
>> + /* System MMU version is 3.x */
>> + __raw_writel((1 << 12) | (2 << 28),
> Can you use the DEFINE instead of hard-code?
Do you think it is required even though it is used nowhere else here?
>> +static int exynos_iommu_attach_device(struct iommu_domain *domain,
>> + struct device *dev)
>> +{
>> + int ret;
>> + struct exynos_iommu_domain *priv = domain->priv;
>> + struct iommu_client *client = NULL;
>> + struct list_head *pos;
>> + unsigned long flags;
>> +
>> + spin_lock_irqsave(&priv->lock, flags);
>> +
>> + list_for_each(pos, &priv->clients) {
> Simply list_for_each_entry.
Variable 'client' must not be used as a loop cursor
because its value must not be changed unless a condition meets.
>> + struct iommu_client *cur;
>> +
>> + cur = list_entry(pos, struct iommu_client, node);
>> + if (cur->dev == dev) {
>> + client = cur;
>> + break;
>> + }
>> + }
>> +
>> + if (client != NULL) {
>> + dev_dbg(dev, "%s: IOMMU with pgtable 0x%lx already attached\n",
>> + __func__, __pa(priv->pgtable));
>> + client->refcnt++;
>> + }
>> +
>> + spin_unlock_irqrestore(&priv->lock, flags);
>> +
>> + if (client != NULL)
>> + return 0;
>> +
>> + client = kmalloc(sizeof(*client), GFP_KERNEL);
> Maybe attach called frequently. how about to use kmem_cache-*?
Thank you for advice.
>> + if (!client)
>> + return -ENOMEM;
>> +
>> + INIT_LIST_HEAD(&client->node);
>> + client->dev = dev;
>> + client->refcnt = 1;
> Dose it possible attach more than one? OMAP has multiple attach codes.
Yes. This function returns earlier than this if client->refcnt is larger than 1.
Please check "if (client != NULL) return 0;" statement in this function.
>> + ret = __exynos_sysmmu_enable(dev, __pa(priv->pgtable), domain);
>> + if (ret) {
>> + kfree(client);
>> + return ret;
>> + }
>> +
>> + spin_lock_irqsave(&priv->lock, flags);
>> + list_add_tail(&client->node, &priv->clients);
>> + spin_unlock_irqrestore(&priv->lock, flags);
>> +
>> + dev_dbg(dev, "%s: Attached new IOMMU with pgtable 0x%lx\n", __func__,
>> + __pa(priv->pgtable));
>> + return 0;
>> +}
>> +
>> +static void exynos_iommu_detach_device(struct iommu_domain *domain,
>> + struct device *dev)
>> +{
>> + struct exynos_iommu_domain *priv = domain->priv;
>> + struct iommu_client *client = NULL;
>> + struct list_head *pos;
>> + unsigned long flags;
>> +
>> + spin_lock_irqsave(&priv->lock, flags);
>> +
>> + list_for_each(pos, &priv->clients) {
>> + struct iommu_client *cur;
>> +
>> + cur = list_entry(pos, struct iommu_client, node);
>> + if (cur->dev == dev) {
>> + cur->refcnt--;
>> + client = cur;
>> + break;
>> + }
>> + }
>> +
>> + spin_unlock_irqrestore(&priv->lock, flags);
>> +
>> + if (WARN_ON(client == NULL))
>> + return;
>> +
>> + if (client->refcnt > 0) {
> It never triggered. as you use true/false scheme. I remember you said
> previous patch. use the refcount but actual meaning is true/false.
I think you are talking about the conversations about v6 patchset.
client->refcnt is really a reference counter.
>> + dev_dbg(dev, "%s: Detaching IOMMU with pgtable 0x%lx delayed\n",
>> + __func__, __pa(priv->pgtable));
>> + return;
>> + }
>> +
>> + BUG_ON(client->refcnt != 0);
> Do you think "minus value"?
Yes. but I think it never be happened logically.
It is just "assert(client->refcnt == 0)".
May it is better to remove the BUG_ON.
Thank you for kind review.
Cho KyongHo.
More information about the linux-arm-kernel
mailing list