Re: [PATCH v9 6/6] drivers: remoteproc: Add Xilinx r5 remoteproc driver

From: Tanmay Shah
Date: Thu Sep 08 2022 - 15:42:46 EST


Hi Mathieu,

Thanks for reviewing this patch-set. Ack to comments and will fix in new revision.

On 9/2/22 12:59 PM, Mathieu Poirier wrote:
[...]

+
+static const struct rproc_ops zynqmp_r5_rproc_ops = {
+ .prepare = zynqmp_r5_rproc_prepare,
+ .unprepare = zynqmp_r5_rproc_unprepare,
+ .start = zynqmp_r5_rproc_start,
+ .stop = zynqmp_r5_rproc_stop,
+ .load = rproc_elf_load_segments,
+ .parse_fw = zynqmp_r5_parse_fw,
+ .find_loaded_rsc_table = rproc_elf_find_loaded_rsc_table,
+ .sanity_check = rproc_elf_sanity_check,
+ .get_boot_addr = rproc_elf_get_boot_addr,
+};
+
+/**
+ * zynqmp_r5_add_rproc_core()
+ * Probes ZynqMP R5 processor device node
+ * this is called for each individual R5 core
+ *
+ * @cdev: Device node of each r5 core
+ *
+ * Return: zynqmp_r5_core object for success, error pointer in case of error.
+ */
+static struct zynqmp_r5_core *zynqmp_r5_add_rproc_core(struct device *cdev)
+{
+ struct zynqmp_r5_core *r5_core;
+ struct rproc *r5_rproc;
+ int ret;
+
+ /* Set up DMA mask */
+ ret = dma_set_coherent_mask(cdev, DMA_BIT_MASK(32));
+ if (ret)
+ return ERR_PTR(ret);
+
+ /* Allocate remoteproc instance */
+ r5_rproc = rproc_alloc(cdev, dev_name(cdev),
+ &zynqmp_r5_rproc_ops,
+ NULL, sizeof(struct zynqmp_r5_core));
+ if (!r5_rproc) {
+ dev_err(cdev, "failed to allocate memory for rproc instance\n");
+ return ERR_PTR(-ENOMEM);
+ }
+
+ r5_rproc->auto_boot = false;
+ r5_core = (struct zynqmp_r5_core *)r5_rproc->priv;
+ r5_core->dev = cdev;
+ r5_core->np = dev_of_node(cdev);
+ if (!r5_core->np) {
+ dev_err(cdev, "can't get device node for r5 core\n");
+ return ERR_PTR(-EINVAL);
Here @r5_rproc is leaked and r5_rproc->rproc never free'ed.
This function is only used in cluster_init.

r5_rproc->rproc is free'ed in release_r5_core label in
zynqmp_r5_cluster_init function if this function fails.
It would be the case if @r5_core was returned, but it is not. A such
@r5_core[i] zynqmp_r5_cluster_init() is still null and the memory leaked.

Ack. Yes you are right. Thanks!!


This avoids multiple free of same resource.

+ }
+
+ /* Add R5 remoteproc core */
+ ret = rproc_add(r5_rproc);
+ if (ret) {
+ dev_err(cdev, "failed to add r5 remoteproc\n");
+ return ERR_PTR(ret);
Same.
Same as above.
Ack. I will fix this.
+ }
+
+ r5_core->rproc = r5_rproc;
+ return r5_core;
+}
+
+/**
+ * zynqmp_r5_get_tcm_node()
+ * Ideally this function should parse tcm node and store information
+ * in r5_core instance. We will use hardcoded TCM information from
+ * driver for now in this function.
+ *
It would be a good idea to explain _why_ hardcoded values are used.
It is explained in function, but will add in description too.

+ * @cluster: pointer to zynqmp_r5_cluster type object
+ *
+ * Return: 0 for success and error code for failure.
+ */
+static int zynqmp_r5_get_tcm_node(struct zynqmp_r5_cluster *cluster)
+{
+ struct device *dev = cluster->dev;
+ struct zynqmp_r5_core *r5_core;
+ int tcm_bank_count, tcm_node;
+ int i, j;
+
+ /*
+ * ToDo: Use predefined TCM address space values from driver while
I'm not sure why there is a "ToDo" here since this is exactly what you are
doing.
Ack. ToDo needs to go.
+ * system-dt spec is not final for TCM
+ */
+ tcm_bank_count = ARRAY_SIZE(zynqmp_tcm_banks);
+
+ /* count per core tcm banks */
+ tcm_bank_count = tcm_bank_count / cluster->core_count;
+
+ if (tcm_bank_count <= 0)
+ return -EINVAL;
As far as I can tell this can't happen and as such should be removed.
I kept this check for future, when we add TCM dt nodes. I will remove it.

+
+ /*
+ * r5 core 0 will use all of TCM banks in lockstep mode.
+ * In split mode, r5 core0 will use 128k and r5 core1 will use another
+ * 128k. Assign TCM banks to each core accordingly
+ */
+ tcm_node = 0;
+ for (i = 0; i < cluster->core_count; i++) {
+ r5_core = cluster->r5_cores[i];
+ r5_core->tcm_banks = devm_kcalloc(dev, tcm_bank_count,
+ sizeof(struct mem_bank_data *),
+ GFP_KERNEL);
+ if (!r5_core->tcm_banks)
+ return -ENOMEM;
+
+ for (j = 0; j < tcm_bank_count; j++) {
+ /*
+ * Use pre-defined TCM reg values.
+ * Eventually this should be replaced by values
+ * parsed from dts.
+ */
+ r5_core->tcm_banks[j] =
+ (struct mem_bank_data *)&zynqmp_tcm_banks[tcm_node];
+ tcm_node++;
+ }
+
+ r5_core->tcm_bank_count = tcm_bank_count;
+ }
+
+ return 0;
+}
+
+/**
+ * zynqmp_r5_get_mem_region_node()
+ * parse memory-region property from dt node and add
+ * memory region carveouts
+ *
+ * @r5_core: pointer to zynqmp_r5_core type object
+ *
+ * Return: 0 for success and error code for failure.
+ */
+static int zynqmp_r5_get_mem_region_node(struct zynqmp_r5_core *r5_core)
+{
+ struct device_node *np, *rmem_np;
+ struct reserved_mem **rmem;
+ int res_mem_count, i;
+ struct device *dev;
+
+ dev = r5_core->dev;
+ np = r5_core->np;
+
+ res_mem_count = of_property_count_elems_of_size(np, "memory-region",
+ sizeof(phandle));
+ if (res_mem_count <= 0) {
+ dev_warn(dev, "failed to get memory-region property %d\n",
+ res_mem_count);
+ r5_core->rmem_count = 0;
+ return 0;
+ }
+
+ rmem = devm_kcalloc(dev, res_mem_count,
+ sizeof(struct reserved_mem *), GFP_KERNEL);
+ if (!rmem)
+ return -ENOMEM;
+
+ for (i = 0; i < res_mem_count; i++) {
+ rmem_np = of_parse_phandle(np, "memory-region", i);
+ if (!rmem_np)
+ goto release_rmem;
+
+ rmem[i] = of_reserved_mem_lookup(rmem_np);
+ if (!rmem[i]) {
+ of_node_put(rmem_np);
+ goto release_rmem;
+ }
+
+ of_node_put(rmem_np);
+ }
+
+ r5_core->rmem_count = res_mem_count;
+ r5_core->rmem = rmem;
+ return 0;
+
+release_rmem:
+ for (i--; i >= 0; i--)
+ kfree(rmem[i]);
Function of_reserved_mem_lookup() doesn't allocate a new reserved_mem structure,
you get a reference to an existing one. As such there is no need to explicitly
free individual elements of the rmem array.
Ack.
+ devm_kfree(dev, rmem);
+ return -ENOMEM;
+}
+
+/*
+ * zynqmp_r5_core_init()
+ * Create and initialize zynqmp_r5_core type object
+ *
+ * @cluster: pointer to zynqmp_r5_cluster type object
+ *
+ * Return: 0 for success and error code for failure.
+ */
+static int zynqmp_r5_core_init(struct zynqmp_r5_cluster *cluster,
+ enum rpu_oper_mode fw_reg_val,
+ enum rpu_tcm_comb tcm_mode)
+{
+ struct device *dev = cluster->dev;
+ struct zynqmp_r5_core *r5_core;
+ int ret, i;
+
+ ret = zynqmp_r5_get_tcm_node(cluster);
+ if (ret < 0) {
+ dev_err(dev, "can't get tcm node, err %d\n", ret);
+ return ret;
+ }
+
+ for (i = 0; i < cluster->core_count; i++) {
+ r5_core = cluster->r5_cores[i];
+
+ ret = zynqmp_r5_get_mem_region_node(r5_core);
+ if (ret)
+ dev_warn(dev, "memory-region prop failed %d\n", ret);
+
+ /* Initialize r5 cores with power-domains parsed from dts */
+ ret = of_property_read_u32_index(r5_core->np, "power-domains",
+ 1, &r5_core->pm_domain_id);
+ if (ret) {
+ dev_err(dev, "failed to get power-domains property\n");
+ return ret;
+ }
+
+ ret = zynqmp_r5_set_mode(r5_core, fw_reg_val, tcm_mode);
+ if (ret) {
+ dev_err(dev, "failed to set r5 cluster mode %d, err %d\n",
+ cluster->mode, ret);
+ return ret;
+ }
+ }
+
+ return 0;
+}
+
+/*
+ * zynqmp_r5_cluster_init()
+ * Create and initialize zynqmp_r5_cluster type object
+ *
+ * @cluster: pointer to zynqmp_r5_cluster type object
+ *
+ * Return: 0 for success and error code for failure.
+ */
+static int zynqmp_r5_cluster_init(struct zynqmp_r5_cluster *cluster)
+{
+ enum zynqmp_r5_cluster_mode cluster_mode = LOCKSTEP_MODE;
+ struct device *dev = cluster->dev;
+ struct device_node *dev_node = dev_of_node(dev);
+ struct platform_device *child_pdev;
+ struct zynqmp_r5_core **r5_cores;
+ enum rpu_oper_mode fw_reg_val;
+ struct device **child_devs;
+ struct device_node *child;
+ enum rpu_tcm_comb tcm_mode;
+ int core_count, ret, i;
+
+ ret = of_property_read_u32(dev_node, "xlnx,cluster-mode", &cluster_mode);
+
+ /*
+ * on success returns 0, if not defined then returns -EINVAL,
+ * In that case, default is LOCKSTEP mode
+ */
+ if (ret != -EINVAL && ret != 0) {
+ dev_err(dev, "Invalid xlnx,cluster-mode property\n");
+ return -EINVAL;
+ }
+
+ /*
+ * For now driver only supports split mode and lockstep mode.
+ * fail driver probe if either of that is not set in dts.
+ */
+ if (cluster_mode == LOCKSTEP_MODE) {
Why is are LOCKSTEP_MODE and LOCKSTEP_MODE defined when the same values were
already added in xlnx-zynqmp.h in patch 3/6?
Actually LOCKSTEP_MODE defined in this driver is taken from dt-bindings of
this driver. However, PM_RPU_MODE_LOCKSTEP defined in firmware represents
value expected by firmware. They are not the same defines. LOCKSTEP_MODE = 1
which is defined as per xlnx,cluster-mode bindings and similar to ti
cluster-mode bindings. PM_RPU_MODE_LOCKSTEP = 0 which is expected from
firmware.
Ah yes, you are correct.


And why not add SINGLE_CPU_MODE
there as well instead of duplicating things in enum zynqmp_r5_cluster_mode?

SINGLE_CPU_MODE just represents xlnx,cluster-mode value from dt-bindings. It
is not supported in firmware.

There are multiple ways to configure r5 cluster in single-cpu mode.

1) Add new EEMI call which asks firmware to configure cluster in single-cpu
mode.

or 2) use existing EEMI calls and configure cores from driver. As
single-cpu-mode is not supported yet, I am not sure, how it will

be implemented or what its design will look like. So, I haven't modified
anything in firmware driver related to single-cpu mode.


That is fine.

+ tcm_mode = PM_RPU_TCM_COMB;
+ fw_reg_val = PM_RPU_MODE_LOCKSTEP;
+ } else if (cluster_mode == SPLIT_MODE) {
+ tcm_mode = PM_RPU_TCM_SPLIT;
+ fw_reg_val = PM_RPU_MODE_SPLIT;
+ } else {
+ dev_err(dev, "driver does not support cluster mode %d\n", cluster_mode);
+ return -EINVAL;
+ }
+
+ /*
+ * Number of cores is decided by number of child nodes of
+ * r5f subsystem node in dts. If Split mode is used in dts
+ * 2 child nodes are expected.
+ * In lockstep mode if two child nodes are available,
+ * only use first child node and consider it as core0
+ * and ignore core1 dt node.
+ */
+ core_count = of_get_available_child_count(dev_node);
+ if (core_count <= 0) {
+ dev_err(dev, "Invalid number of r5 cores %d", core_count);
+ return -EINVAL;
+ } else if (cluster_mode == SPLIT_MODE && core_count != 2) {
+ dev_err(dev, "Invalid number of r5 cores for split mode\n");
+ return -EINVAL;
+ } else if (cluster_mode == LOCKSTEP_MODE && core_count == 2) {
+ dev_warn(dev, "Only r5 core0 will be used\n");
+ core_count = 1;
+ }
+
+ child_devs = kcalloc(core_count, sizeof(struct device *), GFP_KERNEL);
+ if (!child_devs)
+ return -ENOMEM;
+
+ r5_cores = kcalloc(core_count,
+ sizeof(struct zynqmp_r5_core *), GFP_KERNEL);
+ if (!r5_cores) {
+ kfree(child_devs);
+ return -ENOMEM;
+ }
+
+ i = 0;
+ for_each_available_child_of_node(dev_node, child) {
+ child_pdev = of_find_device_by_node(child);
+ if (!child_pdev) {
+ of_node_put(child);
+ ret = -ENODEV;
+ goto release_r5_cores;
+ }
+
+ child_devs[i] = &child_pdev->dev;
+
+ /* create and add remoteproc instance of type struct rproc */
+ r5_cores[i] = zynqmp_r5_add_rproc_core(&child_pdev->dev);
+ if (IS_ERR(r5_cores[i])) {
+ of_node_put(child);
+ ret = PTR_ERR(r5_cores[i]);
+ goto release_r5_cores;
If we are here and i == 0, we are leaking a reference to child_pdev since the
loop in release_r5_cores won't execute.
Ack. I will check if i == 0 and put_device() before jumping to
release_r5_cores.
+ }
+
+ i++;
+
+ /*
+ * If two child nodes are available in dts in lockstep mode,
+ * then ignore second child node.
+ */
+ if (i == core_count) {
+ of_node_put(child);
+ break;
+ }
Instead of using the i to check the cores, simply break if cluster_mode ==
LOCKSTEP_MODE. That will avoid the fragile dance around the manipulation of
variable i and simplify the release of resources in release_r5_cores.

Actually variable name "i" needs to be changed. I need that variable to
track how many cores  are added successfully.

So, if first core is added and second core fails, then I can release
resources allocated for first core successfully.

So, instead of "i" I can use meaningful variable name like "cores_added" or
"cores_allocated".

You can still use variable @i, just don't use it to break out of the loop. As I
pointed out, it would be much easier to simply do that if in lockstep mode.

Ah! makes sense. I will check for lockstep mode instead of @i counts.

Thanks.

+ }
+
+ cluster->mode = cluster_mode;
+ cluster->core_count = core_count;
+ cluster->r5_cores = r5_cores;
+
+ ret = zynqmp_r5_core_init(cluster, fw_reg_val, tcm_mode);
+ if (ret < 0) {
+ dev_err(dev, "failed to init r5 core err %d\n", ret);
+ cluster->core_count = 0;
+ cluster->r5_cores = NULL;
+ goto release_r5_cores;
+ }
+
+ kfree(child_devs);
+ return 0;
+
+release_r5_cores:
+ for (i--; i >= 0; i--) {
+ put_device(child_devs[i]);
+ rproc_del(r5_cores[i]->rproc);
+ rproc_free(r5_cores[i]->rproc);
+ }
+ kfree(r5_cores);
+ kfree(child_devs);
+ return ret;
+}
+
+static void zynqmp_r5_cluster_exit(void *data)
+{
+ struct platform_device *pdev = (struct platform_device *)data;
+ struct zynqmp_r5_cluster *cluster;
+ struct zynqmp_r5_core *r5_core;
+ int i;
+
+ cluster = (struct zynqmp_r5_cluster *)platform_get_drvdata(pdev);
+ if (!cluster)
+ return;
+
+ for (i = 0; i < cluster->core_count; i++) {
+ r5_core = cluster->r5_cores[i];
+ put_device(r5_core->dev);
+ rproc_del(r5_core->rproc);
+ rproc_free(r5_core->rproc);
+ }
+
+ kfree(cluster->r5_cores);
+ kfree(cluster);
+ platform_set_drvdata(pdev, NULL);
+}
+
+/*
+ * zynqmp_r5_remoteproc_probe()
+ *
+ * @pdev: domain platform device for R5 cluster
+ *
+ * called when driver is probed, for each R5 core specified in DT,
+ * setup as needed to do remoteproc-related operations
+ *
+ * Return: 0 for success, negative value for failure.
+ */
+static int zynqmp_r5_remoteproc_probe(struct platform_device *pdev)
+{
+ struct zynqmp_r5_cluster *cluster;
+ struct device *dev = &pdev->dev;
+ int ret;
+
+ cluster = kzalloc(sizeof(*cluster), GFP_KERNEL);
+ if (!cluster)
+ return -ENOMEM;
+
+ cluster->dev = dev;
+
+ ret = devm_of_platform_populate(dev);
+ if (ret) {
+ dev_err_probe(dev, ret, "failed to populate platform dev\n");
+ kfree(cluster);
+ return ret;
+ }
+
+ /* wire in so each core can be cleaned up at driver remove */
+ platform_set_drvdata(pdev, cluster);
+
+ ret = zynqmp_r5_cluster_init(cluster);
+ if (ret) {
+ zynqmp_r5_cluster_exit(pdev);
+ dev_err_probe(dev, ret, "Invalid r5f subsystem device tree\n");
+ return ret;
+ }
+
+ ret = devm_add_action_or_reset(dev, zynqmp_r5_cluster_exit, pdev);
+ if (ret)
+ return ret;
+
+ return 0;
+}
+
+/* Match table for OF platform binding */
+static const struct of_device_id zynqmp_r5_remoteproc_match[] = {
+ { .compatible = "xlnx,zynqmp-r5fss", },
+ { /* end of list */ },
+};
+MODULE_DEVICE_TABLE(of, zynqmp_r5_remoteproc_match);
+
+static struct platform_driver zynqmp_r5_remoteproc_driver = {
+ .probe = zynqmp_r5_remoteproc_probe,
+ .driver = {
+ .name = "zynqmp_r5_remoteproc",
+ .of_match_table = zynqmp_r5_remoteproc_match,
+ },
+};
+module_platform_driver(zynqmp_r5_remoteproc_driver);
+
+MODULE_DESCRIPTION("Xilinx R5F remote processor driver");
+MODULE_AUTHOR("Xilinx Inc.");
+MODULE_LICENSE("GPL");
There is a discrepency between the GPL-2.0 in the SPDS identifier and the above.

More comments tomorrow or Tuesday.

Thanks,
Mathieu

--
2.25.1