Re: [PATCH v5 2/2] Add Intel LGM soc DMA support.

From: Reddy, MallikarjunaX
Date: Thu Aug 27 2020 - 10:46:59 EST


Hi Peter,
Thank you very much for the review and detailed explanation, i will take the inputs and update in the next patch.
My comments inline..

On 8/24/2020 7:24 PM, Peter Ujfalusi wrote:
Hi,

On 24/08/2020 5.30, Reddy, MallikarjunaX wrote:

+static void dma_free_desc_resource(struct virt_dma_desc *vdesc)
+{
+    struct dw2_desc_sw *ds = to_lgm_dma_desc(vdesc);
+    struct ldma_chan *c = ds->chan;
+
+    dma_pool_free(c->desc_pool, ds->desc_hw, ds->desc_phys);
+    kfree(ds);
+    c->ds = NULL;
Is there a chance that c->ds != ds?
No, from the code i don't see any such scenario, let me know if you find
any corner case.
The desc_free callback is used to _free_ up the memory used for the
descriptor. Nothing less, nothing more.
You should not touch the c->ds in this callback, just free up the memory
used for the given vdesc.

More on that a bit later.
Ok.

+}
+
+static struct dw2_desc_sw *
+dma_alloc_desc_resource(int num, struct ldma_chan *c)
+{
+    struct device *dev = c->vchan.chan.device->dev;
+    struct dw2_desc_sw *ds;
+
+    if (num > c->desc_num) {
+        dev_err(dev, "sg num %d exceed max %d\n", num, c->desc_num);
+        return NULL;
+    }
+
+    ds = kzalloc(sizeof(*ds), GFP_NOWAIT);
+    if (!ds)
+        return NULL;
+
+    ds->chan = c;
+
+    ds->desc_hw = dma_pool_zalloc(c->desc_pool, GFP_ATOMIC,
+                      &ds->desc_phys);
+    if (!ds->desc_hw) {
+        dev_dbg(dev, "out of memory for link descriptor\n");
+        kfree(ds);
+        return NULL;
+    }
+    ds->desc_cnt = num;
+
+    return ds;
+}
+
+static void ldma_chan_irq_en(struct ldma_chan *c)
+{
+    struct ldma_dev *d = to_ldma_dev(c->vchan.chan.device);
+    unsigned long flags;
+
+    spin_lock_irqsave(&d->dev_lock, flags);
+    writel(c->nr, d->base + DMA_CS);
+    writel(DMA_CI_EOP, d->base + DMA_CIE);
+    writel(BIT(c->nr), d->base + DMA_IRNEN);
+    spin_unlock_irqrestore(&d->dev_lock, flags);
+}
+
+static void dma_issue_pending(struct dma_chan *chan)
+{
+    struct ldma_chan *c = to_ldma_chan(chan);
+    struct ldma_dev *d = to_ldma_dev(c->vchan.chan.device);
+    unsigned long flags;
+
+    if (d->ver == DMA_VER22) {
+        spin_lock_irqsave(&c->vchan.lock, flags);
+        if (vchan_issue_pending(&c->vchan)) {
+            struct virt_dma_desc *vdesc;
+
+            /* Get the next descriptor */
+            vdesc = vchan_next_desc(&c->vchan);
+            if (!vdesc) {
+                c->ds = NULL;
+                return;
+            }
+            list_del(&vdesc->node);
+            c->ds = to_lgm_dma_desc(vdesc);
you have set c->ds in dma_prep_slave_sg and the only way I can see that
you will not leak memory is that the client must terminate_sync() after
each transfer so that the synchronize callback is invoked between each
prep_sg/issue_pending/competion.
Yes, client must call dmaengine_synchronize after each transfer to make
sure free the memory assoicated with previously issued descriptors if any.
No, client should not need to invoke synchronize between transfers,
clients must be able to do:
dmaengine_prep_slave_single/dmaengine_prep_slave_sg
dma_async_issue_pending
* handle the callback
dmaengine_prep_slave_single/dmaengine_prep_slave_sg
dma_async_issue_pending
* handle the callback

without any terminate_all_sync() in between.
Imagine that the client is preparing/issuing a new transfer in the
completion callback for example.

Clients must be able to do also:
dmaengine_prep_slave_single/dmaengine_prep_slave_sg
dmaengine_prep_slave_single/dmaengine_prep_slave_sg
...
dmaengine_prep_slave_single/dmaengine_prep_slave_sg
dma_async_issue_pending

and then the DMA will complete the transfers in FIFO order and when the
first is completed it will move to the next one. Client will receive
callbacks for each completion (if requested).
Thanks for the detailed explanation peter.

and also from the driver we are freeing up the descriptor from work
queue atfer each transfer.(addressed below comments **)
+            spin_unlock_irqrestore(&c->vchan.lock, flags);
+            ldma_chan_desc_hw_cfg(c, c->ds->desc_phys,
c->ds->desc_cnt);
+            ldma_chan_irq_en(c);
+        }
If there is nothing peding, you will leave the spinlock wide open...
Seems i misplaced the lock. i will fix it in next version.
+    }
+    ldma_chan_on(c);
+}
+
+static void dma_synchronize(struct dma_chan *chan)
+{
+    struct ldma_chan *c = to_ldma_chan(chan);
+
+    /*
+     * clear any pending work if any. In that
+     * case the resource needs to be free here.
+     */
+    cancel_work_sync(&c->work);
+    vchan_synchronize(&c->vchan);
+    if (c->ds)
+        dma_free_desc_resource(&c->ds->vdesc);
+}
+
+static int dma_terminate_all(struct dma_chan *chan)
+{
+    struct ldma_chan *c = to_ldma_chan(chan);
+    unsigned long flags;
+    LIST_HEAD(head);
+
+    spin_lock_irqsave(&c->vchan.lock, flags);
+    vchan_get_all_descriptors(&c->vchan, &head);
+    spin_unlock_irqrestore(&c->vchan.lock, flags);
+    vchan_dma_desc_free_list(&c->vchan, &head);
+
+    return ldma_chan_reset(c);
+}
+
+static int dma_resume_chan(struct dma_chan *chan)
+{
+    struct ldma_chan *c = to_ldma_chan(chan);
+
+    ldma_chan_on(c);
+
+    return 0;
+}
+
+static int dma_pause_chan(struct dma_chan *chan)
+{
+    struct ldma_chan *c = to_ldma_chan(chan);
+
+    return ldma_chan_off(c);
+}
+
+static enum dma_status
+dma_tx_status(struct dma_chan *chan, dma_cookie_t cookie,
+          struct dma_tx_state *txstate)
+{
+    struct ldma_chan *c = to_ldma_chan(chan);
+    struct ldma_dev *d = to_ldma_dev(c->vchan.chan.device);
+    enum dma_status status = DMA_COMPLETE;
+
+    if (d->ver == DMA_VER22)
+        status = dma_cookie_status(chan, cookie, txstate);
+
+    return status;
+}
+
+static void dma_chan_irq(int irq, void *data)
+{
+    struct ldma_chan *c = data;
+    struct ldma_dev *d = to_ldma_dev(c->vchan.chan.device);
+    u32 stat;
+
+    /* Disable channel interrupts  */
+    writel(c->nr, d->base + DMA_CS);
+    stat = readl(d->base + DMA_CIS);
+    if (!stat)
+        return;
+
+    writel(readl(d->base + DMA_CIE) & ~DMA_CI_ALL, d->base + DMA_CIE);
+    writel(stat, d->base + DMA_CIS);
+    queue_work(d->wq, &c->work);
+}
+
+static irqreturn_t dma_interrupt(int irq, void *dev_id)
+{
+    struct ldma_dev *d = dev_id;
+    struct ldma_chan *c;
+    unsigned long irncr;
+    u32 cid;
+
+    irncr = readl(d->base + DMA_IRNCR);
+    if (!irncr) {
+        dev_err(d->dev, "dummy interrupt\n");
+        return IRQ_NONE;
+    }
+
+    for_each_set_bit(cid, &irncr, d->chan_nrs) {
+        /* Mask */
+        writel(readl(d->base + DMA_IRNEN) & ~BIT(cid), d->base +
DMA_IRNEN);
+        /* Ack */
+        writel(readl(d->base + DMA_IRNCR) | BIT(cid), d->base +
DMA_IRNCR);
+
+        c = &d->chans[cid];
+        dma_chan_irq(irq, c);
+    }
+
+    return IRQ_HANDLED;
+}
+
+static struct dma_async_tx_descriptor *
+dma_prep_slave_sg(struct dma_chan *chan, struct scatterlist *sgl,
+          unsigned int sglen, enum dma_transfer_direction dir,
+          unsigned long flags, void *context)
+{
+    struct ldma_chan *c = to_ldma_chan(chan);
+    size_t len, avail, total = 0;
+    struct dw2_desc *hw_ds;
+    struct dw2_desc_sw *ds;
+    struct scatterlist *sg;
+    int num = sglen, i;
+    dma_addr_t addr;
+
+    if (!sgl)
+        return NULL;
+
+    for_each_sg(sgl, sg, sglen, i) {
+        avail = sg_dma_len(sg);
+        if (avail > DMA_MAX_SIZE)
+            num += DIV_ROUND_UP(avail, DMA_MAX_SIZE) - 1;
+    }
+
+    ds = dma_alloc_desc_resource(num, c);
+    if (!ds)
+        return NULL;
+
+    c->ds = ds;
If you still have a transfer running then you are going to get lost that
dscriptor?
No, please let me know if you find any such corner case.
This is the place when you prepare the descriptor, but it is not yet
commited to the hardware. Client might never call issue_pending, client
might prepare another transfer before telling the DMA driver to start
the transfers.
ok

+
+    num = 0;
+    /* sop and eop has to be handled nicely */
+    for_each_sg(sgl, sg, sglen, i) {
+        addr = sg_dma_address(sg);
+        avail = sg_dma_len(sg);
+        total += avail;
+
+        do {
+            len = min_t(size_t, avail, DMA_MAX_SIZE);
+
+            hw_ds = &ds->desc_hw[num];
+            switch (sglen) {
+            case 1:
+                hw_ds->status.field.sop = 1;
+                hw_ds->status.field.eop = 1;
+                break;
+            default:
+                if (num == 0) {
+                    hw_ds->status.field.sop = 1;
+                    hw_ds->status.field.eop = 0;
+                } else if (num == (sglen - 1)) {
+                    hw_ds->status.field.sop = 0;
+                    hw_ds->status.field.eop = 1;
+                } else {
+                    hw_ds->status.field.sop = 0;
+                    hw_ds->status.field.eop = 0;
+                }
+                break;
+            }
+
+            /* Only 32 bit address supported */
+            hw_ds->addr = (u32)addr;
+            hw_ds->status.field.len = len;
+            hw_ds->status.field.c = 0;
+            hw_ds->status.field.bofs = addr & 0x3;
+            /* Ensure data ready before ownership change */
+            wmb();
+            hw_ds->status.field.own = DMA_OWN;
+            /* Ensure ownership changed before moving forward */
+            wmb();
+            num++;
+            addr += len;
+            avail -= len;
+        } while (avail);
+    }
+
+    ds->size = total;
+
+    return vchan_tx_prep(&c->vchan, &ds->vdesc, DMA_CTRL_ACK);
+}
+
+static int
+dma_slave_config(struct dma_chan *chan, struct dma_slave_config *cfg)
+{
+    struct ldma_chan *c = to_ldma_chan(chan);
+    struct ldma_dev *d = to_ldma_dev(c->vchan.chan.device);
+    struct ldma_port *p = c->port;
+    unsigned long flags;
+    u32 bl;
+
+    if ((cfg->direction == DMA_DEV_TO_MEM &&
+         cfg->src_addr_width != DMA_SLAVE_BUSWIDTH_4_BYTES) ||
+        (cfg->direction == DMA_MEM_TO_DEV &&
+         cfg->dst_addr_width != DMA_SLAVE_BUSWIDTH_4_BYTES) ||
According to the probe function these width restrictions are only valid
for DMA_VER22?
YES
Right, hdma_ops does not have device_config specified.

+        !is_slave_direction(cfg->direction))
+        return -EINVAL;
+
+    /* Default setting will be used */
+    if (!cfg->src_maxburst && !cfg->dst_maxburst)
+        return 0;
maxburst == 0 is identical to maxburst == 1, it is just not set
explicitly. Iow 1 word per DMA request.
This is not clear to me. Can you elaborate?
You handle the *_maxburst == 1 and *maxburst == 0 differently while they
are the same thing.
ok, got it. i will fix it in next version.

+
+    /* Must be the same */
+    if (cfg->src_maxburst && cfg->dst_maxburst &&
+        cfg->src_maxburst != cfg->dst_maxburst)
+        return -EINVAL;
+
+    if (cfg->dst_maxburst)
+        cfg->src_maxburst = cfg->dst_maxburst;
+
+    bl = ilog2(cfg->src_maxburst);
+
+    spin_lock_irqsave(&d->dev_lock, flags);
+    writel(p->portid, d->base + DMA_PS);
+    ldma_update_bits(d, DMA_PCTRL_RXBL | DMA_PCTRL_TXBL,
+             FIELD_PREP(DMA_PCTRL_RXBL, bl) |
+             FIELD_PREP(DMA_PCTRL_TXBL, bl), DMA_PCTRL);
+    spin_unlock_irqrestore(&d->dev_lock, flags);
What drivers usually do is to save the cfg and inprepare time take it
into account when setting up the transfer.
Write the change to the HW before the trasnfer is started (if it has
changed from previous settings)

Client drivers usually set the slave config ones, in most cases during
probe, so the slave config rarely changes runtime, but there are cases
for that.
Ok, got it. i will update in the next version.
Thanks

+
+    return 0;
+}
+
+static int dma_alloc_chan_resources(struct dma_chan *chan)
+{
+    struct ldma_chan *c = to_ldma_chan(chan);
+    struct ldma_dev *d = to_ldma_dev(c->vchan.chan.device);
+    struct device *dev = c->vchan.chan.device->dev;
+    size_t    desc_sz;
+
+    if (d->ver > DMA_VER22) {
+        c->flags |= CHAN_IN_USE;
+        return 0;
+    }
+
+    if (c->desc_pool)
+        return c->desc_num;
+
+    desc_sz = c->desc_num * sizeof(struct dw2_desc);
+    c->desc_pool = dma_pool_create(c->name, dev, desc_sz,
+                       __alignof__(struct dw2_desc), 0);
+
+    if (!c->desc_pool) {
+        dev_err(dev, "unable to allocate descriptor pool\n");
+        return -ENOMEM;
+    }
+
+    return c->desc_num;
+}
+
+static void dma_free_chan_resources(struct dma_chan *chan)
+{
+    struct ldma_chan *c = to_ldma_chan(chan);
+    struct ldma_dev *d = to_ldma_dev(c->vchan.chan.device);
+
+    if (d->ver == DMA_VER22) {
+        dma_pool_destroy(c->desc_pool);
+        c->desc_pool = NULL;
+        vchan_free_chan_resources(to_virt_chan(chan));
+        ldma_chan_reset(c);
+    } else {
+        c->flags &= ~CHAN_IN_USE;
+    }
+}
+
+static void dma_work(struct work_struct *work)
+{
+    struct ldma_chan *c = container_of(work, struct ldma_chan, work);
+    struct dma_async_tx_descriptor *tx = &c->ds->vdesc.tx;
+    struct dmaengine_desc_callback cb;
+
+    dmaengine_desc_get_callback(tx, &cb);
+    dma_cookie_complete(tx);
+    dmaengine_desc_callback_invoke(&cb, NULL);
When you are going to free up the descriptor?
**
Seems i missed free up the descriptor here. i will fix and update in the
next version.
This is the place when you can do c->ds = NULL and you should also look
for entries in the issued list to pick up any issued but not commited
transfers.

Let's say the client issued two transfers, in issue_pending you have
started the first one.
Now that is completed, you let the client know about it (you can free up
the descritor as well safely) and then you pick up the next one from the
issued list and start that.
If there is nothing pending then c->ds would be NULL, if there were
pending transfer, you will set c->ds to the next transfer.
c->ds indicates that there is a transfer in progress by the HW.
Thanks peter. I will take your inputs and address the not committed transfers in the next patch.

+
+    dma_dev->device_alloc_chan_resources =
+        d->inst->ops->device_alloc_chan_resources;
+    dma_dev->device_free_chan_resources =
+        d->inst->ops->device_free_chan_resources;
+    dma_dev->device_terminate_all = d->inst->ops->device_terminate_all;
+    dma_dev->device_issue_pending = d->inst->ops->device_issue_pending;
+    dma_dev->device_tx_status = d->inst->ops->device_tx_status;
+    dma_dev->device_resume = d->inst->ops->device_resume;
+    dma_dev->device_pause = d->inst->ops->device_pause;
+    dma_dev->device_config = d->inst->ops->device_config;
+    dma_dev->device_prep_slave_sg = d->inst->ops->device_prep_slave_sg;
+    dma_dev->device_synchronize = d->inst->ops->device_synchronize;
+
+    if (d->ver == DMA_VER22) {
+        dma_dev->src_addr_widths = BIT(DMA_SLAVE_BUSWIDTH_4_BYTES);
+        dma_dev->dst_addr_widths = BIT(DMA_SLAVE_BUSWIDTH_4_BYTES);
+        dma_dev->directions = BIT(DMA_MEM_TO_DEV) |
+                      BIT(DMA_DEV_TO_MEM);
+        dma_dev->residue_granularity =
+                    DMA_RESIDUE_GRANULARITY_DESCRIPTOR;
+    }
So, if version is != DMA_VER22, then you don't support any direction?
Why register the DMA device if it can not do any transfer?
Only dma0 instance (intel,lgm-cdma) is used as a general purpose slave
DMA. we set both control and datapath here.
Other instances we set only control path. data path is taken care by dma
client(GSWIP).
How the client (GSWIP) can request a channel from intel,lgm-* ? Don't
you need some capabilities for the DMA device so core can sort out the
request?
client request channel by name, dma_request_slave_channel(dev, name);

Only thing needs to do is get the channel, set the descriptor and just
on the channel.
How do you 'on' the channel?
we on the channel in issue_pending.

+
+    platform_set_drvdata(pdev, d);
+
+    ldma_dev_init(d);
+
+    ret = dma_async_device_register(dma_dev);
+    if (ret) {
+        dev_err(dev, "Failed to register slave DMA engine device\n");
+        return ret;
+    }
+
+    ret = of_dma_controller_register(pdev->dev.of_node, ldma_xlate, d);
+    if (ret) {
+        dev_err(dev, "Failed to register of DMA controller\n");
+        dma_async_device_unregister(dma_dev);
+        return ret;
+    }
+
+    dev_info(dev, "Init done - rev: %x, ports: %d channels: %d\n",
d->ver,
+         d->port_nrs, d->chan_nrs);
+
+    return 0;
+}
+
+static struct platform_driver intel_ldma_driver = {
+    .probe = intel_ldma_probe,
+    .driver = {
+        .name = DRIVER_NAME,
+        .of_match_table = intel_ldma_match,
+    },
+};
+
+static int __init intel_ldma_init(void)
+{
+    return platform_driver_register(&intel_ldma_driver);
+}
+
+device_initcall(intel_ldma_init);
diff --git a/include/linux/dma/lgm_dma.h b/include/linux/dma/lgm_dma.h
new file mode 100644
index 000000000000..3a2ee6ad0710
--- /dev/null
+++ b/include/linux/dma/lgm_dma.h
@@ -0,0 +1,27 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * Copyright (c) 2016 ~ 2019 Intel Corporation.
+ */
+#ifndef LGM_DMA_H
+#define LGM_DMA_H
+
+#include <linux/types.h>
+#include <linux/dmaengine.h>
+
+/*!
+ * \fn int intel_dma_chan_desc_cfg(struct dma_chan *chan, dma_addr_t
desc_base,
+ *                                 int desc_num)
+ * \brief Configure low level channel descriptors
+ * \param[in] chan   pointer to DMA channel that the client is using
+ * \param[in] desc_base   descriptor base physical address
+ * \param[in] desc_num   number of descriptors
+ * \return   0 on success
+ * \return   kernel bug reported on failure
+ *
+ * This function configure the low level channel descriptors. It
will be
+ * used by CBM whose descriptor is not DDR, actually some registers.
+ */
+int intel_dma_chan_desc_cfg(struct dma_chan *chan, dma_addr_t
desc_base,
+                int desc_num);
+
+#endif /* LGM_DMA_H */

- Péter

Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki

- Péter

Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki