[PATCH] RapidIO: Add mport driver for Tsi721 bridge
Bounine, Alexandre
Alexandre.Bounine at idt.com
Thu Aug 18 06:21:43 EST 2011
Andrew Morton wrote:
> On Fri, 12 Aug 2011 15:45:34 -0400
> Alexandre Bounine <alexandre.bounine at idt.com> wrote:
>
> > Add RapidIO mport driver for IDT TSI721 PCI Express-to-SRIO bridge
device.
>
> What a huge driver.
It is not over yet. We have plans to add more stuff later.
> >
> > ...
> > +config RAPIDIO_TSI721
> > + bool "IDT Tsi721 PCI Express SRIO Controller support"
> > + depends on RAPIDIO && PCI && PCIEPORTBUS
>
> The dependency on PCI is redundant - PCIEPORTBUS already depends on
> PCI. Doesn't matter much though.
Agree. Will clean it for next release.
> > + default "n"
> > + ---help---
> > + Include support for IDT Tsi721 PCI Express Serial RapidIO
> controller.
> >
> > ...
> >
> > + /* Wait until DMA transfer is finished */
> > + while ((ch_stat = ioread32(priv->regs +
> > + TSI721_DMAC_STS(TSI721_DMACH_MAINT))) &
TSI721_DMAC_STS_RUN) {
> > + udelay(10);
> > + i++;
> > + if (i >= 5000000) {
> > + dev_dbg(&priv->pdev->dev,
> > + "%s : DMA[%d] read timeout
ch_status=%x\n",
> > + __func__, TSI721_DMACH_MAINT, ch_stat);
> > + if (!do_wr)
> > + *data = 0xffffffff;
> > + err = -EFAULT;
> > + goto err_out;
> > + }
> > + }
>
> Fifty seconds!?!?
Should be "udelay(1)" - 10 usec is too much here.
> EFAULT seems an inappropriate errno.
Switching to EIO.
> > + if (ch_stat & TSI721_DMAC_STS_ABORT) {
> > + /* If DMA operation aborted due to error,
> > + * reinitialize DMA channel
> > + */
.....
> > + iowrite32(0, priv->regs +
> > + TSI721_DMAC_DWRCNT(TSI721_DMACH_MAINT));
> > + udelay(1);
> > + if (!do_wr)
> > + *data = 0xffffffff;
> > + err = -EFAULT;
Here EIO as well because ABORT error reports more events than just bad
address.
> > ...
> >
> > +static int
> > +tsi721_pw_handler(struct rio_mport *mport)
> > +{
> > + struct tsi721_device *priv = mport->priv;
> > + u32 pw_stat;
> > + u32 pw_buf[TSI721_RIO_PW_MSG_SIZE/sizeof(u32)];
> > +
> > +
> > + pw_stat = ioread32(priv->regs + TSI721_RIO_PW_RX_STAT);
> > +
> > + if (pw_stat & TSI721_RIO_PW_RX_STAT_PW_VAL) {
> > + pw_buf[0] = ioread32(priv->regs +
TSI721_RIO_PW_RX_CAPT(0));
> > + pw_buf[1] = ioread32(priv->regs +
TSI721_RIO_PW_RX_CAPT(1));
> > + pw_buf[2] = ioread32(priv->regs +
TSI721_RIO_PW_RX_CAPT(2));
> > + pw_buf[3] = ioread32(priv->regs +
TSI721_RIO_PW_RX_CAPT(3));
> > +
> > + /* Queue PW message (if there is room in FIFO),
> > + * otherwise discard it.
> > + */
> > + spin_lock(&priv->pw_fifo_lock);
> > + if (kfifo_avail(&priv->pw_fifo) >=
TSI721_RIO_PW_MSG_SIZE)
> > + kfifo_in(&priv->pw_fifo, pw_buf,
> > + TSI721_RIO_PW_MSG_SIZE);
> > + else
> > + priv->pw_discard_count++;
> > + spin_unlock(&priv->pw_fifo_lock);
> > + }
> > +
> > + /* Clear pending PW interrupts */
> > + iowrite32(TSI721_RIO_PW_RX_STAT_PW_DISC |
TSI721_RIO_PW_RX_STAT_PW_VAL,
> > + priv->regs + TSI721_RIO_PW_RX_STAT);
> > +
> > + schedule_work(&priv->pw_work);
>
> I see scheduled work being done, but no flush_scheduled_work() or
> similar. This is often a bug, leading to code being executed after
> device shutdown or even after rmmod.
Possibly my oversight. I will check if/where flush is needed.
> > + return 0;
> > +}
> > +
> > +static void tsi721_pw_dpc(struct work_struct *work)
> > +{
> > + struct tsi721_device *priv = container_of(work, struct
tsi721_device,
> > + pw_work);
> > + unsigned long flags;
> > + u32 msg_buffer[RIO_PW_MSG_SIZE/sizeof(u32)]; /* Use full size PW
message
> > + buffer for RIO
layer */
> > +
> > + /*
> > + * Process port-write messages
> > + */
> > + spin_lock_irqsave(&priv->pw_fifo_lock, flags);
> > + while (kfifo_out(&priv->pw_fifo, (unsigned char *)msg_buffer,
> > + TSI721_RIO_PW_MSG_SIZE)) {
> > + /* Process one message */
> > + spin_unlock_irqrestore(&priv->pw_fifo_lock, flags);
> > +#ifdef DEBUG_PW
> > + {
> > + u32 i;
> > + pr_debug("%s : Port-Write Message:", __func__);
> > + for (i = 0; i < RIO_PW_MSG_SIZE/sizeof(u32); ) {
> > + pr_debug("0x%02x: %08x %08x %08x %08x", i*4,
> > + msg_buffer[i], msg_buffer[i + 1],
> > + msg_buffer[i + 2], msg_buffer[i + 3]);
> > + i += 4;
> > + }
> > + pr_debug("\n");
> > + }
> > +#endif
> > + /* Pass the port-write message to RIO core for
processing */
> > + rio_inb_pwrite_handler((union rio_pw_msg *)msg_buffer);
> > + spin_lock_irqsave(&priv->pw_fifo_lock, flags);
> > + }
> > + spin_unlock_irqrestore(&priv->pw_fifo_lock, flags);
> > +}
>
> The lock handling in this function is pretty ugly-looking. Is it
> correct?
Ugly for sure. I am trying to protect message FIFO and not to lock it
during
execution of rio_inb_pwrite_handler() because it may take relatively
long time.
Inbound port-write controller does not have any HW queue to store PW
messages
and my goal is to get them from device ASAP and to place into the FIFO.
I will redo this function to make locking to look better.
> >
> > ...
> >
> > +static void tsi721_db_dpc(struct work_struct *work)
> > +{
> > + struct tsi721_device *priv = container_of(work, struct
tsi721_device,
> > + idb_work);
> > + struct rio_mport *mport;
> > + struct rio_dbell *dbell;
> > + int found = 0;
> > + u32 wr_ptr, rd_ptr;
> > + u64 *idb_entry;
> > + u32 regval;
> > + union {
> > + u64 msg;
> > + u8 bytes[8];
> > + } idb;
> > +
> > + /*
> > + * Process queued inbound doorbells
> > + */
> > + mport = priv->mport;
> > +
> > + wr_ptr = ioread32(priv->regs + TSI721_IDQ_WP(IDB_QUEUE));
> > + rd_ptr = ioread32(priv->regs + TSI721_IDQ_RP(IDB_QUEUE));
> > +
> > + while (wr_ptr != rd_ptr) {
> > + idb_entry = (u64 *)(priv->idb_base +
> > + (TSI721_IDB_ENTRY_SIZE *
rd_ptr));
> > + rd_ptr++;
> > + idb.msg = *idb_entry;
>
> Is this code correct on both little-endian and big-endian hardware?
Yes. It works on both. Tested on x86 and powerpc.
> > + *idb_entry = 0;
> > +
> > + /* Process one doorbell */
> > + list_for_each_entry(dbell, &mport->dbells, node) {
> > + if ((dbell->res->start <= DBELL_INF(idb.bytes))
&&
> > + (dbell->res->end >= DBELL_INF(idb.bytes))) {
> > ...
> >
> > +static void tsi721_interrupts_init(struct tsi721_device *priv)
> > +{
> > + u32 intr;
> > +
> > + /* Enable IDB interrupts */
> > + iowrite32(TSI721_SR_CHINT_ALL,
> > + priv->regs + TSI721_SR_CHINT(IDB_QUEUE));
> > + iowrite32(TSI721_SR_CHINT_IDBQRCV,
> > + priv->regs + TSI721_SR_CHINTE(IDB_QUEUE));
> > + iowrite32(TSI721_INT_SR2PC_CHAN(IDB_QUEUE),
> > + priv->regs + TSI721_DEV_CHAN_INTE);
> > +
> > + /* Enable SRIO MAC interrupts */
> > + iowrite32(TSI721_RIO_EM_DEV_INT_EN_INT,
> > + priv->regs + TSI721_RIO_EM_DEV_INT_EN);
> > +
> > + if (priv->flags & TSI721_USING_MSIX)
> > + intr = TSI721_DEV_INT_SRIO;
> > + else
> > + intr = TSI721_DEV_INT_SR2PC_CH | TSI721_DEV_INT_SRIO |
> > + TSI721_DEV_INT_SMSG_CH;
> > +
> > + iowrite32(intr, priv->regs + TSI721_DEV_INTE);
> > + (void)ioread32(priv->regs + TSI721_DEV_INTE);
>
> Why include all of these void casts, btw?
Old habit. Used to have compiler warnings in some situations.
Not applicable to this case though. Will clean-up.
> > +}
> > +
> > +/**
> > + * tsi721_request_msix - register interrupt service for MSI-X mode.
> > + * @mport: RapidIO master port structure
> > + *
> > + * Registers MSI-X interrupt service routines for interrupts that
are active
> > + * immediately after mport initialization. Messaging interrupt
service routines
> > + * should be registered during corresponding open requests.
> > + */
> > +static int tsi721_request_msix(struct rio_mport *mport)
> > +{
> > + struct tsi721_device *priv = mport->priv;
> > + int err = 0;
> > +
> > + err = request_irq(priv->msix[TSI721_VECT_IDB].vector,
> > + tsi721_sr2pc_ch_msix, 0,
> > + priv->msix[TSI721_VECT_IDB].irq_name, (void
*)mport);
> > + if (err)
> > + goto out;
> > +
> > + err = request_irq(priv->msix[TSI721_VECT_PWRX].vector,
> > + tsi721_srio_msix, 0,
> > + priv->msix[TSI721_VECT_PWRX].irq_name, (void
> *)mport);
>
> Did we leak the first IRQ here?
Yes we did. Will fix it here and in one more place.
>
> > +out:
> > + return err;
> > +}
> > +
> >
> > ...
> >
> > +static int tsi721_enable_msix(struct tsi721_device *priv)
> > +{
> > + struct msix_entry entries[TSI721_VECT_MAX];
> > + int err;
> > + int i;
> > +
> > + entries[TSI721_VECT_IDB].entry =
TSI721_MSIX_SR2PC_IDBQ_RCV(IDB_QUEUE);
> > + entries[TSI721_VECT_PWRX].entry = TSI721_MSIX_SRIO_MAC_INT;
> > +
> > + /*
> > + * Initialize MSI-X entries for Messaging Engine:
> > + * this driver supports four RIO mailboxes (inbound and
outbound)
> > + * NOTE: Inbound message MBOX 0...4 use IB channels 4...7.
Therefore
> > + * offset +4 is added to IB MBOX number.
> > + */
> > + for (i = 0; i < RIO_MAX_MBOX; i++) {
> > + entries[TSI721_VECT_IMB0_RCV + i].entry =
> > + TSI721_MSIX_IMSG_DQ_RCV(i + 4);
> > + entries[TSI721_VECT_IMB0_INT + i].entry =
> > + TSI721_MSIX_IMSG_INT(i + 4);
> > + entries[TSI721_VECT_OMB0_DONE + i].entry =
> > + TSI721_MSIX_OMSG_DONE(i);
> > + entries[TSI721_VECT_OMB0_INT + i].entry =
> > + TSI721_MSIX_OMSG_INT(i);
> > + }
> > +
> > + err = pci_enable_msix(priv->pdev, entries, ARRAY_SIZE(entries));
> > + if (err) {
> > + if (err > 0)
> > + dev_info(&priv->pdev->dev,
> > + "Only %d MSI-X vectors available, "
> > + "not using MSI-X\n", err);
> > + return err;
> > + }
> > +
> > + /*
> > + * Copy MSI-X vector information into tsi721 private structure
> > + */
> > + priv->msix[TSI721_VECT_IDB].vector =
entries[TSI721_VECT_IDB].vector;
> > + snprintf(priv->msix[TSI721_VECT_IDB].irq_name,
IRQ_DEVICE_NAME_MAX,
> > + DRV_NAME "-idb at pci:%s", pci_name(priv->pdev));
> > + priv->msix[TSI721_VECT_PWRX].vector =
entries[TSI721_VECT_PWRX].vector;
> > + snprintf(priv->msix[TSI721_VECT_PWRX].irq_name,
IRQ_DEVICE_NAME_MAX,
> > + DRV_NAME "-pwrx at pci:%s", pci_name(priv->pdev));
> > +
> > + for (i = 0; i < RIO_MAX_MBOX; i++) {
> > + priv->msix[TSI721_VECT_IMB0_RCV + i].vector =
> > + entries[TSI721_VECT_IMB0_RCV +
i].vector;
> > + snprintf(priv->msix[TSI721_VECT_IMB0_RCV + i].irq_name,
> > + IRQ_DEVICE_NAME_MAX, DRV_NAME "-imbr%d at pci:%s",
> > + i, pci_name(priv->pdev));
> > +
> > + priv->msix[TSI721_VECT_IMB0_INT + i].vector =
> > + entries[TSI721_VECT_IMB0_INT +
i].vector;
> > + snprintf(priv->msix[TSI721_VECT_IMB0_INT + i].irq_name,
> > + IRQ_DEVICE_NAME_MAX, DRV_NAME "-imbi%d at pci:%s",
> > + i, pci_name(priv->pdev));
> > +
> > + priv->msix[TSI721_VECT_OMB0_DONE + i].vector =
> > + entries[TSI721_VECT_OMB0_DONE +
i].vector;
> > + snprintf(priv->msix[TSI721_VECT_OMB0_DONE + i].irq_name,
> > + IRQ_DEVICE_NAME_MAX, DRV_NAME "-ombd%d at pci:%s",
> > + i, pci_name(priv->pdev));
> > +
> > + priv->msix[TSI721_VECT_OMB0_INT + i].vector =
> > + entries[TSI721_VECT_OMB0_INT +
i].vector;
> > + snprintf(priv->msix[TSI721_VECT_OMB0_INT + i].irq_name,
> > + IRQ_DEVICE_NAME_MAX, DRV_NAME "-ombi%d at pci:%s",
> > + i, pci_name(priv->pdev));
> > + }
>
> Did we need a dependency on CONFIG_PCI_MSI?
To reduce a driver footprint yes, but it will not help if
system implements regular MSI only. This may be better handled by using
device specific config option depending on CONFIG_PCI_MSI. This way I
will be able
to remove MSI-X code while keeping MSI function. Marking as TODO.
If code size is not a concern, pci_enable_msix() has dependency on
CONFIG_PCI_MSI
and will fail if MSI support is not enabled.
>
> > + return 0;
> > +}
> > +
> >
> > ...
> >
> > +static int tsi721_bdma_ch_init(struct tsi721_device *priv, int
chnum)
> > +{
> > + struct tsi721_dma_desc *bd_ptr;
> > + u64 *sts_ptr;
> > + dma_addr_t bd_phys, sts_phys;
> > + int sts_size;
> > + int bd_num = priv->bdma[chnum].bd_num;
> > +
> > + dev_dbg(&priv->pdev->dev, "Init Block DMA Engine, CH%d\n",
chnum);
> > +
> > + /*
> > + * Initialize DMA channel for maintenance requests
> > + */
> > +
> > + /* Allocate space for DMA descriptors */
> > + bd_ptr = dma_alloc_coherent(&priv->pdev->dev,
> > + bd_num * sizeof(struct
tsi721_dma_desc),
> > + &bd_phys, GFP_KERNEL);
> > + if (!bd_ptr)
> > + return -ENOMEM;
> > +
> > + priv->bdma[chnum].bd_phys = bd_phys;
> > + priv->bdma[chnum].bd_base = bd_ptr;
> > +
> > + memset(bd_ptr, 0, bd_num * sizeof(struct tsi721_dma_desc));
>
> There it is again. Perhaps we need a dma_zalloc_coherent().
It will be nice match to kzalloc().
> > + dev_dbg(&priv->pdev->dev, "DMA descriptors @ %p (phys =
%llx)\n",
> > + bd_ptr, (unsigned long long)bd_phys);
> > +
> > + /* Allocate space for descriptor status FIFO */
> > + sts_size = (bd_num >= TSI721_DMA_MINSTSSZ) ?
> > + bd_num : TSI721_DMA_MINSTSSZ;
> > + sts_size = roundup_pow_of_two(sts_size);
> > + sts_ptr = dma_alloc_coherent(&priv->pdev->dev,
> > + sts_size * sizeof(struct
tsi721_dma_sts),
> > + &sts_phys, GFP_KERNEL);
> > + if (!sts_ptr) {
> > + /* Free space allocated for DMA descriptors */
> > + dma_free_coherent(&priv->pdev->dev,
> > + bd_num * sizeof(struct
tsi721_dma_desc),
> > + bd_ptr, bd_phys);
> > + priv->bdma[chnum].bd_base = NULL;
> > + return -ENOMEM;
> > + }
> > +
> > + priv->bdma[chnum].sts_phys = sts_phys;
> > + priv->bdma[chnum].sts_base = sts_ptr;
> > + priv->bdma[chnum].sts_size = sts_size;
> > +
> > + memset(sts_ptr, 0, sts_size);
>
> and again.
This memset() may be safely removed. I will review all similar cases.
> >
> > ...
> >
> > +struct tsi721_dma_desc {
> > + __le32 type_id;
> > +
...
> > +
> > + union {
> > + struct { /* if DTYPE == 1 */
> > + __le32 bufptr_lo;
> > + __le32 bufptr_hi;
> > + __le32 s_dist;
> > + __le32 s_size;
> > + } t1;
> > + __le32 data[4]; /* if DTYPE == 2 */
> > + u32 reserved[4]; /* if DTYPE == 3 */
> > + };
> > +} __attribute__((aligned(32)));
>
> We have the __aligned helper for this. (more below)
No excuse to ignore it. Will update.
> >
> > ...
> >
>
> _______________________________________________
> Linuxppc-dev mailing list
> Linuxppc-dev at lists.ozlabs.org
> https://lists.ozlabs.org/listinfo/linuxppc-dev
More information about the Linuxppc-dev
mailing list