[Skiboot] [PATCH V3 3/5] capi: Enable capi mode

Alistair Popple alistair at popple.id.au
Wed May 10 15:40:14 AEST 2017


On Tue, 9 May 2017 01:17:22 PM Frederic Barrat wrote:
>
> Le 04/05/2017 à 17:57, Christophe Lombard a écrit :
> > Enable the Coherently attached processor interface. The PHB is used as
> > a CAPI interface.
> > CAPI Adapters can be connected to whether PEC0 or PEC2. Single port
>
>                                    ^ either?
>
> Also title of the commit should say it's for phb4.
>
>
> > CAPI adapter can be connected to either PEC0 or PEC2, but Dual-Port
> > Adapter can be only connected to PEC2
> >    CAPP0 attached to PHB0(PEC0 - single port)
> >    CAPP1 attached to PHB3(PEC2 - single or dual port)
> > As we did for PHB3, a new specific file 'phb4-capp.h' is created to
> > contain the CAPP register definitions.
> >
> > Signed-off-by: Christophe Lombard <clombard at linux.vnet.ibm.com>
> > ---
> >  core/cpu.c          |  17 +-
> >  hw/phb4.c           | 511 +++++++++++++++++++++++++++++++++++++++++++++++++++-
> >  include/chip.h      |   1 +
> >  include/opal-api.h  |   1 +
> >  include/phb4-capp.h |  60 ++++++
> >  include/phb4-regs.h |   7 +-
> >  include/phb4.h      |   1 +
> >  7 files changed, 591 insertions(+), 7 deletions(-)
> >  create mode 100644 include/phb4-capp.h
> >
> > diff --git a/core/cpu.c b/core/cpu.c
> > index c7e650d..a096d2f 100644
> > --- a/core/cpu.c
> > +++ b/core/cpu.c
> > @@ -1099,12 +1099,27 @@ static int64_t opal_reinit_cpus(uint64_t flags)
> >  opal_call(OPAL_REINIT_CPUS, opal_reinit_cpus, 1);
> >
> >  #define NMMU_XLAT_CTL_PTCR 0xb
> > +#define NMMU_XLAT_CTL_SMCR 0x12
> >  static int64_t nmmu_set_ptcr(uint64_t chip_id, struct dt_node *node, uint64_t ptcr)
> >  {
> >  	uint32_t nmmu_base_addr;
> > +	struct proc_chip *chip;
> > +	uint64_t reg;
> > +	int64_t rc;
> >
> >  	nmmu_base_addr = dt_get_address(node, 0, NULL);
> > -	return xscom_write(chip_id, nmmu_base_addr + NMMU_XLAT_CTL_PTCR, ptcr);
> > +	rc = xscom_write(chip_id, nmmu_base_addr + NMMU_XLAT_CTL_PTCR, ptcr);
> > +	if (!rc) {
> > +		/* Disable for Radix page-walk cache */
> > +		chip = get_chip(chip_id);
> > +		if ((chip->type == PROC_CHIP_P9_NIMBUS) ||
> > +		    (chip->type == PROC_CHIP_P9_CUMULUS)) {
> > +			xscom_read(chip_id, nmmu_base_addr + NMMU_XLAT_CTL_SMCR, &reg);
> > +			reg |= PPC_BIT(52);

Some of the lab folk have mentioned bit 53 although I'm not sure how
it's related to this. Need to follow it up.

> > +			xscom_write(chip_id, nmmu_base_addr + NMMU_XLAT_CTL_SMCR, reg);
> > +		}
> > +	}
> > +	return rc;
> >  }
>
> [ + Alistair for awareness ]
>
> It should be a separate patch on its own, to make it more visible.
> We had discussed it, but for the record: we need to limit the workaround
> to DD1.x and DD2.0, as it will be fixed with DD2.1.
> Stewart can comment, but it looks like it can be tested with
> chip->ec_level. From chip.h, it's encoded like:
> 	/* 0xMm (DD1.0 = 0x10) */
> Also, I think we should drop the test on P9_CUMULUS, since we don't know
> how things will look like there.

Rather than detecting chip ec_levels I think we should read a property
from the device-tree that gets added when creating the nmmu node on
the affected parts. I guess it doesn't make a huge difference but it
does mean the host OS can easily tell if it's been disabled or not.

Thanks.

- Alistair

>
>
>
> >  /*
> > diff --git a/hw/phb4.c b/hw/phb4.c
> > index 0b0c3d8..921e211 100644
> > --- a/hw/phb4.c
> > +++ b/hw/phb4.c
> > @@ -48,7 +48,7 @@
> >  #include <affinity.h>
> >  #include <phb4.h>
> >  #include <phb4-regs.h>
> > -#include <capp.h>
> > +#include <phb4-capp.h>
> >  #include <fsp.h>
> >  #include <chip.h>
> >  #include <chiptod.h>
> > @@ -76,6 +76,11 @@ static void phb4_init_hw(struct phb4 *p, bool first_init);
> >  #define PHBLOGCFG(p, fmt, a...) do {} while (0)
> >  #endif
> >
> > +enum capi_dma_tvt {
> > +	CAPI_DMA_TVT0,
> > +	CAPI_DMA_TVT1,
> > +};
> > +
> >  /* Note: The "ASB" name is historical, practically this means access via
> >   * the XSCOM backdoor
> >   */
> > @@ -2101,6 +2106,119 @@ static int64_t phb4_freset(struct pci_slot *slot)
> >  	return OPAL_HARDWARE;
> >  }
> >
> > +extern struct lock capi_lock;
> > +
> > +static void do_capp_recovery_scoms(struct phb4 *p)
> > +{
> > +	uint64_t reg;
> > +	uint32_t offset;
> > +
> > +	PHBDBG(p, "Doing CAPP recovery scoms\n");
> > +
> > +	offset = PHB4_CAPP_REG_OFFSET(p);
> > +	/* disable snoops */
> > +	xscom_write(p->chip_id, SNOOP_CAPI_CONFIG + offset, 0);
> > +
> > +	/* clear err rpt reg*/
> > +	xscom_write(p->chip_id, CAPP_ERR_RPT_CLR + offset, 0);
> > +	/* clear capp fir */
> > +	xscom_write(p->chip_id, CAPP_FIR + offset, 0);
> > +
> > +	xscom_read(p->chip_id, CAPP_ERR_STATUS_CTRL + offset, &reg);
> > +	reg &= ~(PPC_BIT(0) | PPC_BIT(1));
> > +	xscom_write(p->chip_id, CAPP_ERR_STATUS_CTRL + offset, reg);
> > +}
> > +
> > +/*
> > + * Disable CAPI mode on a PHB.
> > + *
> > + * Must be done while PHB is fenced and in recovery. Leaves CAPP in recovery -
> > + * we can't come out of recovery until the PHB has been reinitialised.
> > + *
> > + * We don't reset generic error registers here - we rely on phb4_init_hw() to
> > + * do that.
> > + *
> > + * Sets PHB4_CAPP_DISABLING flag when complete.
> > + */
> > +static void disable_capi_mode(struct phb4 *p)
> > +{
> > +	struct proc_chip *chip = get_chip(p->chip_id);
> > +	uint64_t reg;
> > +	uint32_t offset = PHB4_CAPP_REG_OFFSET(p);
> > +
> > +	lock(&capi_lock);
> > +
> > +	xscom_read(p->chip_id, p->pe_xscom + 0x7, &reg);
> > +	if (!(reg & PPC_BIT(0))) {
> > +		/* Not in CAPI mode, no action required */
> > +		goto out;
> > +	}
> > +
> > +	PHBDBG(p, "CAPP: Disabling CAPI mode\n");
> > +	if (!(chip->capp_phb4_attached_mask & (1 << p->index)))
> > +		PHBERR(p, "CAPP: CAPP attached mask not set!\n");
> > +
> > +	xscom_read(p->chip_id, CAPP_ERR_STATUS_CTRL + offset, &reg);
> > +	if (!(reg & PPC_BIT(0))) {
> > +		PHBERR(p, "CAPP: not in recovery, can't disable CAPI mode!\n");
> > +		goto out;
> > +	}
> > +
> > +	/* CAPI will not be decoded. */
> > +	out_be64(p->regs + PHB_CAPI_CMPM, 0ull);
> > +
> > +	/* Disable NBW decoding */
> > +	out_be64(p->regs + PHB_PBL_NBW_CMP_MASK, 0ull);
> > +
> > +	/* Disable ASN */
> > +	out_be64(p->regs + PHB_PBL_ASN_CMPM, 0ull);
> > +
> > +	/* APC Master PB Control Register - disable examining cResps */
> > +	xscom_read(p->chip_id, APC_MASTER_PB_CTRL + offset, &reg);
> > +	reg &= ~PPC_BIT(0);
> > +	xscom_write(p->chip_id, APC_MASTER_PB_CTRL + offset, reg);
> > +
> > +	/* APC Master Config Register - de-select PHBs */
> > +	xscom_read(p->chip_id, APC_MASTER_CAPI_CTRL + offset, &reg);
> > +	reg &= ~PPC_BITMASK(2, 3);
> > +	xscom_write(p->chip_id, APC_MASTER_CAPI_CTRL + offset, reg);
> > +
> > +	/* Snoop CAPI Configuration Register - disable snooping */
> > +	xscom_write(p->chip_id, SNOOP_CAPI_CONFIG + offset, 0ull);
> > +
> > +	/* Read Epsilon Tier0 _ default value */
> > +	xscom_write(p->chip_id, SNOOP_CONTROL + offset, 0ull);
> > +
> > +	/* CAPP Transport Control Register */
> > +	xscom_write(p->chip_id, TRANSPORT_CONTROL + offset, 0x0001000000000000);
> > +
> > +	/* Epoch and Recovery Timers Control Register */
> > +	xscom_write(p->chip_id, EPOCH_RECOVERY_TIMERS_CTRL + offset, 0ull);
> > +
> > +	/* Flush SUE State Map Register */
> > +	xscom_write(p->chip_id, FLUSH_SUE_STATE_MAP + offset, 0ull);
> > +
> > +	/* Flush SUE uOP1 Register */
> > +	xscom_write(p->chip_id, FLUSH_SUE_UOP1 + offset, 0ull);
> > +
> > +	/* APC FSM Read Machine Mask Register */
> > +	xscom_write(p->chip_id, APC_FSM_READ_MASK + offset, 0ull);
> > +
> > +	/* Snoop copy of FSM Read Machine Mask Register */
> > +	xscom_write(p->chip_id, XPT_FSM_RMM + offset, 0ull);
> > +
> > +	/* PE Secure CAPP Enable Register - we're all done! Disable CAPP mode! */
> > +	xscom_write(p->chip_id, p->pe_xscom + 0x7, 0ull);
> > +
> > +	/* Trigger CAPP recovery scoms after reinit */
> > +	p->flags |= PHB4_CAPP_DISABLING;
> > +
> > +	chip->capp_phb4_attached_mask &= ~(1 << p->index);
> > +
> > +out:
> > +	unlock(&capi_lock);
> > +}
> > +
> >  static int64_t phb4_creset(struct pci_slot *slot)
> >  {
> >  	struct phb4 *p = phb_to_phb4(slot->phb);
> > @@ -2110,12 +2228,11 @@ static int64_t phb4_creset(struct pci_slot *slot)
> >  	case PHB4_SLOT_CRESET_START:
> >  		PHBDBG(p, "CRESET: Starts\n");
> >
> > -		/* do steps 3-5 of capp recovery procedure */
> > -#if 0
> >  		if (p->flags & PHB4_CAPP_RECOVERY)
> >  			do_capp_recovery_scoms(p);
> > -#endif
> > -		/* XXX TODO XXX */
> > +
> > +		if (!(p->flags & PHB4_CAPP_RECOVERY))
> > +			disable_capi_mode(p);
>
> Are we ready to support this? I guess that's a question to be sorted out
> with the PSL team.
> disable_capi_mode() is mostly useful to support kexec(), which is
> desirable, but I'm wondering if it's not a bit early.
>
>
> >
> >  		pci_slot_set_state(slot, PHB4_SLOT_CRESET_WAIT_CQ);
> >  		slot->retries = 500;
> > @@ -2128,6 +2245,12 @@ static int64_t phb4_creset(struct pci_slot *slot)
> >  		p->flags &= ~PHB4_AIB_FENCED;
> >  		p->flags &= ~PHB4_CAPP_RECOVERY;
> >  		phb4_init_hw(p, false);
> > +
> > +		if (p->flags & PHB4_CAPP_DISABLING) {
> > +			do_capp_recovery_scoms(p);
> > +			p->flags &= ~PHB4_CAPP_DISABLING;
> > +		}
> > +
> >  		pci_slot_set_state(slot, PHB4_SLOT_CRESET_FRESET);
> >  		return pci_slot_set_sm_timeout(slot, msecs_to_tb(100));
> >  	case PHB4_SLOT_CRESET_FRESET:
> > @@ -2499,6 +2622,338 @@ static int64_t phb4_get_diag_data(struct phb *phb,
> >  	return OPAL_SUCCESS;
> >  }
> >
> > +static uint64_t tve_encode_50b_noxlate(uint64_t start_addr, uint64_t end_addr)
> > +{
> > +	uint64_t tve;
> > +
> > +	/*
> > +	 * Put start address bits 49:24 into TVE[52:53]||[0:23]
> > +	 * and end address bits 49:24 into TVE[54:55]||[24:47]
> > +	 * and set TVE[51]
> > +	 */
> > +	tve  = (start_addr << 16) & (0xffffffull << 40);
> > +	tve |= (start_addr >> 38) & (3ull << 10);
> > +	tve |= (end_addr >>  8) & (0xfffffful << 16);
> > +	tve |= (end_addr >> 40) & (3ull << 8);
> > +	tve |= PPC_BIT(51) | IODA3_TVT_NON_TRANSLATE_50;
> > +	return tve;
> > +}
> > +
> > +static void phb4_init_capp_regs(struct phb4 *p)
> > +{
> > +	uint64_t reg;
> > +	uint32_t offset;
> > +
> > +	offset = PHB4_CAPP_REG_OFFSET(p);
> > +
> > +	/* Enable cresp examination by CAPP */
> > +	xscom_read(p->chip_id, APC_MASTER_PB_CTRL + offset, &reg);
> > +	reg |= PPC_BIT(0);
> > +	if (p->rev == PHB4_REV_NIMBUS_DD10) {
> > +		reg |= PPC_BIT(1);
> > +		/* disable vg not sys */
> > +		reg |= PPC_BIT(3);
> > +	}
> > +	xscom_write(p->chip_id, APC_MASTER_PB_CTRL + offset, reg);
> > +
> > +	/* Set PHB mode, HPC Dir State and P9 mode */
> > +	xscom_write(p->chip_id, APC_MASTER_CAPI_CTRL + offset, 0x1772000000000000);
> > +	PHBINF(p, "CAPP: port attached\n");
> > +
> > +	/* Set snoop ttype decoding , dir size to 256k */
> > +	xscom_write(p->chip_id, SNOOP_CAPI_CONFIG + offset, 0xA000000000000000);
> > +
> > +	/* Use Read Epsilon Tier2 for all scopes, Address Pipeline Master
> > +	 * Wait Count to highest(1023) and Number of rpt_hang.data to 3
> > +	 */
> > +	xscom_write(p->chip_id, SNOOP_CONTROL + offset, 0x8000000010072000);
> > +
> > +	/* TLBI Hang Divider = 1.  LPC buffers=0. X16 PCIe(14 buffers) */
> > +	xscom_write(p->chip_id, TRANSPORT_CONTROL + offset, 0x081400000000000A);
>
>
> Mellanox had requested to set the max for the TLBI hang divider, as
> their CX5 could have a pretty long path so we shouldn't timeout too
> fast. The capp team semmed ok with it, it is just more annoying for
> debug (more traces in the buffers when we detect the timeout).
>
>
> > +
> > +	/* Enable epoch timer */
> > +	xscom_write(p->chip_id, EPOCH_RECOVERY_TIMERS_CTRL + offset, 0xC0000000FFF0FFFE);
> > +
> > +	xscom_write(p->chip_id, FLUSH_SUE_STATE_MAP + offset, 0x1DCF5F6600000000);
> > +	xscom_write(p->chip_id, FLUSH_SUE_UOP1 + offset, 0xE310280428000000);
> > +
> > +	/* capp owns PHB read buffers */
> > +	xscom_write(p->chip_id, APC_FSM_READ_MASK + offset, 0xFFFFFFFFFFFF0000);
> > +	xscom_write(p->chip_id, XPT_FSM_RMM + offset, 0xFFFFFFFFFFFF0000);
> > +
> > +	/* Initialize CI Store Buffers */
> > +	xscom_read(p->chip_id, TRANSPORT_CONTROL + offset, &reg);
> > +	reg |= PPC_BIT(63);
> > +	xscom_write(p->chip_id, TRANSPORT_CONTROL + offset, reg);
> > +
> > +	/* Deassert TLBI_FENCED and tlbi_psl_is_dead */
> > +	xscom_write(p->chip_id, CAPP_ERR_STATUS_CTRL + offset, 0);
> > +}
> > +
> > +/* override some inits with CAPI defaults */
> > +static void phb4_init_capp_errors(struct phb4 *p)
> > +{
> > +	/* Init_77: TXE Error AIB Fence Enable Register */
> > +	out_be64(p->regs + 0x0d30,	0xdff7ff0bf7ddfff0ull);
> > +
> > +	/* Init_86: RXE_ARB Error AIB Fence Enable Register */
> > +	out_be64(p->regs + 0x0db0,	0xfbffd7bbff7fbfefull);
> > +
> > +	/* Init_95: RXE_MRG Error AIB Fence Enable Register */
> > +	out_be64(p->regs + 0x0e30,	0xfffffeffff7fff57ull);
> > +
> > +	/* Init_104: RXE_TCE Error AIB Fence Enable Register */
> > +	out_be64(p->regs + 0x0eb0,	0xffaeffafffffffffull);
> > +
> > +	/* Init_113: PHB Error AIB Fence Enable Register */
> > +	out_be64(p->regs + 0x0cb0,	0x35777073ff000000ull);
> > +}
> > +
> > +/* Power Bus Common Queue Registers
> > + * All PBCQ and PBAIB registers are accessed via SCOM
> > + * NestBase = 4010C00 for PEC0
> > + *            4011000 for PEC1
> > + *            4011400 for PEC2
> > + *
> > + * Some registers are shared amongst all of the stacks and will only
> > + * have 1 copy. Other registers are implemented one per stack.
> > + * Registers that are duplicated will have an additional offset
> > + * of “StackBase” so that they have a unique address.
> > + * Stackoffset = 00000040 for Stack0
> > + *             = 00000080 for Stack1
> > + *             = 000000C0 for Stack2
> > + */
> > +static int64_t enable_capi_mode(struct phb4 *p, uint64_t pe_number,
> > +				enum capi_dma_tvt dma_tvt)
> > +{
> > +	uint64_t reg, start_addr, end_addr;
> > +	uint32_t offset;
> > +	int i;
> > +
> > +	xscom_read(p->chip_id, p->pe_xscom + 0x7, &reg);
> > +	if (reg & PPC_BIT(0))
> > +		PHBDBG(p, "Already in CAPP mode\n");
> > +
> > +	/* PEC Phase 3 (PBCQ) registers Init */
> > +	/* poll cqstat
> > +	 * CAPP0 attached to PHB0(PEC0)
> > +	 * CAPP1 attached to PHB3(PEC2)
> > +	 */
> > +	offset = 0x40;
> > +	if (p->index > 0 && p->index < 3)
> > +		offset = 0x80;
> > +	else if (p->index > 2)
> > +		offset = 0xC0;
>
>
> Where do we detect and fail in case of an invalid phb index? I think it
> wouldn't pass the capp ucode load stage, but it seems like we should do
> it here as well. It doesn't make sense to enable capi on a phb index !=
> 0 and 3.
>
>
> > +
> > +	for (i = 0; i < 500000; i++) {
> > +		xscom_read(p->chip_id, p->pe_xscom + offset + 0xC, &reg);
> > +		if (!(reg & 0xC000000000000000))
> > +			break;
> > +		time_wait_us(10);
> > +	}
> > +	if (reg & 0xC000000000000000) {
> > +		PHBERR(p, "CAPP: Timeout waiting for pending transaction\n");
> > +		return OPAL_HARDWARE;
> > +	}
> > +
> > +	/* Enable CAPP Mode , Set 14 CI Store buffers for CAPP,
> > +	 * Set 48 Read machines for CAPP.
> > +	 */
> > +	reg = 0x800EFFFFFFFFFFFF;
> > +	xscom_write(p->chip_id, p->pe_xscom + 0x7, reg);
> > +
> > +	if (p->rev == PHB4_REV_NIMBUS_DD10) {
> > +		/* Ignores the PB init signal */
> > +		xscom_read(p->chip_id, p->pe_xscom + 0x0, &reg);
> > +		reg |= PPC_BIT(12);
> > +		xscom_write(p->chip_id, p->pe_xscom + 0x0, reg);
> > +	}
> > +
> > +	/* PEC Phase 4 (PHB) registers adjustment
> > +	 * Bit [0:7] XSL_DSNCTL[capiind]
> > +	 * Init_25 - CAPI Compare/Mask
> > +	 */
> > +	out_be64(p->regs + PHB_CAPI_CMPM,
> > +		 0x0200FE0000000000Ull | PHB_CAPI_CMPM_ENABLE);
> > +
> > +	if (!(p->rev == PHB4_REV_NIMBUS_DD10)) {
> > +		/* Init_123 :  NBW Compare/Mask Register */
> > +		out_be64(p->regs + PHB_PBL_NBW_CMP_MASK,
> > +			 0x0300FF0000000000Ull | PHB_CAPI_NBW_MASK_ENABLE);
> > +
> > +		/* Init_24 - ASN Compare/Mask */
> > +		out_be64(p->regs + PHB_PBL_ASN_CMPM,
> > +			 0x0400FF0000000000Ull | PHB_CAPI_ASN_ENABLE);
> > +	}
> > +
> > +	/* non-translate/50-bit mode */
> > +	out_be64(p->regs + PHB_XLATE_PREFIX, 0x0000000000000000Ull);
> > +
> > +	/* set tve no translate mode allow mmio window */
> > +	memset(p->tve_cache, 0x0, sizeof(p->tve_cache));
> > +
> > +	/*
> > +	 * In 50-bit non-translate mode, the fields of the TVE are
> > +	 * used to perform an address range check. In this mode TCE
> > +	 * Table Size(0) must be a '1' (TVE[51] = 1)
> > +	 *      PCI Addr(49:24) >= TVE[52:53]+TVE[0:23] and
> > +	 *      PCI Addr(49:24) < TVE[54:55]+TVE[24:47]
> > +	 *
> > +	 * TVE[51] = 1
> > +	 * TVE[56] = 1: 50-bit Non-Translate Mode Enable
> > +	 * TVE[0:23] = 0x000000
> > +	 * TVE[24:47] = 0xFFFFFF
> > +	 *
> > +	 * capi dma mode: CAPP DMA mode needs access to all of memory
> > +	 * capi mode: Allow address range (bit 14 = 1)
> > +	 *            0x0002000000000000: 0x0002FFFFFFFFFFFF
> > +	 *            TVE[52:53] = '10' and TVE[54:55] = '10'
> > +	 *
> > +	 * --> we use capi dma mode by default
> > +	 */
> > +	if (dma_tvt == CAPI_DMA_TVT0) {
> > +		/*
> > +		 * TVT#0: CAPI window + DMA, all memory
> > +		 */
> > +		start_addr = 0ull;
> > +		end_addr   = 0x0003ffffffffffffull;
> > +		p->tve_cache[pe_number * 2] =
> > +			tve_encode_50b_noxlate(start_addr, end_addr);
> > +	} else {
> > +		/*
> > +		 * TVT#0: CAPI window, in bypass mode
> > +		 * TVT#1: DMA, all memory, in bypass mode
> > +		 */
> > +		start_addr = 0x0002000000000000ull;
> > +		end_addr   = 0x0002FFFFFFFFFFFFull;
> > +		p->tve_cache[pe_number * 2] =
> > +			tve_encode_50b_noxlate(start_addr, end_addr);
> > +
> > +		start_addr = (1ull << 59);
> > +		end_addr   = start_addr + 0x0003ffffffffffffull;
> > +		p->tve_cache[pe_number * 2 + 1] =
> > +			tve_encode_50b_noxlate(start_addr, end_addr);
> > +	}
>
> This is my mistake: for Mellanox (CAPI_DMA_TVT1), we should define TVT#0
> like for "normal" capi, i.e. same as CAPI_DMA_TVT0. I'm pretty sure it
> will create problems with the dummy read buffer otherwise (and
> potentially others).
>
>
> > +
> > +	phb4_ioda_sel(p, IODA3_TBL_TVT, 0, true);
> > +	for (i = 0; i < p->tvt_size; i++)
> > +		out_be64(p->regs + PHB_IODA_DATA0, p->tve_cache[i]);
> > +
> > +	/* set mbt bar to pass capi mmio window. First applied cleared
> > +	 * values to HW
> > +	 */
> > +	for (i = 0; i < p->mbt_size; i++) {
> > +		p->mbt_cache[i][0] = 0;
> > +		p->mbt_cache[i][1] = 0;
> > +	}
> > +	phb4_ioda_sel(p, IODA3_TBL_MBT, 0, true);
> > +	for (i = 0; i < p->mbt_size; i++) {
> > +		out_be64(p->regs + PHB_IODA_DATA0, p->mbt_cache[i][0]);
> > +		out_be64(p->regs + PHB_IODA_DATA0, p->mbt_cache[i][1]);
> > +	}
> > +
> > +	p->mbt_cache[0][0] = IODA3_MBT0_ENABLE |
> > +			     IODA3_MBT0_TYPE_M64 |
> > +		SETFIELD(IODA3_MBT0_MODE, 0ull, IODA3_MBT0_MODE_SINGLE_PE) |
> > +		SETFIELD(IODA3_MBT0_MDT_COLUMN, 0ull, 0) |
> > +		(p->mm0_base & IODA3_MBT0_BASE_ADDR);
> > +	p->mbt_cache[0][1] = IODA3_MBT1_ENABLE |
> > +		((~(p->mm0_size - 1)) & IODA3_MBT1_MASK) |
> > +		SETFIELD(IODA3_MBT1_SINGLE_PE_NUM, 0ull, pe_number);
> > +
> > +	p->mbt_cache[1][0] = IODA3_MBT0_ENABLE |
> > +			     IODA3_MBT0_TYPE_M64 |
> > +		SETFIELD(IODA3_MBT0_MODE, 0ull, IODA3_MBT0_MODE_SINGLE_PE) |
> > +		SETFIELD(IODA3_MBT0_MDT_COLUMN, 0ull, 0) |
> > +		(0x0002000000000000ULL & IODA3_MBT0_BASE_ADDR);
> > +	p->mbt_cache[1][1] = IODA3_MBT1_ENABLE |
> > +		(0x00ff000000000000ULL & IODA3_MBT1_MASK) |
> > +		SETFIELD(IODA3_MBT1_SINGLE_PE_NUM, 0ull, pe_number);
> > +
> > +	phb4_ioda_sel(p, IODA3_TBL_MBT, 0, true);
> > +	for (i = 0; i < p->mbt_size; i++) {
> > +		out_be64(p->regs + PHB_IODA_DATA0, p->mbt_cache[i][0]);
> > +		out_be64(p->regs + PHB_IODA_DATA0, p->mbt_cache[i][1]);
> > +	}
> > +
> > +	phb4_init_capp_errors(p);
> > +
> > +	phb4_init_capp_regs(p);
> > +
> > +	if (!chiptod_capp_timebase_sync(p->chip_id, CAPP_TFMR,
> > +					CAPP_TB,
> > +					PHB4_CAPP_REG_OFFSET(p))) {
> > +		PHBERR(p, "CAPP: Failed to sync timebase\n");
> > +		return OPAL_HARDWARE;
> > +	}
> > +	return OPAL_SUCCESS;
> > +}
> > +
> > +static int64_t phb4_set_capi_mode(struct phb *phb, uint64_t mode,
> > +				  uint64_t pe_number)
> > +{
> > +	struct phb4 *p = phb_to_phb4(phb);
> > +	struct proc_chip *chip = get_chip(p->chip_id);
> > +	uint64_t reg;
> > +	uint32_t offset;
> > +
> > +
> > +	lock(&capi_lock);
> > +	chip->capp_phb4_attached_mask |= 1 << p->index;
> > +	unlock(&capi_lock);
> > +
> > +	offset = PHB4_CAPP_REG_OFFSET(p);
> > +	xscom_read(p->chip_id, CAPP_ERR_STATUS_CTRL + offset, &reg);
> > +	if ((reg & PPC_BIT(5))) {
> > +		PHBERR(p, "CAPP: recovery failed (%016llx)\n", reg);
> > +		return OPAL_HARDWARE;
> > +	} else if ((reg & PPC_BIT(0)) && (!(reg & PPC_BIT(1)))) {
> > +		PHBDBG(p, "CAPP: recovery in progress\n");
> > +		return OPAL_BUSY;
> > +	}
> > +
> > +	switch (mode) {
> > +	case OPAL_PHB_CAPI_MODE_PCIE:
> > +		return OPAL_UNSUPPORTED;
> > +
> > +	case OPAL_PHB_CAPI_MODE_CAPI:
> > +		return enable_capi_mode(p, pe_number, CAPI_DMA_TVT0);
> > +
> > +	case OPAL_PHB_CAPI_MODE_DMA:
> > +		/* shouldn't be called, enabled by default on p9 */
> > +		return OPAL_UNSUPPORTED;
> > +
> > +	case OPAL_PHB_CAPI_MODE_DMA_TVT1:
> > +		return enable_capi_mode(p, pe_number, CAPI_DMA_TVT1);
> > +
> > +	case OPAL_PHB_CAPI_MODE_SNOOP_OFF:
> > +		/* nothing to do */
> > +		return OPAL_SUCCESS;
>
> I think we should fail here since we don't support turning snooping off.
> We can revisit if the capp tlbi filtering hw bug is ever fixed.
> An alternative would be to try to *really* deactivate snooping (and not
> rely on capp tlbi filtering), but that needs more investigation as the
> capp team had mentioned it's more complicated to do.
>
>
> > +	case OPAL_PHB_CAPI_MODE_SNOOP_ON:
> > +		/* nothing to do */
> > +		return OPAL_SUCCESS;
> > +	}
> > +
> > +	return OPAL_UNSUPPORTED;
> > +}
> > +
> > +static int64_t phb4_set_capp_recovery(struct phb *phb)
> > +{
> > +	struct phb4 *p = phb_to_phb4(phb);
> > +
> > +	if (p->flags & PHB4_CAPP_RECOVERY)
> > +		return 0;
> > +
> > +	/* set opal event flag to indicate eeh condition */
> > +	opal_update_pending_evt(OPAL_EVENT_PCI_ERROR,
> > +				OPAL_EVENT_PCI_ERROR);
> > +
> > +	p->flags |= PHB4_CAPP_RECOVERY;
> > +
> > +	return 0;
> > +}
> > +
> >  static const struct phb_ops phb4_ops = {
> >  	.cfg_read8		= phb4_pcicfg_read8,
> >  	.cfg_read16		= phb4_pcicfg_read16,
> > @@ -2531,6 +2986,8 @@ static const struct phb_ops phb4_ops = {
> >  	.get_diag_data		= NULL,
> >  	.get_diag_data2		= phb4_get_diag_data,
> >  	.tce_kill		= phb4_tce_kill,
> > +	.set_capi_mode		= phb4_set_capi_mode,
> > +	.set_capp_recovery	= phb4_set_capp_recovery,
> >  };
> >
> >  static void phb4_init_ioda3(struct phb4 *p)
> > @@ -3090,6 +3547,9 @@ static void phb4_add_properties(struct phb4 *p)
> >
> >  	dt_add_property_cells(np, "ibm,phb-diag-data-size",
> >  			      sizeof(struct OpalIoPhb4ErrorData));
> > +
> > +	/* Indicate to Linux that CAPP timebase sync is supported */
> > +	dt_add_property_string(np, "ibm,capp-timebase-sync", NULL);
> >  }
> >
> >  static bool phb4_calculate_windows(struct phb4 *p)
> > @@ -3215,6 +3675,44 @@ static const struct irq_source_ops phb4_lsi_ops = {
> >  	.attributes = phb4_lsi_attributes,
> >  };
> >
> > +/*
> > + * Trigger a creset to disable CAPI mode on kernel shutdown.
> > + *
> > + * This helper is called repeatedly by the host sync notifier mechanism, which
> > + * relies on the kernel to regularly poll the OPAL_SYNC_HOST_REBOOT call as it
> > + * shuts down.
> > + *
> > + * This is a somewhat hacky abuse of the host sync notifier mechanism, but the
> > + * alternatives require a new API call which won't work for older kernels.
> > + */
> > +static bool phb4_host_sync_reset(void *data)
> > +{
> > +	struct phb4 *p = (struct phb4 *)data;
> > +	struct pci_slot *slot = p->phb.slot;
> > +	struct proc_chip *chip = get_chip(p->chip_id);
> > +	int64_t rc;
> > +
> > +	switch (slot->state) {
> > +	case PHB4_SLOT_NORMAL:
> > +		lock(&capi_lock);
> > +		rc = (chip->capp_phb4_attached_mask & (1 << p->index)) ?
> > +			OPAL_PHB_CAPI_MODE_CAPI :
> > +			OPAL_PHB_CAPI_MODE_PCIE;
> > +		unlock(&capi_lock);
> > +
> > +		if (rc == OPAL_PHB_CAPI_MODE_PCIE)
> > +			return true;
> > +
> > +		PHBINF(p, "PHB in CAPI mode, resetting\n");
> > +		p->flags &= ~PHB4_CAPP_RECOVERY;
> > +		phb4_creset(slot);
> > +		return false;
> > +	default:
> > +		rc = slot->ops.poll(slot);
> > +		return rc <= OPAL_SUCCESS;
> > +	}
> > +}
> > +
>
> Like for disable capi mode above, are we ready to support it at this
> stage? Again, I think it's only useful for kexec()
>
>    Fred
>
>
> >  static void phb4_create(struct dt_node *np)
> >  {
> >  	const struct dt_property *prop;
> > @@ -3397,6 +3895,9 @@ static void phb4_create(struct dt_node *np)
> >  				(p->rev == PHB4_REV_NIMBUS_DD10) ?
> >  				&phb4_dd1_lsi_ops : &phb4_lsi_ops);
> >
> > +	/* Trigger a creset to disable CAPI mode on kernel shutdown. */
> > +	opal_add_host_sync_notifier(phb4_host_sync_reset, p);
> > +
> >  	/* Platform additional setup */
> >  	if (platform.pci_setup_phb)
> >  		platform.pci_setup_phb(&p->phb, p->index);
> > diff --git a/include/chip.h b/include/chip.h
> > index b957e45..17f062b 100644
> > --- a/include/chip.h
> > +++ b/include/chip.h
> > @@ -191,6 +191,7 @@ struct proc_chip {
> >
> >  	/* Must hold capi_lock to change */
> >  	uint8_t			capp_phb3_attached_mask;
> > +	uint8_t			capp_phb4_attached_mask;
> >  	uint8_t			capp_ucode_loaded;
> >
> >  	/* Used by hw/centaur.c */
> > diff --git a/include/opal-api.h b/include/opal-api.h
> > index 80033c6..fef9c93 100644
> > --- a/include/opal-api.h
> > +++ b/include/opal-api.h
> > @@ -1124,6 +1124,7 @@ enum {
> >  	OPAL_PHB_CAPI_MODE_SNOOP_OFF    = 2,
> >  	OPAL_PHB_CAPI_MODE_SNOOP_ON	= 3,
> >  	OPAL_PHB_CAPI_MODE_DMA		= 4,
> > +	OPAL_PHB_CAPI_MODE_DMA_TVT1	= 5,
> >  };
> >
> >  /* CAPI feature flags (in device-tree) */
> > diff --git a/include/phb4-capp.h b/include/phb4-capp.h
> > new file mode 100644
> > index 0000000..418838d
> > --- /dev/null
> > +++ b/include/phb4-capp.h
> > @@ -0,0 +1,60 @@
> > +/* Copyright 2013-2017 IBM Corp.
> > + *
> > + * Licensed under the Apache License, Version 2.0 (the "License");
> > + * you may not use this file except in compliance with the License.
> > + * You may obtain a copy of the License at
> > + *
> > + *	http://www.apache.org/licenses/LICENSE-2.0
> > + *
> > + * Unless required by applicable law or agreed to in writing, software
> > + * distributed under the License is distributed on an "AS IS" BASIS,
> > + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
> > + * implied.
> > + * See the License for the specific language governing permissions and
> > + * limitations under the License.
> > + */
> > +
> > +#ifndef __PHB4_CAPP_H
> > +#define __PHB4_CAPP_H
> > +
> > +#define CAPP_SNP_ARRAY_WRITE_REG		0x2010841  /* Satellite 2 */
> > +#define CAPP_SNP_ARRAY_ADDR_REG 		0x2010828
> > +#define CAPP_APC_MASTER_ARRAY_ADDR_REG		0x201082A
> > +#define CAPP_APC_MASTER_ARRAY_WRITE_REG 	0x2010842  /* Satellite 2 */
> > +
> > +#define CAPP_FIR				0x2010800
> > +#define CAPP_ERR_RPT_CLR			0x2010813
> > +#define APC_MASTER_PB_CTRL			0x2010818
> > +#define APC_MASTER_CAPI_CTRL			0x2010819
> > +#define LCO_MASTER_TARGET			0x2010821
> > +#define EPOCH_RECOVERY_TIMERS_CTRL		0x201082C
> > +#define SNOOP_CAPI_CONFIG			0x201081A
> > +#define SNOOP_CONTROL				0x201081B
> > +#define TRANSPORT_CONTROL			0x201081C
> > +#define CAPP_TB 				0x2010826
> > +#define CAPP_TFMR				0x2010827
> > +#define CAPP_ERR_STATUS_CTRL			0x201080E
> > +#define FLUSH_SUE_STATE_MAP			0x201080F
> > +#define FLUSH_CPIG_STATE_MAP			0x2010820
> > +#define FLUSH_SUE_UOP1				0x2010843  /* Satellite 2 */
> > +#define APC_FSM_READ_MASK			0x2010823
> > +#define XPT_FSM_RMM				0x2010831
> > +
> > +/* CAPP0 attached to PHB0(PEC0 - single port)
> > + * CAPP1 attached to PHB3(PEC2 - single or dual port)
> > + */
> > +#define CAPP0_PHB_INDEX 0
> > +#define CAPP1_PHB_INDEX 3
> > +
> > +/* SCOM address Base (Ring = ‘0010’b)
> > + * CAPP Unit Satellite           SCOM address Base
> > + * CAPP 0    S1 (sat = ‘0000’b)  x02010800
> > + * CAPP 0    S2 (sat = ‘0001’b)  x02010840
> > + * CAPP 1    S1 (sat = ‘0000’b)  x04010800
> > + * CAPP 1    S2 (sat = ‘0001’b)  x04010840
> > + */
> > +#define CAPP1_REG_OFFSET 0x2000000
> > +
> > +#define PHB4_CAPP_REG_OFFSET(p) ((p)->index == 0 ? 0x0 : CAPP1_REG_OFFSET)
> > +
> > +#endif /* __PHB4_CAPP_H */
> > diff --git a/include/phb4-regs.h b/include/phb4-regs.h
> > index 5480949..5eeaf18 100644
> > --- a/include/phb4-regs.h
> > +++ b/include/phb4-regs.h
> > @@ -71,8 +71,12 @@
> >  #define PHB_PEST_BAR			0x1a8
> >  #define   PHB_PEST_BAR_ENABLE		PPC_BIT(0)
> >  #define   PHB_PEST_BASE_ADDRESS		PPC_BITMASK(8,51)
> > +#define PHB_PBL_ASN_CMPM		0x1C0
> > +#define   PHB_CAPI_ASN_ENABLE		PPC_BIT(63)
> > +#define PHB_CAPI_CMPM			0x1C8
> > +#define   PHB_CAPI_CMPM_ENABLE		PPC_BIT(63)
> >  #define PHB_M64_UPPER_BITS		0x1f0
> > -#define PHB_INTREP_TIMER		0x1f8
> > +#define PHB_XLATE_PREFIX		0x1f8
> >  #define PHB_DMARD_SYNC			0x200
> >  #define   PHB_DMARD_SYNC_START		PPC_BIT(0)
> >  #define   PHB_DMARD_SYNC_COMPLETE	PPC_BIT(1)
> > @@ -241,6 +245,7 @@
> >  #define PHB_PBL_TIMEOUT_CTRL			0x1810
> >  #define PHB_PBL_NPTAG_ENABLE			0x1820
> >  #define PHB_PBL_NBW_CMP_MASK			0x1830
> > +#define   PHB_CAPI_NBW_MASK_ENABLE		PPC_BIT(63)
> >  #define PHB_PBL_SYS_LINK_INIT			0x1838
> >  #define PHB_PBL_BUF_STATUS			0x1840
> >  #define PHB_PBL_ERR_STATUS			0x1900
> > diff --git a/include/phb4.h b/include/phb4.h
> > index bed6eef..651946c 100644
> > --- a/include/phb4.h
> > +++ b/include/phb4.h
> > @@ -222,6 +222,7 @@ struct phb4_err {
> >  #define PHB4_CFG_USE_ASB	0x00000002
> >  #define PHB4_CFG_BLOCKED	0x00000004
> >  #define PHB4_CAPP_RECOVERY	0x00000008
> > +#define PHB4_CAPP_DISABLING	0x00000010
> >
> >  struct phb4 {
> >  	unsigned int		index;	    /* 0..5 index inside p9 */
> >
>



More information about the Skiboot mailing list