freebsd-dev/sys/dev/ntb/ntb_hw/ntb_hw_intel.c

3167 lines
89 KiB
C
Raw Normal View History

/*-
* Copyright (c) 2016-2017 Alexander Motin <mav@FreeBSD.org>
* Copyright (C) 2013 Intel Corporation
* Copyright (C) 2015 EMC Corporation
* All rights reserved.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*/
/*
* The Non-Transparent Bridge (NTB) is a device that allows you to connect
* two or more systems using a PCI-e links, providing remote memory access.
*
* This module contains a driver for NTB hardware in Intel Xeon/Atom CPUs.
*
* NOTE: Much of the code in this module is shared with Linux. Any patches may
* be picked up and redistributed in Linux with a dual GPL/BSD license.
*/
#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");
#include <sys/param.h>
#include <sys/kernel.h>
#include <sys/systm.h>
#include <sys/bus.h>
#include <sys/endian.h>
#include <sys/interrupt.h>
#include <sys/lock.h>
#include <sys/malloc.h>
#include <sys/module.h>
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
#include <sys/mutex.h>
#include <sys/pciio.h>
#include <sys/queue.h>
#include <sys/rman.h>
#include <sys/sbuf.h>
2015-10-13 03:11:21 +00:00
#include <sys/sysctl.h>
#include <vm/vm.h>
#include <vm/pmap.h>
#include <machine/bus.h>
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
#include <machine/intr_machdep.h>
#include <machine/resource.h>
#include <dev/pci/pcireg.h>
#include <dev/pci/pcivar.h>
#include "ntb_hw_intel.h"
#include "../ntb.h"
#define MAX_MSIX_INTERRUPTS MAX(XEON_DB_COUNT, ATOM_DB_COUNT)
#define NTB_HB_TIMEOUT 1 /* second */
#define ATOM_LINK_RECOVERY_TIME 500 /* ms */
#define BAR_HIGH_MASK (~((1ull << 12) - 1))
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
#define NTB_MSIX_VER_GUARD 0xaabbccdd
#define NTB_MSIX_RECEIVED 0xe0f0e0f0
/*
* PCI constants could be somewhere more generic, but aren't defined/used in
* pci.c.
*/
#define PCI_MSIX_ENTRY_SIZE 16
#define PCI_MSIX_ENTRY_LOWER_ADDR 0
#define PCI_MSIX_ENTRY_UPPER_ADDR 4
#define PCI_MSIX_ENTRY_DATA 8
enum ntb_device_type {
NTB_XEON,
NTB_ATOM
};
/* ntb_conn_type are hardware numbers, cannot change. */
enum ntb_conn_type {
NTB_CONN_TRANSPARENT = 0,
NTB_CONN_B2B = 1,
NTB_CONN_RP = 2,
};
enum ntb_b2b_direction {
NTB_DEV_USD = 0,
NTB_DEV_DSD = 1,
};
enum ntb_bar {
NTB_CONFIG_BAR = 0,
NTB_B2B_BAR_1,
NTB_B2B_BAR_2,
NTB_B2B_BAR_3,
NTB_MAX_BARS
};
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
enum {
NTB_MSIX_GUARD = 0,
NTB_MSIX_DATA0,
NTB_MSIX_DATA1,
NTB_MSIX_DATA2,
NTB_MSIX_OFS0,
NTB_MSIX_OFS1,
NTB_MSIX_OFS2,
NTB_MSIX_DONE,
NTB_MAX_MSIX_SPAD
};
/* Device features and workarounds */
#define HAS_FEATURE(ntb, feature) \
(((ntb)->features & (feature)) != 0)
struct ntb_hw_info {
uint32_t device_id;
const char *desc;
enum ntb_device_type type;
uint32_t features;
};
struct ntb_pci_bar_info {
bus_space_tag_t pci_bus_tag;
bus_space_handle_t pci_bus_handle;
int pci_resource_id;
struct resource *pci_resource;
vm_paddr_t pbase;
caddr_t vbase;
vm_size_t size;
vm_memattr_t map_mode;
/* Configuration register offsets */
uint32_t psz_off;
uint32_t ssz_off;
uint32_t pbarxlat_off;
};
struct ntb_int_info {
struct resource *res;
int rid;
void *tag;
};
struct ntb_vec {
struct ntb_softc *ntb;
uint32_t num;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
unsigned masked;
};
struct ntb_reg {
uint32_t ntb_ctl;
uint32_t lnk_sta;
uint8_t db_size;
unsigned mw_bar[NTB_MAX_BARS];
};
struct ntb_alt_reg {
uint32_t db_bell;
uint32_t db_mask;
uint32_t spad;
};
struct ntb_xlat_reg {
uint32_t bar0_base;
uint32_t bar2_base;
uint32_t bar4_base;
uint32_t bar5_base;
uint32_t bar2_xlat;
uint32_t bar4_xlat;
uint32_t bar5_xlat;
uint32_t bar2_limit;
uint32_t bar4_limit;
uint32_t bar5_limit;
};
struct ntb_b2b_addr {
uint64_t bar0_addr;
uint64_t bar2_addr64;
uint64_t bar4_addr64;
uint64_t bar4_addr32;
uint64_t bar5_addr32;
};
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
struct ntb_msix_data {
uint32_t nmd_ofs;
uint32_t nmd_data;
};
struct ntb_softc {
/* ntb.c context. Do not move! Must go first! */
void *ntb_store;
device_t device;
enum ntb_device_type type;
uint32_t features;
struct ntb_pci_bar_info bar_info[NTB_MAX_BARS];
struct ntb_int_info int_info[MAX_MSIX_INTERRUPTS];
uint32_t allocated_interrupts;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
struct ntb_msix_data peer_msix_data[XEON_NONLINK_DB_MSIX_BITS];
struct ntb_msix_data msix_data[XEON_NONLINK_DB_MSIX_BITS];
bool peer_msix_good;
bool peer_msix_done;
struct ntb_pci_bar_info *peer_lapic_bar;
struct callout peer_msix_work;
struct callout heartbeat_timer;
struct callout lr_timer;
struct ntb_vec *msix_vec;
uint32_t ppd;
enum ntb_conn_type conn_type;
enum ntb_b2b_direction dev_type;
/* Offset of peer bar0 in B2B BAR */
uint64_t b2b_off;
/* Memory window used to access peer bar0 */
#define B2B_MW_DISABLED UINT8_MAX
uint8_t b2b_mw_idx;
uint32_t msix_xlat;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
uint8_t msix_mw_idx;
uint8_t mw_count;
uint8_t spad_count;
uint8_t db_count;
uint8_t db_vec_count;
uint8_t db_vec_shift;
/* Protects local db_mask. */
#define DB_MASK_LOCK(sc) mtx_lock_spin(&(sc)->db_mask_lock)
#define DB_MASK_UNLOCK(sc) mtx_unlock_spin(&(sc)->db_mask_lock)
#define DB_MASK_ASSERT(sc,f) mtx_assert(&(sc)->db_mask_lock, (f))
struct mtx db_mask_lock;
volatile uint32_t ntb_ctl;
volatile uint32_t lnk_sta;
uint64_t db_valid_mask;
uint64_t db_link_mask;
uint64_t db_mask;
uint64_t fake_db; /* NTB_SB01BASE_LOCKUP*/
uint64_t force_db; /* NTB_SB01BASE_LOCKUP*/
int last_ts; /* ticks @ last irq */
const struct ntb_reg *reg;
const struct ntb_alt_reg *self_reg;
const struct ntb_alt_reg *peer_reg;
const struct ntb_xlat_reg *xlat_reg;
};
#ifdef __i386__
static __inline uint64_t
bus_space_read_8(bus_space_tag_t tag, bus_space_handle_t handle,
bus_size_t offset)
{
return (bus_space_read_4(tag, handle, offset) |
((uint64_t)bus_space_read_4(tag, handle, offset + 4)) << 32);
}
static __inline void
bus_space_write_8(bus_space_tag_t tag, bus_space_handle_t handle,
bus_size_t offset, uint64_t val)
{
bus_space_write_4(tag, handle, offset, val);
bus_space_write_4(tag, handle, offset + 4, val >> 32);
}
#endif
#define intel_ntb_bar_read(SIZE, bar, offset) \
bus_space_read_ ## SIZE (ntb->bar_info[(bar)].pci_bus_tag, \
ntb->bar_info[(bar)].pci_bus_handle, (offset))
#define intel_ntb_bar_write(SIZE, bar, offset, val) \
bus_space_write_ ## SIZE (ntb->bar_info[(bar)].pci_bus_tag, \
ntb->bar_info[(bar)].pci_bus_handle, (offset), (val))
#define intel_ntb_reg_read(SIZE, offset) \
intel_ntb_bar_read(SIZE, NTB_CONFIG_BAR, offset)
#define intel_ntb_reg_write(SIZE, offset, val) \
intel_ntb_bar_write(SIZE, NTB_CONFIG_BAR, offset, val)
#define intel_ntb_mw_read(SIZE, offset) \
intel_ntb_bar_read(SIZE, intel_ntb_mw_to_bar(ntb, ntb->b2b_mw_idx), \
offset)
#define intel_ntb_mw_write(SIZE, offset, val) \
intel_ntb_bar_write(SIZE, intel_ntb_mw_to_bar(ntb, ntb->b2b_mw_idx), \
offset, val)
static int intel_ntb_probe(device_t device);
static int intel_ntb_attach(device_t device);
static int intel_ntb_detach(device_t device);
static uint64_t intel_ntb_db_valid_mask(device_t dev);
static void intel_ntb_spad_clear(device_t dev);
static uint64_t intel_ntb_db_vector_mask(device_t dev, uint32_t vector);
static bool intel_ntb_link_is_up(device_t dev, enum ntb_speed *speed,
enum ntb_width *width);
static int intel_ntb_link_enable(device_t dev, enum ntb_speed speed,
enum ntb_width width);
static int intel_ntb_link_disable(device_t dev);
static int intel_ntb_spad_read(device_t dev, unsigned int idx, uint32_t *val);
static int intel_ntb_peer_spad_write(device_t dev, unsigned int idx, uint32_t val);
static unsigned intel_ntb_user_mw_to_idx(struct ntb_softc *, unsigned uidx);
static inline enum ntb_bar intel_ntb_mw_to_bar(struct ntb_softc *, unsigned mw);
static inline bool bar_is_64bit(struct ntb_softc *, enum ntb_bar);
static inline void bar_get_xlat_params(struct ntb_softc *, enum ntb_bar,
uint32_t *base, uint32_t *xlat, uint32_t *lmt);
static int intel_ntb_map_pci_bars(struct ntb_softc *ntb);
static int intel_ntb_mw_set_wc_internal(struct ntb_softc *, unsigned idx,
vm_memattr_t);
static void print_map_success(struct ntb_softc *, struct ntb_pci_bar_info *,
const char *);
static int map_mmr_bar(struct ntb_softc *ntb, struct ntb_pci_bar_info *bar);
static int map_memory_window_bar(struct ntb_softc *ntb,
struct ntb_pci_bar_info *bar);
static void intel_ntb_unmap_pci_bar(struct ntb_softc *ntb);
static int intel_ntb_remap_msix(device_t, uint32_t desired, uint32_t avail);
static int intel_ntb_init_isr(struct ntb_softc *ntb);
static int intel_ntb_setup_legacy_interrupt(struct ntb_softc *ntb);
static int intel_ntb_setup_msix(struct ntb_softc *ntb, uint32_t num_vectors);
static void intel_ntb_teardown_interrupts(struct ntb_softc *ntb);
static inline uint64_t intel_ntb_vec_mask(struct ntb_softc *, uint64_t db_vector);
static void intel_ntb_interrupt(struct ntb_softc *, uint32_t vec);
static void ndev_vec_isr(void *arg);
static void ndev_irq_isr(void *arg);
static inline uint64_t db_ioread(struct ntb_softc *, uint64_t regoff);
static inline void db_iowrite(struct ntb_softc *, uint64_t regoff, uint64_t);
static inline void db_iowrite_raw(struct ntb_softc *, uint64_t regoff, uint64_t);
static int intel_ntb_create_msix_vec(struct ntb_softc *ntb, uint32_t num_vectors);
static void intel_ntb_free_msix_vec(struct ntb_softc *ntb);
static void intel_ntb_get_msix_info(struct ntb_softc *ntb);
static void intel_ntb_exchange_msix(void *);
static struct ntb_hw_info *intel_ntb_get_device_info(uint32_t device_id);
static void intel_ntb_detect_max_mw(struct ntb_softc *ntb);
static int intel_ntb_detect_xeon(struct ntb_softc *ntb);
static int intel_ntb_detect_atom(struct ntb_softc *ntb);
static int intel_ntb_xeon_init_dev(struct ntb_softc *ntb);
static int intel_ntb_atom_init_dev(struct ntb_softc *ntb);
static void intel_ntb_teardown_xeon(struct ntb_softc *ntb);
static void configure_atom_secondary_side_bars(struct ntb_softc *ntb);
static void xeon_reset_sbar_size(struct ntb_softc *, enum ntb_bar idx,
enum ntb_bar regbar);
static void xeon_set_sbar_base_and_limit(struct ntb_softc *,
uint64_t base_addr, enum ntb_bar idx, enum ntb_bar regbar);
static void xeon_set_pbar_xlat(struct ntb_softc *, uint64_t base_addr,
enum ntb_bar idx);
static int xeon_setup_b2b_mw(struct ntb_softc *,
const struct ntb_b2b_addr *addr, const struct ntb_b2b_addr *peer_addr);
static inline bool link_is_up(struct ntb_softc *ntb);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
static inline bool _xeon_link_is_up(struct ntb_softc *ntb);
static inline bool atom_link_is_err(struct ntb_softc *ntb);
static inline enum ntb_speed intel_ntb_link_sta_speed(struct ntb_softc *);
static inline enum ntb_width intel_ntb_link_sta_width(struct ntb_softc *);
static void atom_link_hb(void *arg);
static void recover_atom_link(void *arg);
static bool intel_ntb_poll_link(struct ntb_softc *ntb);
static void save_bar_parameters(struct ntb_pci_bar_info *bar);
static void intel_ntb_sysctl_init(struct ntb_softc *);
static int sysctl_handle_features(SYSCTL_HANDLER_ARGS);
static int sysctl_handle_link_admin(SYSCTL_HANDLER_ARGS);
static int sysctl_handle_link_status_human(SYSCTL_HANDLER_ARGS);
static int sysctl_handle_link_status(SYSCTL_HANDLER_ARGS);
static int sysctl_handle_register(SYSCTL_HANDLER_ARGS);
static unsigned g_ntb_hw_debug_level;
SYSCTL_UINT(_hw_ntb, OID_AUTO, debug_level, CTLFLAG_RWTUN,
&g_ntb_hw_debug_level, 0, "ntb_hw log level -- higher is more verbose");
#define intel_ntb_printf(lvl, ...) do { \
if ((lvl) <= g_ntb_hw_debug_level) { \
device_printf(ntb->device, __VA_ARGS__); \
} \
} while (0)
#define _NTB_PAT_UC 0
#define _NTB_PAT_WC 1
#define _NTB_PAT_WT 4
#define _NTB_PAT_WP 5
#define _NTB_PAT_WB 6
#define _NTB_PAT_UCM 7
static unsigned g_ntb_mw_pat = _NTB_PAT_UC;
SYSCTL_UINT(_hw_ntb, OID_AUTO, default_mw_pat, CTLFLAG_RDTUN,
&g_ntb_mw_pat, 0, "Configure the default memory window cache flags (PAT): "
"UC: " __XSTRING(_NTB_PAT_UC) ", "
"WC: " __XSTRING(_NTB_PAT_WC) ", "
"WT: " __XSTRING(_NTB_PAT_WT) ", "
"WP: " __XSTRING(_NTB_PAT_WP) ", "
"WB: " __XSTRING(_NTB_PAT_WB) ", "
"UC-: " __XSTRING(_NTB_PAT_UCM));
static inline vm_memattr_t
intel_ntb_pat_flags(void)
{
switch (g_ntb_mw_pat) {
case _NTB_PAT_WC:
return (VM_MEMATTR_WRITE_COMBINING);
case _NTB_PAT_WT:
return (VM_MEMATTR_WRITE_THROUGH);
case _NTB_PAT_WP:
return (VM_MEMATTR_WRITE_PROTECTED);
case _NTB_PAT_WB:
return (VM_MEMATTR_WRITE_BACK);
case _NTB_PAT_UCM:
return (VM_MEMATTR_WEAK_UNCACHEABLE);
case _NTB_PAT_UC:
/* FALLTHROUGH */
default:
return (VM_MEMATTR_UNCACHEABLE);
}
}
/*
* Well, this obviously doesn't belong here, but it doesn't seem to exist
* anywhere better yet.
*/
static inline const char *
intel_ntb_vm_memattr_to_str(vm_memattr_t pat)
{
switch (pat) {
case VM_MEMATTR_WRITE_COMBINING:
return ("WRITE_COMBINING");
case VM_MEMATTR_WRITE_THROUGH:
return ("WRITE_THROUGH");
case VM_MEMATTR_WRITE_PROTECTED:
return ("WRITE_PROTECTED");
case VM_MEMATTR_WRITE_BACK:
return ("WRITE_BACK");
case VM_MEMATTR_WEAK_UNCACHEABLE:
return ("UNCACHED");
case VM_MEMATTR_UNCACHEABLE:
return ("UNCACHEABLE");
default:
return ("UNKNOWN");
}
}
static int g_ntb_msix_idx = 1;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
SYSCTL_INT(_hw_ntb, OID_AUTO, msix_mw_idx, CTLFLAG_RDTUN, &g_ntb_msix_idx,
0, "Use this memory window to access the peer MSIX message complex on "
"certain Xeon-based NTB systems, as a workaround for a hardware errata. "
"Like b2b_mw_idx, negative values index from the last available memory "
"window. (Applies on Xeon platforms with SB01BASE_LOCKUP errata.)");
static int g_ntb_mw_idx = -1;
SYSCTL_INT(_hw_ntb, OID_AUTO, b2b_mw_idx, CTLFLAG_RDTUN, &g_ntb_mw_idx,
0, "Use this memory window to access the peer NTB registers. A "
"non-negative value starts from the first MW index; a negative value "
"starts from the last MW index. The default is -1, i.e., the last "
"available memory window. Both sides of the NTB MUST set the same "
"value here! (Applies on Xeon platforms with SDOORBELL_LOCKUP errata.)");
/* Hardware owns the low 16 bits of features. */
#define NTB_BAR_SIZE_4K (1 << 0)
#define NTB_SDOORBELL_LOCKUP (1 << 1)
#define NTB_SB01BASE_LOCKUP (1 << 2)
#define NTB_B2BDOORBELL_BIT14 (1 << 3)
/* Software/configuration owns the top 16 bits. */
#define NTB_SPLIT_BAR (1ull << 16)
#define NTB_FEATURES_STR \
"\20\21SPLIT_BAR4\04B2B_DOORBELL_BIT14\03SB01BASE_LOCKUP" \
"\02SDOORBELL_LOCKUP\01BAR_SIZE_4K"
static struct ntb_hw_info pci_ids[] = {
/* XXX: PS/SS IDs left out until they are supported. */
{ 0x0C4E8086, "BWD Atom Processor S1200 Non-Transparent Bridge B2B",
NTB_ATOM, 0 },
{ 0x37258086, "JSF Xeon C35xx/C55xx Non-Transparent Bridge B2B",
NTB_XEON, NTB_SDOORBELL_LOCKUP | NTB_B2BDOORBELL_BIT14 },
{ 0x3C0D8086, "SNB Xeon E5/Core i7 Non-Transparent Bridge B2B",
NTB_XEON, NTB_SDOORBELL_LOCKUP | NTB_B2BDOORBELL_BIT14 },
{ 0x0E0D8086, "IVT Xeon E5 V2 Non-Transparent Bridge B2B", NTB_XEON,
NTB_SDOORBELL_LOCKUP | NTB_B2BDOORBELL_BIT14 |
NTB_SB01BASE_LOCKUP | NTB_BAR_SIZE_4K },
{ 0x2F0D8086, "HSX Xeon E5 V3 Non-Transparent Bridge B2B", NTB_XEON,
NTB_SDOORBELL_LOCKUP | NTB_B2BDOORBELL_BIT14 |
NTB_SB01BASE_LOCKUP },
{ 0x6F0D8086, "BDX Xeon E5 V4 Non-Transparent Bridge B2B", NTB_XEON,
NTB_SDOORBELL_LOCKUP | NTB_B2BDOORBELL_BIT14 |
NTB_SB01BASE_LOCKUP },
};
static const struct ntb_reg atom_reg = {
.ntb_ctl = ATOM_NTBCNTL_OFFSET,
.lnk_sta = ATOM_LINK_STATUS_OFFSET,
.db_size = sizeof(uint64_t),
.mw_bar = { NTB_B2B_BAR_1, NTB_B2B_BAR_2 },
};
static const struct ntb_alt_reg atom_pri_reg = {
.db_bell = ATOM_PDOORBELL_OFFSET,
.db_mask = ATOM_PDBMSK_OFFSET,
.spad = ATOM_SPAD_OFFSET,
};
static const struct ntb_alt_reg atom_b2b_reg = {
.db_bell = ATOM_B2B_DOORBELL_OFFSET,
.spad = ATOM_B2B_SPAD_OFFSET,
};
static const struct ntb_xlat_reg atom_sec_xlat = {
#if 0
/* "FIXME" says the Linux driver. */
.bar0_base = ATOM_SBAR0BASE_OFFSET,
.bar2_base = ATOM_SBAR2BASE_OFFSET,
.bar4_base = ATOM_SBAR4BASE_OFFSET,
.bar2_limit = ATOM_SBAR2LMT_OFFSET,
.bar4_limit = ATOM_SBAR4LMT_OFFSET,
#endif
.bar2_xlat = ATOM_SBAR2XLAT_OFFSET,
.bar4_xlat = ATOM_SBAR4XLAT_OFFSET,
};
static const struct ntb_reg xeon_reg = {
.ntb_ctl = XEON_NTBCNTL_OFFSET,
.lnk_sta = XEON_LINK_STATUS_OFFSET,
.db_size = sizeof(uint16_t),
.mw_bar = { NTB_B2B_BAR_1, NTB_B2B_BAR_2, NTB_B2B_BAR_3 },
};
static const struct ntb_alt_reg xeon_pri_reg = {
.db_bell = XEON_PDOORBELL_OFFSET,
.db_mask = XEON_PDBMSK_OFFSET,
.spad = XEON_SPAD_OFFSET,
};
static const struct ntb_alt_reg xeon_b2b_reg = {
.db_bell = XEON_B2B_DOORBELL_OFFSET,
.spad = XEON_B2B_SPAD_OFFSET,
};
static const struct ntb_xlat_reg xeon_sec_xlat = {
.bar0_base = XEON_SBAR0BASE_OFFSET,
.bar2_base = XEON_SBAR2BASE_OFFSET,
.bar4_base = XEON_SBAR4BASE_OFFSET,
.bar5_base = XEON_SBAR5BASE_OFFSET,
.bar2_limit = XEON_SBAR2LMT_OFFSET,
.bar4_limit = XEON_SBAR4LMT_OFFSET,
.bar5_limit = XEON_SBAR5LMT_OFFSET,
.bar2_xlat = XEON_SBAR2XLAT_OFFSET,
.bar4_xlat = XEON_SBAR4XLAT_OFFSET,
.bar5_xlat = XEON_SBAR5XLAT_OFFSET,
};
static struct ntb_b2b_addr xeon_b2b_usd_addr = {
.bar0_addr = XEON_B2B_BAR0_ADDR,
.bar2_addr64 = XEON_B2B_BAR2_ADDR64,
.bar4_addr64 = XEON_B2B_BAR4_ADDR64,
.bar4_addr32 = XEON_B2B_BAR4_ADDR32,
.bar5_addr32 = XEON_B2B_BAR5_ADDR32,
};
static struct ntb_b2b_addr xeon_b2b_dsd_addr = {
.bar0_addr = XEON_B2B_BAR0_ADDR,
.bar2_addr64 = XEON_B2B_BAR2_ADDR64,
.bar4_addr64 = XEON_B2B_BAR4_ADDR64,
.bar4_addr32 = XEON_B2B_BAR4_ADDR32,
.bar5_addr32 = XEON_B2B_BAR5_ADDR32,
};
SYSCTL_NODE(_hw_ntb, OID_AUTO, xeon_b2b, CTLFLAG_RW, 0,
"B2B MW segment overrides -- MUST be the same on both sides");
SYSCTL_UQUAD(_hw_ntb_xeon_b2b, OID_AUTO, usd_bar2_addr64, CTLFLAG_RDTUN,
&xeon_b2b_usd_addr.bar2_addr64, 0, "If using B2B topology on Xeon "
"hardware, use this 64-bit address on the bus between the NTB devices for "
"the window at BAR2, on the upstream side of the link. MUST be the same "
"address on both sides.");
SYSCTL_UQUAD(_hw_ntb_xeon_b2b, OID_AUTO, usd_bar4_addr64, CTLFLAG_RDTUN,
&xeon_b2b_usd_addr.bar4_addr64, 0, "See usd_bar2_addr64, but BAR4.");
SYSCTL_UQUAD(_hw_ntb_xeon_b2b, OID_AUTO, usd_bar4_addr32, CTLFLAG_RDTUN,
&xeon_b2b_usd_addr.bar4_addr32, 0, "See usd_bar2_addr64, but BAR4 "
"(split-BAR mode).");
SYSCTL_UQUAD(_hw_ntb_xeon_b2b, OID_AUTO, usd_bar5_addr32, CTLFLAG_RDTUN,
&xeon_b2b_usd_addr.bar5_addr32, 0, "See usd_bar2_addr64, but BAR5 "
"(split-BAR mode).");
SYSCTL_UQUAD(_hw_ntb_xeon_b2b, OID_AUTO, dsd_bar2_addr64, CTLFLAG_RDTUN,
&xeon_b2b_dsd_addr.bar2_addr64, 0, "If using B2B topology on Xeon "
"hardware, use this 64-bit address on the bus between the NTB devices for "
"the window at BAR2, on the downstream side of the link. MUST be the same"
" address on both sides.");
SYSCTL_UQUAD(_hw_ntb_xeon_b2b, OID_AUTO, dsd_bar4_addr64, CTLFLAG_RDTUN,
&xeon_b2b_dsd_addr.bar4_addr64, 0, "See dsd_bar2_addr64, but BAR4.");
SYSCTL_UQUAD(_hw_ntb_xeon_b2b, OID_AUTO, dsd_bar4_addr32, CTLFLAG_RDTUN,
&xeon_b2b_dsd_addr.bar4_addr32, 0, "See dsd_bar2_addr64, but BAR4 "
"(split-BAR mode).");
SYSCTL_UQUAD(_hw_ntb_xeon_b2b, OID_AUTO, dsd_bar5_addr32, CTLFLAG_RDTUN,
&xeon_b2b_dsd_addr.bar5_addr32, 0, "See dsd_bar2_addr64, but BAR5 "
"(split-BAR mode).");
/*
* OS <-> Driver interface structures
*/
MALLOC_DEFINE(M_NTB, "ntb_hw", "ntb_hw driver memory allocations");
/*
* OS <-> Driver linkage functions
*/
static int
intel_ntb_probe(device_t device)
{
2015-10-13 03:12:55 +00:00
struct ntb_hw_info *p;
p = intel_ntb_get_device_info(pci_get_devid(device));
2015-10-13 03:12:55 +00:00
if (p == NULL)
return (ENXIO);
2015-10-13 03:12:55 +00:00
device_set_desc(device, p->desc);
return (0);
}
static int
intel_ntb_attach(device_t device)
{
2015-10-13 03:12:55 +00:00
struct ntb_softc *ntb;
struct ntb_hw_info *p;
int error;
ntb = device_get_softc(device);
p = intel_ntb_get_device_info(pci_get_devid(device));
2015-10-13 03:12:55 +00:00
ntb->device = device;
ntb->type = p->type;
ntb->features = p->features;
ntb->b2b_mw_idx = B2B_MW_DISABLED;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
ntb->msix_mw_idx = B2B_MW_DISABLED;
/* Heartbeat timer for NTB_ATOM since there is no link interrupt */
callout_init(&ntb->heartbeat_timer, 1);
callout_init(&ntb->lr_timer, 1);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
callout_init(&ntb->peer_msix_work, 1);
mtx_init(&ntb->db_mask_lock, "ntb hw bits", NULL, MTX_SPIN);
if (ntb->type == NTB_ATOM)
error = intel_ntb_detect_atom(ntb);
else
error = intel_ntb_detect_xeon(ntb);
if (error != 0)
goto out;
intel_ntb_detect_max_mw(ntb);
pci_enable_busmaster(ntb->device);
error = intel_ntb_map_pci_bars(ntb);
if (error != 0)
2015-10-13 03:12:55 +00:00
goto out;
if (ntb->type == NTB_ATOM)
error = intel_ntb_atom_init_dev(ntb);
else
error = intel_ntb_xeon_init_dev(ntb);
if (error != 0)
2015-10-13 03:12:55 +00:00
goto out;
intel_ntb_spad_clear(device);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
intel_ntb_poll_link(ntb);
intel_ntb_sysctl_init(ntb);
/* Attach children to this controller */
error = ntb_register_device(device);
2015-10-13 03:12:55 +00:00
out:
if (error != 0)
intel_ntb_detach(device);
return (error);
}
static int
intel_ntb_detach(device_t device)
{
2015-10-13 03:12:55 +00:00
struct ntb_softc *ntb;
ntb = device_get_softc(device);
/* Detach & delete all children */
ntb_unregister_device(device);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
if (ntb->self_reg != NULL) {
DB_MASK_LOCK(ntb);
db_iowrite(ntb, ntb->self_reg->db_mask, ntb->db_valid_mask);
DB_MASK_UNLOCK(ntb);
}
callout_drain(&ntb->heartbeat_timer);
callout_drain(&ntb->lr_timer);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
callout_drain(&ntb->peer_msix_work);
pci_disable_busmaster(ntb->device);
if (ntb->type == NTB_XEON)
intel_ntb_teardown_xeon(ntb);
intel_ntb_teardown_interrupts(ntb);
mtx_destroy(&ntb->db_mask_lock);
intel_ntb_unmap_pci_bar(ntb);
return (0);
}
/*
* Driver internal routines
*/
static inline enum ntb_bar
intel_ntb_mw_to_bar(struct ntb_softc *ntb, unsigned mw)
{
KASSERT(mw < ntb->mw_count,
("%s: mw:%u > count:%u", __func__, mw, (unsigned)ntb->mw_count));
KASSERT(ntb->reg->mw_bar[mw] != 0, ("invalid mw"));
return (ntb->reg->mw_bar[mw]);
}
static inline bool
bar_is_64bit(struct ntb_softc *ntb, enum ntb_bar bar)
{
/* XXX This assertion could be stronger. */
KASSERT(bar < NTB_MAX_BARS, ("bogus bar"));
return (bar < NTB_B2B_BAR_2 || !HAS_FEATURE(ntb, NTB_SPLIT_BAR));
}
static inline void
bar_get_xlat_params(struct ntb_softc *ntb, enum ntb_bar bar, uint32_t *base,
uint32_t *xlat, uint32_t *lmt)
{
uint32_t basev, lmtv, xlatv;
switch (bar) {
case NTB_B2B_BAR_1:
basev = ntb->xlat_reg->bar2_base;
lmtv = ntb->xlat_reg->bar2_limit;
xlatv = ntb->xlat_reg->bar2_xlat;
break;
case NTB_B2B_BAR_2:
basev = ntb->xlat_reg->bar4_base;
lmtv = ntb->xlat_reg->bar4_limit;
xlatv = ntb->xlat_reg->bar4_xlat;
break;
case NTB_B2B_BAR_3:
basev = ntb->xlat_reg->bar5_base;
lmtv = ntb->xlat_reg->bar5_limit;
xlatv = ntb->xlat_reg->bar5_xlat;
break;
default:
KASSERT(bar >= NTB_B2B_BAR_1 && bar < NTB_MAX_BARS,
("bad bar"));
basev = lmtv = xlatv = 0;
break;
}
if (base != NULL)
*base = basev;
if (xlat != NULL)
*xlat = xlatv;
if (lmt != NULL)
*lmt = lmtv;
}
static int
intel_ntb_map_pci_bars(struct ntb_softc *ntb)
{
int rc;
ntb->bar_info[NTB_CONFIG_BAR].pci_resource_id = PCIR_BAR(0);
rc = map_mmr_bar(ntb, &ntb->bar_info[NTB_CONFIG_BAR]);
if (rc != 0)
goto out;
2015-10-13 03:12:55 +00:00
ntb->bar_info[NTB_B2B_BAR_1].pci_resource_id = PCIR_BAR(2);
rc = map_memory_window_bar(ntb, &ntb->bar_info[NTB_B2B_BAR_1]);
if (rc != 0)
goto out;
ntb->bar_info[NTB_B2B_BAR_1].psz_off = XEON_PBAR23SZ_OFFSET;
ntb->bar_info[NTB_B2B_BAR_1].ssz_off = XEON_SBAR23SZ_OFFSET;
ntb->bar_info[NTB_B2B_BAR_1].pbarxlat_off = XEON_PBAR2XLAT_OFFSET;
2015-10-13 03:12:55 +00:00
ntb->bar_info[NTB_B2B_BAR_2].pci_resource_id = PCIR_BAR(4);
rc = map_memory_window_bar(ntb, &ntb->bar_info[NTB_B2B_BAR_2]);
if (rc != 0)
goto out;
ntb->bar_info[NTB_B2B_BAR_2].psz_off = XEON_PBAR4SZ_OFFSET;
ntb->bar_info[NTB_B2B_BAR_2].ssz_off = XEON_SBAR4SZ_OFFSET;
ntb->bar_info[NTB_B2B_BAR_2].pbarxlat_off = XEON_PBAR4XLAT_OFFSET;
if (!HAS_FEATURE(ntb, NTB_SPLIT_BAR))
goto out;
ntb->bar_info[NTB_B2B_BAR_3].pci_resource_id = PCIR_BAR(5);
rc = map_memory_window_bar(ntb, &ntb->bar_info[NTB_B2B_BAR_3]);
ntb->bar_info[NTB_B2B_BAR_3].psz_off = XEON_PBAR5SZ_OFFSET;
ntb->bar_info[NTB_B2B_BAR_3].ssz_off = XEON_SBAR5SZ_OFFSET;
ntb->bar_info[NTB_B2B_BAR_3].pbarxlat_off = XEON_PBAR5XLAT_OFFSET;
out:
2015-10-13 03:12:55 +00:00
if (rc != 0)
device_printf(ntb->device,
"unable to allocate pci resource\n");
return (rc);
}
static void
print_map_success(struct ntb_softc *ntb, struct ntb_pci_bar_info *bar,
const char *kind)
{
device_printf(ntb->device,
"Mapped BAR%d v:[%p-%p] p:[%p-%p] (0x%jx bytes) (%s)\n",
PCI_RID2BAR(bar->pci_resource_id), bar->vbase,
(char *)bar->vbase + bar->size - 1,
(void *)bar->pbase, (void *)(bar->pbase + bar->size - 1),
(uintmax_t)bar->size, kind);
}
static int
map_mmr_bar(struct ntb_softc *ntb, struct ntb_pci_bar_info *bar)
{
bar->pci_resource = bus_alloc_resource_any(ntb->device, SYS_RES_MEMORY,
2015-10-13 03:12:55 +00:00
&bar->pci_resource_id, RF_ACTIVE);
if (bar->pci_resource == NULL)
return (ENXIO);
2015-10-13 03:12:55 +00:00
save_bar_parameters(bar);
bar->map_mode = VM_MEMATTR_UNCACHEABLE;
print_map_success(ntb, bar, "mmr");
2015-10-13 03:12:55 +00:00
return (0);
}
static int
map_memory_window_bar(struct ntb_softc *ntb, struct ntb_pci_bar_info *bar)
{
int rc;
vm_memattr_t mapmode;
uint8_t bar_size_bits = 0;
2015-10-13 03:12:55 +00:00
bar->pci_resource = bus_alloc_resource_any(ntb->device, SYS_RES_MEMORY,
&bar->pci_resource_id, RF_ACTIVE);
if (bar->pci_resource == NULL)
return (ENXIO);
2015-10-13 03:12:55 +00:00
save_bar_parameters(bar);
/*
* Ivytown NTB BAR sizes are misreported by the hardware due to a
* hardware issue. To work around this, query the size it should be
* configured to by the device and modify the resource to correspond to
* this new size. The BIOS on systems with this problem is required to
* provide enough address space to allow the driver to make this change
* safely.
*
* Ideally I could have just specified the size when I allocated the
* resource like:
* bus_alloc_resource(ntb->device,
* SYS_RES_MEMORY, &bar->pci_resource_id, 0ul, ~0ul,
* 1ul << bar_size_bits, RF_ACTIVE);
* but the PCI driver does not honor the size in this call, so we have
* to modify it after the fact.
*/
if (HAS_FEATURE(ntb, NTB_BAR_SIZE_4K)) {
2015-10-13 03:12:55 +00:00
if (bar->pci_resource_id == PCIR_BAR(2))
bar_size_bits = pci_read_config(ntb->device,
XEON_PBAR23SZ_OFFSET, 1);
else
bar_size_bits = pci_read_config(ntb->device,
XEON_PBAR45SZ_OFFSET, 1);
rc = bus_adjust_resource(ntb->device, SYS_RES_MEMORY,
bar->pci_resource, bar->pbase,
bar->pbase + (1ul << bar_size_bits) - 1);
if (rc != 0) {
2015-10-13 03:12:55 +00:00
device_printf(ntb->device,
"unable to resize bar\n");
return (rc);
}
2015-10-13 03:12:55 +00:00
save_bar_parameters(bar);
}
bar->map_mode = VM_MEMATTR_UNCACHEABLE;
print_map_success(ntb, bar, "mw");
/*
* Optionally, mark MW BARs as anything other than UC to improve
* performance.
*/
mapmode = intel_ntb_pat_flags();
if (mapmode == bar->map_mode)
return (0);
rc = pmap_change_attr((vm_offset_t)bar->vbase, bar->size, mapmode);
if (rc == 0) {
bar->map_mode = mapmode;
2015-10-13 03:12:55 +00:00
device_printf(ntb->device,
"Marked BAR%d v:[%p-%p] p:[%p-%p] as "
"%s.\n",
PCI_RID2BAR(bar->pci_resource_id), bar->vbase,
(char *)bar->vbase + bar->size - 1,
(void *)bar->pbase, (void *)(bar->pbase + bar->size - 1),
intel_ntb_vm_memattr_to_str(mapmode));
} else
device_printf(ntb->device,
"Unable to mark BAR%d v:[%p-%p] p:[%p-%p] as "
"%s: %d\n",
PCI_RID2BAR(bar->pci_resource_id), bar->vbase,
(char *)bar->vbase + bar->size - 1,
(void *)bar->pbase, (void *)(bar->pbase + bar->size - 1),
intel_ntb_vm_memattr_to_str(mapmode), rc);
/* Proceed anyway */
return (0);
}
static void
intel_ntb_unmap_pci_bar(struct ntb_softc *ntb)
{
struct ntb_pci_bar_info *current_bar;
int i;
for (i = 0; i < NTB_MAX_BARS; i++) {
current_bar = &ntb->bar_info[i];
if (current_bar->pci_resource != NULL)
bus_release_resource(ntb->device, SYS_RES_MEMORY,
current_bar->pci_resource_id,
current_bar->pci_resource);
}
}
static int
intel_ntb_setup_msix(struct ntb_softc *ntb, uint32_t num_vectors)
{
uint32_t i;
int rc;
for (i = 0; i < num_vectors; i++) {
ntb->int_info[i].rid = i + 1;
ntb->int_info[i].res = bus_alloc_resource_any(ntb->device,
SYS_RES_IRQ, &ntb->int_info[i].rid, RF_ACTIVE);
if (ntb->int_info[i].res == NULL) {
device_printf(ntb->device,
"bus_alloc_resource failed\n");
return (ENOMEM);
}
ntb->int_info[i].tag = NULL;
ntb->allocated_interrupts++;
rc = bus_setup_intr(ntb->device, ntb->int_info[i].res,
INTR_MPSAFE | INTR_TYPE_MISC, NULL, ndev_vec_isr,
&ntb->msix_vec[i], &ntb->int_info[i].tag);
if (rc != 0) {
device_printf(ntb->device, "bus_setup_intr failed\n");
return (ENXIO);
}
}
return (0);
}
/*
* The Linux NTB driver drops from MSI-X to legacy INTx if a unique vector
* cannot be allocated for each MSI-X message. JHB seems to think remapping
* should be okay. This tunable should enable us to test that hypothesis
* when someone gets their hands on some Xeon hardware.
*/
static int ntb_force_remap_mode;
SYSCTL_INT(_hw_ntb, OID_AUTO, force_remap_mode, CTLFLAG_RDTUN,
&ntb_force_remap_mode, 0, "If enabled, force MSI-X messages to be remapped"
" to a smaller number of ithreads, even if the desired number are "
"available");
/*
* In case it is NOT ok, give consumers an abort button.
*/
static int ntb_prefer_intx;
SYSCTL_INT(_hw_ntb, OID_AUTO, prefer_intx_to_remap, CTLFLAG_RDTUN,
&ntb_prefer_intx, 0, "If enabled, prefer to use legacy INTx mode rather "
"than remapping MSI-X messages over available slots (match Linux driver "
"behavior)");
/*
* Remap the desired number of MSI-X messages to available ithreads in a simple
* round-robin fashion.
*/
static int
intel_ntb_remap_msix(device_t dev, uint32_t desired, uint32_t avail)
{
u_int *vectors;
uint32_t i;
int rc;
if (ntb_prefer_intx != 0)
return (ENXIO);
vectors = malloc(desired * sizeof(*vectors), M_NTB, M_ZERO | M_WAITOK);
for (i = 0; i < desired; i++)
vectors[i] = (i % avail) + 1;
rc = pci_remap_msix(dev, desired, vectors);
free(vectors, M_NTB);
return (rc);
}
static int
intel_ntb_init_isr(struct ntb_softc *ntb)
{
uint32_t desired_vectors, num_vectors;
int rc;
ntb->allocated_interrupts = 0;
ntb->last_ts = ticks;
/*
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
* Mask all doorbell interrupts. (Except link events!)
*/
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
DB_MASK_LOCK(ntb);
ntb->db_mask = ntb->db_valid_mask;
db_iowrite(ntb, ntb->self_reg->db_mask, ntb->db_mask);
DB_MASK_UNLOCK(ntb);
num_vectors = desired_vectors = MIN(pci_msix_count(ntb->device),
ntb->db_count);
if (desired_vectors >= 1) {
rc = pci_alloc_msix(ntb->device, &num_vectors);
if (ntb_force_remap_mode != 0 && rc == 0 &&
num_vectors == desired_vectors)
num_vectors--;
if (rc == 0 && num_vectors < desired_vectors) {
rc = intel_ntb_remap_msix(ntb->device, desired_vectors,
num_vectors);
if (rc == 0)
num_vectors = desired_vectors;
else
pci_release_msi(ntb->device);
}
if (rc != 0)
num_vectors = 1;
} else
num_vectors = 1;
if (ntb->type == NTB_XEON && num_vectors < ntb->db_vec_count) {
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP)) {
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
device_printf(ntb->device,
"Errata workaround does not support MSI or INTX\n");
return (EINVAL);
}
ntb->db_vec_count = 1;
ntb->db_vec_shift = XEON_DB_TOTAL_SHIFT;
rc = intel_ntb_setup_legacy_interrupt(ntb);
} else {
if (num_vectors - 1 != XEON_NONLINK_DB_MSIX_BITS &&
HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP)) {
device_printf(ntb->device,
"Errata workaround expects %d doorbell bits\n",
XEON_NONLINK_DB_MSIX_BITS);
return (EINVAL);
}
intel_ntb_create_msix_vec(ntb, num_vectors);
rc = intel_ntb_setup_msix(ntb, num_vectors);
}
if (rc != 0) {
device_printf(ntb->device,
"Error allocating interrupts: %d\n", rc);
intel_ntb_free_msix_vec(ntb);
}
return (rc);
}
static int
intel_ntb_setup_legacy_interrupt(struct ntb_softc *ntb)
{
int rc;
ntb->int_info[0].rid = 0;
ntb->int_info[0].res = bus_alloc_resource_any(ntb->device, SYS_RES_IRQ,
&ntb->int_info[0].rid, RF_SHAREABLE|RF_ACTIVE);
if (ntb->int_info[0].res == NULL) {
device_printf(ntb->device, "bus_alloc_resource failed\n");
return (ENOMEM);
}
ntb->int_info[0].tag = NULL;
ntb->allocated_interrupts = 1;
rc = bus_setup_intr(ntb->device, ntb->int_info[0].res,
INTR_MPSAFE | INTR_TYPE_MISC, NULL, ndev_irq_isr,
ntb, &ntb->int_info[0].tag);
if (rc != 0) {
device_printf(ntb->device, "bus_setup_intr failed\n");
return (ENXIO);
}
return (0);
}
static void
intel_ntb_teardown_interrupts(struct ntb_softc *ntb)
{
struct ntb_int_info *current_int;
int i;
2015-10-13 03:12:55 +00:00
for (i = 0; i < ntb->allocated_interrupts; i++) {
current_int = &ntb->int_info[i];
if (current_int->tag != NULL)
bus_teardown_intr(ntb->device, current_int->res,
current_int->tag);
if (current_int->res != NULL)
bus_release_resource(ntb->device, SYS_RES_IRQ,
rman_get_rid(current_int->res), current_int->res);
}
intel_ntb_free_msix_vec(ntb);
pci_release_msi(ntb->device);
}
/*
* Doorbell register and mask are 64-bit on Atom, 16-bit on Xeon. Abstract it
* out to make code clearer.
*/
static inline uint64_t
db_ioread(struct ntb_softc *ntb, uint64_t regoff)
{
if (ntb->type == NTB_ATOM)
return (intel_ntb_reg_read(8, regoff));
KASSERT(ntb->type == NTB_XEON, ("bad ntb type"));
return (intel_ntb_reg_read(2, regoff));
}
static inline void
db_iowrite(struct ntb_softc *ntb, uint64_t regoff, uint64_t val)
{
KASSERT((val & ~ntb->db_valid_mask) == 0,
("%s: Invalid bits 0x%jx (valid: 0x%jx)", __func__,
(uintmax_t)(val & ~ntb->db_valid_mask),
(uintmax_t)ntb->db_valid_mask));
if (regoff == ntb->self_reg->db_mask)
DB_MASK_ASSERT(ntb, MA_OWNED);
db_iowrite_raw(ntb, regoff, val);
}
static inline void
db_iowrite_raw(struct ntb_softc *ntb, uint64_t regoff, uint64_t val)
{
if (ntb->type == NTB_ATOM) {
intel_ntb_reg_write(8, regoff, val);
return;
}
KASSERT(ntb->type == NTB_XEON, ("bad ntb type"));
intel_ntb_reg_write(2, regoff, (uint16_t)val);
}
static void
intel_ntb_db_set_mask(device_t dev, uint64_t bits)
{
struct ntb_softc *ntb = device_get_softc(dev);
DB_MASK_LOCK(ntb);
ntb->db_mask |= bits;
if (!HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP))
db_iowrite(ntb, ntb->self_reg->db_mask, ntb->db_mask);
DB_MASK_UNLOCK(ntb);
}
static void
intel_ntb_db_clear_mask(device_t dev, uint64_t bits)
{
struct ntb_softc *ntb = device_get_softc(dev);
uint64_t ibits;
int i;
KASSERT((bits & ~ntb->db_valid_mask) == 0,
("%s: Invalid bits 0x%jx (valid: 0x%jx)", __func__,
(uintmax_t)(bits & ~ntb->db_valid_mask),
(uintmax_t)ntb->db_valid_mask));
DB_MASK_LOCK(ntb);
ibits = ntb->fake_db & ntb->db_mask & bits;
ntb->db_mask &= ~bits;
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP)) {
/* Simulate fake interrupts if unmasked DB bits are set. */
ntb->force_db |= ibits;
for (i = 0; i < XEON_NONLINK_DB_MSIX_BITS; i++) {
if ((ibits & intel_ntb_db_vector_mask(dev, i)) != 0)
swi_sched(ntb->int_info[i].tag, 0);
}
} else {
db_iowrite(ntb, ntb->self_reg->db_mask, ntb->db_mask);
}
DB_MASK_UNLOCK(ntb);
}
static uint64_t
intel_ntb_db_read(device_t dev)
{
struct ntb_softc *ntb = device_get_softc(dev);
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP))
return (ntb->fake_db);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
return (db_ioread(ntb, ntb->self_reg->db_bell));
}
static void
intel_ntb_db_clear(device_t dev, uint64_t bits)
{
struct ntb_softc *ntb = device_get_softc(dev);
KASSERT((bits & ~ntb->db_valid_mask) == 0,
("%s: Invalid bits 0x%jx (valid: 0x%jx)", __func__,
(uintmax_t)(bits & ~ntb->db_valid_mask),
(uintmax_t)ntb->db_valid_mask));
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP)) {
DB_MASK_LOCK(ntb);
ntb->fake_db &= ~bits;
DB_MASK_UNLOCK(ntb);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
return;
}
db_iowrite(ntb, ntb->self_reg->db_bell, bits);
}
static inline uint64_t
intel_ntb_vec_mask(struct ntb_softc *ntb, uint64_t db_vector)
{
uint64_t shift, mask;
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP)) {
/*
* Remap vectors in custom way to make at least first
* three doorbells to not generate stray events.
* This breaks Linux compatibility (if one existed)
* when more then one DB is used (not by if_ntb).
*/
if (db_vector < XEON_NONLINK_DB_MSIX_BITS - 1)
return (1 << db_vector);
if (db_vector == XEON_NONLINK_DB_MSIX_BITS - 1)
return (0x7ffc);
}
shift = ntb->db_vec_shift;
mask = (1ull << shift) - 1;
return (mask << (shift * db_vector));
}
static void
intel_ntb_interrupt(struct ntb_softc *ntb, uint32_t vec)
{
uint64_t vec_mask;
ntb->last_ts = ticks;
vec_mask = intel_ntb_vec_mask(ntb, vec);
if ((vec_mask & ntb->db_link_mask) != 0) {
if (intel_ntb_poll_link(ntb))
ntb_link_event(ntb->device);
}
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP) &&
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
(vec_mask & ntb->db_link_mask) == 0) {
DB_MASK_LOCK(ntb);
/*
* Do not report same DB events again if not cleared yet,
* unless the mask was just cleared for them and this
* interrupt handler call can be the consequence of it.
*/
vec_mask &= ~ntb->fake_db | ntb->force_db;
ntb->force_db &= ~vec_mask;
/* Update our internal doorbell register. */
ntb->fake_db |= vec_mask;
/* Do not report masked DB events. */
vec_mask &= ~ntb->db_mask;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
DB_MASK_UNLOCK(ntb);
}
if ((vec_mask & ntb->db_valid_mask) != 0)
ntb_db_event(ntb->device, vec);
}
static void
ndev_vec_isr(void *arg)
{
struct ntb_vec *nvec = arg;
intel_ntb_interrupt(nvec->ntb, nvec->num);
}
static void
ndev_irq_isr(void *arg)
{
/* If we couldn't set up MSI-X, we only have the one vector. */
intel_ntb_interrupt(arg, 0);
}
static int
intel_ntb_create_msix_vec(struct ntb_softc *ntb, uint32_t num_vectors)
{
uint32_t i;
ntb->msix_vec = malloc(num_vectors * sizeof(*ntb->msix_vec), M_NTB,
M_ZERO | M_WAITOK);
for (i = 0; i < num_vectors; i++) {
ntb->msix_vec[i].num = i;
ntb->msix_vec[i].ntb = ntb;
}
return (0);
}
static void
intel_ntb_free_msix_vec(struct ntb_softc *ntb)
{
if (ntb->msix_vec == NULL)
return;
free(ntb->msix_vec, M_NTB);
ntb->msix_vec = NULL;
}
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
static void
intel_ntb_get_msix_info(struct ntb_softc *ntb)
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
{
struct pci_devinfo *dinfo;
struct pcicfg_msix *msix;
uint32_t laddr, data, i, offset;
dinfo = device_get_ivars(ntb->device);
msix = &dinfo->cfg.msix;
CTASSERT(XEON_NONLINK_DB_MSIX_BITS == nitems(ntb->msix_data));
for (i = 0; i < XEON_NONLINK_DB_MSIX_BITS; i++) {
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
offset = msix->msix_table_offset + i * PCI_MSIX_ENTRY_SIZE;
laddr = bus_read_4(msix->msix_table_res, offset +
PCI_MSIX_ENTRY_LOWER_ADDR);
intel_ntb_printf(2, "local MSIX addr(%u): 0x%x\n", i, laddr);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
KASSERT((laddr & MSI_INTEL_ADDR_BASE) == MSI_INTEL_ADDR_BASE,
("local MSIX addr 0x%x not in MSI base 0x%x", laddr,
MSI_INTEL_ADDR_BASE));
ntb->msix_data[i].nmd_ofs = laddr;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
data = bus_read_4(msix->msix_table_res, offset +
PCI_MSIX_ENTRY_DATA);
intel_ntb_printf(2, "local MSIX data(%u): 0x%x\n", i, data);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
ntb->msix_data[i].nmd_data = data;
}
}
static struct ntb_hw_info *
intel_ntb_get_device_info(uint32_t device_id)
{
struct ntb_hw_info *ep;
for (ep = pci_ids; ep < &pci_ids[nitems(pci_ids)]; ep++) {
if (ep->device_id == device_id)
return (ep);
}
return (NULL);
}
static void
intel_ntb_teardown_xeon(struct ntb_softc *ntb)
{
if (ntb->reg != NULL)
intel_ntb_link_disable(ntb->device);
}
static void
intel_ntb_detect_max_mw(struct ntb_softc *ntb)
{
if (ntb->type == NTB_ATOM) {
ntb->mw_count = ATOM_MW_COUNT;
return;
}
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR))
ntb->mw_count = XEON_HSX_SPLIT_MW_COUNT;
else
ntb->mw_count = XEON_SNB_MW_COUNT;
}
static int
intel_ntb_detect_xeon(struct ntb_softc *ntb)
{
uint8_t ppd, conn_type;
ppd = pci_read_config(ntb->device, NTB_PPD_OFFSET, 1);
ntb->ppd = ppd;
if ((ppd & XEON_PPD_DEV_TYPE) != 0)
ntb->dev_type = NTB_DEV_DSD;
else
ntb->dev_type = NTB_DEV_USD;
if ((ppd & XEON_PPD_SPLIT_BAR) != 0)
ntb->features |= NTB_SPLIT_BAR;
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP) &&
!HAS_FEATURE(ntb, NTB_SPLIT_BAR)) {
device_printf(ntb->device,
"Can not apply SB01BASE_LOCKUP workaround "
"with split BARs disabled!\n");
device_printf(ntb->device,
"Expect system hangs under heavy NTB traffic!\n");
ntb->features &= ~NTB_SB01BASE_LOCKUP;
}
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
/*
* SDOORBELL errata workaround gets in the way of SB01BASE_LOCKUP
* errata workaround; only do one at a time.
*/
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP))
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
ntb->features &= ~NTB_SDOORBELL_LOCKUP;
conn_type = ppd & XEON_PPD_CONN_TYPE;
switch (conn_type) {
case NTB_CONN_B2B:
ntb->conn_type = conn_type;
break;
case NTB_CONN_RP:
case NTB_CONN_TRANSPARENT:
default:
device_printf(ntb->device, "Unsupported connection type: %u\n",
(unsigned)conn_type);
return (ENXIO);
}
return (0);
}
static int
intel_ntb_detect_atom(struct ntb_softc *ntb)
{
uint32_t ppd, conn_type;
ppd = pci_read_config(ntb->device, NTB_PPD_OFFSET, 4);
ntb->ppd = ppd;
if ((ppd & ATOM_PPD_DEV_TYPE) != 0)
ntb->dev_type = NTB_DEV_DSD;
else
ntb->dev_type = NTB_DEV_USD;
conn_type = (ppd & ATOM_PPD_CONN_TYPE) >> 8;
switch (conn_type) {
case NTB_CONN_B2B:
ntb->conn_type = conn_type;
break;
default:
device_printf(ntb->device, "Unsupported NTB configuration\n");
return (ENXIO);
}
return (0);
}
static int
intel_ntb_xeon_init_dev(struct ntb_softc *ntb)
{
int rc;
ntb->spad_count = XEON_SPAD_COUNT;
ntb->db_count = XEON_DB_COUNT;
ntb->db_link_mask = XEON_DB_LINK_BIT;
ntb->db_vec_count = XEON_DB_MSIX_VECTOR_COUNT;
ntb->db_vec_shift = XEON_DB_MSIX_VECTOR_SHIFT;
if (ntb->conn_type != NTB_CONN_B2B) {
device_printf(ntb->device, "Connection type %d not supported\n",
ntb->conn_type);
return (ENXIO);
}
ntb->reg = &xeon_reg;
ntb->self_reg = &xeon_pri_reg;
ntb->peer_reg = &xeon_b2b_reg;
ntb->xlat_reg = &xeon_sec_xlat;
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP)) {
ntb->force_db = ntb->fake_db = 0;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
ntb->msix_mw_idx = (ntb->mw_count + g_ntb_msix_idx) %
ntb->mw_count;
intel_ntb_printf(2, "Setting up MSIX mw idx %d means %u\n",
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
g_ntb_msix_idx, ntb->msix_mw_idx);
rc = intel_ntb_mw_set_wc_internal(ntb, ntb->msix_mw_idx,
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
VM_MEMATTR_UNCACHEABLE);
KASSERT(rc == 0, ("shouldn't fail"));
} else if (HAS_FEATURE(ntb, NTB_SDOORBELL_LOCKUP)) {
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
/*
* There is a Xeon hardware errata related to writes to SDOORBELL or
* B2BDOORBELL in conjunction with inbound access to NTB MMIO space,
* which may hang the system. To workaround this, use a memory
* window to access the interrupt and scratch pad registers on the
* remote system.
*/
ntb->b2b_mw_idx = (ntb->mw_count + g_ntb_mw_idx) %
ntb->mw_count;
intel_ntb_printf(2, "Setting up b2b mw idx %d means %u\n",
g_ntb_mw_idx, ntb->b2b_mw_idx);
rc = intel_ntb_mw_set_wc_internal(ntb, ntb->b2b_mw_idx,
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
VM_MEMATTR_UNCACHEABLE);
KASSERT(rc == 0, ("shouldn't fail"));
} else if (HAS_FEATURE(ntb, NTB_B2BDOORBELL_BIT14))
/*
* HW Errata on bit 14 of b2bdoorbell register. Writes will not be
* mirrored to the remote system. Shrink the number of bits by one,
* since bit 14 is the last bit.
*
* On REGS_THRU_MW errata mode, we don't use the b2bdoorbell register
* anyway. Nor for non-B2B connection types.
*/
ntb->db_count = XEON_DB_COUNT - 1;
ntb->db_valid_mask = (1ull << ntb->db_count) - 1;
if (ntb->dev_type == NTB_DEV_USD)
rc = xeon_setup_b2b_mw(ntb, &xeon_b2b_dsd_addr,
&xeon_b2b_usd_addr);
else
rc = xeon_setup_b2b_mw(ntb, &xeon_b2b_usd_addr,
&xeon_b2b_dsd_addr);
if (rc != 0)
return (rc);
2015-10-13 03:12:55 +00:00
/* Enable Bus Master and Memory Space on the secondary side */
intel_ntb_reg_write(2, XEON_SPCICMD_OFFSET,
PCIM_CMD_MEMEN | PCIM_CMD_BUSMASTEREN);
/*
* Mask all doorbell interrupts.
*/
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
DB_MASK_LOCK(ntb);
ntb->db_mask = ntb->db_valid_mask;
db_iowrite(ntb, ntb->self_reg->db_mask, ntb->db_mask);
DB_MASK_UNLOCK(ntb);
rc = intel_ntb_init_isr(ntb);
return (rc);
}
static int
intel_ntb_atom_init_dev(struct ntb_softc *ntb)
{
int error;
KASSERT(ntb->conn_type == NTB_CONN_B2B,
("Unsupported NTB configuration (%d)\n", ntb->conn_type));
ntb->spad_count = ATOM_SPAD_COUNT;
ntb->db_count = ATOM_DB_COUNT;
ntb->db_vec_count = ATOM_DB_MSIX_VECTOR_COUNT;
ntb->db_vec_shift = ATOM_DB_MSIX_VECTOR_SHIFT;
ntb->db_valid_mask = (1ull << ntb->db_count) - 1;
ntb->reg = &atom_reg;
ntb->self_reg = &atom_pri_reg;
ntb->peer_reg = &atom_b2b_reg;
ntb->xlat_reg = &atom_sec_xlat;
/*
* FIXME - MSI-X bug on early Atom HW, remove once internal issue is
* resolved. Mask transaction layer internal parity errors.
*/
pci_write_config(ntb->device, 0xFC, 0x4, 4);
configure_atom_secondary_side_bars(ntb);
/* Enable Bus Master and Memory Space on the secondary side */
intel_ntb_reg_write(2, ATOM_SPCICMD_OFFSET,
PCIM_CMD_MEMEN | PCIM_CMD_BUSMASTEREN);
2015-10-13 03:12:55 +00:00
error = intel_ntb_init_isr(ntb);
if (error != 0)
return (error);
/* Initiate PCI-E link training */
intel_ntb_link_enable(ntb->device, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
callout_reset(&ntb->heartbeat_timer, 0, atom_link_hb, ntb);
return (0);
}
/* XXX: Linux driver doesn't seem to do any of this for Atom. */
static void
configure_atom_secondary_side_bars(struct ntb_softc *ntb)
{
if (ntb->dev_type == NTB_DEV_USD) {
intel_ntb_reg_write(8, ATOM_PBAR2XLAT_OFFSET,
XEON_B2B_BAR2_ADDR64);
intel_ntb_reg_write(8, ATOM_PBAR4XLAT_OFFSET,
XEON_B2B_BAR4_ADDR64);
intel_ntb_reg_write(8, ATOM_MBAR23_OFFSET, XEON_B2B_BAR2_ADDR64);
intel_ntb_reg_write(8, ATOM_MBAR45_OFFSET, XEON_B2B_BAR4_ADDR64);
} else {
intel_ntb_reg_write(8, ATOM_PBAR2XLAT_OFFSET,
XEON_B2B_BAR2_ADDR64);
intel_ntb_reg_write(8, ATOM_PBAR4XLAT_OFFSET,
XEON_B2B_BAR4_ADDR64);
intel_ntb_reg_write(8, ATOM_MBAR23_OFFSET, XEON_B2B_BAR2_ADDR64);
intel_ntb_reg_write(8, ATOM_MBAR45_OFFSET, XEON_B2B_BAR4_ADDR64);
}
}
/*
* When working around Xeon SDOORBELL errata by remapping remote registers in a
* MW, limit the B2B MW to half a MW. By sharing a MW, half the shared MW
* remains for use by a higher layer.
*
* Will only be used if working around SDOORBELL errata and the BIOS-configured
* MW size is sufficiently large.
*/
static unsigned int ntb_b2b_mw_share;
SYSCTL_UINT(_hw_ntb, OID_AUTO, b2b_mw_share, CTLFLAG_RDTUN, &ntb_b2b_mw_share,
0, "If enabled (non-zero), prefer to share half of the B2B peer register "
"MW with higher level consumers. Both sides of the NTB MUST set the same "
"value here.");
static void
xeon_reset_sbar_size(struct ntb_softc *ntb, enum ntb_bar idx,
enum ntb_bar regbar)
{
struct ntb_pci_bar_info *bar;
uint8_t bar_sz;
if (!HAS_FEATURE(ntb, NTB_SPLIT_BAR) && idx >= NTB_B2B_BAR_3)
return;
bar = &ntb->bar_info[idx];
bar_sz = pci_read_config(ntb->device, bar->psz_off, 1);
if (idx == regbar) {
if (ntb->b2b_off != 0)
bar_sz--;
else
bar_sz = 0;
}
pci_write_config(ntb->device, bar->ssz_off, bar_sz, 1);
bar_sz = pci_read_config(ntb->device, bar->ssz_off, 1);
(void)bar_sz;
}
static void
xeon_set_sbar_base_and_limit(struct ntb_softc *ntb, uint64_t bar_addr,
enum ntb_bar idx, enum ntb_bar regbar)
{
uint64_t reg_val;
uint32_t base_reg, lmt_reg;
bar_get_xlat_params(ntb, idx, &base_reg, NULL, &lmt_reg);
if (idx == regbar) {
if (ntb->b2b_off)
bar_addr += ntb->b2b_off;
else
bar_addr = 0;
}
if (!bar_is_64bit(ntb, idx)) {
intel_ntb_reg_write(4, base_reg, bar_addr);
reg_val = intel_ntb_reg_read(4, base_reg);
(void)reg_val;
intel_ntb_reg_write(4, lmt_reg, bar_addr);
reg_val = intel_ntb_reg_read(4, lmt_reg);
(void)reg_val;
} else {
intel_ntb_reg_write(8, base_reg, bar_addr);
reg_val = intel_ntb_reg_read(8, base_reg);
(void)reg_val;
intel_ntb_reg_write(8, lmt_reg, bar_addr);
reg_val = intel_ntb_reg_read(8, lmt_reg);
(void)reg_val;
}
}
static void
xeon_set_pbar_xlat(struct ntb_softc *ntb, uint64_t base_addr, enum ntb_bar idx)
{
struct ntb_pci_bar_info *bar;
bar = &ntb->bar_info[idx];
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR) && idx >= NTB_B2B_BAR_2) {
intel_ntb_reg_write(4, bar->pbarxlat_off, base_addr);
base_addr = intel_ntb_reg_read(4, bar->pbarxlat_off);
} else {
intel_ntb_reg_write(8, bar->pbarxlat_off, base_addr);
base_addr = intel_ntb_reg_read(8, bar->pbarxlat_off);
}
(void)base_addr;
}
static int
xeon_setup_b2b_mw(struct ntb_softc *ntb, const struct ntb_b2b_addr *addr,
const struct ntb_b2b_addr *peer_addr)
{
struct ntb_pci_bar_info *b2b_bar;
vm_size_t bar_size;
uint64_t bar_addr;
enum ntb_bar b2b_bar_num, i;
if (ntb->b2b_mw_idx == B2B_MW_DISABLED) {
b2b_bar = NULL;
b2b_bar_num = NTB_CONFIG_BAR;
ntb->b2b_off = 0;
} else {
b2b_bar_num = intel_ntb_mw_to_bar(ntb, ntb->b2b_mw_idx);
KASSERT(b2b_bar_num > 0 && b2b_bar_num < NTB_MAX_BARS,
("invalid b2b mw bar"));
b2b_bar = &ntb->bar_info[b2b_bar_num];
bar_size = b2b_bar->size;
if (ntb_b2b_mw_share != 0 &&
(bar_size >> 1) >= XEON_B2B_MIN_SIZE)
ntb->b2b_off = bar_size >> 1;
else if (bar_size >= XEON_B2B_MIN_SIZE) {
ntb->b2b_off = 0;
} else {
device_printf(ntb->device,
"B2B bar size is too small!\n");
return (EIO);
}
}
/*
* Reset the secondary bar sizes to match the primary bar sizes.
* (Except, disable or halve the size of the B2B secondary bar.)
*/
for (i = NTB_B2B_BAR_1; i < NTB_MAX_BARS; i++)
xeon_reset_sbar_size(ntb, i, b2b_bar_num);
bar_addr = 0;
if (b2b_bar_num == NTB_CONFIG_BAR)
bar_addr = addr->bar0_addr;
else if (b2b_bar_num == NTB_B2B_BAR_1)
bar_addr = addr->bar2_addr64;
else if (b2b_bar_num == NTB_B2B_BAR_2 && !HAS_FEATURE(ntb, NTB_SPLIT_BAR))
bar_addr = addr->bar4_addr64;
else if (b2b_bar_num == NTB_B2B_BAR_2)
bar_addr = addr->bar4_addr32;
else if (b2b_bar_num == NTB_B2B_BAR_3)
bar_addr = addr->bar5_addr32;
else
KASSERT(false, ("invalid bar"));
intel_ntb_reg_write(8, XEON_SBAR0BASE_OFFSET, bar_addr);
/*
* Other SBARs are normally hit by the PBAR xlat, except for the b2b
* register BAR. The B2B BAR is either disabled above or configured
* half-size. It starts at PBAR xlat + offset.
*
* Also set up incoming BAR limits == base (zero length window).
*/
xeon_set_sbar_base_and_limit(ntb, addr->bar2_addr64, NTB_B2B_BAR_1,
b2b_bar_num);
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR)) {
xeon_set_sbar_base_and_limit(ntb, addr->bar4_addr32,
NTB_B2B_BAR_2, b2b_bar_num);
xeon_set_sbar_base_and_limit(ntb, addr->bar5_addr32,
NTB_B2B_BAR_3, b2b_bar_num);
} else
xeon_set_sbar_base_and_limit(ntb, addr->bar4_addr64,
NTB_B2B_BAR_2, b2b_bar_num);
/* Zero incoming translation addrs */
intel_ntb_reg_write(8, XEON_SBAR2XLAT_OFFSET, 0);
intel_ntb_reg_write(8, XEON_SBAR4XLAT_OFFSET, 0);
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP)) {
uint32_t xlat_reg, lmt_reg;
enum ntb_bar bar_num;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
/*
* We point the chosen MSIX MW BAR xlat to remote LAPIC for
* workaround
*/
bar_num = intel_ntb_mw_to_bar(ntb, ntb->msix_mw_idx);
bar_get_xlat_params(ntb, bar_num, NULL, &xlat_reg, &lmt_reg);
if (bar_is_64bit(ntb, bar_num)) {
intel_ntb_reg_write(8, xlat_reg, MSI_INTEL_ADDR_BASE);
ntb->msix_xlat = intel_ntb_reg_read(8, xlat_reg);
intel_ntb_reg_write(8, lmt_reg, 0);
} else {
intel_ntb_reg_write(4, xlat_reg, MSI_INTEL_ADDR_BASE);
ntb->msix_xlat = intel_ntb_reg_read(4, xlat_reg);
intel_ntb_reg_write(4, lmt_reg, 0);
}
ntb->peer_lapic_bar = &ntb->bar_info[bar_num];
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
}
(void)intel_ntb_reg_read(8, XEON_SBAR2XLAT_OFFSET);
(void)intel_ntb_reg_read(8, XEON_SBAR4XLAT_OFFSET);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
/* Zero outgoing translation limits (whole bar size windows) */
intel_ntb_reg_write(8, XEON_PBAR2LMT_OFFSET, 0);
intel_ntb_reg_write(8, XEON_PBAR4LMT_OFFSET, 0);
/* Set outgoing translation offsets */
xeon_set_pbar_xlat(ntb, peer_addr->bar2_addr64, NTB_B2B_BAR_1);
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR)) {
xeon_set_pbar_xlat(ntb, peer_addr->bar4_addr32, NTB_B2B_BAR_2);
xeon_set_pbar_xlat(ntb, peer_addr->bar5_addr32, NTB_B2B_BAR_3);
} else
xeon_set_pbar_xlat(ntb, peer_addr->bar4_addr64, NTB_B2B_BAR_2);
/* Set the translation offset for B2B registers */
bar_addr = 0;
if (b2b_bar_num == NTB_CONFIG_BAR)
bar_addr = peer_addr->bar0_addr;
else if (b2b_bar_num == NTB_B2B_BAR_1)
bar_addr = peer_addr->bar2_addr64;
else if (b2b_bar_num == NTB_B2B_BAR_2 && !HAS_FEATURE(ntb, NTB_SPLIT_BAR))
bar_addr = peer_addr->bar4_addr64;
else if (b2b_bar_num == NTB_B2B_BAR_2)
bar_addr = peer_addr->bar4_addr32;
else if (b2b_bar_num == NTB_B2B_BAR_3)
bar_addr = peer_addr->bar5_addr32;
else
KASSERT(false, ("invalid bar"));
/*
* B2B_XLAT_OFFSET is a 64-bit register but can only be written 32 bits
* at a time.
*/
intel_ntb_reg_write(4, XEON_B2B_XLAT_OFFSETL, bar_addr & 0xffffffff);
intel_ntb_reg_write(4, XEON_B2B_XLAT_OFFSETU, bar_addr >> 32);
return (0);
}
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
static inline bool
_xeon_link_is_up(struct ntb_softc *ntb)
{
if (ntb->conn_type == NTB_CONN_TRANSPARENT)
return (true);
return ((ntb->lnk_sta & NTB_LINK_STATUS_ACTIVE) != 0);
}
static inline bool
link_is_up(struct ntb_softc *ntb)
{
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
if (ntb->type == NTB_XEON)
return (_xeon_link_is_up(ntb) && (ntb->peer_msix_good ||
!HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP)));
KASSERT(ntb->type == NTB_ATOM, ("ntb type"));
return ((ntb->ntb_ctl & ATOM_CNTL_LINK_DOWN) == 0);
}
static inline bool
atom_link_is_err(struct ntb_softc *ntb)
{
uint32_t status;
KASSERT(ntb->type == NTB_ATOM, ("ntb type"));
status = intel_ntb_reg_read(4, ATOM_LTSSMSTATEJMP_OFFSET);
if ((status & ATOM_LTSSMSTATEJMP_FORCEDETECT) != 0)
return (true);
status = intel_ntb_reg_read(4, ATOM_IBSTERRRCRVSTS0_OFFSET);
return ((status & ATOM_IBIST_ERR_OFLOW) != 0);
}
/* Atom does not have link status interrupt, poll on that platform */
static void
atom_link_hb(void *arg)
{
struct ntb_softc *ntb = arg;
sbintime_t timo, poll_ts;
timo = NTB_HB_TIMEOUT * hz;
poll_ts = ntb->last_ts + timo;
/*
* Delay polling the link status if an interrupt was received, unless
* the cached link status says the link is down.
*/
if ((sbintime_t)ticks - poll_ts < 0 && link_is_up(ntb)) {
timo = poll_ts - ticks;
goto out;
}
if (intel_ntb_poll_link(ntb))
ntb_link_event(ntb->device);
if (!link_is_up(ntb) && atom_link_is_err(ntb)) {
/* Link is down with error, proceed with recovery */
callout_reset(&ntb->lr_timer, 0, recover_atom_link, ntb);
return;
}
out:
callout_reset(&ntb->heartbeat_timer, timo, atom_link_hb, ntb);
}
static void
atom_perform_link_restart(struct ntb_softc *ntb)
{
uint32_t status;
/* Driver resets the NTB ModPhy lanes - magic! */
intel_ntb_reg_write(1, ATOM_MODPHY_PCSREG6, 0xe0);
intel_ntb_reg_write(1, ATOM_MODPHY_PCSREG4, 0x40);
intel_ntb_reg_write(1, ATOM_MODPHY_PCSREG4, 0x60);
intel_ntb_reg_write(1, ATOM_MODPHY_PCSREG6, 0x60);
/* Driver waits 100ms to allow the NTB ModPhy to settle */
pause("ModPhy", hz / 10);
/* Clear AER Errors, write to clear */
status = intel_ntb_reg_read(4, ATOM_ERRCORSTS_OFFSET);
status &= PCIM_AER_COR_REPLAY_ROLLOVER;
intel_ntb_reg_write(4, ATOM_ERRCORSTS_OFFSET, status);
/* Clear unexpected electrical idle event in LTSSM, write to clear */
status = intel_ntb_reg_read(4, ATOM_LTSSMERRSTS0_OFFSET);
status |= ATOM_LTSSMERRSTS0_UNEXPECTEDEI;
intel_ntb_reg_write(4, ATOM_LTSSMERRSTS0_OFFSET, status);
/* Clear DeSkew Buffer error, write to clear */
status = intel_ntb_reg_read(4, ATOM_DESKEWSTS_OFFSET);
status |= ATOM_DESKEWSTS_DBERR;
intel_ntb_reg_write(4, ATOM_DESKEWSTS_OFFSET, status);
status = intel_ntb_reg_read(4, ATOM_IBSTERRRCRVSTS0_OFFSET);
status &= ATOM_IBIST_ERR_OFLOW;
intel_ntb_reg_write(4, ATOM_IBSTERRRCRVSTS0_OFFSET, status);
/* Releases the NTB state machine to allow the link to retrain */
status = intel_ntb_reg_read(4, ATOM_LTSSMSTATEJMP_OFFSET);
status &= ~ATOM_LTSSMSTATEJMP_FORCEDETECT;
intel_ntb_reg_write(4, ATOM_LTSSMSTATEJMP_OFFSET, status);
}
static int
intel_ntb_port_number(device_t dev)
{
struct ntb_softc *ntb = device_get_softc(dev);
return (ntb->dev_type == NTB_DEV_USD ? 0 : 1);
}
static int
intel_ntb_peer_port_count(device_t dev)
{
return (1);
}
static int
intel_ntb_peer_port_number(device_t dev, int pidx)
{
struct ntb_softc *ntb = device_get_softc(dev);
if (pidx != 0)
return (-EINVAL);
return (ntb->dev_type == NTB_DEV_USD ? 1 : 0);
}
static int
intel_ntb_peer_port_idx(device_t dev, int port)
{
int peer_port;
peer_port = intel_ntb_peer_port_number(dev, 0);
if (peer_port == -EINVAL || port != peer_port)
return (-EINVAL);
return (0);
}
static int
intel_ntb_link_enable(device_t dev, enum ntb_speed speed __unused,
enum ntb_width width __unused)
{
struct ntb_softc *ntb = device_get_softc(dev);
uint32_t cntl;
intel_ntb_printf(2, "%s\n", __func__);
if (ntb->type == NTB_ATOM) {
pci_write_config(ntb->device, NTB_PPD_OFFSET,
ntb->ppd | ATOM_PPD_INIT_LINK, 4);
return (0);
}
if (ntb->conn_type == NTB_CONN_TRANSPARENT) {
ntb_link_event(dev);
return (0);
}
cntl = intel_ntb_reg_read(4, ntb->reg->ntb_ctl);
cntl &= ~(NTB_CNTL_LINK_DISABLE | NTB_CNTL_CFG_LOCK);
cntl |= NTB_CNTL_P2S_BAR23_SNOOP | NTB_CNTL_S2P_BAR23_SNOOP;
cntl |= NTB_CNTL_P2S_BAR4_SNOOP | NTB_CNTL_S2P_BAR4_SNOOP;
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR))
cntl |= NTB_CNTL_P2S_BAR5_SNOOP | NTB_CNTL_S2P_BAR5_SNOOP;
intel_ntb_reg_write(4, ntb->reg->ntb_ctl, cntl);
return (0);
}
static int
intel_ntb_link_disable(device_t dev)
{
struct ntb_softc *ntb = device_get_softc(dev);
uint32_t cntl;
intel_ntb_printf(2, "%s\n", __func__);
if (ntb->conn_type == NTB_CONN_TRANSPARENT) {
ntb_link_event(dev);
return (0);
}
cntl = intel_ntb_reg_read(4, ntb->reg->ntb_ctl);
cntl &= ~(NTB_CNTL_P2S_BAR23_SNOOP | NTB_CNTL_S2P_BAR23_SNOOP);
cntl &= ~(NTB_CNTL_P2S_BAR4_SNOOP | NTB_CNTL_S2P_BAR4_SNOOP);
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR))
cntl &= ~(NTB_CNTL_P2S_BAR5_SNOOP | NTB_CNTL_S2P_BAR5_SNOOP);
cntl |= NTB_CNTL_LINK_DISABLE | NTB_CNTL_CFG_LOCK;
intel_ntb_reg_write(4, ntb->reg->ntb_ctl, cntl);
return (0);
}
static bool
intel_ntb_link_enabled(device_t dev)
{
struct ntb_softc *ntb = device_get_softc(dev);
uint32_t cntl;
if (ntb->type == NTB_ATOM) {
cntl = pci_read_config(ntb->device, NTB_PPD_OFFSET, 4);
return ((cntl & ATOM_PPD_INIT_LINK) != 0);
}
if (ntb->conn_type == NTB_CONN_TRANSPARENT)
return (true);
cntl = intel_ntb_reg_read(4, ntb->reg->ntb_ctl);
return ((cntl & NTB_CNTL_LINK_DISABLE) == 0);
}
static void
recover_atom_link(void *arg)
{
struct ntb_softc *ntb = arg;
unsigned speed, width, oldspeed, oldwidth;
uint32_t status32;
atom_perform_link_restart(ntb);
/*
* There is a potential race between the 2 NTB devices recovering at
* the same time. If the times are the same, the link will not recover
* and the driver will be stuck in this loop forever. Add a random
* interval to the recovery time to prevent this race.
*/
status32 = arc4random() % ATOM_LINK_RECOVERY_TIME;
pause("Link", (ATOM_LINK_RECOVERY_TIME + status32) * hz / 1000);
if (atom_link_is_err(ntb))
goto retry;
status32 = intel_ntb_reg_read(4, ntb->reg->ntb_ctl);
if ((status32 & ATOM_CNTL_LINK_DOWN) != 0)
goto out;
status32 = intel_ntb_reg_read(4, ntb->reg->lnk_sta);
width = NTB_LNK_STA_WIDTH(status32);
speed = status32 & NTB_LINK_SPEED_MASK;
oldwidth = NTB_LNK_STA_WIDTH(ntb->lnk_sta);
oldspeed = ntb->lnk_sta & NTB_LINK_SPEED_MASK;
if (oldwidth != width || oldspeed != speed)
goto retry;
out:
callout_reset(&ntb->heartbeat_timer, NTB_HB_TIMEOUT * hz, atom_link_hb,
ntb);
return;
retry:
callout_reset(&ntb->lr_timer, NTB_HB_TIMEOUT * hz, recover_atom_link,
ntb);
}
/*
* Polls the HW link status register(s); returns true if something has changed.
*/
static bool
intel_ntb_poll_link(struct ntb_softc *ntb)
{
uint32_t ntb_cntl;
uint16_t reg_val;
if (ntb->type == NTB_ATOM) {
ntb_cntl = intel_ntb_reg_read(4, ntb->reg->ntb_ctl);
if (ntb_cntl == ntb->ntb_ctl)
return (false);
ntb->ntb_ctl = ntb_cntl;
ntb->lnk_sta = intel_ntb_reg_read(4, ntb->reg->lnk_sta);
} else {
db_iowrite_raw(ntb, ntb->self_reg->db_bell, ntb->db_link_mask);
reg_val = pci_read_config(ntb->device, ntb->reg->lnk_sta, 2);
if (reg_val == ntb->lnk_sta)
return (false);
ntb->lnk_sta = reg_val;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP)) {
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
if (_xeon_link_is_up(ntb)) {
if (!ntb->peer_msix_good) {
callout_reset(&ntb->peer_msix_work, 0,
intel_ntb_exchange_msix, ntb);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
return (false);
}
} else {
ntb->peer_msix_good = false;
ntb->peer_msix_done = false;
}
}
}
return (true);
}
static inline enum ntb_speed
intel_ntb_link_sta_speed(struct ntb_softc *ntb)
{
if (!link_is_up(ntb))
return (NTB_SPEED_NONE);
return (ntb->lnk_sta & NTB_LINK_SPEED_MASK);
}
static inline enum ntb_width
intel_ntb_link_sta_width(struct ntb_softc *ntb)
{
if (!link_is_up(ntb))
return (NTB_WIDTH_NONE);
return (NTB_LNK_STA_WIDTH(ntb->lnk_sta));
}
SYSCTL_NODE(_hw_ntb, OID_AUTO, debug_info, CTLFLAG_RW, 0,
"Driver state, statistics, and HW registers");
#define NTB_REGSZ_MASK (3ul << 30)
#define NTB_REG_64 (1ul << 30)
#define NTB_REG_32 (2ul << 30)
#define NTB_REG_16 (3ul << 30)
#define NTB_REG_8 (0ul << 30)
#define NTB_DB_READ (1ul << 29)
#define NTB_PCI_REG (1ul << 28)
#define NTB_REGFLAGS_MASK (NTB_REGSZ_MASK | NTB_DB_READ | NTB_PCI_REG)
static void
intel_ntb_sysctl_init(struct ntb_softc *ntb)
{
struct sysctl_oid_list *globals, *tree_par, *regpar, *statpar, *errpar;
struct sysctl_ctx_list *ctx;
struct sysctl_oid *tree, *tmptree;
ctx = device_get_sysctl_ctx(ntb->device);
globals = SYSCTL_CHILDREN(device_get_sysctl_tree(ntb->device));
SYSCTL_ADD_PROC(ctx, globals, OID_AUTO, "link_status",
CTLFLAG_RD | CTLTYPE_STRING, ntb, 0,
sysctl_handle_link_status_human, "A",
"Link status (human readable)");
SYSCTL_ADD_PROC(ctx, globals, OID_AUTO, "active",
CTLFLAG_RD | CTLTYPE_UINT, ntb, 0, sysctl_handle_link_status,
"IU", "Link status (1=active, 0=inactive)");
SYSCTL_ADD_PROC(ctx, globals, OID_AUTO, "admin_up",
CTLFLAG_RW | CTLTYPE_UINT, ntb, 0, sysctl_handle_link_admin,
"IU", "Set/get interface status (1=UP, 0=DOWN)");
tree = SYSCTL_ADD_NODE(ctx, globals, OID_AUTO, "debug_info",
CTLFLAG_RD, NULL, "Driver state, statistics, and HW registers");
tree_par = SYSCTL_CHILDREN(tree);
SYSCTL_ADD_UINT(ctx, tree_par, OID_AUTO, "conn_type", CTLFLAG_RD,
&ntb->conn_type, 0, "0 - Transparent; 1 - B2B; 2 - Root Port");
SYSCTL_ADD_UINT(ctx, tree_par, OID_AUTO, "dev_type", CTLFLAG_RD,
&ntb->dev_type, 0, "0 - USD; 1 - DSD");
SYSCTL_ADD_UINT(ctx, tree_par, OID_AUTO, "ppd", CTLFLAG_RD,
&ntb->ppd, 0, "Raw PPD register (cached)");
if (ntb->b2b_mw_idx != B2B_MW_DISABLED) {
SYSCTL_ADD_U8(ctx, tree_par, OID_AUTO, "b2b_idx", CTLFLAG_RD,
&ntb->b2b_mw_idx, 0,
"Index of the MW used for B2B remote register access");
SYSCTL_ADD_UQUAD(ctx, tree_par, OID_AUTO, "b2b_off",
CTLFLAG_RD, &ntb->b2b_off,
"If non-zero, offset of B2B register region in shared MW");
}
SYSCTL_ADD_PROC(ctx, tree_par, OID_AUTO, "features",
CTLFLAG_RD | CTLTYPE_STRING, ntb, 0, sysctl_handle_features, "A",
"Features/errata of this NTB device");
SYSCTL_ADD_UINT(ctx, tree_par, OID_AUTO, "ntb_ctl", CTLFLAG_RD,
__DEVOLATILE(uint32_t *, &ntb->ntb_ctl), 0,
"NTB CTL register (cached)");
SYSCTL_ADD_UINT(ctx, tree_par, OID_AUTO, "lnk_sta", CTLFLAG_RD,
__DEVOLATILE(uint32_t *, &ntb->lnk_sta), 0,
"LNK STA register (cached)");
SYSCTL_ADD_U8(ctx, tree_par, OID_AUTO, "mw_count", CTLFLAG_RD,
&ntb->mw_count, 0, "MW count");
SYSCTL_ADD_U8(ctx, tree_par, OID_AUTO, "spad_count", CTLFLAG_RD,
&ntb->spad_count, 0, "Scratchpad count");
SYSCTL_ADD_U8(ctx, tree_par, OID_AUTO, "db_count", CTLFLAG_RD,
&ntb->db_count, 0, "Doorbell count");
SYSCTL_ADD_U8(ctx, tree_par, OID_AUTO, "db_vec_count", CTLFLAG_RD,
&ntb->db_vec_count, 0, "Doorbell vector count");
SYSCTL_ADD_U8(ctx, tree_par, OID_AUTO, "db_vec_shift", CTLFLAG_RD,
&ntb->db_vec_shift, 0, "Doorbell vector shift");
SYSCTL_ADD_UQUAD(ctx, tree_par, OID_AUTO, "db_valid_mask", CTLFLAG_RD,
&ntb->db_valid_mask, "Doorbell valid mask");
SYSCTL_ADD_UQUAD(ctx, tree_par, OID_AUTO, "db_link_mask", CTLFLAG_RD,
&ntb->db_link_mask, "Doorbell link mask");
SYSCTL_ADD_UQUAD(ctx, tree_par, OID_AUTO, "db_mask", CTLFLAG_RD,
&ntb->db_mask, "Doorbell mask (cached)");
tmptree = SYSCTL_ADD_NODE(ctx, tree_par, OID_AUTO, "registers",
CTLFLAG_RD, NULL, "Raw HW registers (big-endian)");
regpar = SYSCTL_CHILDREN(tmptree);
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "ntbcntl",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb, NTB_REG_32 |
ntb->reg->ntb_ctl, sysctl_handle_register, "IU",
"NTB Control register");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "lnkcap",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb, NTB_REG_32 |
0x19c, sysctl_handle_register, "IU",
"NTB Link Capabilities");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "lnkcon",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb, NTB_REG_32 |
0x1a0, sysctl_handle_register, "IU",
"NTB Link Control register");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "db_mask",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | NTB_DB_READ | ntb->self_reg->db_mask,
sysctl_handle_register, "QU", "Doorbell mask register");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "db_bell",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | NTB_DB_READ | ntb->self_reg->db_bell,
sysctl_handle_register, "QU", "Doorbell register");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "incoming_xlat23",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | ntb->xlat_reg->bar2_xlat,
sysctl_handle_register, "QU", "Incoming XLAT23 register");
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR)) {
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "incoming_xlat4",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | ntb->xlat_reg->bar4_xlat,
sysctl_handle_register, "IU", "Incoming XLAT4 register");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "incoming_xlat5",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | ntb->xlat_reg->bar5_xlat,
sysctl_handle_register, "IU", "Incoming XLAT5 register");
} else {
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "incoming_xlat45",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | ntb->xlat_reg->bar4_xlat,
sysctl_handle_register, "QU", "Incoming XLAT45 register");
}
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "incoming_lmt23",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | ntb->xlat_reg->bar2_limit,
sysctl_handle_register, "QU", "Incoming LMT23 register");
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR)) {
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "incoming_lmt4",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | ntb->xlat_reg->bar4_limit,
sysctl_handle_register, "IU", "Incoming LMT4 register");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "incoming_lmt5",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | ntb->xlat_reg->bar5_limit,
sysctl_handle_register, "IU", "Incoming LMT5 register");
} else {
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "incoming_lmt45",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | ntb->xlat_reg->bar4_limit,
sysctl_handle_register, "QU", "Incoming LMT45 register");
}
if (ntb->type == NTB_ATOM)
return;
tmptree = SYSCTL_ADD_NODE(ctx, regpar, OID_AUTO, "xeon_stats",
CTLFLAG_RD, NULL, "Xeon HW statistics");
statpar = SYSCTL_CHILDREN(tmptree);
SYSCTL_ADD_PROC(ctx, statpar, OID_AUTO, "upstream_mem_miss",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_16 | XEON_USMEMMISS_OFFSET,
sysctl_handle_register, "SU", "Upstream Memory Miss");
tmptree = SYSCTL_ADD_NODE(ctx, regpar, OID_AUTO, "xeon_hw_err",
CTLFLAG_RD, NULL, "Xeon HW errors");
errpar = SYSCTL_CHILDREN(tmptree);
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "ppd",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_8 | NTB_PCI_REG | NTB_PPD_OFFSET,
sysctl_handle_register, "CU", "PPD");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "pbar23_sz",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_8 | NTB_PCI_REG | XEON_PBAR23SZ_OFFSET,
sysctl_handle_register, "CU", "PBAR23 SZ (log2)");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "pbar4_sz",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_8 | NTB_PCI_REG | XEON_PBAR4SZ_OFFSET,
sysctl_handle_register, "CU", "PBAR4 SZ (log2)");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "pbar5_sz",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_8 | NTB_PCI_REG | XEON_PBAR5SZ_OFFSET,
sysctl_handle_register, "CU", "PBAR5 SZ (log2)");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "sbar23_sz",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_8 | NTB_PCI_REG | XEON_SBAR23SZ_OFFSET,
sysctl_handle_register, "CU", "SBAR23 SZ (log2)");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "sbar4_sz",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_8 | NTB_PCI_REG | XEON_SBAR4SZ_OFFSET,
sysctl_handle_register, "CU", "SBAR4 SZ (log2)");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "sbar5_sz",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_8 | NTB_PCI_REG | XEON_SBAR5SZ_OFFSET,
sysctl_handle_register, "CU", "SBAR5 SZ (log2)");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "devsts",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_16 | NTB_PCI_REG | XEON_DEVSTS_OFFSET,
sysctl_handle_register, "SU", "DEVSTS");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "lnksts",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_16 | NTB_PCI_REG | XEON_LINK_STATUS_OFFSET,
sysctl_handle_register, "SU", "LNKSTS");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "slnksts",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_16 | NTB_PCI_REG | XEON_SLINK_STATUS_OFFSET,
sysctl_handle_register, "SU", "SLNKSTS");
SYSCTL_ADD_PROC(ctx, errpar, OID_AUTO, "uncerrsts",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | NTB_PCI_REG | XEON_UNCERRSTS_OFFSET,
sysctl_handle_register, "IU", "UNCERRSTS");
SYSCTL_ADD_PROC(ctx, errpar, OID_AUTO, "corerrsts",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | NTB_PCI_REG | XEON_CORERRSTS_OFFSET,
sysctl_handle_register, "IU", "CORERRSTS");
if (ntb->conn_type != NTB_CONN_B2B)
return;
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "outgoing_xlat23",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | ntb->bar_info[NTB_B2B_BAR_1].pbarxlat_off,
sysctl_handle_register, "QU", "Outgoing XLAT23 register");
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR)) {
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "outgoing_xlat4",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | ntb->bar_info[NTB_B2B_BAR_2].pbarxlat_off,
sysctl_handle_register, "IU", "Outgoing XLAT4 register");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "outgoing_xlat5",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | ntb->bar_info[NTB_B2B_BAR_3].pbarxlat_off,
sysctl_handle_register, "IU", "Outgoing XLAT5 register");
} else {
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "outgoing_xlat45",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | ntb->bar_info[NTB_B2B_BAR_2].pbarxlat_off,
sysctl_handle_register, "QU", "Outgoing XLAT45 register");
}
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "outgoing_lmt23",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | XEON_PBAR2LMT_OFFSET,
sysctl_handle_register, "QU", "Outgoing LMT23 register");
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR)) {
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "outgoing_lmt4",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | XEON_PBAR4LMT_OFFSET,
sysctl_handle_register, "IU", "Outgoing LMT4 register");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "outgoing_lmt5",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | XEON_PBAR5LMT_OFFSET,
sysctl_handle_register, "IU", "Outgoing LMT5 register");
} else {
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "outgoing_lmt45",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | XEON_PBAR4LMT_OFFSET,
sysctl_handle_register, "QU", "Outgoing LMT45 register");
}
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "sbar01_base",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | ntb->xlat_reg->bar0_base,
sysctl_handle_register, "QU", "Secondary BAR01 base register");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "sbar23_base",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | ntb->xlat_reg->bar2_base,
sysctl_handle_register, "QU", "Secondary BAR23 base register");
if (HAS_FEATURE(ntb, NTB_SPLIT_BAR)) {
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "sbar4_base",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | ntb->xlat_reg->bar4_base,
sysctl_handle_register, "IU",
"Secondary BAR4 base register");
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "sbar5_base",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_32 | ntb->xlat_reg->bar5_base,
sysctl_handle_register, "IU",
"Secondary BAR5 base register");
} else {
SYSCTL_ADD_PROC(ctx, regpar, OID_AUTO, "sbar45_base",
CTLFLAG_RD | CTLTYPE_OPAQUE, ntb,
NTB_REG_64 | ntb->xlat_reg->bar4_base,
sysctl_handle_register, "QU",
"Secondary BAR45 base register");
}
}
static int
sysctl_handle_features(SYSCTL_HANDLER_ARGS)
{
struct ntb_softc *ntb = arg1;
struct sbuf sb;
int error;
sbuf_new_for_sysctl(&sb, NULL, 256, req);
sbuf_printf(&sb, "%b", ntb->features, NTB_FEATURES_STR);
error = sbuf_finish(&sb);
sbuf_delete(&sb);
if (error || !req->newptr)
return (error);
return (EINVAL);
}
static int
sysctl_handle_link_admin(SYSCTL_HANDLER_ARGS)
{
struct ntb_softc *ntb = arg1;
unsigned old, new;
int error;
old = intel_ntb_link_enabled(ntb->device);
error = SYSCTL_OUT(req, &old, sizeof(old));
if (error != 0 || req->newptr == NULL)
return (error);
error = SYSCTL_IN(req, &new, sizeof(new));
if (error != 0)
return (error);
intel_ntb_printf(0, "Admin set interface state to '%sabled'\n",
(new != 0)? "en" : "dis");
if (new != 0)
error = intel_ntb_link_enable(ntb->device, NTB_SPEED_AUTO, NTB_WIDTH_AUTO);
else
error = intel_ntb_link_disable(ntb->device);
return (error);
}
static int
sysctl_handle_link_status_human(SYSCTL_HANDLER_ARGS)
{
struct ntb_softc *ntb = arg1;
struct sbuf sb;
enum ntb_speed speed;
enum ntb_width width;
int error;
sbuf_new_for_sysctl(&sb, NULL, 32, req);
if (intel_ntb_link_is_up(ntb->device, &speed, &width))
sbuf_printf(&sb, "up / PCIe Gen %u / Width x%u",
(unsigned)speed, (unsigned)width);
else
sbuf_printf(&sb, "down");
error = sbuf_finish(&sb);
sbuf_delete(&sb);
if (error || !req->newptr)
return (error);
return (EINVAL);
}
static int
sysctl_handle_link_status(SYSCTL_HANDLER_ARGS)
{
struct ntb_softc *ntb = arg1;
unsigned res;
int error;
res = intel_ntb_link_is_up(ntb->device, NULL, NULL);
error = SYSCTL_OUT(req, &res, sizeof(res));
if (error || !req->newptr)
return (error);
return (EINVAL);
}
static int
sysctl_handle_register(SYSCTL_HANDLER_ARGS)
{
struct ntb_softc *ntb;
const void *outp;
uintptr_t sz;
uint64_t umv;
char be[sizeof(umv)];
size_t outsz;
uint32_t reg;
bool db, pci;
int error;
ntb = arg1;
reg = arg2 & ~NTB_REGFLAGS_MASK;
sz = arg2 & NTB_REGSZ_MASK;
db = (arg2 & NTB_DB_READ) != 0;
pci = (arg2 & NTB_PCI_REG) != 0;
KASSERT(!(db && pci), ("bogus"));
if (db) {
KASSERT(sz == NTB_REG_64, ("bogus"));
umv = db_ioread(ntb, reg);
outsz = sizeof(uint64_t);
} else {
switch (sz) {
case NTB_REG_64:
if (pci)
umv = pci_read_config(ntb->device, reg, 8);
else
umv = intel_ntb_reg_read(8, reg);
outsz = sizeof(uint64_t);
break;
case NTB_REG_32:
if (pci)
umv = pci_read_config(ntb->device, reg, 4);
else
umv = intel_ntb_reg_read(4, reg);
outsz = sizeof(uint32_t);
break;
case NTB_REG_16:
if (pci)
umv = pci_read_config(ntb->device, reg, 2);
else
umv = intel_ntb_reg_read(2, reg);
outsz = sizeof(uint16_t);
break;
case NTB_REG_8:
if (pci)
umv = pci_read_config(ntb->device, reg, 1);
else
umv = intel_ntb_reg_read(1, reg);
outsz = sizeof(uint8_t);
break;
default:
panic("bogus");
break;
}
}
/* Encode bigendian so that sysctl -x is legible. */
be64enc(be, umv);
outp = ((char *)be) + sizeof(umv) - outsz;
error = SYSCTL_OUT(req, outp, outsz);
if (error || !req->newptr)
return (error);
return (EINVAL);
}
static unsigned
intel_ntb_user_mw_to_idx(struct ntb_softc *ntb, unsigned uidx)
{
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
if ((ntb->b2b_mw_idx != B2B_MW_DISABLED && ntb->b2b_off == 0 &&
uidx >= ntb->b2b_mw_idx) ||
(ntb->msix_mw_idx != B2B_MW_DISABLED && uidx >= ntb->msix_mw_idx))
uidx++;
if ((ntb->b2b_mw_idx != B2B_MW_DISABLED && ntb->b2b_off == 0 &&
uidx >= ntb->b2b_mw_idx) &&
(ntb->msix_mw_idx != B2B_MW_DISABLED && uidx >= ntb->msix_mw_idx))
uidx++;
return (uidx);
}
#ifndef EARLY_AP_STARTUP
static int msix_ready;
static void
intel_ntb_msix_ready(void *arg __unused)
{
msix_ready = 1;
}
SYSINIT(intel_ntb_msix_ready, SI_SUB_SMP, SI_ORDER_ANY,
intel_ntb_msix_ready, NULL);
#endif
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
static void
intel_ntb_exchange_msix(void *ctx)
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
{
struct ntb_softc *ntb;
uint32_t val;
unsigned i;
ntb = ctx;
if (ntb->peer_msix_good)
goto msix_good;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
if (ntb->peer_msix_done)
goto msix_done;
#ifndef EARLY_AP_STARTUP
/* Block MSIX negotiation until SMP started and IRQ reshuffled. */
if (!msix_ready)
goto reschedule;
#endif
intel_ntb_get_msix_info(ntb);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
for (i = 0; i < XEON_NONLINK_DB_MSIX_BITS; i++) {
intel_ntb_peer_spad_write(ntb->device, NTB_MSIX_DATA0 + i,
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
ntb->msix_data[i].nmd_data);
intel_ntb_peer_spad_write(ntb->device, NTB_MSIX_OFS0 + i,
ntb->msix_data[i].nmd_ofs - ntb->msix_xlat);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
}
intel_ntb_peer_spad_write(ntb->device, NTB_MSIX_GUARD, NTB_MSIX_VER_GUARD);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
intel_ntb_spad_read(ntb->device, NTB_MSIX_GUARD, &val);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
if (val != NTB_MSIX_VER_GUARD)
goto reschedule;
for (i = 0; i < XEON_NONLINK_DB_MSIX_BITS; i++) {
intel_ntb_spad_read(ntb->device, NTB_MSIX_DATA0 + i, &val);
intel_ntb_printf(2, "remote MSIX data(%u): 0x%x\n", i, val);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
ntb->peer_msix_data[i].nmd_data = val;
intel_ntb_spad_read(ntb->device, NTB_MSIX_OFS0 + i, &val);
intel_ntb_printf(2, "remote MSIX addr(%u): 0x%x\n", i, val);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
ntb->peer_msix_data[i].nmd_ofs = val;
}
ntb->peer_msix_done = true;
msix_done:
intel_ntb_peer_spad_write(ntb->device, NTB_MSIX_DONE, NTB_MSIX_RECEIVED);
intel_ntb_spad_read(ntb->device, NTB_MSIX_DONE, &val);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
if (val != NTB_MSIX_RECEIVED)
goto reschedule;
intel_ntb_spad_clear(ntb->device);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
ntb->peer_msix_good = true;
/* Give peer time to see our NTB_MSIX_RECEIVED. */
goto reschedule;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
msix_good:
intel_ntb_poll_link(ntb);
ntb_link_event(ntb->device);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
return;
reschedule:
ntb->lnk_sta = pci_read_config(ntb->device, ntb->reg->lnk_sta, 2);
if (_xeon_link_is_up(ntb)) {
callout_reset(&ntb->peer_msix_work,
hz * (ntb->peer_msix_good ? 2 : 1) / 10,
intel_ntb_exchange_msix, ntb);
} else
intel_ntb_spad_clear(ntb->device);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
}
/*
* Public API to the rest of the OS
*/
static uint8_t
intel_ntb_spad_count(device_t dev)
{
struct ntb_softc *ntb = device_get_softc(dev);
return (ntb->spad_count);
}
static uint8_t
intel_ntb_mw_count(device_t dev)
{
struct ntb_softc *ntb = device_get_softc(dev);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
uint8_t res;
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
res = ntb->mw_count;
if (ntb->b2b_mw_idx != B2B_MW_DISABLED && ntb->b2b_off == 0)
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
res--;
if (ntb->msix_mw_idx != B2B_MW_DISABLED)
res--;
return (res);
}
static int
intel_ntb_spad_write(device_t dev, unsigned int idx, uint32_t val)
{
struct ntb_softc *ntb = device_get_softc(dev);
if (idx >= ntb->spad_count)
return (EINVAL);
intel_ntb_reg_write(4, ntb->self_reg->spad + idx * 4, val);
return (0);
}
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
/*
* Zeros the local scratchpad.
*/
static void
intel_ntb_spad_clear(device_t dev)
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
{
struct ntb_softc *ntb = device_get_softc(dev);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
unsigned i;
for (i = 0; i < ntb->spad_count; i++)
intel_ntb_spad_write(dev, i, 0);
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
}
static int
intel_ntb_spad_read(device_t dev, unsigned int idx, uint32_t *val)
{
struct ntb_softc *ntb = device_get_softc(dev);
if (idx >= ntb->spad_count)
return (EINVAL);
*val = intel_ntb_reg_read(4, ntb->self_reg->spad + idx * 4);
return (0);
}
static int
intel_ntb_peer_spad_write(device_t dev, unsigned int idx, uint32_t val)
{
struct ntb_softc *ntb = device_get_softc(dev);
if (idx >= ntb->spad_count)
return (EINVAL);
if (HAS_FEATURE(ntb, NTB_SDOORBELL_LOCKUP))
intel_ntb_mw_write(4, XEON_SPAD_OFFSET + idx * 4, val);
else
intel_ntb_reg_write(4, ntb->peer_reg->spad + idx * 4, val);
return (0);
}
static int
intel_ntb_peer_spad_read(device_t dev, unsigned int idx, uint32_t *val)
{
struct ntb_softc *ntb = device_get_softc(dev);
if (idx >= ntb->spad_count)
return (EINVAL);
if (HAS_FEATURE(ntb, NTB_SDOORBELL_LOCKUP))
*val = intel_ntb_mw_read(4, XEON_SPAD_OFFSET + idx * 4);
else
*val = intel_ntb_reg_read(4, ntb->peer_reg->spad + idx * 4);
return (0);
}
static int
intel_ntb_mw_get_range(device_t dev, unsigned mw_idx, vm_paddr_t *base,
caddr_t *vbase, size_t *size, size_t *align, size_t *align_size,
bus_addr_t *plimit)
{
struct ntb_softc *ntb = device_get_softc(dev);
struct ntb_pci_bar_info *bar;
bus_addr_t limit;
size_t bar_b2b_off;
enum ntb_bar bar_num;
if (mw_idx >= intel_ntb_mw_count(dev))
return (EINVAL);
mw_idx = intel_ntb_user_mw_to_idx(ntb, mw_idx);
bar_num = intel_ntb_mw_to_bar(ntb, mw_idx);
bar = &ntb->bar_info[bar_num];
bar_b2b_off = 0;
if (mw_idx == ntb->b2b_mw_idx) {
KASSERT(ntb->b2b_off != 0,
("user shouldn't get non-shared b2b mw"));
bar_b2b_off = ntb->b2b_off;
}
if (bar_is_64bit(ntb, bar_num))
limit = BUS_SPACE_MAXADDR;
else
limit = BUS_SPACE_MAXADDR_32BIT;
if (base != NULL)
*base = bar->pbase + bar_b2b_off;
if (vbase != NULL)
*vbase = bar->vbase + bar_b2b_off;
if (size != NULL)
*size = bar->size - bar_b2b_off;
if (align != NULL)
*align = bar->size;
if (align_size != NULL)
*align_size = 1;
if (plimit != NULL)
*plimit = limit;
return (0);
}
static int
intel_ntb_mw_set_trans(device_t dev, unsigned idx, bus_addr_t addr, size_t size)
{
struct ntb_softc *ntb = device_get_softc(dev);
struct ntb_pci_bar_info *bar;
uint64_t base, limit, reg_val;
size_t bar_size, mw_size;
uint32_t base_reg, xlat_reg, limit_reg;
enum ntb_bar bar_num;
if (idx >= intel_ntb_mw_count(dev))
return (EINVAL);
idx = intel_ntb_user_mw_to_idx(ntb, idx);
bar_num = intel_ntb_mw_to_bar(ntb, idx);
bar = &ntb->bar_info[bar_num];
bar_size = bar->size;
if (idx == ntb->b2b_mw_idx)
mw_size = bar_size - ntb->b2b_off;
else
mw_size = bar_size;
/* Hardware requires that addr is aligned to bar size */
if ((addr & (bar_size - 1)) != 0)
return (EINVAL);
if (size > mw_size)
return (EINVAL);
bar_get_xlat_params(ntb, bar_num, &base_reg, &xlat_reg, &limit_reg);
limit = 0;
if (bar_is_64bit(ntb, bar_num)) {
base = intel_ntb_reg_read(8, base_reg) & BAR_HIGH_MASK;
if (limit_reg != 0 && size != mw_size)
limit = base + size;
/* Set and verify translation address */
intel_ntb_reg_write(8, xlat_reg, addr);
reg_val = intel_ntb_reg_read(8, xlat_reg) & BAR_HIGH_MASK;
if (reg_val != addr) {
intel_ntb_reg_write(8, xlat_reg, 0);
return (EIO);
}
/* Set and verify the limit */
intel_ntb_reg_write(8, limit_reg, limit);
reg_val = intel_ntb_reg_read(8, limit_reg) & BAR_HIGH_MASK;
if (reg_val != limit) {
intel_ntb_reg_write(8, limit_reg, base);
intel_ntb_reg_write(8, xlat_reg, 0);
return (EIO);
}
} else {
/* Configure 32-bit (split) BAR MW */
if ((addr & UINT32_MAX) != addr)
return (ERANGE);
if (((addr + size) & UINT32_MAX) != (addr + size))
return (ERANGE);
base = intel_ntb_reg_read(4, base_reg) & BAR_HIGH_MASK;
if (limit_reg != 0 && size != mw_size)
limit = base + size;
/* Set and verify translation address */
intel_ntb_reg_write(4, xlat_reg, addr);
reg_val = intel_ntb_reg_read(4, xlat_reg) & BAR_HIGH_MASK;
if (reg_val != addr) {
intel_ntb_reg_write(4, xlat_reg, 0);
return (EIO);
}
/* Set and verify the limit */
intel_ntb_reg_write(4, limit_reg, limit);
reg_val = intel_ntb_reg_read(4, limit_reg) & BAR_HIGH_MASK;
if (reg_val != limit) {
intel_ntb_reg_write(4, limit_reg, base);
intel_ntb_reg_write(4, xlat_reg, 0);
return (EIO);
}
}
return (0);
}
static int
intel_ntb_mw_clear_trans(device_t dev, unsigned mw_idx)
{
return (intel_ntb_mw_set_trans(dev, mw_idx, 0, 0));
}
static int
intel_ntb_mw_get_wc(device_t dev, unsigned idx, vm_memattr_t *mode)
{
struct ntb_softc *ntb = device_get_softc(dev);
struct ntb_pci_bar_info *bar;
if (idx >= intel_ntb_mw_count(dev))
return (EINVAL);
idx = intel_ntb_user_mw_to_idx(ntb, idx);
bar = &ntb->bar_info[intel_ntb_mw_to_bar(ntb, idx)];
*mode = bar->map_mode;
return (0);
}
static int
intel_ntb_mw_set_wc(device_t dev, unsigned idx, vm_memattr_t mode)
{
struct ntb_softc *ntb = device_get_softc(dev);
if (idx >= intel_ntb_mw_count(dev))
return (EINVAL);
idx = intel_ntb_user_mw_to_idx(ntb, idx);
return (intel_ntb_mw_set_wc_internal(ntb, idx, mode));
}
static int
intel_ntb_mw_set_wc_internal(struct ntb_softc *ntb, unsigned idx, vm_memattr_t mode)
{
struct ntb_pci_bar_info *bar;
int rc;
bar = &ntb->bar_info[intel_ntb_mw_to_bar(ntb, idx)];
if (bar->map_mode == mode)
return (0);
rc = pmap_change_attr((vm_offset_t)bar->vbase, bar->size, mode);
if (rc == 0)
bar->map_mode = mode;
return (rc);
}
static void
intel_ntb_peer_db_set(device_t dev, uint64_t bit)
{
struct ntb_softc *ntb = device_get_softc(dev);
if (HAS_FEATURE(ntb, NTB_SB01BASE_LOCKUP)) {
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
struct ntb_pci_bar_info *lapic;
unsigned i;
lapic = ntb->peer_lapic_bar;
for (i = 0; i < XEON_NONLINK_DB_MSIX_BITS; i++) {
if ((bit & intel_ntb_db_vector_mask(dev, i)) != 0)
NTB: workaround for high traffic hardware hang This patch comes from Dave Jiang's Linux tree, davejiang/ntb. It hasn't been accepted into Linus' tree, so I do not have an authoritative SHA1 to point at. Original commit log: ===================================================================== A hardware errata causes the NTB to hang when heavy bi-directional traffic in addition to the usage of BAR0/1 (where the registers reside, including the doorbell registers to trigger interrupts). This workaround is only available on Haswell and Broadwell platform. The workaround is to enable split BAR in the BIOS to allow the 64bit BAR4 to be split into two 32bit BAR4 and BAR5. The BAR4 shall be pointed to LAPIC region of the remote host. We will bypass the db mechanism and directly trigger the MSIX interrupts. The offsets and vectors are exchanged during transport scratch pad negotiation. The scratch pads are now overloaded in order to allow the exchange of the information. This gets around using the doorbell and prevents the lockup with additional pcode changes in BIOS. Signed-off-by: Dave Jiang <dave.jiang@intel.com> ===================================================================== Notable changes in the FreeBSD version of this patch: * The MSIX BAR is configurable, like hw.ntb.b2b_mw_idx (msix_mw_idx). The Linux version of the patch only uses BAR4. * MSIX negotiation aborts if the link goes down. Obtained from: Linux (Dual BSD/GPL driver) Sponsored by: EMC / Isilon Storage Division
2016-02-14 22:37:28 +00:00
bus_space_write_4(lapic->pci_bus_tag,
lapic->pci_bus_handle,
ntb->peer_msix_data[i].nmd_ofs,
ntb->peer_msix_data[i].nmd_data);
}
return;
}
if (HAS_FEATURE(ntb, NTB_SDOORBELL_LOCKUP)) {
intel_ntb_mw_write(2, XEON_PDOORBELL_OFFSET, bit);
return;
2015-10-13 03:12:55 +00:00
}
db_iowrite(ntb, ntb->peer_reg->db_bell, bit);
}
static int
intel_ntb_peer_db_addr(device_t dev, bus_addr_t *db_addr, vm_size_t *db_size)
{
struct ntb_softc *ntb = device_get_softc(dev);
struct ntb_pci_bar_info *bar;
uint64_t regoff;
KASSERT((db_addr != NULL && db_size != NULL), ("must be non-NULL"));
if (!HAS_FEATURE(ntb, NTB_SDOORBELL_LOCKUP)) {
bar = &ntb->bar_info[NTB_CONFIG_BAR];
regoff = ntb->peer_reg->db_bell;
} else {
KASSERT(ntb->b2b_mw_idx != B2B_MW_DISABLED,
("invalid b2b idx"));
bar = &ntb->bar_info[intel_ntb_mw_to_bar(ntb, ntb->b2b_mw_idx)];
regoff = XEON_PDOORBELL_OFFSET;
}
KASSERT(bar->pci_bus_tag != X86_BUS_SPACE_IO, ("uh oh"));
/* HACK: Specific to current x86 bus implementation. */
*db_addr = ((uint64_t)bar->pci_bus_handle + regoff);
*db_size = ntb->reg->db_size;
return (0);
}
static uint64_t
intel_ntb_db_valid_mask(device_t dev)
{
struct ntb_softc *ntb = device_get_softc(dev);
return (ntb->db_valid_mask);
}
static int
intel_ntb_db_vector_count(device_t dev)
{
struct ntb_softc *ntb = device_get_softc(dev);
return (ntb->db_vec_count);
}
static uint64_t
intel_ntb_db_vector_mask(device_t dev, uint32_t vector)
{
struct ntb_softc *ntb = device_get_softc(dev);
if (vector > ntb->db_vec_count)
return (0);
return (ntb->db_valid_mask & intel_ntb_vec_mask(ntb, vector));
}
static bool
intel_ntb_link_is_up(device_t dev, enum ntb_speed *speed, enum ntb_width *width)
{
struct ntb_softc *ntb = device_get_softc(dev);
if (speed != NULL)
*speed = intel_ntb_link_sta_speed(ntb);
if (width != NULL)
*width = intel_ntb_link_sta_width(ntb);
return (link_is_up(ntb));
}
static void
save_bar_parameters(struct ntb_pci_bar_info *bar)
{
2015-10-13 03:12:55 +00:00
bar->pci_bus_tag = rman_get_bustag(bar->pci_resource);
bar->pci_bus_handle = rman_get_bushandle(bar->pci_resource);
bar->pbase = rman_get_start(bar->pci_resource);
bar->size = rman_get_size(bar->pci_resource);
bar->vbase = rman_get_virtual(bar->pci_resource);
}
static device_method_t ntb_intel_methods[] = {
/* Device interface */
DEVMETHOD(device_probe, intel_ntb_probe),
DEVMETHOD(device_attach, intel_ntb_attach),
DEVMETHOD(device_detach, intel_ntb_detach),
/* Bus interface */
DEVMETHOD(bus_child_location_str, ntb_child_location_str),
DEVMETHOD(bus_print_child, ntb_print_child),
/* NTB interface */
DEVMETHOD(ntb_port_number, intel_ntb_port_number),
DEVMETHOD(ntb_peer_port_count, intel_ntb_peer_port_count),
DEVMETHOD(ntb_peer_port_number, intel_ntb_peer_port_number),
DEVMETHOD(ntb_peer_port_idx, intel_ntb_peer_port_idx),
DEVMETHOD(ntb_link_is_up, intel_ntb_link_is_up),
DEVMETHOD(ntb_link_enable, intel_ntb_link_enable),
DEVMETHOD(ntb_link_disable, intel_ntb_link_disable),
DEVMETHOD(ntb_link_enabled, intel_ntb_link_enabled),
DEVMETHOD(ntb_mw_count, intel_ntb_mw_count),
DEVMETHOD(ntb_mw_get_range, intel_ntb_mw_get_range),
DEVMETHOD(ntb_mw_set_trans, intel_ntb_mw_set_trans),
DEVMETHOD(ntb_mw_clear_trans, intel_ntb_mw_clear_trans),
DEVMETHOD(ntb_mw_get_wc, intel_ntb_mw_get_wc),
DEVMETHOD(ntb_mw_set_wc, intel_ntb_mw_set_wc),
DEVMETHOD(ntb_spad_count, intel_ntb_spad_count),
DEVMETHOD(ntb_spad_clear, intel_ntb_spad_clear),
DEVMETHOD(ntb_spad_write, intel_ntb_spad_write),
DEVMETHOD(ntb_spad_read, intel_ntb_spad_read),
DEVMETHOD(ntb_peer_spad_write, intel_ntb_peer_spad_write),
DEVMETHOD(ntb_peer_spad_read, intel_ntb_peer_spad_read),
DEVMETHOD(ntb_db_valid_mask, intel_ntb_db_valid_mask),
DEVMETHOD(ntb_db_vector_count, intel_ntb_db_vector_count),
DEVMETHOD(ntb_db_vector_mask, intel_ntb_db_vector_mask),
DEVMETHOD(ntb_db_clear, intel_ntb_db_clear),
DEVMETHOD(ntb_db_clear_mask, intel_ntb_db_clear_mask),
DEVMETHOD(ntb_db_read, intel_ntb_db_read),
DEVMETHOD(ntb_db_set_mask, intel_ntb_db_set_mask),
DEVMETHOD(ntb_peer_db_addr, intel_ntb_peer_db_addr),
DEVMETHOD(ntb_peer_db_set, intel_ntb_peer_db_set),
DEVMETHOD_END
};
static DEFINE_CLASS_0(ntb_hw, ntb_intel_driver, ntb_intel_methods,
sizeof(struct ntb_softc));
DRIVER_MODULE(ntb_hw_intel, pci, ntb_intel_driver, ntb_hw_devclass, NULL, NULL);
MODULE_DEPEND(ntb_hw_intel, ntb, 1, 1, 1);
MODULE_VERSION(ntb_hw_intel, 1);
MODULE_PNP_INFO("W32:vendor/device;D:#", pci, ntb_hw_intel, pci_ids,
nitems(pci_ids));