From 6e768f0682e26e7683c4af5b0de04a2e08bc67db Mon Sep 17 00:00:00 2001 From: Sebastian Sanchez Date: Mon, 17 Oct 2016 04:19:35 -0700 Subject: [PATCH] IB/hfi1: Optimize devdata cachelines Profiling shows hot path struct members that need to be in a minimum set of cachelines. Group these struct member in the same cacheline: sc2vl_lock sc2vl rhf_rcv_function_map rcv_limit rhf_offset Group these struct member in the same cacheline: process_pio_send process_dma_send pport rcd int_counter flags num_pports first_user_ctxt Fill holes in struct hfi1_devdata revealed by pahole. Reviewed-by: Mike Marciniszyn Signed-off-by: Sebastian Sanchez Signed-off-by: Dennis Dalessandro Signed-off-by: Doug Ledford --- drivers/infiniband/hw/hfi1/hfi.h | 108 ++++++++++++++++--------------- 1 file changed, 55 insertions(+), 53 deletions(-) diff --git a/drivers/infiniband/hw/hfi1/hfi.h b/drivers/infiniband/hw/hfi1/hfi.h index a2ea643a79fc..25e44c4af995 100644 --- a/drivers/infiniband/hw/hfi1/hfi.h +++ b/drivers/infiniband/hw/hfi1/hfi.h @@ -852,32 +852,29 @@ struct hfi1_devdata { u8 __iomem *kregend; /* physical address of chip for io_remap, etc. */ resource_size_t physaddr; - /* receive context data */ - struct hfi1_ctxtdata **rcd; + /* Per VL data. Enough for all VLs but not all elements are set/used. */ + struct per_vl_data vld[PER_VL_SEND_CONTEXTS]; /* send context data */ struct send_context_info *send_contexts; /* map hardware send contexts to software index */ u8 *hw_to_sw; /* spinlock for allocating and releasing send context resources */ spinlock_t sc_lock; - /* Per VL data. Enough for all VLs but not all elements are set/used. */ - struct per_vl_data vld[PER_VL_SEND_CONTEXTS]; /* lock for pio_map */ spinlock_t pio_map_lock; + /* Send Context initialization lock. */ + spinlock_t sc_init_lock; + /* lock for sdma_map */ + spinlock_t sde_map_lock; /* array of kernel send contexts */ struct send_context **kernel_send_context; /* array of vl maps */ struct pio_vl_map __rcu *pio_map; - /* seqlock for sc2vl */ - seqlock_t sc2vl_lock; - u64 sc2vl[4]; - /* Send Context initialization lock. */ - spinlock_t sc_init_lock; + /* default flags to last descriptor */ + u64 default_desc1; /* fields common to all SDMA engines */ - /* default flags to last descriptor */ - u64 default_desc1; volatile __le64 *sdma_heads_dma; /* DMA'ed by chip */ dma_addr_t sdma_heads_phys; void *sdma_pad_dma; /* DMA'ed by chip */ @@ -888,8 +885,6 @@ struct hfi1_devdata { u32 chip_sdma_engines; /* num used */ u32 num_sdma; - /* lock for sdma_map */ - spinlock_t sde_map_lock; /* array of engines sized by num_sdma */ struct sdma_engine *per_sdma; /* array of vl maps */ @@ -898,14 +893,11 @@ struct hfi1_devdata { wait_queue_head_t sdma_unfreeze_wq; atomic_t sdma_unfreeze_count; + u32 lcb_access_count; /* count of LCB users */ + /* common data between shared ASIC HFIs in this OS */ struct hfi1_asic_data *asic_data; - /* hfi1_pportdata, points to array of (physical) port-specific - * data structs, indexed by pidx (0..n-1) - */ - struct hfi1_pportdata *pport; - /* mem-mapped pointer to base of PIO buffers */ void __iomem *piobase; /* @@ -922,20 +914,13 @@ struct hfi1_devdata { /* send context numbers and sizes for each type */ struct sc_config_sizes sc_sizes[SC_MAX]; - u32 lcb_access_count; /* count of LCB users */ - char *boardname; /* human readable board info */ - /* device (not port) flags, basically device capabilities */ - u32 flags; - /* reset value */ u64 z_int_counter; u64 z_rcv_limit; u64 z_send_schedule; - /* percpu int_counter */ - u64 __percpu *int_counter; - u64 __percpu *rcv_limit; + u64 __percpu *send_schedule; /* number of receive contexts in use by the driver */ u32 num_rcv_contexts; @@ -950,6 +935,7 @@ struct hfi1_devdata { /* base receive interrupt timeout, in CSR units */ u32 rcv_intr_timeout_csr; + u32 freezelen; /* max length of freezemsg */ u64 __iomem *egrtidbase; spinlock_t sendctrl_lock; /* protect changes to SendCtrl */ spinlock_t rcvctrl_lock; /* protect changes to RcvCtrl */ @@ -971,7 +957,6 @@ struct hfi1_devdata { * IB link status cheaply */ struct hfi1_status *status; - u32 freezelen; /* max length of freezemsg */ /* revision register shadow */ u64 revision; @@ -999,6 +984,8 @@ struct hfi1_devdata { u16 rcvegrbufsize_shift; /* both sides of the PCIe link are gen3 capable */ u8 link_gen3_capable; + /* default link down value (poll/sleep) */ + u8 link_default; /* localbus width (1, 2,4,8,16,32) from config space */ u32 lbus_width; /* localbus speed in MHz */ @@ -1034,8 +1021,6 @@ struct hfi1_devdata { u8 hfi1_id; /* implementation code */ u8 icode; - /* default link down value (poll/sleep) */ - u8 link_default; /* vAU of this device */ u8 vau; /* vCU of this device */ @@ -1046,27 +1031,17 @@ struct hfi1_devdata { u16 vl15_init; /* Misc small ints */ - /* Number of physical ports available */ - u8 num_pports; - /* Lowest context number which can be used by user processes */ - u8 first_user_ctxt; u8 n_krcv_queues; u8 qos_shift; - u8 qpn_mask; - u16 rhf_offset; /* offset of RHF within receive header entry */ u16 irev; /* implementation revision */ u16 dc8051_ver; /* 8051 firmware version */ + spinlock_t hfi1_diag_trans_lock; /* protect diag observer ops */ struct platform_config platform_config; struct platform_config_cache pcfg_cache; struct diag_client *diag_client; - spinlock_t hfi1_diag_trans_lock; /* protect diag observer ops */ - - u8 psxmitwait_supported; - /* cycle length of PS* counters in HW (in picoseconds) */ - u16 psxmitwait_check_rate; /* MSI-X information */ struct hfi1_msix_entry *msix_entries; @@ -1081,6 +1056,9 @@ struct hfi1_devdata { struct rcv_array_data rcv_entries; + /* cycle length of PS* counters in HW (in picoseconds) */ + u16 psxmitwait_check_rate; + /* * 64 bit synthetic counters */ @@ -1113,11 +1091,11 @@ struct hfi1_devdata { struct err_info_rcvport err_info_rcvport; struct err_info_constraint err_info_rcv_constraint; struct err_info_constraint err_info_xmit_constraint; - u8 err_info_uncorrectable; - u8 err_info_fmconfig; atomic_t drop_packet; u8 do_drop; + u8 err_info_uncorrectable; + u8 err_info_fmconfig; /* * Software counters for the status bits defined by the @@ -1140,47 +1118,71 @@ struct hfi1_devdata { u64 sw_cce_err_status_aggregate; /* Software counter that aggregates all bypass packet rcv errors */ u64 sw_rcv_bypass_packet_errors; - /* receive interrupt functions */ - rhf_rcv_function_ptr *rhf_rcv_function_map; + /* receive interrupt function */ rhf_rcv_function_ptr normal_rhf_rcv_functions[8]; + /* Save the enabled LCB error bits */ + u64 lcb_err_en; + /* * Handlers for outgoing data so that snoop/capture does not * have to have its hooks in the send path */ - send_routine process_pio_send; + send_routine process_pio_send ____cacheline_aligned_in_smp; send_routine process_dma_send; void (*pio_inline_send)(struct hfi1_devdata *dd, struct pio_buf *pbuf, u64 pbc, const void *from, size_t count); + /* hfi1_pportdata, points to array of (physical) port-specific + * data structs, indexed by pidx (0..n-1) + */ + struct hfi1_pportdata *pport; + /* receive context data */ + struct hfi1_ctxtdata **rcd; + u64 __percpu *int_counter; + /* device (not port) flags, basically device capabilities */ + u16 flags; + /* Number of physical ports available */ + u8 num_pports; + /* Lowest context number which can be used by user processes */ + u8 first_user_ctxt; + /* adding a new field here would make it part of this cacheline */ + + /* seqlock for sc2vl */ + seqlock_t sc2vl_lock ____cacheline_aligned_in_smp; + u64 sc2vl[4]; + /* receive interrupt functions */ + rhf_rcv_function_ptr *rhf_rcv_function_map; + u64 __percpu *rcv_limit; + u16 rhf_offset; /* offset of RHF within receive header entry */ + /* adding a new field here would make it part of this cacheline */ /* OUI comes from the HW. Used everywhere as 3 separate bytes. */ u8 oui1; u8 oui2; u8 oui3; + u8 dc_shutdown; + /* Timer and counter used to detect RcvBufOvflCnt changes */ struct timer_list rcverr_timer; - u32 rcv_ovfl_cnt; wait_queue_head_t event_queue; - /* Save the enabled LCB error bits */ - u64 lcb_err_en; - u8 dc_shutdown; - /* receive context tail dummy address */ __le64 *rcvhdrtail_dummy_kvaddr; dma_addr_t rcvhdrtail_dummy_dma; - bool eprom_available; /* true if EPROM is available for this device */ - bool aspm_supported; /* Does HW support ASPM */ - bool aspm_enabled; /* ASPM state: enabled/disabled */ + u32 rcv_ovfl_cnt; /* Serialize ASPM enable/disable between multiple verbs contexts */ spinlock_t aspm_lock; /* Number of verbs contexts which have disabled ASPM */ atomic_t aspm_disabled_cnt; struct hfi1_affinity *affinity; + bool eprom_available; /* true if EPROM is available for this device */ + bool aspm_supported; /* Does HW support ASPM */ + bool aspm_enabled; /* ASPM state: enabled/disabled */ struct rhashtable sdma_rht; + struct kobject kobj; }; -- 2.20.1