[PATCH] powerpc: Add Initiate Coprocessor Store Word (icswx) support
Benjamin Herrenschmidt
benh at kernel.crashing.org
Tue May 3 16:47:17 EST 2011
On Tue, 2011-05-03 at 16:43 +1000, Anton Blanchard wrote:
> From: Tseng-Hui (Frank) Lin <thlin at linux.vnet.ibm.com>
>
> Icswx is a PowerPC instruction to send data to a co-processor. On Book-S
> processors the LPAR_ID and process ID (PID) of the owning process are
> registered in the window context of the co-processor at initialization
> time. When the icswx instruction is executed the L2 generates a cop-reg
> transaction on PowerBus. The transaction has no address and the
> processor does not perform an MMU access to authenticate the transaction.
> The co-processor compares the LPAR_ID and the PID included in the
> transaction and the LPAR_ID and PID held in the window context to
> determine if the process is authorized to generate the transaction.
>
> The OS needs to assign a 16-bit PID for the process. This cop-PID needs
> to be updated during context switch. The cop-PID needs to be destroyed
> when the context is destroyed.
> Signed-off-by: Sonny Rao <sonnyrao at linux.vnet.ibm.com>
> Signed-off-by: Tseng-Hui (Frank) Lin <thlin at linux.vnet.ibm.com>
> Signed-off-by: Anton Blanchard <anton at samba.org>
> ---
>
> Changes:
>
> - Go back to dynamically allocating the spinlock to avoid the mmu_context.h
> and spinlock_types.h include mess.
My only comment (sorry Anton :-) would have been that we could lazily
allocate the spinlock on the first use_cop() ... or do we have that
potentially called in the wrong context ?
Cheers,
Ben.
> Index: linux-powerpc/arch/powerpc/include/asm/cputable.h
> ===================================================================
> --- linux-powerpc.orig/arch/powerpc/include/asm/cputable.h 2011-05-03 16:39:21.779140935 +1000
> +++ linux-powerpc/arch/powerpc/include/asm/cputable.h 2011-05-03 16:39:26.059220272 +1000
> @@ -197,6 +197,7 @@ extern const char *powerpc_base_platform
> #define CPU_FTR_STCX_CHECKS_ADDRESS LONG_ASM_CONST(0x0200000000000000)
> #define CPU_FTR_POPCNTB LONG_ASM_CONST(0x0400000000000000)
> #define CPU_FTR_POPCNTD LONG_ASM_CONST(0x0800000000000000)
> +#define CPU_FTR_ICSWX LONG_ASM_CONST(0x1000000000000000)
>
> #ifndef __ASSEMBLY__
>
> @@ -418,7 +419,8 @@ extern const char *powerpc_base_platform
> CPU_FTR_COHERENT_ICACHE | \
> CPU_FTR_PURR | CPU_FTR_SPURR | CPU_FTR_REAL_LE | \
> CPU_FTR_DSCR | CPU_FTR_SAO | CPU_FTR_ASYM_SMT | \
> - CPU_FTR_STCX_CHECKS_ADDRESS | CPU_FTR_POPCNTB | CPU_FTR_POPCNTD)
> + CPU_FTR_STCX_CHECKS_ADDRESS | CPU_FTR_POPCNTB | CPU_FTR_POPCNTD | \
> + CPU_FTR_ICSWX)
> #define CPU_FTRS_CELL (CPU_FTR_USE_TB | CPU_FTR_LWSYNC | \
> CPU_FTR_PPCAS_ARCH_V2 | CPU_FTR_CTRL | \
> CPU_FTR_ALTIVEC_COMP | CPU_FTR_MMCRA | CPU_FTR_SMT | \
> Index: linux-powerpc/arch/powerpc/include/asm/mmu-hash64.h
> ===================================================================
> --- linux-powerpc.orig/arch/powerpc/include/asm/mmu-hash64.h 2011-05-03 16:39:21.789141120 +1000
> +++ linux-powerpc/arch/powerpc/include/asm/mmu-hash64.h 2011-05-03 16:39:26.059220272 +1000
> @@ -408,6 +408,7 @@ static inline void subpage_prot_init_new
> #endif /* CONFIG_PPC_SUBPAGE_PROT */
>
> typedef unsigned long mm_context_id_t;
> +struct spinlock;
>
> typedef struct {
> mm_context_id_t id;
> @@ -423,6 +424,11 @@ typedef struct {
> #ifdef CONFIG_PPC_SUBPAGE_PROT
> struct subpage_prot_table spt;
> #endif /* CONFIG_PPC_SUBPAGE_PROT */
> +#ifdef CONFIG_PPC_ICSWX
> + struct spinlock *cop_lockp; /* guard acop and cop_pid */
> + unsigned long acop; /* mask of enabled coprocessor types */
> + unsigned int cop_pid; /* pid value used with coprocessors */
> +#endif /* CONFIG_PPC_ICSWX */
> } mm_context_t;
>
>
> Index: linux-powerpc/arch/powerpc/include/asm/mmu_context.h
> ===================================================================
> --- linux-powerpc.orig/arch/powerpc/include/asm/mmu_context.h 2011-05-03 16:39:21.779140935 +1000
> +++ linux-powerpc/arch/powerpc/include/asm/mmu_context.h 2011-05-03 16:39:26.059220272 +1000
> @@ -32,6 +32,10 @@ extern void __destroy_context(unsigned l
> extern void mmu_context_init(void);
> #endif
>
> +extern void switch_cop(struct mm_struct *next);
> +extern int use_cop(unsigned long acop, struct mm_struct *mm);
> +extern void drop_cop(unsigned long acop, struct mm_struct *mm);
> +
> /*
> * switch_mm is the entry point called from the architecture independent
> * code in kernel/sched.c
> @@ -55,6 +59,12 @@ static inline void switch_mm(struct mm_s
> if (prev == next)
> return;
>
> +#ifdef CONFIG_PPC_ICSWX
> + /* Switch coprocessor context only if prev or next uses a coprocessor */
> + if (prev->context.acop || next->context.acop)
> + switch_cop(next);
> +#endif /* CONFIG_PPC_ICSWX */
> +
> /* We must stop all altivec streams before changing the HW
> * context
> */
> Index: linux-powerpc/arch/powerpc/include/asm/reg.h
> ===================================================================
> --- linux-powerpc.orig/arch/powerpc/include/asm/reg.h 2011-05-03 16:39:21.799141306 +1000
> +++ linux-powerpc/arch/powerpc/include/asm/reg.h 2011-05-03 16:39:26.059220272 +1000
> @@ -188,6 +188,7 @@
>
> #define SPRN_CTR 0x009 /* Count Register */
> #define SPRN_DSCR 0x11
> +#define SPRN_ACOP 0x1F /* Available Coprocessor Register */
> #define SPRN_CTRLF 0x088
> #define SPRN_CTRLT 0x098
> #define CTRL_CT 0xc0000000 /* current thread */
> Index: linux-powerpc/arch/powerpc/mm/mmu_context_hash64.c
> ===================================================================
> --- linux-powerpc.orig/arch/powerpc/mm/mmu_context_hash64.c 2011-05-03 16:39:21.759140565 +1000
> +++ linux-powerpc/arch/powerpc/mm/mmu_context_hash64.c 2011-05-03 16:39:26.059220272 +1000
> @@ -20,9 +20,205 @@
> #include <linux/idr.h>
> #include <linux/module.h>
> #include <linux/gfp.h>
> +#include <linux/slab.h>
>
> #include <asm/mmu_context.h>
>
> +#ifdef CONFIG_PPC_ICSWX
> +/*
> + * The processor and its L2 cache cause the icswx instruction to
> + * generate a COP_REQ transaction on PowerBus. The transaction has
> + * no address, and the processor does not perform an MMU access
> + * to authenticate the transaction. The command portion of the
> + * PowerBus COP_REQ transaction includes the LPAR_ID (LPID) and
> + * the coprocessor Process ID (PID), which the coprocessor compares
> + * to the authorized LPID and PID held in the coprocessor, to determine
> + * if the process is authorized to generate the transaction.
> + * The data of the COP_REQ transaction is 128-byte or less and is
> + * placed in cacheable memory on a 128-byte cache line boundary.
> + *
> + * The task to use a coprocessor should use use_cop() to allocate
> + * a coprocessor PID before executing icswx instruction. use_cop()
> + * also enables the coprocessor context switching. Drop_cop() is
> + * used to free the coprocessor PID.
> + *
> + * Example:
> + * Host Fabric Interface (HFI) is a PowerPC network coprocessor.
> + * Each HFI have multiple windows. Each HFI window serves as a
> + * network device sending to and receiving from HFI network.
> + * HFI immediate send function uses icswx instruction. The immediate
> + * send function allows small (single cache-line) packets be sent
> + * without using the regular HFI send FIFO and doorbell, which are
> + * much slower than immediate send.
> + *
> + * For each task intending to use HFI immediate send, the HFI driver
> + * calls use_cop() to obtain a coprocessor PID for the task.
> + * The HFI driver then allocate a free HFI window and save the
> + * coprocessor PID to the HFI window to allow the task to use the
> + * HFI window.
> + *
> + * The HFI driver repeatedly creates immediate send packets and
> + * issues icswx instruction to send data through the HFI window.
> + * The HFI compares the coprocessor PID in the CPU PID register
> + * to the PID held in the HFI window to determine if the transaction
> + * is allowed.
> + *
> + * When the task to release the HFI window, the HFI driver calls
> + * drop_cop() to release the coprocessor PID.
> + */
> +
> +#define COP_PID_NONE 0
> +#define COP_PID_MIN (COP_PID_NONE + 1)
> +#define COP_PID_MAX (0xFFFF)
> +
> +static DEFINE_SPINLOCK(mmu_context_acop_lock);
> +static DEFINE_IDA(cop_ida);
> +
> +void switch_cop(struct mm_struct *next)
> +{
> + mtspr(SPRN_PID, next->context.cop_pid);
> + mtspr(SPRN_ACOP, next->context.acop);
> +}
> +
> +static int new_cop_pid(struct ida *ida, int min_id, int max_id,
> + spinlock_t *lock)
> +{
> + int index;
> + int err;
> +
> +again:
> + if (!ida_pre_get(ida, GFP_KERNEL))
> + return -ENOMEM;
> +
> + spin_lock(lock);
> + err = ida_get_new_above(ida, min_id, &index);
> + spin_unlock(lock);
> +
> + if (err == -EAGAIN)
> + goto again;
> + else if (err)
> + return err;
> +
> + if (index > max_id) {
> + spin_lock(lock);
> + ida_remove(ida, index);
> + spin_unlock(lock);
> + return -ENOMEM;
> + }
> +
> + return index;
> +}
> +
> +static void sync_cop(void *arg)
> +{
> + struct mm_struct *mm = arg;
> +
> + if (mm == current->active_mm)
> + switch_cop(current->active_mm);
> +}
> +
> +/**
> + * Start using a coprocessor.
> + * @acop: mask of coprocessor to be used.
> + * @mm: The mm the coprocessor to associate with. Most likely current mm.
> + *
> + * Return a positive PID if successful. Negative errno otherwise.
> + * The returned PID will be fed to the coprocessor to determine if an
> + * icswx transaction is authenticated.
> + */
> +int use_cop(unsigned long acop, struct mm_struct *mm)
> +{
> + int ret;
> +
> + if (!cpu_has_feature(CPU_FTR_ICSWX))
> + return -ENODEV;
> +
> + if (!mm || !acop)
> + return -EINVAL;
> +
> + /* We need to make sure mm_users doesn't change */
> + down_read(&mm->mmap_sem);
> + spin_lock(mm->context.cop_lockp);
> +
> + if (mm->context.cop_pid == COP_PID_NONE) {
> + ret = new_cop_pid(&cop_ida, COP_PID_MIN, COP_PID_MAX,
> + &mmu_context_acop_lock);
> + if (ret < 0)
> + goto out;
> +
> + mm->context.cop_pid = ret;
> + }
> + mm->context.acop |= acop;
> +
> + sync_cop(mm);
> +
> + /*
> + * If this is a threaded process then there might be other threads
> + * running. We need to send an IPI to force them to pick up any
> + * change in PID and ACOP.
> + */
> + if (atomic_read(&mm->mm_users) > 1)
> + smp_call_function(sync_cop, mm, 1);
> +
> + ret = mm->context.cop_pid;
> +
> +out:
> + spin_unlock(mm->context.cop_lockp);
> + up_read(&mm->mmap_sem);
> +
> + return ret;
> +}
> +EXPORT_SYMBOL_GPL(use_cop);
> +
> +/**
> + * Stop using a coprocessor.
> + * @acop: mask of coprocessor to be stopped.
> + * @mm: The mm the coprocessor associated with.
> + */
> +void drop_cop(unsigned long acop, struct mm_struct *mm)
> +{
> + int free_pid = COP_PID_NONE;
> +
> + if (!cpu_has_feature(CPU_FTR_ICSWX))
> + return;
> +
> + if (WARN_ON_ONCE(!mm))
> + return;
> +
> + /* We need to make sure mm_users doesn't change */
> + down_read(&mm->mmap_sem);
> + spin_lock(mm->context.cop_lockp);
> +
> + mm->context.acop &= ~acop;
> +
> + if ((!mm->context.acop) && (mm->context.cop_pid != COP_PID_NONE)) {
> + free_pid = mm->context.cop_pid;
> + mm->context.cop_pid = COP_PID_NONE;
> + }
> +
> + sync_cop(mm);
> +
> + /*
> + * If this is a threaded process then there might be other threads
> + * running. We need to send an IPI to force them to pick up any
> + * change in PID and ACOP.
> + */
> + if (atomic_read(&mm->mm_users) > 1)
> + smp_call_function(sync_cop, mm, 1);
> +
> + if (free_pid != COP_PID_NONE) {
> + spin_lock(&mmu_context_acop_lock);
> + ida_remove(&cop_ida, free_pid);
> + spin_unlock(&mmu_context_acop_lock);
> + }
> +
> + spin_unlock(mm->context.cop_lockp);
> + up_read(&mm->mmap_sem);
> +}
> +EXPORT_SYMBOL_GPL(drop_cop);
> +
> +#endif /* CONFIG_PPC_ICSWX */
> +
> static DEFINE_SPINLOCK(mmu_context_lock);
> static DEFINE_IDA(mmu_context_ida);
>
> @@ -78,6 +274,16 @@ int init_new_context(struct task_struct
> slice_set_user_psize(mm, mmu_virtual_psize);
> subpage_prot_init_new_context(mm);
> mm->context.id = index;
> +#ifdef CONFIG_PPC_ICSWX
> + mm->context.cop_lockp = kmalloc(sizeof(spinlock_t), GFP_KERNEL);
> + if (!mm->context.cop_lockp) {
> + __destroy_context(index);
> + subpage_prot_free(mm);
> + mm->context.id = NO_CONTEXT;
> + return -ENOMEM;
> + }
> + spin_lock_init(mm->context.cop_lockp);
> +#endif /* CONFIG_PPC_ICSWX */
>
> return 0;
> }
> @@ -92,6 +298,11 @@ EXPORT_SYMBOL_GPL(__destroy_context);
>
> void destroy_context(struct mm_struct *mm)
> {
> +#ifdef CONFIG_PPC_ICSWX
> + drop_cop(mm->context.acop, mm);
> + kfree(mm->context.cop_lockp);
> + mm->context.cop_lockp = NULL;
> +#endif /* CONFIG_PPC_ICSWX */
> __destroy_context(mm->context.id);
> subpage_prot_free(mm);
> mm->context.id = MMU_NO_CONTEXT;
> Index: linux-powerpc/arch/powerpc/platforms/Kconfig.cputype
> ===================================================================
> --- linux-powerpc.orig/arch/powerpc/platforms/Kconfig.cputype 2011-05-03 16:39:21.809141492 +1000
> +++ linux-powerpc/arch/powerpc/platforms/Kconfig.cputype 2011-05-03 16:39:26.059220272 +1000
> @@ -230,6 +230,24 @@ config VSX
>
> If in doubt, say Y here.
>
> +config PPC_ICSWX
> + bool "Support for PowerPC icswx coprocessor instruction"
> + depends on POWER4
> + default n
> + ---help---
> +
> + This option enables kernel support for the PowerPC Initiate
> + Coprocessor Store Word (icswx) coprocessor instruction on POWER7
> + or newer processors.
> +
> + This option is only useful if you have a processor that supports
> + the icswx coprocessor instruction. It does not have any effect
> + on processors without the icswx coprocessor instruction.
> +
> + This option slightly increases kernel memory usage.
> +
> + If in doubt, say N here.
> +
> config SPE
> bool "SPE Support"
> depends on E200 || (E500 && !PPC_E500MC)
More information about the Linuxppc-dev
mailing list