[PATCH 1/3 v2] powerpc: Split ICSWX ACOP and PID processing

Benjamin Herrenschmidt benh at kernel.crashing.org
Thu Sep 29 17:02:53 EST 2011


On Tue, 2011-09-27 at 20:05 +1000, Anton Blanchard wrote:
> Hi Jimi,
> 
> > Some processors, like embedded, that already have a PID register that
> > is managed by the system.  This patch separates the ACOP and PID
> > processing into separate files so that the ACOP code can be shared.
> > 
> > Signed-off-by: Jimi Xenidis <jimix at pobox.com>
> 
> Looks good.
> 
> Acked-by: Anton Blanchard <anton at samba.org>

Please, update the patch so that it applies :-)

(IE. Anton's patch to fix deadlocks in the icswx code broke it, you
need to rebase and apply Anton's fix to your new copy of the code)

Cheers,
Ben.

> Anton
> 
> > ---
> > Re: galak at kernel.crashing.org
> >     Fix typo in arch/powerpc/mm/Makefile
> > 
> > Re: anton at samba.org
> >     merge in: powerpc: Fix deadlock in icswx code
> > ---
> >  arch/powerpc/mm/Makefile               |    2 +
> >  arch/powerpc/mm/icswx.c                |  162
> > ++++++++++++++++++++++++++ arch/powerpc/mm/icswx.h                |
> > 34 ++++++ arch/powerpc/mm/icswx_pid.c            |   87 ++++++++++++++
> >  arch/powerpc/mm/mmu_context_hash64.c   |  195
> > --------------------------------
> > arch/powerpc/platforms/Kconfig.cputype |   10 ++- 6 files changed,
> > 294 insertions(+), 196 deletions(-) create mode 100644
> > arch/powerpc/mm/icswx.c create mode 100644 arch/powerpc/mm/icswx.h
> >  create mode 100644 arch/powerpc/mm/icswx_pid.c
> > 
> > diff --git a/arch/powerpc/mm/Makefile b/arch/powerpc/mm/Makefile
> > index bdca46e..fb7976f 100644
> > --- a/arch/powerpc/mm/Makefile
> > +++ b/arch/powerpc/mm/Makefile
> > @@ -21,6 +21,8 @@ obj-$(CONFIG_PPC_STD_MMU_32)	+= ppc_mmu_32.o
> >  obj-$(CONFIG_PPC_STD_MMU)	+= hash_low_$(CONFIG_WORD_SIZE).o \
> >  				   tlb_hash$(CONFIG_WORD_SIZE).o \
> >  				   mmu_context_hash$(CONFIG_WORD_SIZE).o
> > +obj-$(CONFIG_PPC_ICSWX)		+= icswx.o
> > +obj-$(CONFIG_PPC_ICSWX_PID)	+= icswx_pid.o
> >  obj-$(CONFIG_40x)		+= 40x_mmu.o
> >  obj-$(CONFIG_44x)		+= 44x_mmu.o
> >  obj-$(CONFIG_PPC_FSL_BOOK3E)	+= fsl_booke_mmu.o
> > diff --git a/arch/powerpc/mm/icswx.c b/arch/powerpc/mm/icswx.c
> > new file mode 100644
> > index 0000000..2f1dd29
> > --- /dev/null
> > +++ b/arch/powerpc/mm/icswx.c
> > @@ -0,0 +1,162 @@
> > +/*
> > + *  ICSWX and ACOP Management
> > + *
> > + *  Copyright (C) 2011 Anton Blanchard, IBM Corp. <anton at samba.org>
> > + *
> > + *  This program is free software; you can redistribute it and/or
> > + *  modify it under the terms of the GNU General Public License
> > + *  as published by the Free Software Foundation; either version
> > + *  2 of the License, or (at your option) any later version.
> > + *
> > + */
> > +
> > +#include <linux/sched.h>
> > +#include <linux/kernel.h>
> > +#include <linux/errno.h>
> > +#include <linux/types.h>
> > +#include <linux/mm.h>
> > +#include <linux/spinlock.h>
> > +#include <linux/module.h>
> > +#include "icswx.h"
> > +
> > +
> > +/*
> > + * 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 cache block or less, typically 64 or 128 bytes in
> > + * size, and is placed in cacheable memory on a 128-byte boundary
> > + * _always_.
> > + *
> > + * The task to use a coprocessor should use use_cop() mark the use of
> > + * the coprocessor type (CT) and context swithing.  On a server
> > + * processor the PID register is used only for coprocessor management
> > + * and so a coprocessor PID is allocated before executing icswx
> > + * instruction. Drop_cop() is used to free the resources created by
> > + * use_cop().
> > + *
> > + * 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.
> > + */
> > +
> > +void switch_cop(struct mm_struct *next)
> > +{
> > +#ifdef CONFIG_ICSWX_PID
> > +	mtspr(SPRN_PID, next->context.cop_pid);
> > +#endif
> > +	mtspr(SPRN_ACOP, next->context.acop);
> > +}
> > +
> > +/**
> > + * 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;
> > +
> > +	/* The page_table_lock ensures mm_users won't change under
> > us */
> > +	spin_lock(&mm->page_table_lock);
> > +	spin_lock(mm->context.cop_lockp);
> > +
> > +	ret = get_cop_pid(mm);
> > +	if (ret < 0)
> > +		goto out;
> > +
> > +	/* update acop */
> > +	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);
> > +
> > +out:
> > +	spin_unlock(mm->context.cop_lockp);
> > +	spin_unlock(&mm->page_table_lock);
> > +
> > +	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;
> > +
> > +	if (!cpu_has_feature(CPU_FTR_ICSWX))
> > +		return;
> > +
> > +	if (WARN_ON_ONCE(!mm))
> > +		return;
> > +
> > +	/* The page_table_lock ensures mm_users won't change under
> > us */
> > +	spin_lock(&mm->page_table_lock);
> > +	spin_lock(mm->context.cop_lockp);
> > +
> > +	mm->context.acop &= ~acop;
> > +
> > +	free_pid = disable_cop_pid(mm);
> > +	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)
> > +		free_cop_pid(free_pid);
> > +
> > +	spin_unlock(mm->context.cop_lockp);
> > +	spin_unlock(&mm->page_table_lock);
> > +}
> > +EXPORT_SYMBOL_GPL(drop_cop);
> > diff --git a/arch/powerpc/mm/icswx.h b/arch/powerpc/mm/icswx.h
> > new file mode 100644
> > index 0000000..5121ddd
> > --- /dev/null
> > +++ b/arch/powerpc/mm/icswx.h
> > @@ -0,0 +1,34 @@
> > +/*
> > + *  ICSWX and ACOP Management
> > + *
> > + *  Copyright (C) 2011 Anton Blanchard, IBM Corp. <anton at samba.org>
> > + *
> > + *  This program is free software; you can redistribute it and/or
> > + *  modify it under the terms of the GNU General Public License
> > + *  as published by the Free Software Foundation; either version
> > + *  2 of the License, or (at your option) any later version.
> > + *
> > + */
> > +
> > +#include <asm/mmu_context.h>
> > +
> > +/* also used to denote that PIDs are not used */
> > +#define COP_PID_NONE 0
> > +
> > +static inline void sync_cop(void *arg)
> > +{
> > +	struct mm_struct *mm = arg;
> > +
> > +	if (mm == current->active_mm)
> > +		switch_cop(current->active_mm);
> > +}
> > +
> > +#ifdef CONFIG_PPC_ICSWX_PID
> > +extern int get_cop_pid(struct mm_struct *mm);
> > +extern int disable_cop_pid(struct mm_struct *mm);
> > +extern void free_cop_pid(int free_pid);
> > +#else
> > +#define get_cop_pid(m) (COP_PID_NONE)
> > +#define disable_cop_pid(m) (COP_PID_NONE)
> > +#define free_cop_pid(p)
> > +#endif
> > diff --git a/arch/powerpc/mm/icswx_pid.c b/arch/powerpc/mm/icswx_pid.c
> > new file mode 100644
> > index 0000000..91e30eb
> > --- /dev/null
> > +++ b/arch/powerpc/mm/icswx_pid.c
> > @@ -0,0 +1,87 @@
> > +/*
> > + *  ICSWX and ACOP/PID Management
> > + *
> > + *  Copyright (C) 2011 Anton Blanchard, IBM Corp. <anton at samba.org>
> > + *
> > + *  This program is free software; you can redistribute it and/or
> > + *  modify it under the terms of the GNU General Public License
> > + *  as published by the Free Software Foundation; either version
> > + *  2 of the License, or (at your option) any later version.
> > + *
> > + */
> > +
> > +#include <linux/sched.h>
> > +#include <linux/kernel.h>
> > +#include <linux/errno.h>
> > +#include <linux/types.h>
> > +#include <linux/mm.h>
> > +#include <linux/spinlock.h>
> > +#include <linux/idr.h>
> > +#include <linux/module.h>
> > +#include "icswx.h"
> > +
> > +#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);
> > +
> > +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;
> > +}
> > +
> > +int get_cop_pid(struct mm_struct *mm)
> > +{
> > +	int pid;
> > +
> > +	if (mm->context.cop_pid == COP_PID_NONE) {
> > +		pid = new_cop_pid(&cop_ida, COP_PID_MIN, COP_PID_MAX,
> > +				  &mmu_context_acop_lock);
> > +		if (pid >= 0)
> > +			mm->context.cop_pid = pid;
> > +	}
> > +	return mm->context.cop_pid;
> > +}
> > +
> > +int disable_cop_pid(struct mm_struct *mm)
> > +{
> > +	int free_pid = COP_PID_NONE;
> > +
> > +	if ((!mm->context.acop) && (mm->context.cop_pid !=
> > COP_PID_NONE)) {
> > +		free_pid = mm->context.cop_pid;
> > +		mm->context.cop_pid = COP_PID_NONE;
> > +	}
> > +	return free_pid;
> > +}
> > +
> > +void free_cop_pid(int free_pid)
> > +{
> > +	spin_lock(&mmu_context_acop_lock);
> > +	ida_remove(&cop_ida, free_pid);
> > +	spin_unlock(&mmu_context_acop_lock);
> > +}
> > diff --git a/arch/powerpc/mm/mmu_context_hash64.c
> > b/arch/powerpc/mm/mmu_context_hash64.c index 3bafc3d..a75832c 100644
> > --- a/arch/powerpc/mm/mmu_context_hash64.c
> > +++ b/arch/powerpc/mm/mmu_context_hash64.c
> > @@ -24,201 +24,6 @@
> >  
> >  #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);
> >  
> > diff --git a/arch/powerpc/platforms/Kconfig.cputype
> > b/arch/powerpc/platforms/Kconfig.cputype index e06e395..3cd22e5 100644
> > --- a/arch/powerpc/platforms/Kconfig.cputype
> > +++ b/arch/powerpc/platforms/Kconfig.cputype
> > @@ -234,7 +234,7 @@ config VSX
> >  
> >  config PPC_ICSWX
> >  	bool "Support for PowerPC icswx coprocessor instruction"
> > -	depends on POWER4
> > +	depends on POWER4 || PPC_A2
> >  	default n
> >  	---help---
> >  
> > @@ -250,6 +250,14 @@ config PPC_ICSWX
> >  
> >  	  If in doubt, say N here.
> >  
> > +config PPC_ICSWX_PID
> > +	bool "icswx requires direct PID management"
> > +	depends on PPC_ICSWX && POWER4
> > +	default y
> > +	---help---
> > +	  PID register in server is used explicitly for ICSWX.  In
> > +	  embedded systems PID managment is done by the system.
> > +
> >  config SPE
> >  	bool "SPE Support"
> >  	depends on E200 || (E500 && !PPC_E500MC)




More information about the Linuxppc-dev mailing list