Re: [HMM v13 08/18] mm/hmm: heterogeneous memory management (HMM for short)

[Jerome Glisse]

On Mon, Nov 21, 2016 at 01:29:23PM +1100, Balbir Singh wrote:
> On 19/11/16 05:18, Jérôme Glisse wrote:
> > HMM provides 3 separate functionality :
> >     - Mirroring: synchronize CPU page table and device page table
> >     - Device memory: allocating struct page for device memory
> >     - Migration: migrating regular memory to device memory
> > 
> > This patch introduces some common helpers and definitions to all of
> > those 3 functionality.
> > 

[...]

> > +/*
> > + * HMM provides 3 separate functionality :
> > + *   - Mirroring: synchronize CPU page table and device page table
> > + *   - Device memory: allocating struct page for device memory
> > + *   - Migration: migrating regular memory to device memory
> > + *
> > + * Each can be use independently from the others.
> > + *
> > + *
> > + * Mirroring:
> > + *
> > + * HMM provide helpers to mirror process address space on a device. For this it
> > + * provides several helpers to order device page table update in respect to CPU
> > + * page table update. Requirement is that for any given virtual address the CPU
> > + * and device page table can not point to different physical page. It uses the
> > + * mmu_notifier API and introduce virtual address range lock which block CPU
> > + * page table update for a range while the device page table is being updated.
> > + * Usage pattern is:
> > + *
> > + *      hmm_vma_range_lock(vma, start, end);
> > + *      // snap shot CPU page table
> > + *      // update device page table from snapshot
> > + *      hmm_vma_range_unlock(vma, start, end);
> > + *
> > + * Any CPU page table update that conflict with a range lock will wait until
> > + * range is unlock. This garanty proper serialization of CPU and device page
> > + * table update.
> > + *
> > + *
> > + * Device memory:
> > + *
> > + * HMM provides helpers to help leverage device memory either addressable like
> > + * regular memory by the CPU or un-addressable at all. In both case the device
> > + * memory is associated to dedicated structs page (which are allocated like for
> > + * hotplug memory). Device memory management is under the responsability of the
> > + * device driver. HMM only allocate and initialize the struct pages associated
> > + * with the device memory.
> > + *
> > + * Allocating struct page for device memory allow to use device memory allmost
> > + * like any regular memory. Unlike regular memory it can not be added to the
> > + * lru, nor can any memory allocation can use device memory directly. Device
> > + * memory will only end up to be use in a process if device driver migrate some
> 				   in use 
> > + * of the process memory from regular memory to device memory.
> > + *
> 
> A process can never directly allocate device memory?

Well yes and no, if the device driver is first to trigger a page fault on some
memory then it can decide to directly allocate device memory. But usual CPU page
fault would not trigger allocation of device memory. A new mechanism can be added
to achieve that if that make sense but for my main target (x86/pcie) it does not.

> > + *
> > + * Migration:
> > + *
> > + * Existing memory migration mechanism (mm/migrate.c) does not allow to use
> > + * something else than the CPU to copy from source to destination memory. More
> > + * over existing code is not tailor to drive migration from process virtual
> 				tailored
> > + * address rather than from list of pages. Finaly the migration flow does not
> 					      Finally 
> > + * allow for graceful failure at different step of the migration process.
> > + *
> > + * HMM solves all of the above though simple API :
> > + *
> > + *      hmm_vma_migrate(vma, start, end, ops);
> > + *
> > + * With ops struct providing 2 callback alloc_and_copy() which allocated the
> > + * destination memory and initialize it using source memory. Migration can fail
> > + * after this step and thus last callback finalize_and_map() allow the device
> > + * driver to know which page were successfully migrated and which were not.
> > + *
> > + * This can easily be use outside of HMM intended use case.
> > + *
> 
> I think it is a good API to have
> 
> > + *
> > + * This header file contain all the API related to this 3 functionality and
> > + * each functions and struct are more thouroughly documented in below comments.
> > + */
> > +#ifndef LINUX_HMM_H
> > +#define LINUX_HMM_H
> > +
> > +#include <linux/kconfig.h>
> > +
> > +#if IS_ENABLED(CONFIG_HMM)
> > +
> > +
> > +/*
> > + * hmm_pfn_t - HMM use its own pfn type to keep several flags per page
> 		      uses
> > + *
> > + * Flags:
> > + * HMM_PFN_VALID: pfn is valid
> > + * HMM_PFN_WRITE: CPU page table have the write permission set
> 				    has
> > + */
> > +typedef unsigned long hmm_pfn_t;
> > +
> > +#define HMM_PFN_VALID (1 << 0)
> > +#define HMM_PFN_WRITE (1 << 1)
> > +#define HMM_PFN_SHIFT 2
> > +
> > +static inline struct page *hmm_pfn_to_page(hmm_pfn_t pfn)
> > +{
> > +	if (!(pfn & HMM_PFN_VALID))
> > +		return NULL;
> > +	return pfn_to_page(pfn >> HMM_PFN_SHIFT);
> > +}
> > +
> > +static inline unsigned long hmm_pfn_to_pfn(hmm_pfn_t pfn)
> > +{
> > +	if (!(pfn & HMM_PFN_VALID))
> > +		return -1UL;
> > +	return (pfn >> HMM_PFN_SHIFT);
> > +}
> > +
> 
> What is pfn_to_pfn? I presume it means CPU PFN to device PFN
> or is it the reverse? Please add some comments

It is hmm_pfn_t to pfn value as an unsigned long. The memory the pfn
points to can be anything (regular system memory, device memory, ...).

hmm_pfn_t is just a pfn with a set of flags.

> 
> > +static inline hmm_pfn_t hmm_pfn_from_page(struct page *page)
> > +{
> > +	return (page_to_pfn(page) << HMM_PFN_SHIFT) | HMM_PFN_VALID;
> > +}
> > +
> > +static inline hmm_pfn_t hmm_pfn_from_pfn(unsigned long pfn)
> > +{
> > +	return (pfn << HMM_PFN_SHIFT) | HMM_PFN_VALID;
> > +}
> > +
> 
> Same as above
> 
> > +
> > +/* Below are for HMM internal use only ! Not to be use by device driver ! */
> > +void hmm_mm_destroy(struct mm_struct *mm);
> > +
> > +#else /* IS_ENABLED(CONFIG_HMM) */
> > +
> > +/* Below are for HMM internal use only ! Not to be use by device driver ! */
> > +static inline void hmm_mm_destroy(struct mm_struct *mm) {}
> > +
> > +#endif /* IS_ENABLED(CONFIG_HMM) */
> > +#endif /* LINUX_HMM_H */
> > diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> > index 4a8aced..4effdbf 100644
> > --- a/include/linux/mm_types.h
> > +++ b/include/linux/mm_types.h
> > @@ -23,6 +23,7 @@
> >  
> >  struct address_space;
> >  struct mem_cgroup;
> > +struct hmm;
> >  
> >  #define USE_SPLIT_PTE_PTLOCKS	(NR_CPUS >= CONFIG_SPLIT_PTLOCK_CPUS)
> >  #define USE_SPLIT_PMD_PTLOCKS	(USE_SPLIT_PTE_PTLOCKS && \
> > @@ -516,6 +517,10 @@ struct mm_struct {
> >  	atomic_long_t hugetlb_usage;
> >  #endif
> >  	struct work_struct async_put_work;
> > +#if IS_ENABLED(CONFIG_HMM)
> > +	/* HMM need to track few things per mm */
> > +	struct hmm *hmm;
> > +#endif
> >  };
> >  
> >  static inline void mm_init_cpumask(struct mm_struct *mm)
> > diff --git a/kernel/fork.c b/kernel/fork.c
> > index 690a1aad..af0eec8 100644
> > --- a/kernel/fork.c
> > +++ b/kernel/fork.c
> > @@ -27,6 +27,7 @@
> >  #include <linux/binfmts.h>
> >  #include <linux/mman.h>
> >  #include <linux/mmu_notifier.h>
> > +#include <linux/hmm.h>
> >  #include <linux/fs.h>
> >  #include <linux/mm.h>
> >  #include <linux/vmacache.h>
> > @@ -702,6 +703,7 @@ void __mmdrop(struct mm_struct *mm)
> >  	BUG_ON(mm == &init_mm);
> >  	mm_free_pgd(mm);
> >  	destroy_context(mm);
> > +	hmm_mm_destroy(mm);
> >  	mmu_notifier_mm_destroy(mm);
> >  	check_mm(mm);
> >  	free_mm(mm);
> > diff --git a/mm/Kconfig b/mm/Kconfig
> > index 0a21411..be18cc2 100644
> > --- a/mm/Kconfig
> > +++ b/mm/Kconfig
> > @@ -289,6 +289,17 @@ config MIGRATION
> >  config ARCH_ENABLE_HUGEPAGE_MIGRATION
> >  	bool
> >  
> > +config HMM
> > +	bool "Heterogeneous memory management (HMM)"
> > +	depends on MMU
> > +	default n
> > +	help
> > +	  Heterogeneous memory management, set of helpers for:
> > +	    - mirroring of process address space on a device
> > +	    - using device memory transparently inside a process
> > +
> > +	  If unsure, say N to disable HMM.
> > +
> 
> It would be nice to split this into HMM, HMM_MIGRATE and HMM_MIRROR
> 
> >  config PHYS_ADDR_T_64BIT
> >  	def_bool 64BIT || ARCH_PHYS_ADDR_T_64BIT
> >  
> > diff --git a/mm/Makefile b/mm/Makefile
> > index 2ca1faf..6ac1284 100644
> > --- a/mm/Makefile
> > +++ b/mm/Makefile
> > @@ -76,6 +76,7 @@ obj-$(CONFIG_FAILSLAB) += failslab.o
> >  obj-$(CONFIG_MEMORY_HOTPLUG) += memory_hotplug.o
> >  obj-$(CONFIG_MEMTEST)		+= memtest.o
> >  obj-$(CONFIG_MIGRATION) += migrate.o
> > +obj-$(CONFIG_HMM) += hmm.o
> >  obj-$(CONFIG_QUICKLIST) += quicklist.o
> >  obj-$(CONFIG_TRANSPARENT_HUGEPAGE) += huge_memory.o khugepaged.o
> >  obj-$(CONFIG_PAGE_COUNTER) += page_counter.o
> > diff --git a/mm/hmm.c b/mm/hmm.c
> > new file mode 100644
> > index 0000000..342b596
> > --- /dev/null
> > +++ b/mm/hmm.c
> > @@ -0,0 +1,86 @@
> > +/*
> > + * Copyright 2013 Red Hat Inc.
> > + *
> > + * 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.
> > + *
> > + * This program is distributed in the hope that it will be useful,
> > + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> > + * GNU General Public License for more details.
> > + *
> > + * Authors: Jérôme Glisse <jglisse@xxxxxxxxxx>
> > + */
> > +/*
> > + * Refer to include/linux/hmm.h for informations about heterogeneous memory
> > + * management or HMM for short.
> > + */
> > +#include <linux/mm.h>
> > +#include <linux/hmm.h>
> > +#include <linux/slab.h>
> > +#include <linux/sched.h>
> > +
> > +/*
> > + * struct hmm - HMM per mm struct
> > + *
> > + * @mm: mm struct this HMM struct is bound to
> > + */
> > +struct hmm {
> > +	struct mm_struct	*mm;
> > +};
> > +
> > +/*
> > + * hmm_register - register HMM against an mm (HMM internal)
> > + *
> > + * @mm: mm struct to attach to
> > + *
> > + * This is not intended to be use directly by device driver but by other HMM
> > + * component. It allocates an HMM struct if mm does not have one and initialize
> > + * it.
> > + */
> > +static struct hmm *hmm_register(struct mm_struct *mm)
> > +{
> > +	struct hmm *hmm = NULL;
> > +
> > +	if (!mm->hmm) {
> > +		hmm = kmalloc(sizeof(*hmm), GFP_KERNEL);
> > +		if (!hmm)
> > +			return NULL;
> > +		hmm->mm = mm;
> > +	}
> > +
> > +	spin_lock(&mm->page_table_lock);
> > +	if (!mm->hmm)
> > +		/*
> > +		 * The hmm struct can only be free once mm_struct goes away
> > +		 * hence we should always have pre-allocated an new hmm struct
> > +		 * above.
> > +		 */
> > +		mm->hmm = hmm;
> > +	else if (hmm)
> > +		kfree(hmm);
> > +	hmm = mm->hmm;
> > +	spin_unlock(&mm->page_table_lock);
> > +
> > +	return hmm;
> > +}
> > +
> > +void hmm_mm_destroy(struct mm_struct *mm)
> > +{
> > +	struct hmm *hmm;
> > +
> > +	/*
> > +	 * We should not need to lock here as no one should be able to register
> > +	 * a new HMM while an mm is being destroy. But just to be safe ...
> > +	 */
> > +	spin_lock(&mm->page_table_lock);
> > +	hmm = mm->hmm;
> > +	mm->hmm = NULL;
> > +	spin_unlock(&mm->page_table_lock);
> > +	if (!hmm)
> > +		return;
> > +
> 
> kfree can deal with NULL pointers, you can remove the if check

Yeah.

Cheers,
Jérôme