cachepc-linux

Fork of AMDESE/linux with modifications for CachePC side-channel attack
git clone https://git.sinitax.com/sinitax/cachepc-linux
Log | Files | Refs | README | LICENSE | sfeed.txt

this_cpu_ops.rst (11458B)


      1===================
      2this_cpu operations
      3===================
      4
      5:Author: Christoph Lameter, August 4th, 2014
      6:Author: Pranith Kumar, Aug 2nd, 2014
      7
      8this_cpu operations are a way of optimizing access to per cpu
      9variables associated with the *currently* executing processor. This is
     10done through the use of segment registers (or a dedicated register where
     11the cpu permanently stored the beginning of the per cpu	area for a
     12specific processor).
     13
     14this_cpu operations add a per cpu variable offset to the processor
     15specific per cpu base and encode that operation in the instruction
     16operating on the per cpu variable.
     17
     18This means that there are no atomicity issues between the calculation of
     19the offset and the operation on the data. Therefore it is not
     20necessary to disable preemption or interrupts to ensure that the
     21processor is not changed between the calculation of the address and
     22the operation on the data.
     23
     24Read-modify-write operations are of particular interest. Frequently
     25processors have special lower latency instructions that can operate
     26without the typical synchronization overhead, but still provide some
     27sort of relaxed atomicity guarantees. The x86, for example, can execute
     28RMW (Read Modify Write) instructions like inc/dec/cmpxchg without the
     29lock prefix and the associated latency penalty.
     30
     31Access to the variable without the lock prefix is not synchronized but
     32synchronization is not necessary since we are dealing with per cpu
     33data specific to the currently executing processor. Only the current
     34processor should be accessing that variable and therefore there are no
     35concurrency issues with other processors in the system.
     36
     37Please note that accesses by remote processors to a per cpu area are
     38exceptional situations and may impact performance and/or correctness
     39(remote write operations) of local RMW operations via this_cpu_*.
     40
     41The main use of the this_cpu operations has been to optimize counter
     42operations.
     43
     44The following this_cpu() operations with implied preemption protection
     45are defined. These operations can be used without worrying about
     46preemption and interrupts::
     47
     48	this_cpu_read(pcp)
     49	this_cpu_write(pcp, val)
     50	this_cpu_add(pcp, val)
     51	this_cpu_and(pcp, val)
     52	this_cpu_or(pcp, val)
     53	this_cpu_add_return(pcp, val)
     54	this_cpu_xchg(pcp, nval)
     55	this_cpu_cmpxchg(pcp, oval, nval)
     56	this_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
     57	this_cpu_sub(pcp, val)
     58	this_cpu_inc(pcp)
     59	this_cpu_dec(pcp)
     60	this_cpu_sub_return(pcp, val)
     61	this_cpu_inc_return(pcp)
     62	this_cpu_dec_return(pcp)
     63
     64
     65Inner working of this_cpu operations
     66------------------------------------
     67
     68On x86 the fs: or the gs: segment registers contain the base of the
     69per cpu area. It is then possible to simply use the segment override
     70to relocate a per cpu relative address to the proper per cpu area for
     71the processor. So the relocation to the per cpu base is encoded in the
     72instruction via a segment register prefix.
     73
     74For example::
     75
     76	DEFINE_PER_CPU(int, x);
     77	int z;
     78
     79	z = this_cpu_read(x);
     80
     81results in a single instruction::
     82
     83	mov ax, gs:[x]
     84
     85instead of a sequence of calculation of the address and then a fetch
     86from that address which occurs with the per cpu operations. Before
     87this_cpu_ops such sequence also required preempt disable/enable to
     88prevent the kernel from moving the thread to a different processor
     89while the calculation is performed.
     90
     91Consider the following this_cpu operation::
     92
     93	this_cpu_inc(x)
     94
     95The above results in the following single instruction (no lock prefix!)::
     96
     97	inc gs:[x]
     98
     99instead of the following operations required if there is no segment
    100register::
    101
    102	int *y;
    103	int cpu;
    104
    105	cpu = get_cpu();
    106	y = per_cpu_ptr(&x, cpu);
    107	(*y)++;
    108	put_cpu();
    109
    110Note that these operations can only be used on per cpu data that is
    111reserved for a specific processor. Without disabling preemption in the
    112surrounding code this_cpu_inc() will only guarantee that one of the
    113per cpu counters is correctly incremented. However, there is no
    114guarantee that the OS will not move the process directly before or
    115after the this_cpu instruction is executed. In general this means that
    116the value of the individual counters for each processor are
    117meaningless. The sum of all the per cpu counters is the only value
    118that is of interest.
    119
    120Per cpu variables are used for performance reasons. Bouncing cache
    121lines can be avoided if multiple processors concurrently go through
    122the same code paths.  Since each processor has its own per cpu
    123variables no concurrent cache line updates take place. The price that
    124has to be paid for this optimization is the need to add up the per cpu
    125counters when the value of a counter is needed.
    126
    127
    128Special operations
    129------------------
    130
    131::
    132
    133	y = this_cpu_ptr(&x)
    134
    135Takes the offset of a per cpu variable (&x !) and returns the address
    136of the per cpu variable that belongs to the currently executing
    137processor.  this_cpu_ptr avoids multiple steps that the common
    138get_cpu/put_cpu sequence requires. No processor number is
    139available. Instead, the offset of the local per cpu area is simply
    140added to the per cpu offset.
    141
    142Note that this operation is usually used in a code segment when
    143preemption has been disabled. The pointer is then used to
    144access local per cpu data in a critical section. When preemption
    145is re-enabled this pointer is usually no longer useful since it may
    146no longer point to per cpu data of the current processor.
    147
    148
    149Per cpu variables and offsets
    150-----------------------------
    151
    152Per cpu variables have *offsets* to the beginning of the per cpu
    153area. They do not have addresses although they look like that in the
    154code. Offsets cannot be directly dereferenced. The offset must be
    155added to a base pointer of a per cpu area of a processor in order to
    156form a valid address.
    157
    158Therefore the use of x or &x outside of the context of per cpu
    159operations is invalid and will generally be treated like a NULL
    160pointer dereference.
    161
    162::
    163
    164	DEFINE_PER_CPU(int, x);
    165
    166In the context of per cpu operations the above implies that x is a per
    167cpu variable. Most this_cpu operations take a cpu variable.
    168
    169::
    170
    171	int __percpu *p = &x;
    172
    173&x and hence p is the *offset* of a per cpu variable. this_cpu_ptr()
    174takes the offset of a per cpu variable which makes this look a bit
    175strange.
    176
    177
    178Operations on a field of a per cpu structure
    179--------------------------------------------
    180
    181Let's say we have a percpu structure::
    182
    183	struct s {
    184		int n,m;
    185	};
    186
    187	DEFINE_PER_CPU(struct s, p);
    188
    189
    190Operations on these fields are straightforward::
    191
    192	this_cpu_inc(p.m)
    193
    194	z = this_cpu_cmpxchg(p.m, 0, 1);
    195
    196
    197If we have an offset to struct s::
    198
    199	struct s __percpu *ps = &p;
    200
    201	this_cpu_dec(ps->m);
    202
    203	z = this_cpu_inc_return(ps->n);
    204
    205
    206The calculation of the pointer may require the use of this_cpu_ptr()
    207if we do not make use of this_cpu ops later to manipulate fields::
    208
    209	struct s *pp;
    210
    211	pp = this_cpu_ptr(&p);
    212
    213	pp->m--;
    214
    215	z = pp->n++;
    216
    217
    218Variants of this_cpu ops
    219------------------------
    220
    221this_cpu ops are interrupt safe. Some architectures do not support
    222these per cpu local operations. In that case the operation must be
    223replaced by code that disables interrupts, then does the operations
    224that are guaranteed to be atomic and then re-enable interrupts. Doing
    225so is expensive. If there are other reasons why the scheduler cannot
    226change the processor we are executing on then there is no reason to
    227disable interrupts. For that purpose the following __this_cpu operations
    228are provided.
    229
    230These operations have no guarantee against concurrent interrupts or
    231preemption. If a per cpu variable is not used in an interrupt context
    232and the scheduler cannot preempt, then they are safe. If any interrupts
    233still occur while an operation is in progress and if the interrupt too
    234modifies the variable, then RMW actions can not be guaranteed to be
    235safe::
    236
    237	__this_cpu_read(pcp)
    238	__this_cpu_write(pcp, val)
    239	__this_cpu_add(pcp, val)
    240	__this_cpu_and(pcp, val)
    241	__this_cpu_or(pcp, val)
    242	__this_cpu_add_return(pcp, val)
    243	__this_cpu_xchg(pcp, nval)
    244	__this_cpu_cmpxchg(pcp, oval, nval)
    245	__this_cpu_cmpxchg_double(pcp1, pcp2, oval1, oval2, nval1, nval2)
    246	__this_cpu_sub(pcp, val)
    247	__this_cpu_inc(pcp)
    248	__this_cpu_dec(pcp)
    249	__this_cpu_sub_return(pcp, val)
    250	__this_cpu_inc_return(pcp)
    251	__this_cpu_dec_return(pcp)
    252
    253
    254Will increment x and will not fall-back to code that disables
    255interrupts on platforms that cannot accomplish atomicity through
    256address relocation and a Read-Modify-Write operation in the same
    257instruction.
    258
    259
    260&this_cpu_ptr(pp)->n vs this_cpu_ptr(&pp->n)
    261--------------------------------------------
    262
    263The first operation takes the offset and forms an address and then
    264adds the offset of the n field. This may result in two add
    265instructions emitted by the compiler.
    266
    267The second one first adds the two offsets and then does the
    268relocation.  IMHO the second form looks cleaner and has an easier time
    269with (). The second form also is consistent with the way
    270this_cpu_read() and friends are used.
    271
    272
    273Remote access to per cpu data
    274------------------------------
    275
    276Per cpu data structures are designed to be used by one cpu exclusively.
    277If you use the variables as intended, this_cpu_ops() are guaranteed to
    278be "atomic" as no other CPU has access to these data structures.
    279
    280There are special cases where you might need to access per cpu data
    281structures remotely. It is usually safe to do a remote read access
    282and that is frequently done to summarize counters. Remote write access
    283something which could be problematic because this_cpu ops do not
    284have lock semantics. A remote write may interfere with a this_cpu
    285RMW operation.
    286
    287Remote write accesses to percpu data structures are highly discouraged
    288unless absolutely necessary. Please consider using an IPI to wake up
    289the remote CPU and perform the update to its per cpu area.
    290
    291To access per-cpu data structure remotely, typically the per_cpu_ptr()
    292function is used::
    293
    294
    295	DEFINE_PER_CPU(struct data, datap);
    296
    297	struct data *p = per_cpu_ptr(&datap, cpu);
    298
    299This makes it explicit that we are getting ready to access a percpu
    300area remotely.
    301
    302You can also do the following to convert the datap offset to an address::
    303
    304	struct data *p = this_cpu_ptr(&datap);
    305
    306but, passing of pointers calculated via this_cpu_ptr to other cpus is
    307unusual and should be avoided.
    308
    309Remote access are typically only for reading the status of another cpus
    310per cpu data. Write accesses can cause unique problems due to the
    311relaxed synchronization requirements for this_cpu operations.
    312
    313One example that illustrates some concerns with write operations is
    314the following scenario that occurs because two per cpu variables
    315share a cache-line but the relaxed synchronization is applied to
    316only one process updating the cache-line.
    317
    318Consider the following example::
    319
    320
    321	struct test {
    322		atomic_t a;
    323		int b;
    324	};
    325
    326	DEFINE_PER_CPU(struct test, onecacheline);
    327
    328There is some concern about what would happen if the field 'a' is updated
    329remotely from one processor and the local processor would use this_cpu ops
    330to update field b. Care should be taken that such simultaneous accesses to
    331data within the same cache line are avoided. Also costly synchronization
    332may be necessary. IPIs are generally recommended in such scenarios instead
    333of a remote write to the per cpu area of another processor.
    334
    335Even in cases where the remote writes are rare, please bear in
    336mind that a remote write will evict the cache line from the processor
    337that most likely will access it. If the processor wakes up and finds a
    338missing local cache line of a per cpu area, its performance and hence
    339the wake up times will be affected.