1On Fri, 2 Jan 1998, Doug Ledford wrote: 2> 3> I'm working on making the aic7xxx driver more SMP friendly (as well as 4> importing the latest FreeBSD sequencer code to have 7895 support) and wanted 5> to get some info from you. The goal here is to make the various routines 6> SMP safe as well as UP safe during interrupts and other manipulating 7> routines. So far, I've added a spin_lock variable to things like my queue 8> structs. Now, from what I recall, there are some spin lock functions I can 9> use to lock these spin locks from other use as opposed to a (nasty) 10> save_flags(); cli(); stuff; restore_flags(); construct. Where do I find 11> these routines and go about making use of them? Do they only lock on a 12> per-processor basis or can they also lock say an interrupt routine from 13> mucking with a queue if the queue routine was manipulating it when the 14> interrupt occurred, or should I still use a cli(); based construct on that 15> one? 16 17See <asm/spinlock.h>. The basic version is: 18 19 spinlock_t xxx_lock = SPIN_LOCK_UNLOCKED; 20 21 22 unsigned long flags; 23 24 spin_lock_irqsave(&xxx_lock, flags); 25 ... critical section here .. 26 spin_unlock_irqrestore(&xxx_lock, flags); 27 28and the above is always safe. It will disable interrupts _locally_, but the 29spinlock itself will guarantee the global lock, so it will guarantee that 30there is only one thread-of-control within the region(s) protected by that 31lock. 32 33Note that it works well even under UP - the above sequence under UP 34essentially is just the same as doing a 35 36 unsigned long flags; 37 38 save_flags(flags); cli(); 39 ... critical section ... 40 restore_flags(flags); 41 42so the code does _not_ need to worry about UP vs SMP issues: the spinlocks 43work correctly under both (and spinlocks are actually more efficient on 44architectures that allow doing the "save_flags + cli" in one go because I 45don't export that interface normally). 46 47NOTE NOTE NOTE! The reason the spinlock is so much faster than a global 48interrupt lock under SMP is exactly because it disables interrupts only on 49the local CPU. The spin-lock is safe only when you _also_ use the lock 50itself to do locking across CPU's, which implies that EVERYTHING that 51touches a shared variable has to agree about the spinlock they want to 52use. 53 54The above is usually pretty simple (you usually need and want only one 55spinlock for most things - using more than one spinlock can make things a 56lot more complex and even slower and is usually worth it only for 57sequences that you _know_ need to be split up: avoid it at all cost if you 58aren't sure). HOWEVER, it _does_ mean that if you have some code that does 59 60 cli(); 61 .. critical section .. 62 sti(); 63 64and another sequence that does 65 66 spin_lock_irqsave(flags); 67 .. critical section .. 68 spin_unlock_irqrestore(flags); 69 70then they are NOT mutually exclusive, and the critical regions can happen 71at the same time on two different CPU's. That's fine per se, but the 72critical regions had better be critical for different things (ie they 73can't stomp on each other). 74 75The above is a problem mainly if you end up mixing code - for example the 76routines in ll_rw_block() tend to use cli/sti to protect the atomicity of 77their actions, and if a driver uses spinlocks instead then you should 78think about issues like the above.. 79 80This is really the only really hard part about spinlocks: once you start 81using spinlocks they tend to expand to areas you might not have noticed 82before, because you have to make sure the spinlocks correctly protect the 83shared data structures _everywhere_ they are used. The spinlocks are most 84easily added to places that are completely independent of other code (ie 85internal driver data structures that nobody else ever touches, for 86example). 87 88---- 89 90Lesson 2: reader-writer spinlocks. 91 92If your data accesses have a very natural pattern where you usually tend 93to mostly read from the shared variables, the reader-writer locks 94(rw_lock) versions of the spinlocks are often nicer. They allow multiple 95readers to be in the same critical region at once, but if somebody wants 96to change the variables it has to get an exclusive write lock. The 97routines look the same as above: 98 99 rwlock_t xxx_lock = RW_LOCK_UNLOCKED; 100 101 102 unsigned long flags; 103 104 read_lock_irqsave(&xxx_lock, flags); 105 .. critical section that only reads the info ... 106 read_unlock_irqrestore(&xxx_lock, flags); 107 108 write_lock_irqsave(&xxx_lock, flags); 109 .. read and write exclusive access to the info ... 110 write_unlock_irqrestore(&xxx_lock, flags); 111 112The above kind of lock is useful for complex data structures like linked 113lists etc, especially when you know that most of the work is to just 114traverse the list searching for entries without changing the list itself, 115for example. Then you can use the read lock for that kind of list 116traversal, which allows many concurrent readers. Anything that _changes_ 117the list will have to get the write lock. 118 119Note: you cannot "upgrade" a read-lock to a write-lock, so if you at _any_ 120time need to do any changes (even if you don't do it every time), you have 121to get the write-lock at the very beginning. I could fairly easily add a 122primitive to create a "upgradeable" read-lock, but it hasn't been an issue 123yet. Tell me if you'd want one. 124 125---- 126 127Lesson 3: spinlocks revisited. 128 129The single spin-lock primitives above are by no means the only ones. They 130are the most safe ones, and the ones that work under all circumstances, 131but partly _because_ they are safe they are also fairly slow. They are 132much faster than a generic global cli/sti pair, but slower than they'd 133need to be, because they do have to disable interrupts (which is just a 134single instruction on a x86, but it's an expensive one - and on other 135architectures it can be worse). 136 137If you have a case where you have to protect a data structure across 138several CPU's and you want to use spinlocks you can potentially use 139cheaper versions of the spinlocks. IFF you know that the spinlocks are 140never used in interrupt handlers, you can use the non-irq versions: 141 142 spin_lock(&lock); 143 ... 144 spin_unlock(&lock); 145 146(and the equivalent read-write versions too, of course). The spinlock will 147guarantee the same kind of exclusive access, and it will be much faster. 148This is useful if you know that the data in question is only ever 149manipulated from a "process context", ie no interrupts involved. 150 151The reasons you mustn't use these versions if you have interrupts that 152play with the spinlock is that you can get deadlocks: 153 154 spin_lock(&lock); 155 ... 156 <- interrupt comes in: 157 spin_lock(&lock); 158 159where an interrupt tries to lock an already locked variable. This is ok if 160the other interrupt happens on another CPU, but it is _not_ ok if the 161interrupt happens on the same CPU that already holds the lock, because the 162lock will obviously never be released (because the interrupt is waiting 163for the lock, and the lock-holder is interrupted by the interrupt and will 164not continue until the interrupt has been processed). 165 166(This is also the reason why the irq-versions of the spinlocks only need 167to disable the _local_ interrupts - it's ok to use spinlocks in interrupts 168on other CPU's, because an interrupt on another CPU doesn't interrupt the 169CPU that holds the lock, so the lock-holder can continue and eventually 170releases the lock). 171 172Note that you can be clever with read-write locks and interrupts. For 173example, if you know that the interrupt only ever gets a read-lock, then 174you can use a non-irq version of read locks everywhere - because they 175don't block on each other (and thus there is no dead-lock wrt interrupts. 176But when you do the write-lock, you have to use the irq-safe version. 177 178For an example of being clever with rw-locks, see the "waitqueue_lock" 179handling in kernel/sched.c - nothing ever _changes_ a wait-queue from 180within an interrupt, they only read the queue in order to know whom to 181wake up. So read-locks are safe (which is good: they are very common 182indeed), while write-locks need to protect themselves against interrupts. 183 184 Linus 185 186 187