Messages in this thread | | | Date | Wed, 24 Nov 2010 09:09:11 -0500 | From | Chris Metcalf <> | Subject | Re: [PATCH] arch/tile: fix rwlock so would-be write lockers don't block new readers |
| |
On 11/23/2010 9:53 PM, Cypher Wu wrote: > 2010/11/24 Chris Metcalf <cmetcalf@tilera.com>: >> On 11/22/2010 8:36 PM, Cypher Wu wrote: >>> Say, if core A try to write_lock() rwlock and current_ticket_ is 0 and >>> it write next_ticket_ to 1, when it processing the lock, core B try to >>> write_lock() again and write next_ticket_ to 2, then when A >>> write_unlock() it seen that (current_ticket_+1) is not equal to >>> next_ticket_, so it increment current_ticket_, and core B get the >>> lock. If core A try write_lock again before core B write_unlock, it >>> will increment next_ticket_ to 3. And so on. >>> This may rarely happened, I've tested it yesterday for several hours >>> it goes very well under pressure. >> This should be OK when it happens (other than starving out the readers, but >> that was the decision made by doing a ticket lock in the first place). >> Even if we wrap around 255 back to zero on the tickets, the ticket queue >> will work correctly. The key is not to need more than 256 concurrent write >> lock waiters, which we don't. > If we count on that, should we make 'my_ticket_ = (val >> > WR_NEXT_SHIFT) & WR_MASK;'
No, it's OK. As the comment for the declaration of "my_ticket_" says, the trailing underscore reminds us that the high bits are garbage, and when we use the value, we do the mask: "((my_ticket_ - curr_) & WR_MASK)". It turned out doing the mask here made the most sense from a code-generation point of view, partly just because of the possibility of the counter wrapping.
-- Chris Metcalf, Tilera Corp. http://www.tilera.com
| |