> 2023年5月3日 12:44，Paul E. McKenney <paulmck@xxxxxxxxxx> 写道： > > On Wed, May 03, 2023 at 12:31:05PM +0800, Alan Huang wrote: >> >> >>> 2023年5月3日 01:28，Paul E. McKenney <paulmck@xxxxxxxxxx> 写道： >>> >>> On Tue, May 02, 2023 at 11:51:33PM +0800, Alan Huang wrote: >>>> >>>> >>>>> 2023年5月2日 03:02，Paul E. McKenney <paulmck@xxxxxxxxxx> 写道： >>>>> >>>>> On Mon, May 01, 2023 at 07:09:49PM +0800, Alan Huang wrote: >>>>>> >>>>>>> 2023年5月1日 下午7:05，Alan Huang <mmpgouride@xxxxxxxxx> 写道： >>>>>>> >>>>>>> Hi, >>>>>>> >>>>>>> I’m reading section 9.5.4.9, in the second to the last paragraph, there is: >>>>>>> >>>>>>> Before the first horizontal line, all readers see the leftmost myconfig structure, >>>>>>> and after the second horizontal line, all readers will see the rightmost structure. >>>>>>> Between the two lines, that is, during the grace period, different readers might see different objects >>>>>>> >>>>>>> I’m a little confused, in Figure 9.29, the xchg is before the first horizontal line, in my understanding, if the >>>>>>> reader’s reading is before the xchg, it will read the leftmost myconfg structure. >>>>>> >>>>>> And if reader’s reading is after the xchg, it will see the rightmost structure, and because xchg is before the first horizontal line, >>>>>> Readers before the first horizontal line may also see the rightmost structure. >>>>>> >>>>>>> >>>>>>> So, are there some conflicts between the paragraph and the figure? >>>>>>> >>>>>>> Please correct me if I’m wrong. >>>>> >>>>> Consider the point in time just before the left-hand reader's >>>>> rcu_read_unlock() and just after the right-hand reader's rcu_read_lock(). >>>>> These two readers are both running at that time, and one of those readers >>>>> sees the old state and the other sees the new state. >>>> >>>> The sentence that may lead to conflict is this one from the original text: >>>> >>>> Before the first horizontal line, all readers see the leftmost myconfig structure. >>>> >>>> There may be a situation where: >>>> >>>> mcp = xchg(&curconfig, mcp); >>>> rcu_read_lock(); >>>> mcp = … >>>> *cur_a = mcp->a; >>>> -------------------------------------------------first horizontal line------------------------------------------------- >>>> >>>> So, the reader will see the rightmost myconfig structure although it is before the first horizontal line. >>> >>> Ah, I understand now, good catch and thank you! I updated that paragraph >>> to the following, with your Reported-by: >>> >>> The two horizontal grey dashed lines represent the period of time >>> during which different readers get different results, however, >>> each reader will see one and only one of the two objects. All >>> readers that end before the first horizontal line will see the >> >> >> How about this: >> >> mcp = xchg(&curconfig, mcp); >> rcu_read_lock(); >> mcp = … >> *cur_a = mcp->a; >> rcu_read_unlock(); >> -------------------------------------------------first horizontal line------------------------------------------------- >> >> The reader above end before the first horizontal line, but still get the rightmost myconfig structure. >> >> The point is that there is a time period between xchg and the first horizontal line during which a reader >> will see the rightmost myconfig structure. >> >> One possible solution is to move xchg behind the first horizontal line, assuming that I understand correctly. > > Yes, it would in some sense be more accurate to run that horizontal line > right through the "mcp = xchg(&curconfig, mcp);". > > But can you do that so that the diagram looks OK and is legible? I will do my best to do it Thanks, Alan > > Thanx, Paul > >>> leftmost myconfig structure, and all readers that start after the >>> second horizontal line will see the rightmost structure. Between >>> the two lines, that is, during the grace period, different readers >>> might see different objects, but as long as each reader loads the >>> curconfig pointer only once, each reader will see a consistent >>> view of its myconfig structure. >>> >>> Does that help? >>> >>>> BTW, the original text says that black dotted arrows pointing from an older event to a newer event, >>>> but there exists two black dotted arrows in the wrong direction: >>>> >>>> 1.The arrow pointing from synchronize_rcu() to left-hand rcu_read_lock() >>>> 2.The arrow pointing right-hand rcu_read_lock() to synchronize_rcu() >>>> >>>> The directions of the two arrows above may need to be adjusted. I tried to modify this, but after a few hours, I still can’t figure out the svg... >>> >>> The trick is to use an SVG editor. For better or for worse, I use >>> inkscape. But easy for me to change, so I did, again, with your >>> Reported-by. >>> >>> The trick is to flip each arrow horizontally and then vertically, which >>> reverses the direction without having to test one's manual dexterity, >>> which I lack in abundance. ;-) >> >> >> Thanks for the tricks, will try to use it! >> >> Thanks, >> Alan >> >>> >>> Thanx, Paul >>> >>>> Thanks, >>>> Alan >>>> >>>>> >>>>> Does that help, or am I missing the point of your question? >>>>> >>>>> Thanx, Paul
